Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

Brain tumor classification using the diffusion tensor image segmentation (D-SEG) technique

Brain tumor classification using the diffusion tensor image segmentation (D-SEG) technique Neuro-Oncology Neuro-Oncology 17(3), 466 – 476, 2015 doi:10.1093/neuonc/nou159 Advance Access date 13 August 2014 Brain tumor classification using the diffusion tensor image segmentation (D-SEG) technique Timothy L. Jones, Tiernan J. Byrnes, Guang Yang, Franklyn A. Howe, B. Anthony Bell, and Thomas R. Barrick Academic Neurosurgery Unit, St. Georges, University of London, London, UK (T.L.J., T.J.B., B.A.B.); Neurosciences Research Centre, Cardio-vascular and Cell Sciences Institute, St. George’s, University of London, London, UK (G.Y., F.A.H., T.R.B.) Corresponding Author: Timothy L. Jones, PhD, Academic Neurosurgery Unit, St George’s, University of London, Cranmer Terrace, London, SW17 0RE, UK ([email protected]). Background. There is an increasing demand for noninvasive brain tumor biomarkers to guide surgery and subsequent oncother- apy. We present a novel whole-brain diffusion tensor imaging (DTI) segmentation (D-SEG) to delineate tumor volumes of interest (VOIs) for subsequent classification of tumor type. D-SEG uses isotropic ( p) and anisotropic (q) components of the diffusion tensor to segment regions with similar diffusion characteristics. Methods. DTI scans were acquired from 95 patients with low- and high-grade glioma, metastases, and meningioma and from 29 healthy subjects. D-SEG uses k-means clustering of the 2D ( p,q) space to generate segments with different isotropic and aniso- tropic diffusion characteristics. Results. Our results are visualized using a novel RGB color scheme incorporating p, q and T2-weighted information within each seg- ment. The volumetric contribution of each segment to gray matter, white matter, and cerebrospinal fluid spaces was used to generate healthy tissue D-SEG spectra. Tumor VOIs were extracted using a semiautomated flood-filling technique and D-SEG spectra were computed within the VOI. Classification of tumor type using D-SEG spectra was performed using support vector machines. D-SEG was computationally fast and stable and delineated regions of healthy tissue from tumor and edema. D-SEG spectra were consistent for each tumor type, with constituent diffusion characteristics potentially reflecting regional differences in tissue microstructure. Sup- port vector machines classified tumor type with an overall accuracy of 94.7%, providing better classification than previously reported. Conclusions. D-SEG presents a user-friendly, semiautomated biomarker that may provide a valuable adjunct in noninvasive brain tumor diagnosis and treatment planning. Keywords: biomarker, brain tumor, diffusion tensor imaging, glioblastoma, segmentation. The imaging characteristics of newly identified brain tumors sequences yield a correct diagnosis in the majority of cases. may indicate the likely diagnosis and treatment strategy. However, there remains a lack of specificity in challenging sce- Until recently, certain cases of malignant glioma (glioblastoma) narios, such as differentiating: (i) malignant (World Health Orga- and metastatic brain tumors were often considered untreat- nization [WHO] grades III and IV) glial tumors from low-grade 1,2 6 able. Advances in chemotherapeutic and radiotherapy regi- glioma (WHO grades I and II), (ii) malignant glioma from soli- 3 7 mens and appreciation of the role of surgical resection in tary necrotic or cystic cerebral metastasis, and (iii) benign survival resulted in more patients being recommended for en-plaque meningeal tumors (eg, meningioma) from durally treatment. Histological confirmation is usually necessary prior based metastatic deposits. to commencing therapy, yet there remain risks associated Quantifying microscopic diffusion of water molecules with surgery. Noninvasive, accurate, and reproducible bio- using MRI is a proposed surrogate marker of tissue micro- markers are required to assist with decision making. structure. Brain tumors alter regional brain architecture Typical “preoperative” tumor MR protocols include due to differences in cell structure, size, and density and T2-weighted, diffusion-weighted, and gadolinium enhanced the presence of necrosis and edema. Consequently, tumor T1-weighted imaging to evaluate lesion cellularity, vascularity, MR diffusion properties may identify diagnostic intertumoral and blood – brain barrier integrity. These “conventional” differences. Whole-brain maps of diffusion metrics can be Received 25 February 2014; accepted 7 July 2014 # The Author(s) 2014. Published by Oxford University Press on behalf of the Society for Neuro-Oncology. This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] 466 Jones et al.: Tumor classification by a novel DTI segmentation 10,11 generated from diffusion tensor imaging (DTI) data. (mean age 56.3+16.1 y) and 29 healthy subjects (mean age Mean diffusivity (MD) provides a magnitude of isotropic dif- 27.4+7.3 y) were prospectively recruited over an 18-month pe- 2 21 fusion (in mm s ), and fractional anisotropy (FA) provides riod. Patient inclusion criteria were: a radiologically diagnosed a scalar value of diffusion directionality. Differences in MD lesion occupying intracranial space due to undergo surgery and FA among tumor types and grades of malignancy with subsequent histopathological confirmation of tumor 12 – 16 have been investigated with mixed success. type; age over 16 years; and ability to lie flat for 1 h. Tumor An alternative decomposition of the diffusion tensor is into types were: 11 WHO grade I meningiomas, 26 metastases, 31 isotropic ( p)and anisotropic (q)components, where p is a solid grade IV glioblastomas, 7 cystic grade IV glioblastomas, 1 scaled measure of MD, and q is a measure of deviation of the grade III anaplastic astrocytoma, and 19 grade II low-grade gli- 2 21 principal diffusivities from isotropy, both in units of mm s : omas. Of the 95 patients, 82 underwent lesion debulking/resec- tion (11 meningiomas, 26 metastases, 28 high grade gliomas, p = 3MD (1) 16 low-grade gliomas) and 13 had a stereotactic biopsy (10 grade IV glioblastomas, 3 low-grade gliomas). All cases of glio- 2 2 2 q = (l − MD) +(l − MD) +(l − MD) (2) blastoma displayed contrast enhancement on T1-weighted im- 1 2 3 aging, and of the 19 cases of low-grade glioma, 16 did not where l , l , and l are the principal diffusivities of the diffu- 1 2 3 enhance and 3 displayed a faint blush of enhancement. Target- sion tensor and MD ¼ (l + l + l )/3. Each image voxel 1 2 3 ed biopsy of the enhancing region in the low-grade glioma pa- from a DTI dataset can be represented as a coordinate in a tients did not reveal focal cellular anaplasia. Of the 26 2D Cartesian plane referred to as ( p,q) space. metastases studied, 10 originated from lung carcinoma, 7 The majority of studies investigating DTI metrics in tumor from breast carcinoma, 3 melanoma, 3 renal, 2 bowel adenocar- diagnosis utilize manually determined regions of interest cinoma, and 1 prostate. Tumors were all intra-axial and supra- (ROIs) subjectively placed within tumor regions (eg, solid/ne- tentorial; 14 were solitary and 12 were multiple lesions. crotic tumor component, normal-appearing brain, perilesional tissue). ROI placement guided by intensity boundaries on con- Image Acquisition ventional MR images is generally performed on a single image slice, yielding an ROI smaller than the entire lesion. DTIs were acquired using 2 similar 1.5T scanners (termed A and Automated lesion segmentation is an alternative ROI selec- B). Although scanner acquisitions differed, echo times (TEs) tion technique but has been applied mostly to conventional were similar, and repetition times (TRs) were long enough to 19 – 22 MRI, with few examples of tumor segmentation from avoid T1-relaxation effects. Voxel sizes and DTI signal to noise 23,24 diffusion-weighted imaging (DWI) or DTI datasets. Ideally, were similar on each scanner due to acquisition of 12 and 61 tumor segmentation requires minimal user input, is computa- diffusion gradient directions, with 4 and 1 average(s), respec- tionally efficient, and classifies images into regions with different tively. Whole-brain coverage was achievable in a single acquisi- pathological microstructures. In whole-brain DTI datasets, this tion using scanner B, reducing total acquisition time. corresponds to segmenting regions sharing similar diffusion characteristics to reflect similar tissue microstructure. Scanner A We present a novel diffusion segmentation (D-SEG) algo- rithm applied to ( p,q) space. D-SEG automatically segments MRIs were acquired for 41 patients (6 meningiomas, 11 metas- and visualizes regions of similar diffusion characteristics. Pat- tases, 13 glioblastomas, 11 grade II gliomas) and 16 young tern recognition by k-means clustering is used to iteratively healthy subjects (1.5T General Electric Signa LX, quadrature segment ( p,q)space into K nonoverlapping clusters. The num- head coil, maximum gradient strength 22 mT m ). Axial DTIs ber, K, of initial centroids is specified a priori according to the were acquired using a single-shot spin echo planar imaging number of desired clusters as determined by functional and (EPI) sequence. Following acquisition at b¼ 0smm (repeated anatomical considerations. Tumor tissue boundaries identified 10 times), DWIs were acquired (b¼ 1000 s mm )withdiffusion on D-SEG maps are used to semiautomatically delineate vol- gradients applied in 12 directions (TE, 88 ms; TR, 8000 ms; field umes of interest (VOIs). The relative proportion of each ( p,q) of view ¼ 240× 240 mm ; matrix size ¼ 96× 96; slice gap, segment within the VOI reflects the composition of isotropic 2.8 mm; slice thickness, 2.8 mm), providing near isotropic voxels and anisotropic diffusion within the lesion, providing a “signa- 2.5× 2.5× 2.8 mm . Two interleaved acquisitions were acquired, ture”referred toasaD-SEGspectrum. D-SEGisappliedtoa providing contiguous whole-brain coverage over 50 slices and re- cohort of young healthy subjects and a large cohort of peated 4 times to improve signal to noise. The T2-weighted EPI tumor patients to investigate lesion-specific diffusion signa- b ¼ 0s mm images are subsequently referred to as b ¼ 0 tures as surrogate markers of tumor microstructure. Classifi- maps. cation of D-SEG spectra into tumor types is then performed using support vector machines (SVMs). Scanner B MRIs were acquired for 54 patients (5 meningiomas, 15 metas- Materials and Methods tases, 25 glioblastomas, 1 anaplastic astrocytoma, and 8 grade II gliomas) and 13 young healthy subjects (1.5T GE Signa HDx, Patients 8-channel head coil, maximum gradient strength 33 mT m ). All patients participating in this study signed a consent form ap- Differences in DTI acquisition on scanner B compared with proved by the research ethics committee. Ninety-five patients scanner A were that whole-brain DWIs were acquired at a Neuro-Oncology 467 Jones et al.: Tumor classification by a novel DTI segmentation higher angular resolution in 61 noncollinear diffusion gradient cluster, shown for iteration t + 1, directions (TE, 94 ms; TR, 14 000 ms; slice thickness, 2.5 mm; (t+1) no slice gap), providing 2.5 mm isotropic voxels over 55 slices. m = x . i (t) |S | (t) x [S Image Preprocessing An iterative exponential decrease in the number of voxels DWIs were realigned to remove eddy current distortions using changing cluster was observed. D-SEG was terminated after eddy correct (FMRIB Software Library, http://www.fmrib.ox.ac. 250 iterations, after which convergence was achieved. Final uk/fsl) prior to generating p and q maps. Images were skull segmentation of ( p,q) space is displayed as a Voronoi tessella- stripped using Brain Extraction Tool (FMRIB Software). tion (Fig. 1E). Reproducibility of DTI Data Between Scanners Selection of K Between-scanner reproducibility was estimated with 5 healthy We tested our segmentation technique using a range of differ- subjects. For each subject, b ¼ 0 maps acquired on scanner B ent K values (K ¼ 4, 9, 16, and 25). K ¼ 16 was selected because were coregistered to those obtained on scanner A using an it provided the optimum computation time and allowed identi- affine transformation (Statistical Parametric Mapping [SPM]8, fication of our a priori postulated regions within a tumor- http://www.fil.ion.ucl.ac.uk/SPM8) and were used to coregister affected brain, namely: (i) healthy brain GM, (ii) heterogeneous p and q maps. Tissue probability maps of gray matter (GM), WM, (iii) CSF, (iv) solid tumor, (v) regional necrosis, (vi) tumor- white matter (WM), and CSF were computed from each b ¼ 0 associated cystic regions, (vii) perilesional edema, (viii) perile- map SPM8. Hard segmentation maps were computed for sional tumor infiltration, and (ix) distant edema while also iden- GM, WM, and CSF (eg, for GM, p(GM) .p(WM) + p(CSF) at each tifying differences among the 5 tumor types studied. voxel). Voxel-wise comparison of p and q values yielded intra- class correlation coefficients for GM and WM. D-SEG Color Visualization Technique A novel RGB coloring scheme was developed to illustrate the DTI Segmentation Algorithm relative magnitude of p and q diffusion and T2-weighting Histograms of p and q were computed across all brain voxels in (from the b ¼ 0 map) within each D-SEG cluster. A histogram all subjects (n ¼ 123). High intensity noise was removed from of T2-weighted intensities was computed for each subject, the p and q distributions by computing 99.99 percentiles and the 99.99 percentile was discarded, and resultant values were assigning values above this threshold to 1.0. Remaining voxels scaled between 0 and 1. Median p, q, and T2-weighted values were scaled between 0 and 1, generating dataset-wide non- for cluster centroids were ranked from 1 (lowest median) to 16 Gaussian p and q histograms (Fig. 1A and B). (highest median). Rank scores were used to generate an RGB The p and q maps are a set of observations (x , x , .. . , x ) 1 2 color by assigning T2-weighting, p, and q to the red, green, where each observation is a 2D real vector in ( p,q)space and blue channels, respectively (Fig. 2). Color maps were visu- (Fig. 1C). Clustering by k-means partitions the n data points alized using MRIcro. into K disjoint subsets S , where j ={1, 2, ... , 16} by minimizing the within-cluster sum of squares objective function, D-SEG in Healthy Subjects J = ||x − m || , Hard segmentation of b ¼ 0 maps into GM, WM, and CSF was j=1 n[S computed using SPM8 as described above to mask each of the D-SEG maps. The proportion of each segment within each where x is a vector representing the nth data point, and m is n j tissue type was determined and plotted to provide average the geometric centroid of the data points in S .Centroids of D-SEG spectra across all healthy subjects. the initial clusters (m , m , ... , m ) were selected by separating 1 2 n ( p,q) space into K segments of roughly equal size according to median and quartile values of p and q (Fig. 1D, Table 1). These Tumor and Edema Volume of Interest Delineation initial conditions preserve the non-Gaussian structure of the p and q histograms in the cluster initialization. As the data are A combined tumor and edema VOI was semiautomatically de- non-Gaussian, the centroid was defined to be the median lineated for each patient using a 4-voxel neighborhood recur- ( p,q) coordinate of each cluster. The following 2 steps of the al- sive flood-filling algorithm on a slice-by-slice basis. Seed gorithm were repeated: voxel(s) were placed within tumor and edema by a neurosur- Step 1. Assignment step: Assign each voxel to the cluster geon (T.J.) with 6 years of training and 4 years of clinical and whose centroid is closest in ( p,q) space, thus partitioning the research experience of lesion delineation. T.J. was blinded to voxels into K clusters, shown here at the tth iteration, the histopathological diagnosis, and the semiautomated seg- mentation was performed directly from the D-SEG maps with (t) (t) (t) ∗ S ={x :||x − m || ≤ ||x − m || for all i , i [ {1, ... , K}}. ∗ conventional T2-weighted and T1-weighted images (+/2 con- j j j i i i trast) as additional visual guides. No manual editing of the VOI Step 2. Update step: Calculate the new centroids for each was performed post hoc. 468 Jones et al.: Tumor classification by a novel DTI segmentation Fig. 1. D-SEG clustering technique. Normalized histograms of (A) p and (B) q across all subjects (n ¼ 123). (C) The normalized 2D histogram in ( p,q) space for all subjects. (D) Initial clusters with medians in ( p,q) space. (E) Voronoi plot of final clusters (after 250 iterations of the k-means algorithm). All clusters are colored using the D-SEG color mapping technique after 250 iterations (Fig. 2). Cluster numbers in (E) were assigned based on median rank of p in each cluster. Specific segments are associated with increasing anisotropic diffusion (1 to 6), increasing isotropic diffusion (1, 7, 9, 11, 13, 15, and 16), and increasing intermediate diffusivity (1, 8, 10, 12, and 14). Ellipses in (E) show the ( p,q)range of healthy tissue diffusivities (blue ¼ WM, yellow ¼ GM, green ¼ CSF). Neuro-Oncology 469 Jones et al.: Tumor classification by a novel DTI segmentation Table 1. Number and percent of voxels in each D-SEG segment at initialization and termination of the k-means algorithm Segment Number Initial Conditions Algorithm Termination (250 iterations) 2 21 23 2 21 24 Number of Total Voxels, % Number of Total Voxels, % p (mm s × 10 ) q (mm s × 10 ) Constituent Voxels Constituent Voxels 1 4116043 4.66 9732194 11.01 1.22 1.85 2 5720616 6.47 9102275 10.30 1.30 2.86 3 6380942 7.22 6440093 7.29 1.32 3.85 4 5870246 6.64 4277553 4.84 1.34 5.00 5 5260481 5.95 9109123 10.31 1.36 6.59 6 5478436 6.20 7865938 8.90 1.45 9.41 7 5628352 6.37 7741129 8.76 1.45 1.19 8 5720574 6.47 2124115 2.40 1.57 2.11 9 8059367 9.12 7155477 8.10 1.94 1.40 10 6044389 6.84 4351102 4.92 2.03 3.14 11 4416182 5.00 3426063 3.88 2.54 1.73 12 3567915 4.04 1956397 2.21 2.79 4.68 13 4651951 5.27 5434604 6.15 3.26 2.05 14 4844407 5.48 4128078 4.67 3.85 7.73 15 5662375 6.41 3329037 3.77 4.20 3.03 16 6929118 7.84 2178216 2.47 5.48 4.96 Median coordinates in ( p,q) space quantify diffusion characteristics for each segment. Fig. 2. D-SEG color mapping technique. Ranked T2-weighted (red channel), p (green channel), and q (blue channel) maps are shown to the left of D-SEG color maps for 2 axial slices of a healthy subject. D-SEG Tumor Spectra glioblastoma multiforme (GBM), cystic GBM, metastases, and meningioma. Group spectra and classification were not per- The volumetric proportion of each ( p,q) segment to the VOI was formed for the anaplastic astrocytoma case due to insufficient calculated for each case and averaged across tumor type group size (n ¼ 1). to generate D-SEG tumor spectra for low-grade glioma, 470 Jones et al.: Tumor classification by a novel DTI segmentation Table 2. Cross-validated diagnostic results (n ¼ 94), SVM analysis of D-SEG spectra Tumor Type LGG GBM cGBM MET MEN Total Sens. Spec. Accu. 95% CI BER A Confusion matrix—61 direction DTI LGG 10 0 0 0 1 11 90.9 97.5 96.1 (86.5 – 99.5) GBM 0 14 0 0 0 14 100 100 cGBM 0 0 6 0 0 6 100 100 MET 0 0 0 16 0 16 100 100 MEN 1 0 0 0 3 4 75 97.9 B Confusion matrix—12 direction DTI LGG 7 0 0 0 1 8 87.5 100 93.0 (80.9 – 98.5) GBM 0 17 0 0 0 17 100 96.2 cGBM 0 0 1 0 0 1 100 100 MET 0 1 0 9 0 10 90 97.2 MEN 0 0 0 1 6 7 85.7 97.2 C Confusion matrix—combined datasets LGG 19 0 0 0 0 19 100.0 97.3 94.7 (88.0 – 98.3) 6.9 GBM 0 30 1 0 0 31 96.8 98.4 cGBM 0 1 6 0 0 7 85.7 98.9 MET 2 0 0 24 0 26 92.3 98.5 MEN 0 0 0 1 10 11 90.9 100.0 Abbreviations: cGBM, cystic GBM; LGG, low-grade glioma; MEN, meningioma; MET, metastasis. Sens., sensitivity (%); Spec., specificity (%); Accu., accuracy (%); BER, balanced error rate (%). ( p,q) coordinates. The Voronoi plot (Fig. 1E) shows 3 radial lines Tumor Classification of segments through ( p,q) space with unique diffusion charac- The ability of D-SEG spectra to classify tumor type was tested teristics that include: tissue with mostly anisotropic diffusivity across all patients using SVMs. SVM predictions depend on (with q increasing from segment 1 to 6) but with lowest isotro- only a subset of the training data (ie, the support vectors). pic diffusivity, isotropic diffusivity (with p increasing from seg- The technique finds the hyperplane with the largest margin of ment 1 through 7, 9, 11, 13, 15, and 16), and intermediate difference between classes. We used the Gaussian radial diffusivity (with p and q increasing from segment 1 through 8 basis function kernel (s ¼ 1) to map feature vectors into a non- and 12 to 14). linear feature space where an optimal hyperplane was con- structed separating tumor classes. Tenfold cross-validation D-SEG Color Mapping was used to test classification accuracy and reproducibility. To test the integrity of combining tumor DTI from 2 different scan- D-SEG color mapping is shown in Fig. 2 for a healthy subject. The ners, separate SVM classifications of D-SEG spectra acquired for color mapping technique provides visually distinct colors based each acquisition protocol were performed (Table 2A and B). on the diffusion and T2-weighted properties of the tissue in each voxel. White matter regions with high anisotropic diffusion are colored blue. Gray matter regions with low anisotropic and Results isotropic diffusivities are yellow-brown with CSF colored pale yellow. Between-Scanner Reproducibility Mean and standard deviation for intraclass correlation coeffi- D-SEG Spectra in Healthy Subjects cients for GM (0.915+0.097) and WM (0.890+0.110) in Gray matter, WM, and CSF voxels occupy different regions of healthy volunteers showed good interscanner reproducibility ( p,q) space, as shown schematically by the ellipses in Fig. 1E, of p and q diffusion metrics for healthy tissue. and proportionately include different segment amounts result- ing in characteristic D-SEG tissue spectra (Fig. 4A). Gray matter D-SEG Algorithm predominantly includes segments 1, 2, 7, 8, and 9, representing The D-SEG algorithm was computationally fast and reached low isotropic and anisotropic diffusivities, whereas WM almost steady state by 50 iterations. Non-Gaussian characteristics exclusively includes segments 1 to 6, representing low isotropic were apparent in p and q histograms (Fig. 1Aand B) andin diffusion and increasing levels of anisotropic diffusion. CSF the histogram of ( p,q) space. The initial 16 segments assigned spaces include high isotropic diffusion characteristics (seg- to the ( p,q) distribution and the final segmentation after 250 ments 14, 15, and 16). Tissue partial volume effects will be pre- iterations are shown in Fig. 1D and E. Table 1 provides the initial sent in D-SEG segments because tissue class was not used to and final numbers of voxels in each segment and their median define the segmentation. Neuro-Oncology 471 Jones et al.: Tumor classification by a novel DTI segmentation Fig. 3. Individual patient images. From left to right: grade II glioma, glioblastoma with cystic component, cerebral metastasis, and meningioma examples. (A) Fluid attenuated inversion recovery images, (B) T1-weighted postcontrast images, (C) D-SEG color maps, and (d) tumor volumes of interest. All images are illustrated using the radiological convention. Tumor Volume of Interest Delineation Tumor D-SEG Spectra Examples of the VOI extraction technique are shown in Fig. 3 for Figure 4B – F illustrates average D-SEG spectra obtained within the low-grade glioma, glioblastoma, metastases, and meningio- VOIs for each tumor type. The low-grade glioma spectrum con- ma. Conventional fluid attenuated inversion recovery (row A) sisted mostly of segments 9, 11, and 13, representing a lower an- and postcontrast T1-weighted images (row B) indicate the isotropic and higher isotropic diffusion relative to healthy WM. tumor core, cystic, and edematous regions. D-SEG color maps High proportions of intermediate diffusivity segments 10 and show the isotropic and anisotropic diffusion characteristics of 12 potentially represent partial volume effects between tumor the tumor cases (row C) with the extracted VOIs (row D). and WM tissue and were located at the tumor boundary. The D-SEG color images show a visually apparent boundary be- glioblastoma spectrum contained segments of low anisotropic tween healthy and abnormal tissue (solid tumor, necrosis, and isotropic diffusivity (segments 7 and 9, likely corresponding cyst, and edema) that relates to differences in diffusion charac- to solid tumor)aswellassegmentswithhighisotropic andlow teristics located in the regions identified as abnormal in the anisotropic diffusivity (segments 12 and 13, likely corresponding conventional images. While lesion margins on conventional to necrotic regions). High proportions of segments 8, 10, and 11 MRI can be visually indistinct and rely on a subjective choice with greater isotropic diffusivities potentially represent edema re- of thresholding level, the colored segments obtained by gions. Cystic glioblastoma spectra shared such diffusion charac- D-SEG provide a more objective boundary for semiautomatic teristics but with high proportions of segment 15 corresponding lesion delineation. to the cystic region. The D-SEG spectrum of metastases contains 472 Jones et al.: Tumor classification by a novel DTI segmentation Fig. 4. D-SEG spectra. Average proportion of D-SEG segments within VOIs (standard error shown) for: (A) healthy tissue, (B) grade II glioma, (C) glioblastoma, (D) glioblastoma with cystic component, (E) cerebral metastasis, and (F) meningioma. segments 1, 7, 8, and 9 (low isotropic and low anisotropic diffu- balanced error rate of 6.9% after cross-validation (Table 2C). sivity), corresponding to the solid tumor component. Segments Sensitivity and specificity of tumor classification was .90% 10 and 12 likely represent perilesional edema with isotropic diffu- and 97%, respectively, for all tumor types except cystic glio- sivities greater than for glioblastoma. The D-SEG meningioma blastoma. Separate SVM analysis of tumor spectra from the dif- spectrum is markedly different from the other tumor types, ferent DTI acquisitions reveals comparable accuracies (96.1% with a large contribution from segments 1, 2, 3, and 4 (low iso- CI: 86.5% – 99.5% for 61-direction DTI vs 93.0% CI: 80.9% – tropic and increasing anisotropic diffusion), representing the solid 98.5% for 12-direction DTI; Table 2A and B). tumor component. In common with the metastases spectrum, segments 10 and 12 represent the edema region. Discussion We present D-SEG, a fast segmentation and visualization tech- Classification of Tumor Type nique that employs k-means clustering of ( p,q) space to provide SVM analysis of the D-SEG spectra classified tumor type with tissue segments with different isotropic and anisotropic diffu- high overall accuracy (95% CI: 88.0% – 98.3%) and low sion properties. D-SEG maps were colored according to ranked Neuro-Oncology 473 Jones et al.: Tumor classification by a novel DTI segmentation T2-weighted, p and q segment median values to provide a sim- contrast, D-SEG generates tissue type boundaries based on an ple visualization of diffusion characteristics throughout the en- objective clustering of the isotropic and anisotropic diffusivities tire brain that was then used to semiautomatically extract VOIs in (p,q) space. Such segmentation may reflect underlying differ- of abnormal tissue. Distinct D-SEG tumor spectra representing ences in tissue microstructure and potentially relevant patholog- the proportion of diffusion segments within the VOI were com- ical boundaries. However, partial volume effects may result in puted, and SVMs provided exceptionally high classification ac- D-SEG boundaries that do not accurately represent the precise curacy among brain tumor types and grades. difference between pathological and healthy tissue, and further Difficulties arise in multicenter studies incorporating MR dif- work is required to determine the histological ground truth of fusion metrics due to variability in scanner magnetic field, gra- D-SEG boundaries. dient strength, coil channels, and acquisition protocols. Brain tumors are characterized by their heterogeneity in Despite the use of two 1.5T MR scanners with different maxi- size, location, and extent of perilesional edema. Limitations mum gradient strengths and acquisition protocols, the inter- of previous brain tumor diffusion studies are twofold: (i) place- scanner reproducibility of p and q metrics was comparable to ment of ROIs significantly smaller than the lesion potentially 33,34 previous studies. This led to consistent D-SEG spectral pat- excludes relevant diffusion information; (ii) computation of terns in healthy tissue and tumor VOIs for data acquired from 2 average information over whole-lesion ROIs obscures hetero- MR systems. A separate SVM subanalysis of tumor VOI D-SEG geneous diffusion characteristics within the tumor. Spectral spectra generated from the 2 different DTI acquisitions re- comparison using D-SEG overcomes these limitations by pro- vealed comparable diagnostic accuracies, confirming that the viding a pattern of diffusivity across the entire region of abnor- datasets may be combined for the presented analysis. mal tissue. The D-SEG technique separates ( p,q) space into segments D-SEG spectra differ among tumor types in both their con- with distinct isotropic and anisotropic diffusion properties. stituent segment numbers and their proportional contribution Simultaneous application of D-SEG to all healthy and patient to the VOI. Spectra are consistent within the tumor type, con- data ensured that segments contained voxels with the same firmed by small standard errors for segments despite variability diffusion properties in each individual. This allowed meaningful in size, location, natural history, and, in the case of metastases, between-subject comparison of D-SEG spectra. Nevertheless, cellular origin of each lesion. Possible reasons for differences in further work is required to evaluate stability of the final D-SEG diffusion characteristics between tumor types include presence result due to perturbation of the initial algorithmic conditions of necrotic or cystic regions or volumetric proportion of tumor and for different numbers of tumor datasets. In this study, ini- and edema, solid tumor microstructure, and pathophysiology tial conditions were chosen that reflected the local density of of perilesional edema. ( p,q) space and consequently provided similar voxel numbers Malignant tumors are characterized by rapid growth and per segment. Alternative segmentation techniques could be neovascularity. When tumor rate of growth exceeds its blood applied, but an algorithmic investigation of optimality is beyond supply, cell death and regional necrosis result. The loss of cel- the scope of this study. Interestingly (p,q) space is characterized lular structure and boundaries to diffusion results in higher iso- by a non-Gaussian distribution that does not contain explicit tropic diffusion and is observed in glioblastoma D-SEG data clusters. Nevertheless, D-SEG provides a discrete mapping spectra. Tumor cysts may result from necrotic degeneration, of this space dependent on local voxel density generating an in- central hemorrhage, liquefaction, entrapment of CSF, and plas- tuitive separation of isotropic and anisotropic diffusion. ma fluid leaking from a disrupted blood – brain barrier. Cysts D-SEG provides reproducible segmentation of GM, WM, and have high isotropic diffusivity in the D-SEG spectra, reflecting CSF. Although D-SEG does not define exclusive segments for the fluid nature of these regions. Glioblastoma cysts exhibit each tissue type, it provides a spectrum of diffusion properties lower isotropic diffusivity than normal CSF spaces, potentially supporting previous findings of similar isotropic diffusivity in GM reflecting their proteinaceous constituents. and WM, and heterogeneous anisotropic diffusion in WM. The solid component of tumors consists of disorganized In this study, CSF spaces exhibited high magnitudes of isotropic pleomorphic, hypercellular cells with hyperchromatic nuclei, diffusion but also greater anisotropic diffusion than did GM. This lacking the organized structure of nascent neural tissue. In effect was caused by the use of q to quantify anisotropic diffu- common with previous studies, D-SEG spectra indicate that iso- sion, which, unlike FA, is not scaled by the overall magnitude of tropic diffusion within the solid tumor contributes to differenti- 17 41,44,45 diffusion within a voxel. ating among tumor types. In particular, differences in We prospectively recruited 94 patients with histopathologi- cell density of the solid tumor may be responsible for observed cally confirmed lesions occupying intracranial space over an spectral differences in the proportion of isotropic and anisotrop- 18-month period representing a cross section of brain tumors ic segments among tumor types. Our results confirm that iso- encountered in our neurosurgical practice. D-SEG analysis of tropic diffusion is smaller in glioblastoma than in low-grade 37,46,47 these data show that combined tumor and edema VOIs deter- glioma, agreeing with previous studies, and contrasts mined by D-SEG correspond visually with the extent of tumor on the high cellularity of glioblastoma with cellularity that is only standard MRI; however, their complex margins are indistinct on moderately increasedcomparedwithnormalbrain in low- conventional MRI or within p and q maps. A range of region grade glioma. D-SEG spectra confirm a lower isotropic diffusion drawing techniques have been used to determine diffusion char- within the solid component of metastases than glioblastoma, 7,36 acteristics of tumors, such as from manually drawn lesion agreeing with previous studies. This contrasts the densely 36 37 edges or from within ROIs placed in specified brain regions. packed and restricted diffusion within secondary tumors with These techniques are time-consuming and user dependent and the irregular cellular arrangement of microscopic necrosis in are subjective interpretations of the tumor boundary. In glioblastoma. Meningioma D-SEG spectra are markedly 474 Jones et al.: Tumor classification by a novel DTI segmentation different from the other tumor types. The solid tumor compo- these results and determine the additional utility of D-SEG in nent has the lowest isotropic diffusion and an anisotropic diffu- these and other clinical scenarios. sion component that likely reflects interdigitating cellular processes, tight intercellular junctions, and formation of fascic- ular and lobular tissue in association with whorls and psam- Funding moma bodies. This work was supported by Cancer Research UK, grant nos. C8807/ D-SEG spectra differ in segments containing perilesional A3870 and C1459/A13303; the EU, grant no. LSHC-CT-2004-503094 edema. Tumors with vasogenic edema, such as metastases (eTUMOUR); and T.L.J. acknowledges a Royal College of Surgeons of En- and meningioma, comprise a greater proportion of segments gland Research Fellowship. with higher isotropic diffusion than infiltrative cellular edema in glioblastoma. These findings agree with previous studies and reflect the differences in pathophysiology of the 36,46 Acknowledgments edema. Due to the ability of D-SEG to capture differences among tu- We thank Chris Clark, Dominick McIntyre, and Alan Wright for their help with the MR imaging. mors, it provides exceptional accuracy in tackling clinically rel- evant diagnostic scenarios, including differentiating glioma grade and discriminating between isolated metastases and glioblastoma. Although D-SEG accurately distinguishes metas- Conflict of interest statement. We have no conflicts of interest to tases and meningioma, this has limited clinical use except for declare. differentiating durally based metastatic deposits from benign meningioma. D-SEG delineated a clear boundary between tumor and References normal-appearing brain tissue for all except a small number 1. Darefsky AS, King JT, Dubrow R. Adult glioblastoma multiforme of meningioma cases. In these cases the meningioma dis- survival in the temozolomide era: a population-based analysis of played anisotropic diffusion properties similar to healthy WM. Surveillance, Epidemiology, and End Results registries. Cancer. 2012;118(8):2163 – 2172. Here it was necessary to manually draw lesion boundaries for the VOI guided by conventional postcontrast T1-weighted im- 2. Koay E, Sulman EP. Management of brain metastasis: past lessons, ages. A further limitation is that diffusion properties identified modern management, and future considerations. Curr Oncol Rep. by D-SEG are not exclusively representative of healthy or path- 2012;14(1):70 – 78. ological tissue types due to tissue partial volume effects, ef- 3. Stupp R, Mason WP, van den Bent MJ, et al. Radiotherapy plus fects of noise, and certain tumor tissue types sharing similar concomitant and adjuvant temozolomide for glioblastoma. N Engl J Med. 2005;352(10):987 – 996. diffusivities to healthy brain tissue. The use of D-SEG spectra overcomes these problems by providing proportions of seg- 4. Stummer W, Pichlmeier U, Meinel T, et al. Fluorescence-guided ments within VOIs. The 94 brain tumor cases acquired in this surgery with 5-aminolevulinic acid for resection of malignant study represent the largest reported cohort of tumor DTI glioma: a randomised controlled multicentre phase III trial. Lancet Oncol. 2006;7(5):392 – 401. data. Further evaluation of the role of D-SEG in diagnosis re- quires comparing the accuracy of this method with the accura- 5. McGirt MJ, Woodworth GF, Coon AL, et al. Independent predictors of morbidity after image-guided stereotactic brain biopsy: a risk cy of reporting of conventional imaging, as well as testing with assessment of 270 cases. J Neurosurg. 2005;102(5):897 – 901. other tumor types and grades and a range of different metas- tases, as the site of primary disease and thus cell lineage may 6. Kondziolka D, Lunsford LD, Martinez AJ. Unreliability of influence microstructure and thus tumor p-q diffusion contemporary neurodiagnostic imaging in evaluating suspected adult supratentorial (low-grade) astrocytoma. J Neurosurg. 1993; characteristics. 79(4):533 – 536. 7. Wang S, Kim S, Chawla S, et al. Differentiation between Conclusion glioblastomas and solitary brain metastases using diffusion tensor imaging. Neuroimage. 2009;44(3):653 – 660. Advances in imaging of newly identified lesions occupying in- tracranial space have not eliminated the requirement for histo- 8. Tagle P, Villanueva P, Torrealba GHI. Intracranial metastasis or meningioma? Surg Neurol. 2002;58(3 – 4):241 – 245. pathological “tissue” diagnosis in the majority of cases. With an increasing elderly population, improved survival in systemic 9. Basser PJ. Inferring microstructural features and the physiological malignancy, and increased detection of lesions (potentially in- state of tissues from diffusion-weighted images. NMR Biomed. cidentally) at an early stage, there exists a need for accurate 1995;8(7 – 8):333 – 344. and reproducible noninvasive diagnostic tools. In this study, 10. Basser PJ, Mattiello J, LeBihan D. Estimation of the effective we present a technique capable of classifying brain tumors self-diffusion tensor from the NMR spin echo. JMagnReson B. using biomarkers obtained from DTI with an accuracy of 1994;103(3):247 – 254. 94.7%. The potential roles for D-SEG include: (i) delineation of 11. Pierpaoli C, Basser PJ. Toward a quantitative assessment of lesion margins for optimal surgical resection or radiotherapy diffusion anisotropy. Magn Reson Med. 1996;36(6):893 – 906. treatment, (ii) serial volumetric analysis to monitor changes 12. Stadlbauer A, Ganslandt O, Buslei R, et al. Gliomas: in low-grade glioma over time or evaluate response to onco- histopathologic evaluation of changes in directionality and therapy, and (iii) regional assessment of biomarkers for diagno- magnitude of water diffusion at diffusion-tensor MR imaging. sis or surveillance. Further studies are required to replicate Radiology. 2006;240(3):803 – 810. Neuro-Oncology 475 Jones et al.: Tumor classification by a novel DTI segmentation 13. Beppu T, Inoue T, Shibata Y, et al. Fractional anisotropy value by 31. Murphy K. Machine Learning: a Probabilistic Perspective. diffusion tensor magnetic resonance imaging as a predictor of Cambridge, MA: MIT Press; 2012. cell density and proliferation activity of glioblastomas. Surg 32. Pfefferbaum A, Adalsteinsson E, Sullivan EV. Replicability of Neurol. 2005;63(1):56 – 61. diffusion tensor imaging measurements of fractional anisotropy 14. Lu S, Ahn D, Johnson G, et al. Peritumoral diffusion tensor imaging and trace in brain. J Magn Reson Imaging. 2003;18(4):427 – 433. of high-grade gliomas and metastatic brain tumors. AJNR Am J 33. Vollmar C, O’Muircheartaigh J, Barker GJ, et al. Identical, but not Neuroradiol. 2003;24(5):937 – 941. the same: intra-site and inter-site reproducibility of fractional 15. Provenzale JM, Mcgraw P, Mhatre P, et al. Peritumoral brain regions anisotropy measures on two 3.0T scanners. Neuroimage. 2010; in gliomas and meningiomas: investigation with isotropic imaging 51(4):1384 – 1394. and diffusion-tensor. Radiology. 2004;232(2):451 – 460. 34. Bonekamp D, Nagae LM, Degaonkar M, et al. Diffusion 16. Tsuchiya K, Fujikawa A, Nakajima M, et al. Differentiation between tensor imaging in children and adolescents: reproducibility, solitary brain metastasis and high-grade glioma by diffusion hemispheric, and age-related differences. Neuroimage. 2007; tensor imaging. Br J Radiol. 2005;78(930):533 – 537. 34(2):733 – 742. 17. Pen ˜ a A, Green HAL, Carpenter TA, et al. Enhanced visualization and 35. Pierpaoli C, Jezzard P, Basser PJ, et al. Diffusion tensor MR imaging quantification of magnetic resonance diffusion tensor imaging of the human brain. Radiology. 1996;201(3):637 – 648. using the p:q tensor decomposition. Br J Radiol. 2006;79(938): 36. ByrnesTJD,Barrick TR,BellBA, et al. Diffusion tensor imaging 101 – 109. discriminates between glioblastoma and cerebral metastases in 18. Gordillo N, Montseny E, Sobrevilla P. State of the art survey on MRI vivo. NMR Biomed. 2011;24(1):54 – 60. brain tumor segmentation. Magn Reson Imaging. 2013;31(8): 37. Price SJ, Pen ˜ a A, Burnet NG, et al. Tissue signature characterisation 1426 – 1438. of diffusion tensor abnormalities in cerebral gliomas. Eur Radiol. 19. Fletcher-Heath LM, Hall LO, Goldgof DB, et al. Automatic 2004;14(10):1909 – 1917. segmentation of non-enhancing brain tumors in magnetic 38. Kaus MR, Warfield SK, Nabavi A, et al. Automated segmentation of resonance images. Artif Intell Med. 2001;21(1 – 3):43 – 63. MR images of brain tumors. Radiology. 2001;218(2):586 – 591. 20. Clark MC, Hall LO, Goldgof DB, et al. Automatic tumor 39. Kleihues P, Louis DN, Scheithauer BW, et al. The WHO classification segmentation using knowledge-based techniques. IEEE Trans of tumors of the nervous system. J Neuropathol Exp Neurol. 2002; Med Imaging. 1998;17(2):187 – 201. 61(3):215 – 225. 21. Liu J, Udupa JK, Odhner D, et al. A system for brain tumor volume 40. Raza SM, Lang FF, Aggarwal BB, et al. Necrosis and glioblastoma: a estimation via MR imaging and fuzzy connectedness. Comput Med friend or a foe? A review and a hypothesis. Neurosurgery.2002; Imaging Graph. 2005;29(1):21 – 34. 51(1):2 – 12. 22. Clarke LP, Velthuizen RP, Phuphanich S, et al. MRI: stability of three 41. Krabbe K, Gideon P, Wagn P, et al. MR diffusion imaging of human supervised segmentation techniques. Magn Reson. 1993; intracranial tumours. Neuroradiology. 1997;39(7):483 – 489. 11(0730 – 725X (Print)):95 – 106. 42. Poisson M, Philippon J, van Effenterre R, et al. Cerebral pseudocysts 23. Vijayakumar C, Damayanti G, Pant R, et al. Segmentation and following chemotherapy of glioblastomas. Acta Neurochir (Wien). grading of brain tumors on apparent diffusion coefficient 1977;39(3 – 4):143 – 149. images using self-organizing maps. Comput Med Imaging Graph. 2007;31(7):473 – 484. 43. Louis DN, Ohgaki H, Wiestler OD, et al. The 2007 WHO classification of tumours of the central nervous system. Acta 24. Simon D, Fritzsche KH, Thieke C, et al. Diffusion-weighted Neuropathol. 2007;114(2):97 – 109. imaging-based probabilistic segmentation of high- and low- proliferative areas in high-grade gliomas. Cancer Imaging. 2012; 44. Bulakbasi N, Kocaoglu M, Ors F, et al. Combination of single-voxel 12:89 – 99. proton MR spectroscopy and apparent diffusion coefficient 25. MacQueen JB. Kmeans Some methods for classification and calculation in the evaluation of common brain tumors. AJNR Am analysis of multivariate observations. 5th Berkeley Symp Math J Neuroradiol. 2003;24(2):225 – 233. Stat Probab 1967. 1967;1:281 – 297. 45. Kono K, Inoue Y, Nakayama K, et al. The role of diffusion-weighted 26. Nanetti L, Cerliani L, Gazzola V, et al. Group analyses of imaging in patients with brain tumors. AJNR Am J Neuroradiol. connectivity-based cortical parcellation using repeated k-means 2001;22(6):1081 – 1088. clustering. Neuroimage. 2009;47(4):1666 – 1677. 46. Lu S, Ahn D, Johnson G, et al. Diffusion-tensor MR imaging 27. Ashburner J, Friston KJ. Unified segmentation. Neuroimage. 2005; of intracranial neoplasia and associated peritumoral edema: 26(3):839 – 851. introduction of the tumor infiltration index. Radiology. 2004; 232(1):221 – 228. 28. Aurenhammer F. Voronoi diagrams. ACM Comput Surv. 1991;23(3): 345 – 405. 47. Inoue T, Ogasawara K, Beppu T, et al. Diffusion tensor imaging for 29. Rorden C, Brett M. Stereotaxic display of brain lesions. Behav preoperative evaluation of tumor grade in gliomas. Clin Neurol Neurol. 2000;12(4):191 – 200. Neurosurg. 2005;107(3):174 – 180. 48. Lu S, Ahn D, Johnson G, Cha S. Peritumoral diffusion tensor 30. Cristianini N, Shawe-Taylor J. An Introduction to Support Vector Machines and Other Kernel-based Learning Methods. Cambridge: imaging of high-grade gliomas and metastatic brain tumours. Cambridge University Press; 2000. AJNR Am J Neuroradiol. 2003;24(5):937 – 941. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Neuro-Oncology Pubmed Central

Brain tumor classification using the diffusion tensor image segmentation (D-SEG) technique

Download PDF
11 pages

Loading next page...
 
/lp/pubmed-central/brain-tumor-classification-using-the-diffusion-tensor-image-u2uIfVOpwT

References (51)

Publisher
Pubmed Central
Copyright
© The Author(s) 2014. Published by Oxford University Press on behalf of the Society for Neuro-Oncology.
ISSN
1522-8517
eISSN
1523-5866
DOI
10.1093/neuonc/nou159
Publisher site
See Article on Publisher Site

Abstract

Neuro-Oncology Neuro-Oncology 17(3), 466 – 476, 2015 doi:10.1093/neuonc/nou159 Advance Access date 13 August 2014 Brain tumor classification using the diffusion tensor image segmentation (D-SEG) technique Timothy L. Jones, Tiernan J. Byrnes, Guang Yang, Franklyn A. Howe, B. Anthony Bell, and Thomas R. Barrick Academic Neurosurgery Unit, St. Georges, University of London, London, UK (T.L.J., T.J.B., B.A.B.); Neurosciences Research Centre, Cardio-vascular and Cell Sciences Institute, St. George’s, University of London, London, UK (G.Y., F.A.H., T.R.B.) Corresponding Author: Timothy L. Jones, PhD, Academic Neurosurgery Unit, St George’s, University of London, Cranmer Terrace, London, SW17 0RE, UK ([email protected]). Background. There is an increasing demand for noninvasive brain tumor biomarkers to guide surgery and subsequent oncother- apy. We present a novel whole-brain diffusion tensor imaging (DTI) segmentation (D-SEG) to delineate tumor volumes of interest (VOIs) for subsequent classification of tumor type. D-SEG uses isotropic ( p) and anisotropic (q) components of the diffusion tensor to segment regions with similar diffusion characteristics. Methods. DTI scans were acquired from 95 patients with low- and high-grade glioma, metastases, and meningioma and from 29 healthy subjects. D-SEG uses k-means clustering of the 2D ( p,q) space to generate segments with different isotropic and aniso- tropic diffusion characteristics. Results. Our results are visualized using a novel RGB color scheme incorporating p, q and T2-weighted information within each seg- ment. The volumetric contribution of each segment to gray matter, white matter, and cerebrospinal fluid spaces was used to generate healthy tissue D-SEG spectra. Tumor VOIs were extracted using a semiautomated flood-filling technique and D-SEG spectra were computed within the VOI. Classification of tumor type using D-SEG spectra was performed using support vector machines. D-SEG was computationally fast and stable and delineated regions of healthy tissue from tumor and edema. D-SEG spectra were consistent for each tumor type, with constituent diffusion characteristics potentially reflecting regional differences in tissue microstructure. Sup- port vector machines classified tumor type with an overall accuracy of 94.7%, providing better classification than previously reported. Conclusions. D-SEG presents a user-friendly, semiautomated biomarker that may provide a valuable adjunct in noninvasive brain tumor diagnosis and treatment planning. Keywords: biomarker, brain tumor, diffusion tensor imaging, glioblastoma, segmentation. The imaging characteristics of newly identified brain tumors sequences yield a correct diagnosis in the majority of cases. may indicate the likely diagnosis and treatment strategy. However, there remains a lack of specificity in challenging sce- Until recently, certain cases of malignant glioma (glioblastoma) narios, such as differentiating: (i) malignant (World Health Orga- and metastatic brain tumors were often considered untreat- nization [WHO] grades III and IV) glial tumors from low-grade 1,2 6 able. Advances in chemotherapeutic and radiotherapy regi- glioma (WHO grades I and II), (ii) malignant glioma from soli- 3 7 mens and appreciation of the role of surgical resection in tary necrotic or cystic cerebral metastasis, and (iii) benign survival resulted in more patients being recommended for en-plaque meningeal tumors (eg, meningioma) from durally treatment. Histological confirmation is usually necessary prior based metastatic deposits. to commencing therapy, yet there remain risks associated Quantifying microscopic diffusion of water molecules with surgery. Noninvasive, accurate, and reproducible bio- using MRI is a proposed surrogate marker of tissue micro- markers are required to assist with decision making. structure. Brain tumors alter regional brain architecture Typical “preoperative” tumor MR protocols include due to differences in cell structure, size, and density and T2-weighted, diffusion-weighted, and gadolinium enhanced the presence of necrosis and edema. Consequently, tumor T1-weighted imaging to evaluate lesion cellularity, vascularity, MR diffusion properties may identify diagnostic intertumoral and blood – brain barrier integrity. These “conventional” differences. Whole-brain maps of diffusion metrics can be Received 25 February 2014; accepted 7 July 2014 # The Author(s) 2014. Published by Oxford University Press on behalf of the Society for Neuro-Oncology. This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] 466 Jones et al.: Tumor classification by a novel DTI segmentation 10,11 generated from diffusion tensor imaging (DTI) data. (mean age 56.3+16.1 y) and 29 healthy subjects (mean age Mean diffusivity (MD) provides a magnitude of isotropic dif- 27.4+7.3 y) were prospectively recruited over an 18-month pe- 2 21 fusion (in mm s ), and fractional anisotropy (FA) provides riod. Patient inclusion criteria were: a radiologically diagnosed a scalar value of diffusion directionality. Differences in MD lesion occupying intracranial space due to undergo surgery and FA among tumor types and grades of malignancy with subsequent histopathological confirmation of tumor 12 – 16 have been investigated with mixed success. type; age over 16 years; and ability to lie flat for 1 h. Tumor An alternative decomposition of the diffusion tensor is into types were: 11 WHO grade I meningiomas, 26 metastases, 31 isotropic ( p)and anisotropic (q)components, where p is a solid grade IV glioblastomas, 7 cystic grade IV glioblastomas, 1 scaled measure of MD, and q is a measure of deviation of the grade III anaplastic astrocytoma, and 19 grade II low-grade gli- 2 21 principal diffusivities from isotropy, both in units of mm s : omas. Of the 95 patients, 82 underwent lesion debulking/resec- tion (11 meningiomas, 26 metastases, 28 high grade gliomas, p = 3MD (1) 16 low-grade gliomas) and 13 had a stereotactic biopsy (10 grade IV glioblastomas, 3 low-grade gliomas). All cases of glio- 2 2 2 q = (l − MD) +(l − MD) +(l − MD) (2) blastoma displayed contrast enhancement on T1-weighted im- 1 2 3 aging, and of the 19 cases of low-grade glioma, 16 did not where l , l , and l are the principal diffusivities of the diffu- 1 2 3 enhance and 3 displayed a faint blush of enhancement. Target- sion tensor and MD ¼ (l + l + l )/3. Each image voxel 1 2 3 ed biopsy of the enhancing region in the low-grade glioma pa- from a DTI dataset can be represented as a coordinate in a tients did not reveal focal cellular anaplasia. Of the 26 2D Cartesian plane referred to as ( p,q) space. metastases studied, 10 originated from lung carcinoma, 7 The majority of studies investigating DTI metrics in tumor from breast carcinoma, 3 melanoma, 3 renal, 2 bowel adenocar- diagnosis utilize manually determined regions of interest cinoma, and 1 prostate. Tumors were all intra-axial and supra- (ROIs) subjectively placed within tumor regions (eg, solid/ne- tentorial; 14 were solitary and 12 were multiple lesions. crotic tumor component, normal-appearing brain, perilesional tissue). ROI placement guided by intensity boundaries on con- Image Acquisition ventional MR images is generally performed on a single image slice, yielding an ROI smaller than the entire lesion. DTIs were acquired using 2 similar 1.5T scanners (termed A and Automated lesion segmentation is an alternative ROI selec- B). Although scanner acquisitions differed, echo times (TEs) tion technique but has been applied mostly to conventional were similar, and repetition times (TRs) were long enough to 19 – 22 MRI, with few examples of tumor segmentation from avoid T1-relaxation effects. Voxel sizes and DTI signal to noise 23,24 diffusion-weighted imaging (DWI) or DTI datasets. Ideally, were similar on each scanner due to acquisition of 12 and 61 tumor segmentation requires minimal user input, is computa- diffusion gradient directions, with 4 and 1 average(s), respec- tionally efficient, and classifies images into regions with different tively. Whole-brain coverage was achievable in a single acquisi- pathological microstructures. In whole-brain DTI datasets, this tion using scanner B, reducing total acquisition time. corresponds to segmenting regions sharing similar diffusion characteristics to reflect similar tissue microstructure. Scanner A We present a novel diffusion segmentation (D-SEG) algo- rithm applied to ( p,q) space. D-SEG automatically segments MRIs were acquired for 41 patients (6 meningiomas, 11 metas- and visualizes regions of similar diffusion characteristics. Pat- tases, 13 glioblastomas, 11 grade II gliomas) and 16 young tern recognition by k-means clustering is used to iteratively healthy subjects (1.5T General Electric Signa LX, quadrature segment ( p,q)space into K nonoverlapping clusters. The num- head coil, maximum gradient strength 22 mT m ). Axial DTIs ber, K, of initial centroids is specified a priori according to the were acquired using a single-shot spin echo planar imaging number of desired clusters as determined by functional and (EPI) sequence. Following acquisition at b¼ 0smm (repeated anatomical considerations. Tumor tissue boundaries identified 10 times), DWIs were acquired (b¼ 1000 s mm )withdiffusion on D-SEG maps are used to semiautomatically delineate vol- gradients applied in 12 directions (TE, 88 ms; TR, 8000 ms; field umes of interest (VOIs). The relative proportion of each ( p,q) of view ¼ 240× 240 mm ; matrix size ¼ 96× 96; slice gap, segment within the VOI reflects the composition of isotropic 2.8 mm; slice thickness, 2.8 mm), providing near isotropic voxels and anisotropic diffusion within the lesion, providing a “signa- 2.5× 2.5× 2.8 mm . Two interleaved acquisitions were acquired, ture”referred toasaD-SEGspectrum. D-SEGisappliedtoa providing contiguous whole-brain coverage over 50 slices and re- cohort of young healthy subjects and a large cohort of peated 4 times to improve signal to noise. The T2-weighted EPI tumor patients to investigate lesion-specific diffusion signa- b ¼ 0s mm images are subsequently referred to as b ¼ 0 tures as surrogate markers of tumor microstructure. Classifi- maps. cation of D-SEG spectra into tumor types is then performed using support vector machines (SVMs). Scanner B MRIs were acquired for 54 patients (5 meningiomas, 15 metas- Materials and Methods tases, 25 glioblastomas, 1 anaplastic astrocytoma, and 8 grade II gliomas) and 13 young healthy subjects (1.5T GE Signa HDx, Patients 8-channel head coil, maximum gradient strength 33 mT m ). All patients participating in this study signed a consent form ap- Differences in DTI acquisition on scanner B compared with proved by the research ethics committee. Ninety-five patients scanner A were that whole-brain DWIs were acquired at a Neuro-Oncology 467 Jones et al.: Tumor classification by a novel DTI segmentation higher angular resolution in 61 noncollinear diffusion gradient cluster, shown for iteration t + 1, directions (TE, 94 ms; TR, 14 000 ms; slice thickness, 2.5 mm; (t+1) no slice gap), providing 2.5 mm isotropic voxels over 55 slices. m = x . i (t) |S | (t) x [S Image Preprocessing An iterative exponential decrease in the number of voxels DWIs were realigned to remove eddy current distortions using changing cluster was observed. D-SEG was terminated after eddy correct (FMRIB Software Library, http://www.fmrib.ox.ac. 250 iterations, after which convergence was achieved. Final uk/fsl) prior to generating p and q maps. Images were skull segmentation of ( p,q) space is displayed as a Voronoi tessella- stripped using Brain Extraction Tool (FMRIB Software). tion (Fig. 1E). Reproducibility of DTI Data Between Scanners Selection of K Between-scanner reproducibility was estimated with 5 healthy We tested our segmentation technique using a range of differ- subjects. For each subject, b ¼ 0 maps acquired on scanner B ent K values (K ¼ 4, 9, 16, and 25). K ¼ 16 was selected because were coregistered to those obtained on scanner A using an it provided the optimum computation time and allowed identi- affine transformation (Statistical Parametric Mapping [SPM]8, fication of our a priori postulated regions within a tumor- http://www.fil.ion.ucl.ac.uk/SPM8) and were used to coregister affected brain, namely: (i) healthy brain GM, (ii) heterogeneous p and q maps. Tissue probability maps of gray matter (GM), WM, (iii) CSF, (iv) solid tumor, (v) regional necrosis, (vi) tumor- white matter (WM), and CSF were computed from each b ¼ 0 associated cystic regions, (vii) perilesional edema, (viii) perile- map SPM8. Hard segmentation maps were computed for sional tumor infiltration, and (ix) distant edema while also iden- GM, WM, and CSF (eg, for GM, p(GM) .p(WM) + p(CSF) at each tifying differences among the 5 tumor types studied. voxel). Voxel-wise comparison of p and q values yielded intra- class correlation coefficients for GM and WM. D-SEG Color Visualization Technique A novel RGB coloring scheme was developed to illustrate the DTI Segmentation Algorithm relative magnitude of p and q diffusion and T2-weighting Histograms of p and q were computed across all brain voxels in (from the b ¼ 0 map) within each D-SEG cluster. A histogram all subjects (n ¼ 123). High intensity noise was removed from of T2-weighted intensities was computed for each subject, the p and q distributions by computing 99.99 percentiles and the 99.99 percentile was discarded, and resultant values were assigning values above this threshold to 1.0. Remaining voxels scaled between 0 and 1. Median p, q, and T2-weighted values were scaled between 0 and 1, generating dataset-wide non- for cluster centroids were ranked from 1 (lowest median) to 16 Gaussian p and q histograms (Fig. 1A and B). (highest median). Rank scores were used to generate an RGB The p and q maps are a set of observations (x , x , .. . , x ) 1 2 color by assigning T2-weighting, p, and q to the red, green, where each observation is a 2D real vector in ( p,q)space and blue channels, respectively (Fig. 2). Color maps were visu- (Fig. 1C). Clustering by k-means partitions the n data points alized using MRIcro. into K disjoint subsets S , where j ={1, 2, ... , 16} by minimizing the within-cluster sum of squares objective function, D-SEG in Healthy Subjects J = ||x − m || , Hard segmentation of b ¼ 0 maps into GM, WM, and CSF was j=1 n[S computed using SPM8 as described above to mask each of the D-SEG maps. The proportion of each segment within each where x is a vector representing the nth data point, and m is n j tissue type was determined and plotted to provide average the geometric centroid of the data points in S .Centroids of D-SEG spectra across all healthy subjects. the initial clusters (m , m , ... , m ) were selected by separating 1 2 n ( p,q) space into K segments of roughly equal size according to median and quartile values of p and q (Fig. 1D, Table 1). These Tumor and Edema Volume of Interest Delineation initial conditions preserve the non-Gaussian structure of the p and q histograms in the cluster initialization. As the data are A combined tumor and edema VOI was semiautomatically de- non-Gaussian, the centroid was defined to be the median lineated for each patient using a 4-voxel neighborhood recur- ( p,q) coordinate of each cluster. The following 2 steps of the al- sive flood-filling algorithm on a slice-by-slice basis. Seed gorithm were repeated: voxel(s) were placed within tumor and edema by a neurosur- Step 1. Assignment step: Assign each voxel to the cluster geon (T.J.) with 6 years of training and 4 years of clinical and whose centroid is closest in ( p,q) space, thus partitioning the research experience of lesion delineation. T.J. was blinded to voxels into K clusters, shown here at the tth iteration, the histopathological diagnosis, and the semiautomated seg- mentation was performed directly from the D-SEG maps with (t) (t) (t) ∗ S ={x :||x − m || ≤ ||x − m || for all i , i [ {1, ... , K}}. ∗ conventional T2-weighted and T1-weighted images (+/2 con- j j j i i i trast) as additional visual guides. No manual editing of the VOI Step 2. Update step: Calculate the new centroids for each was performed post hoc. 468 Jones et al.: Tumor classification by a novel DTI segmentation Fig. 1. D-SEG clustering technique. Normalized histograms of (A) p and (B) q across all subjects (n ¼ 123). (C) The normalized 2D histogram in ( p,q) space for all subjects. (D) Initial clusters with medians in ( p,q) space. (E) Voronoi plot of final clusters (after 250 iterations of the k-means algorithm). All clusters are colored using the D-SEG color mapping technique after 250 iterations (Fig. 2). Cluster numbers in (E) were assigned based on median rank of p in each cluster. Specific segments are associated with increasing anisotropic diffusion (1 to 6), increasing isotropic diffusion (1, 7, 9, 11, 13, 15, and 16), and increasing intermediate diffusivity (1, 8, 10, 12, and 14). Ellipses in (E) show the ( p,q)range of healthy tissue diffusivities (blue ¼ WM, yellow ¼ GM, green ¼ CSF). Neuro-Oncology 469 Jones et al.: Tumor classification by a novel DTI segmentation Table 1. Number and percent of voxels in each D-SEG segment at initialization and termination of the k-means algorithm Segment Number Initial Conditions Algorithm Termination (250 iterations) 2 21 23 2 21 24 Number of Total Voxels, % Number of Total Voxels, % p (mm s × 10 ) q (mm s × 10 ) Constituent Voxels Constituent Voxels 1 4116043 4.66 9732194 11.01 1.22 1.85 2 5720616 6.47 9102275 10.30 1.30 2.86 3 6380942 7.22 6440093 7.29 1.32 3.85 4 5870246 6.64 4277553 4.84 1.34 5.00 5 5260481 5.95 9109123 10.31 1.36 6.59 6 5478436 6.20 7865938 8.90 1.45 9.41 7 5628352 6.37 7741129 8.76 1.45 1.19 8 5720574 6.47 2124115 2.40 1.57 2.11 9 8059367 9.12 7155477 8.10 1.94 1.40 10 6044389 6.84 4351102 4.92 2.03 3.14 11 4416182 5.00 3426063 3.88 2.54 1.73 12 3567915 4.04 1956397 2.21 2.79 4.68 13 4651951 5.27 5434604 6.15 3.26 2.05 14 4844407 5.48 4128078 4.67 3.85 7.73 15 5662375 6.41 3329037 3.77 4.20 3.03 16 6929118 7.84 2178216 2.47 5.48 4.96 Median coordinates in ( p,q) space quantify diffusion characteristics for each segment. Fig. 2. D-SEG color mapping technique. Ranked T2-weighted (red channel), p (green channel), and q (blue channel) maps are shown to the left of D-SEG color maps for 2 axial slices of a healthy subject. D-SEG Tumor Spectra glioblastoma multiforme (GBM), cystic GBM, metastases, and meningioma. Group spectra and classification were not per- The volumetric proportion of each ( p,q) segment to the VOI was formed for the anaplastic astrocytoma case due to insufficient calculated for each case and averaged across tumor type group size (n ¼ 1). to generate D-SEG tumor spectra for low-grade glioma, 470 Jones et al.: Tumor classification by a novel DTI segmentation Table 2. Cross-validated diagnostic results (n ¼ 94), SVM analysis of D-SEG spectra Tumor Type LGG GBM cGBM MET MEN Total Sens. Spec. Accu. 95% CI BER A Confusion matrix—61 direction DTI LGG 10 0 0 0 1 11 90.9 97.5 96.1 (86.5 – 99.5) GBM 0 14 0 0 0 14 100 100 cGBM 0 0 6 0 0 6 100 100 MET 0 0 0 16 0 16 100 100 MEN 1 0 0 0 3 4 75 97.9 B Confusion matrix—12 direction DTI LGG 7 0 0 0 1 8 87.5 100 93.0 (80.9 – 98.5) GBM 0 17 0 0 0 17 100 96.2 cGBM 0 0 1 0 0 1 100 100 MET 0 1 0 9 0 10 90 97.2 MEN 0 0 0 1 6 7 85.7 97.2 C Confusion matrix—combined datasets LGG 19 0 0 0 0 19 100.0 97.3 94.7 (88.0 – 98.3) 6.9 GBM 0 30 1 0 0 31 96.8 98.4 cGBM 0 1 6 0 0 7 85.7 98.9 MET 2 0 0 24 0 26 92.3 98.5 MEN 0 0 0 1 10 11 90.9 100.0 Abbreviations: cGBM, cystic GBM; LGG, low-grade glioma; MEN, meningioma; MET, metastasis. Sens., sensitivity (%); Spec., specificity (%); Accu., accuracy (%); BER, balanced error rate (%). ( p,q) coordinates. The Voronoi plot (Fig. 1E) shows 3 radial lines Tumor Classification of segments through ( p,q) space with unique diffusion charac- The ability of D-SEG spectra to classify tumor type was tested teristics that include: tissue with mostly anisotropic diffusivity across all patients using SVMs. SVM predictions depend on (with q increasing from segment 1 to 6) but with lowest isotro- only a subset of the training data (ie, the support vectors). pic diffusivity, isotropic diffusivity (with p increasing from seg- The technique finds the hyperplane with the largest margin of ment 1 through 7, 9, 11, 13, 15, and 16), and intermediate difference between classes. We used the Gaussian radial diffusivity (with p and q increasing from segment 1 through 8 basis function kernel (s ¼ 1) to map feature vectors into a non- and 12 to 14). linear feature space where an optimal hyperplane was con- structed separating tumor classes. Tenfold cross-validation D-SEG Color Mapping was used to test classification accuracy and reproducibility. To test the integrity of combining tumor DTI from 2 different scan- D-SEG color mapping is shown in Fig. 2 for a healthy subject. The ners, separate SVM classifications of D-SEG spectra acquired for color mapping technique provides visually distinct colors based each acquisition protocol were performed (Table 2A and B). on the diffusion and T2-weighted properties of the tissue in each voxel. White matter regions with high anisotropic diffusion are colored blue. Gray matter regions with low anisotropic and Results isotropic diffusivities are yellow-brown with CSF colored pale yellow. Between-Scanner Reproducibility Mean and standard deviation for intraclass correlation coeffi- D-SEG Spectra in Healthy Subjects cients for GM (0.915+0.097) and WM (0.890+0.110) in Gray matter, WM, and CSF voxels occupy different regions of healthy volunteers showed good interscanner reproducibility ( p,q) space, as shown schematically by the ellipses in Fig. 1E, of p and q diffusion metrics for healthy tissue. and proportionately include different segment amounts result- ing in characteristic D-SEG tissue spectra (Fig. 4A). Gray matter D-SEG Algorithm predominantly includes segments 1, 2, 7, 8, and 9, representing The D-SEG algorithm was computationally fast and reached low isotropic and anisotropic diffusivities, whereas WM almost steady state by 50 iterations. Non-Gaussian characteristics exclusively includes segments 1 to 6, representing low isotropic were apparent in p and q histograms (Fig. 1Aand B) andin diffusion and increasing levels of anisotropic diffusion. CSF the histogram of ( p,q) space. The initial 16 segments assigned spaces include high isotropic diffusion characteristics (seg- to the ( p,q) distribution and the final segmentation after 250 ments 14, 15, and 16). Tissue partial volume effects will be pre- iterations are shown in Fig. 1D and E. Table 1 provides the initial sent in D-SEG segments because tissue class was not used to and final numbers of voxels in each segment and their median define the segmentation. Neuro-Oncology 471 Jones et al.: Tumor classification by a novel DTI segmentation Fig. 3. Individual patient images. From left to right: grade II glioma, glioblastoma with cystic component, cerebral metastasis, and meningioma examples. (A) Fluid attenuated inversion recovery images, (B) T1-weighted postcontrast images, (C) D-SEG color maps, and (d) tumor volumes of interest. All images are illustrated using the radiological convention. Tumor Volume of Interest Delineation Tumor D-SEG Spectra Examples of the VOI extraction technique are shown in Fig. 3 for Figure 4B – F illustrates average D-SEG spectra obtained within the low-grade glioma, glioblastoma, metastases, and meningio- VOIs for each tumor type. The low-grade glioma spectrum con- ma. Conventional fluid attenuated inversion recovery (row A) sisted mostly of segments 9, 11, and 13, representing a lower an- and postcontrast T1-weighted images (row B) indicate the isotropic and higher isotropic diffusion relative to healthy WM. tumor core, cystic, and edematous regions. D-SEG color maps High proportions of intermediate diffusivity segments 10 and show the isotropic and anisotropic diffusion characteristics of 12 potentially represent partial volume effects between tumor the tumor cases (row C) with the extracted VOIs (row D). and WM tissue and were located at the tumor boundary. The D-SEG color images show a visually apparent boundary be- glioblastoma spectrum contained segments of low anisotropic tween healthy and abnormal tissue (solid tumor, necrosis, and isotropic diffusivity (segments 7 and 9, likely corresponding cyst, and edema) that relates to differences in diffusion charac- to solid tumor)aswellassegmentswithhighisotropic andlow teristics located in the regions identified as abnormal in the anisotropic diffusivity (segments 12 and 13, likely corresponding conventional images. While lesion margins on conventional to necrotic regions). High proportions of segments 8, 10, and 11 MRI can be visually indistinct and rely on a subjective choice with greater isotropic diffusivities potentially represent edema re- of thresholding level, the colored segments obtained by gions. Cystic glioblastoma spectra shared such diffusion charac- D-SEG provide a more objective boundary for semiautomatic teristics but with high proportions of segment 15 corresponding lesion delineation. to the cystic region. The D-SEG spectrum of metastases contains 472 Jones et al.: Tumor classification by a novel DTI segmentation Fig. 4. D-SEG spectra. Average proportion of D-SEG segments within VOIs (standard error shown) for: (A) healthy tissue, (B) grade II glioma, (C) glioblastoma, (D) glioblastoma with cystic component, (E) cerebral metastasis, and (F) meningioma. segments 1, 7, 8, and 9 (low isotropic and low anisotropic diffu- balanced error rate of 6.9% after cross-validation (Table 2C). sivity), corresponding to the solid tumor component. Segments Sensitivity and specificity of tumor classification was .90% 10 and 12 likely represent perilesional edema with isotropic diffu- and 97%, respectively, for all tumor types except cystic glio- sivities greater than for glioblastoma. The D-SEG meningioma blastoma. Separate SVM analysis of tumor spectra from the dif- spectrum is markedly different from the other tumor types, ferent DTI acquisitions reveals comparable accuracies (96.1% with a large contribution from segments 1, 2, 3, and 4 (low iso- CI: 86.5% – 99.5% for 61-direction DTI vs 93.0% CI: 80.9% – tropic and increasing anisotropic diffusion), representing the solid 98.5% for 12-direction DTI; Table 2A and B). tumor component. In common with the metastases spectrum, segments 10 and 12 represent the edema region. Discussion We present D-SEG, a fast segmentation and visualization tech- Classification of Tumor Type nique that employs k-means clustering of ( p,q) space to provide SVM analysis of the D-SEG spectra classified tumor type with tissue segments with different isotropic and anisotropic diffu- high overall accuracy (95% CI: 88.0% – 98.3%) and low sion properties. D-SEG maps were colored according to ranked Neuro-Oncology 473 Jones et al.: Tumor classification by a novel DTI segmentation T2-weighted, p and q segment median values to provide a sim- contrast, D-SEG generates tissue type boundaries based on an ple visualization of diffusion characteristics throughout the en- objective clustering of the isotropic and anisotropic diffusivities tire brain that was then used to semiautomatically extract VOIs in (p,q) space. Such segmentation may reflect underlying differ- of abnormal tissue. Distinct D-SEG tumor spectra representing ences in tissue microstructure and potentially relevant patholog- the proportion of diffusion segments within the VOI were com- ical boundaries. However, partial volume effects may result in puted, and SVMs provided exceptionally high classification ac- D-SEG boundaries that do not accurately represent the precise curacy among brain tumor types and grades. difference between pathological and healthy tissue, and further Difficulties arise in multicenter studies incorporating MR dif- work is required to determine the histological ground truth of fusion metrics due to variability in scanner magnetic field, gra- D-SEG boundaries. dient strength, coil channels, and acquisition protocols. Brain tumors are characterized by their heterogeneity in Despite the use of two 1.5T MR scanners with different maxi- size, location, and extent of perilesional edema. Limitations mum gradient strengths and acquisition protocols, the inter- of previous brain tumor diffusion studies are twofold: (i) place- scanner reproducibility of p and q metrics was comparable to ment of ROIs significantly smaller than the lesion potentially 33,34 previous studies. This led to consistent D-SEG spectral pat- excludes relevant diffusion information; (ii) computation of terns in healthy tissue and tumor VOIs for data acquired from 2 average information over whole-lesion ROIs obscures hetero- MR systems. A separate SVM subanalysis of tumor VOI D-SEG geneous diffusion characteristics within the tumor. Spectral spectra generated from the 2 different DTI acquisitions re- comparison using D-SEG overcomes these limitations by pro- vealed comparable diagnostic accuracies, confirming that the viding a pattern of diffusivity across the entire region of abnor- datasets may be combined for the presented analysis. mal tissue. The D-SEG technique separates ( p,q) space into segments D-SEG spectra differ among tumor types in both their con- with distinct isotropic and anisotropic diffusion properties. stituent segment numbers and their proportional contribution Simultaneous application of D-SEG to all healthy and patient to the VOI. Spectra are consistent within the tumor type, con- data ensured that segments contained voxels with the same firmed by small standard errors for segments despite variability diffusion properties in each individual. This allowed meaningful in size, location, natural history, and, in the case of metastases, between-subject comparison of D-SEG spectra. Nevertheless, cellular origin of each lesion. Possible reasons for differences in further work is required to evaluate stability of the final D-SEG diffusion characteristics between tumor types include presence result due to perturbation of the initial algorithmic conditions of necrotic or cystic regions or volumetric proportion of tumor and for different numbers of tumor datasets. In this study, ini- and edema, solid tumor microstructure, and pathophysiology tial conditions were chosen that reflected the local density of of perilesional edema. ( p,q) space and consequently provided similar voxel numbers Malignant tumors are characterized by rapid growth and per segment. Alternative segmentation techniques could be neovascularity. When tumor rate of growth exceeds its blood applied, but an algorithmic investigation of optimality is beyond supply, cell death and regional necrosis result. The loss of cel- the scope of this study. Interestingly (p,q) space is characterized lular structure and boundaries to diffusion results in higher iso- by a non-Gaussian distribution that does not contain explicit tropic diffusion and is observed in glioblastoma D-SEG data clusters. Nevertheless, D-SEG provides a discrete mapping spectra. Tumor cysts may result from necrotic degeneration, of this space dependent on local voxel density generating an in- central hemorrhage, liquefaction, entrapment of CSF, and plas- tuitive separation of isotropic and anisotropic diffusion. ma fluid leaking from a disrupted blood – brain barrier. Cysts D-SEG provides reproducible segmentation of GM, WM, and have high isotropic diffusivity in the D-SEG spectra, reflecting CSF. Although D-SEG does not define exclusive segments for the fluid nature of these regions. Glioblastoma cysts exhibit each tissue type, it provides a spectrum of diffusion properties lower isotropic diffusivity than normal CSF spaces, potentially supporting previous findings of similar isotropic diffusivity in GM reflecting their proteinaceous constituents. and WM, and heterogeneous anisotropic diffusion in WM. The solid component of tumors consists of disorganized In this study, CSF spaces exhibited high magnitudes of isotropic pleomorphic, hypercellular cells with hyperchromatic nuclei, diffusion but also greater anisotropic diffusion than did GM. This lacking the organized structure of nascent neural tissue. In effect was caused by the use of q to quantify anisotropic diffu- common with previous studies, D-SEG spectra indicate that iso- sion, which, unlike FA, is not scaled by the overall magnitude of tropic diffusion within the solid tumor contributes to differenti- 17 41,44,45 diffusion within a voxel. ating among tumor types. In particular, differences in We prospectively recruited 94 patients with histopathologi- cell density of the solid tumor may be responsible for observed cally confirmed lesions occupying intracranial space over an spectral differences in the proportion of isotropic and anisotrop- 18-month period representing a cross section of brain tumors ic segments among tumor types. Our results confirm that iso- encountered in our neurosurgical practice. D-SEG analysis of tropic diffusion is smaller in glioblastoma than in low-grade 37,46,47 these data show that combined tumor and edema VOIs deter- glioma, agreeing with previous studies, and contrasts mined by D-SEG correspond visually with the extent of tumor on the high cellularity of glioblastoma with cellularity that is only standard MRI; however, their complex margins are indistinct on moderately increasedcomparedwithnormalbrain in low- conventional MRI or within p and q maps. A range of region grade glioma. D-SEG spectra confirm a lower isotropic diffusion drawing techniques have been used to determine diffusion char- within the solid component of metastases than glioblastoma, 7,36 acteristics of tumors, such as from manually drawn lesion agreeing with previous studies. This contrasts the densely 36 37 edges or from within ROIs placed in specified brain regions. packed and restricted diffusion within secondary tumors with These techniques are time-consuming and user dependent and the irregular cellular arrangement of microscopic necrosis in are subjective interpretations of the tumor boundary. In glioblastoma. Meningioma D-SEG spectra are markedly 474 Jones et al.: Tumor classification by a novel DTI segmentation different from the other tumor types. The solid tumor compo- these results and determine the additional utility of D-SEG in nent has the lowest isotropic diffusion and an anisotropic diffu- these and other clinical scenarios. sion component that likely reflects interdigitating cellular processes, tight intercellular junctions, and formation of fascic- ular and lobular tissue in association with whorls and psam- Funding moma bodies. This work was supported by Cancer Research UK, grant nos. C8807/ D-SEG spectra differ in segments containing perilesional A3870 and C1459/A13303; the EU, grant no. LSHC-CT-2004-503094 edema. Tumors with vasogenic edema, such as metastases (eTUMOUR); and T.L.J. acknowledges a Royal College of Surgeons of En- and meningioma, comprise a greater proportion of segments gland Research Fellowship. with higher isotropic diffusion than infiltrative cellular edema in glioblastoma. These findings agree with previous studies and reflect the differences in pathophysiology of the 36,46 Acknowledgments edema. Due to the ability of D-SEG to capture differences among tu- We thank Chris Clark, Dominick McIntyre, and Alan Wright for their help with the MR imaging. mors, it provides exceptional accuracy in tackling clinically rel- evant diagnostic scenarios, including differentiating glioma grade and discriminating between isolated metastases and glioblastoma. Although D-SEG accurately distinguishes metas- Conflict of interest statement. We have no conflicts of interest to tases and meningioma, this has limited clinical use except for declare. differentiating durally based metastatic deposits from benign meningioma. D-SEG delineated a clear boundary between tumor and References normal-appearing brain tissue for all except a small number 1. Darefsky AS, King JT, Dubrow R. Adult glioblastoma multiforme of meningioma cases. In these cases the meningioma dis- survival in the temozolomide era: a population-based analysis of played anisotropic diffusion properties similar to healthy WM. Surveillance, Epidemiology, and End Results registries. Cancer. 2012;118(8):2163 – 2172. Here it was necessary to manually draw lesion boundaries for the VOI guided by conventional postcontrast T1-weighted im- 2. Koay E, Sulman EP. Management of brain metastasis: past lessons, ages. A further limitation is that diffusion properties identified modern management, and future considerations. Curr Oncol Rep. by D-SEG are not exclusively representative of healthy or path- 2012;14(1):70 – 78. ological tissue types due to tissue partial volume effects, ef- 3. Stupp R, Mason WP, van den Bent MJ, et al. Radiotherapy plus fects of noise, and certain tumor tissue types sharing similar concomitant and adjuvant temozolomide for glioblastoma. N Engl J Med. 2005;352(10):987 – 996. diffusivities to healthy brain tissue. The use of D-SEG spectra overcomes these problems by providing proportions of seg- 4. Stummer W, Pichlmeier U, Meinel T, et al. Fluorescence-guided ments within VOIs. The 94 brain tumor cases acquired in this surgery with 5-aminolevulinic acid for resection of malignant study represent the largest reported cohort of tumor DTI glioma: a randomised controlled multicentre phase III trial. Lancet Oncol. 2006;7(5):392 – 401. data. Further evaluation of the role of D-SEG in diagnosis re- quires comparing the accuracy of this method with the accura- 5. McGirt MJ, Woodworth GF, Coon AL, et al. Independent predictors of morbidity after image-guided stereotactic brain biopsy: a risk cy of reporting of conventional imaging, as well as testing with assessment of 270 cases. J Neurosurg. 2005;102(5):897 – 901. other tumor types and grades and a range of different metas- tases, as the site of primary disease and thus cell lineage may 6. Kondziolka D, Lunsford LD, Martinez AJ. Unreliability of influence microstructure and thus tumor p-q diffusion contemporary neurodiagnostic imaging in evaluating suspected adult supratentorial (low-grade) astrocytoma. J Neurosurg. 1993; characteristics. 79(4):533 – 536. 7. Wang S, Kim S, Chawla S, et al. Differentiation between Conclusion glioblastomas and solitary brain metastases using diffusion tensor imaging. Neuroimage. 2009;44(3):653 – 660. Advances in imaging of newly identified lesions occupying in- tracranial space have not eliminated the requirement for histo- 8. Tagle P, Villanueva P, Torrealba GHI. Intracranial metastasis or meningioma? Surg Neurol. 2002;58(3 – 4):241 – 245. pathological “tissue” diagnosis in the majority of cases. With an increasing elderly population, improved survival in systemic 9. Basser PJ. Inferring microstructural features and the physiological malignancy, and increased detection of lesions (potentially in- state of tissues from diffusion-weighted images. NMR Biomed. cidentally) at an early stage, there exists a need for accurate 1995;8(7 – 8):333 – 344. and reproducible noninvasive diagnostic tools. In this study, 10. Basser PJ, Mattiello J, LeBihan D. Estimation of the effective we present a technique capable of classifying brain tumors self-diffusion tensor from the NMR spin echo. JMagnReson B. using biomarkers obtained from DTI with an accuracy of 1994;103(3):247 – 254. 94.7%. The potential roles for D-SEG include: (i) delineation of 11. Pierpaoli C, Basser PJ. Toward a quantitative assessment of lesion margins for optimal surgical resection or radiotherapy diffusion anisotropy. Magn Reson Med. 1996;36(6):893 – 906. treatment, (ii) serial volumetric analysis to monitor changes 12. Stadlbauer A, Ganslandt O, Buslei R, et al. Gliomas: in low-grade glioma over time or evaluate response to onco- histopathologic evaluation of changes in directionality and therapy, and (iii) regional assessment of biomarkers for diagno- magnitude of water diffusion at diffusion-tensor MR imaging. sis or surveillance. Further studies are required to replicate Radiology. 2006;240(3):803 – 810. Neuro-Oncology 475 Jones et al.: Tumor classification by a novel DTI segmentation 13. Beppu T, Inoue T, Shibata Y, et al. Fractional anisotropy value by 31. Murphy K. Machine Learning: a Probabilistic Perspective. diffusion tensor magnetic resonance imaging as a predictor of Cambridge, MA: MIT Press; 2012. cell density and proliferation activity of glioblastomas. Surg 32. Pfefferbaum A, Adalsteinsson E, Sullivan EV. Replicability of Neurol. 2005;63(1):56 – 61. diffusion tensor imaging measurements of fractional anisotropy 14. Lu S, Ahn D, Johnson G, et al. Peritumoral diffusion tensor imaging and trace in brain. J Magn Reson Imaging. 2003;18(4):427 – 433. of high-grade gliomas and metastatic brain tumors. AJNR Am J 33. Vollmar C, O’Muircheartaigh J, Barker GJ, et al. Identical, but not Neuroradiol. 2003;24(5):937 – 941. the same: intra-site and inter-site reproducibility of fractional 15. Provenzale JM, Mcgraw P, Mhatre P, et al. Peritumoral brain regions anisotropy measures on two 3.0T scanners. Neuroimage. 2010; in gliomas and meningiomas: investigation with isotropic imaging 51(4):1384 – 1394. and diffusion-tensor. Radiology. 2004;232(2):451 – 460. 34. Bonekamp D, Nagae LM, Degaonkar M, et al. Diffusion 16. Tsuchiya K, Fujikawa A, Nakajima M, et al. Differentiation between tensor imaging in children and adolescents: reproducibility, solitary brain metastasis and high-grade glioma by diffusion hemispheric, and age-related differences. Neuroimage. 2007; tensor imaging. Br J Radiol. 2005;78(930):533 – 537. 34(2):733 – 742. 17. Pen ˜ a A, Green HAL, Carpenter TA, et al. Enhanced visualization and 35. Pierpaoli C, Jezzard P, Basser PJ, et al. Diffusion tensor MR imaging quantification of magnetic resonance diffusion tensor imaging of the human brain. Radiology. 1996;201(3):637 – 648. using the p:q tensor decomposition. Br J Radiol. 2006;79(938): 36. ByrnesTJD,Barrick TR,BellBA, et al. Diffusion tensor imaging 101 – 109. discriminates between glioblastoma and cerebral metastases in 18. Gordillo N, Montseny E, Sobrevilla P. State of the art survey on MRI vivo. NMR Biomed. 2011;24(1):54 – 60. brain tumor segmentation. Magn Reson Imaging. 2013;31(8): 37. Price SJ, Pen ˜ a A, Burnet NG, et al. Tissue signature characterisation 1426 – 1438. of diffusion tensor abnormalities in cerebral gliomas. Eur Radiol. 19. Fletcher-Heath LM, Hall LO, Goldgof DB, et al. Automatic 2004;14(10):1909 – 1917. segmentation of non-enhancing brain tumors in magnetic 38. Kaus MR, Warfield SK, Nabavi A, et al. Automated segmentation of resonance images. Artif Intell Med. 2001;21(1 – 3):43 – 63. MR images of brain tumors. Radiology. 2001;218(2):586 – 591. 20. Clark MC, Hall LO, Goldgof DB, et al. Automatic tumor 39. Kleihues P, Louis DN, Scheithauer BW, et al. The WHO classification segmentation using knowledge-based techniques. IEEE Trans of tumors of the nervous system. J Neuropathol Exp Neurol. 2002; Med Imaging. 1998;17(2):187 – 201. 61(3):215 – 225. 21. Liu J, Udupa JK, Odhner D, et al. A system for brain tumor volume 40. Raza SM, Lang FF, Aggarwal BB, et al. Necrosis and glioblastoma: a estimation via MR imaging and fuzzy connectedness. Comput Med friend or a foe? A review and a hypothesis. Neurosurgery.2002; Imaging Graph. 2005;29(1):21 – 34. 51(1):2 – 12. 22. Clarke LP, Velthuizen RP, Phuphanich S, et al. MRI: stability of three 41. Krabbe K, Gideon P, Wagn P, et al. MR diffusion imaging of human supervised segmentation techniques. Magn Reson. 1993; intracranial tumours. Neuroradiology. 1997;39(7):483 – 489. 11(0730 – 725X (Print)):95 – 106. 42. Poisson M, Philippon J, van Effenterre R, et al. Cerebral pseudocysts 23. Vijayakumar C, Damayanti G, Pant R, et al. Segmentation and following chemotherapy of glioblastomas. Acta Neurochir (Wien). grading of brain tumors on apparent diffusion coefficient 1977;39(3 – 4):143 – 149. images using self-organizing maps. Comput Med Imaging Graph. 2007;31(7):473 – 484. 43. Louis DN, Ohgaki H, Wiestler OD, et al. The 2007 WHO classification of tumours of the central nervous system. Acta 24. Simon D, Fritzsche KH, Thieke C, et al. Diffusion-weighted Neuropathol. 2007;114(2):97 – 109. imaging-based probabilistic segmentation of high- and low- proliferative areas in high-grade gliomas. Cancer Imaging. 2012; 44. Bulakbasi N, Kocaoglu M, Ors F, et al. Combination of single-voxel 12:89 – 99. proton MR spectroscopy and apparent diffusion coefficient 25. MacQueen JB. Kmeans Some methods for classification and calculation in the evaluation of common brain tumors. AJNR Am analysis of multivariate observations. 5th Berkeley Symp Math J Neuroradiol. 2003;24(2):225 – 233. Stat Probab 1967. 1967;1:281 – 297. 45. Kono K, Inoue Y, Nakayama K, et al. The role of diffusion-weighted 26. Nanetti L, Cerliani L, Gazzola V, et al. Group analyses of imaging in patients with brain tumors. AJNR Am J Neuroradiol. connectivity-based cortical parcellation using repeated k-means 2001;22(6):1081 – 1088. clustering. Neuroimage. 2009;47(4):1666 – 1677. 46. Lu S, Ahn D, Johnson G, et al. Diffusion-tensor MR imaging 27. Ashburner J, Friston KJ. Unified segmentation. Neuroimage. 2005; of intracranial neoplasia and associated peritumoral edema: 26(3):839 – 851. introduction of the tumor infiltration index. Radiology. 2004; 232(1):221 – 228. 28. Aurenhammer F. Voronoi diagrams. ACM Comput Surv. 1991;23(3): 345 – 405. 47. Inoue T, Ogasawara K, Beppu T, et al. Diffusion tensor imaging for 29. Rorden C, Brett M. Stereotaxic display of brain lesions. Behav preoperative evaluation of tumor grade in gliomas. Clin Neurol Neurol. 2000;12(4):191 – 200. Neurosurg. 2005;107(3):174 – 180. 48. Lu S, Ahn D, Johnson G, Cha S. Peritumoral diffusion tensor 30. Cristianini N, Shawe-Taylor J. An Introduction to Support Vector Machines and Other Kernel-based Learning Methods. Cambridge: imaging of high-grade gliomas and metastatic brain tumours. Cambridge University Press; 2000. AJNR Am J Neuroradiol. 2003;24(5):937 – 941.

Journal

Neuro-Oncology – Pubmed Central

Published: Aug 13, 2014

There are no references for this article.