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

Learn More →

A multi-sequence and habitat-based MRI radiomics signature for preoperative prediction of MGMT promoter methylation in astrocytomas with prognostic implication

A multi-sequence and habitat-based MRI radiomics signature for preoperative prediction of MGMT... Objectives Oxygen 6-methylguanine-DNA methyltransferase (MGMT) promoter methylation is a significant prognostic biomarker in astrocytomas, especially for temozolomide (TMZ) chemotherapy. This study aimed to preoperatively predict MGMT methylation status based on magnetic resonance imaging (MRI) radiomics and validate its value for evaluation of TMZ chemotherapy effect. Methods We retrospectively reviewed a cohort of 105 patients with grade II-IV astrocytomas. Radiomic features were extracted from the tumour and peritumoral oedema habitats on contrast-enhanced T1-weighted images, T2-weighted fluid-attenuated inver- sion recovery images and apparent diffusion coefficient (ADC) maps. The following radiomics analysis was structured in three phases: feature reduction, signature construction and discrimination statistics. A fusion radiomics signature was finally developed using logistic regression modelling. Predictive performance was compared between the radiomics signature, previously reported clinical factors and ADC parameters. Validation was additionally performed on a time-independent cohort (n = 31). The prognostic value of the signature on overall survival for TMZ chemotherapy was explored using Kaplan Meier estimation. Results The fusion radiomics signature exhibited supreme power for predicting MGMT promoter methylation, with area under the curve values of 0.925 in the training cohort and 0.902 in the validation cohort. Performance of the radiomics signature surpassed that of clinical factors and ADC parameters. Moreover, the radiomics approach successfully divided patients into high- risk and low-risk groups for overall survival after TMZ chemotherapy (p =0.03). Conclusions The proposed radiomics signature accurately predicted MGMT promoter methylation in patients with astrocytomas, and achieved survival stratification for TMZ chemotherapy, thus providing a preoperative basis for individualised treatment planning. Key Points � Radiomics using magnetic resonance imaging can preoperatively perform satisfactory prediction of MGMT methylation in grade II-IV astrocytomas. � Habitat-based radiomics can improve efficacy in predicting MGMT methylation status. � Multi-sequence radiomics signature has the power to evaluate TMZ chemotherapy effect. . . . . Keywords Astrocytoma Methylation Prognosis Diagnostic imaging ROC curve Jingwei Wei and Guoqiang Yang contributed equally to this work. Electronic supplementary material The online version of this article (https://doi.org/10.1007/s00330-018-5575-z) contains supplementary material, which is available to authorized users. * Hui Zhang Beijing Key Laboratory of Molecular Imaging, Beijing 100190, [email protected] China * Jie Tian University of Chinese Academy of Sciences, Beijing 100049, China [email protected] Department of Radiology, First Clinical Medical College, Shanxi Medical University, Taiyuan 030001, Shanxi Province, China Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China 878 Eur Radiol (2019) 29:877–888 Abbreviations feature data to targeted clinical outcomes [13, 14]. ADC Apparent diffusion coefficient Regions of interest (ROIs) are delineated on the tumour AUC Area under the curve and sub-regions of the lesion known as habitats. Thus, DCA Decision curve analysis radiomics not only offers holistic imaging information, GBM Glioblastoma but also explores the microenvironment of the tumour IDH Isocitrate dehydrogenase by analysing explicit sub-regional features that describe MGMT Oxygen 6-methylguanine- genetic heterogeneity more granularly [15]. For gliomas, DNA methyltransferase radiomics studies on molecular subtype prediction have MRI Magnetic resonance imaging demonstrated sufficient predictive accuracy for isocitrate OS Overall survival dehydrogenase (IDH) and 1p19q codeletion [16, 17]. ROC Receiver operating characteristic Habitat-based radiomics have also been shown to have ROI Region of interest the capability to identify survival stratification in glioblas- T1-WI T1-weighted imaging tomas (GBMs) [18]. These studies suggest that habitat- T2-FLAIR T2-weighted fluid-attenuated inversion recovery based radiomics may be similarly useful for the preoper- images ative prediction of MGMT promoter methylation in pa- TMZ Temozolomide tients with astrocytomas. In this study, we investigated the utility of a multi-sequence and multi-habitat MR radiomics signature as a preoperative Introduction and non-invasive biomarker of MGMT methylation predic- tion in patients with grade II–IV astrocytomas, and discuss Astrocytoma is the most common type of glioma, and carries a the prognostic implications for survival stratification on poor prognosis [1, 2]. The average survival time ranges from 17 TMZ chemotherapy response. weeks to 3 years [2, 3]. Fortunately, a subgroup of grade II-IV astrocytoma patients with oxygen 6-methylguanine-DNA methyltransferase (MGMT) promoter methylation show good responses to temozolomide (TMZ) chemotherapy and im- Materials and methods proved survival after treatment, which underscored the role of MGMT as a judicious molecular biomarker with a prognostic Patients implication [4–7]. Preoperative identification of MGMT pro- moter methylation would be of great clinical significance in This retrospective study was approved by the institutional selecting potential patients benefiting from TMZ chemothera- review board. We reviewed 105 patients who were newly py, thus assisting with planning the therapy regime. However, diagnosed with grade II–IV astrocytoma from October 2011 the standard approach for MGMT status confirmation requires to March 2017. The inclusion and exclusion criteria are given a large tissue sample that is usually obtained through surgery. in the Online Supplemental Appendix E1. The patient recruit- For patients with unresectable tumours, biopsy runs the risk of ment pathway is shown in Online Supplemental Fig. 1. neurological deficits and can accordingly jeopardise the quality A total of 105 patients met the study criteria and were divid- of life of the patient [8, 9]. Thus, there is an urgent need in ed into a training dataset (31 December 2011 to 2 November clinical practice for preoperative and non-invasive prediction 2015, n = 74) and a time-independent validation dataset (16 of MGMT promoter methylation in grade II-IV astrocytomas. November 2015 to 21 March 2017, n = 31). Demographic Magnetic resonance imaging (MRI), as a powerful non- and clinical data were acquired from medical reports, including invasive diagnostic imaging tool for astrocytoma manage- sex, age, grade and radiological characteristics (Table 1). ment [10], opens up the possibility of having this preop- For evaluation of the TMZ chemotherapy effect, the erative prediction. Previous studies have verified that cer- inclusion criteria were further expanded in the 105 pa- tain radiological characteristics on MR images such as tients: (1) patients receiving adjuvant treatment following tumour necrosis, enhancement patterns and tumour loca- surgery consisting of either (a) concomitant radiation plus tion are associated with MGMT promoter methylation TMZ followed by adjuvant TMZ or (b) TMZ; (2) overall [11, 12]. However, subjective judgements by radiologists survival (OS) was categorised according to death or sur- are vulnerable to inter-observer variability and generally vival exceeding the median survival of patients with each lack power and accuracy. In contrast, a newly emerging tumour grade (605 days for grade II, 398 days for grade technology termed radiomics could resolve this problem III and 169 days for grade IV). Twenty-two patients met by quantitative imaging analysis. Radiomics converts the inclusion criteria for the TMZ survival analysis. encrypted medical images into usable data by extracting Demographic and clinical data of the 22 patients are high-throughput imaging features and relating imaging shown in Table 2. Eur Radiol (2019) 29:877–888 879 Table 1 Clinical characteristics in the training and validation cohorts Characteristic Training cohort Validation cohort p (Inter) N=74 N=31 MGMT (+) MGMT (-) p (Intra) MGMT (+) MGMT (-) p (Intra) Gender 0.674 0.139 0.902 Male 32 (43.2) 10 (13.5) 8 (25.8) 10 (32.3) Female 23 (31.1) 9 (12.2) 10 (32.3) 3 (9.7) Age, y 0.225 0.701 0.195 ≤52 32 (43.2) 8 (10.8) 13 (41.9) 8 (25.8) >52 23 (31.1) 11 (14.9) 5 (16.1) 5 (16.1) Grade 0.258 0.509 0.905 II 24 (32.4) 5 (6.8) 7 (22.6) 4 (12.9) III 22 (29.7) 8 (10.8) 9 (29.0) 5 (16.1) IV 9 (12.2) 6 (8.1) 2 (6.5) 4 (12.9) Tumour size 0.508 1.000 0.073 ≤6cm 27 (36.5) 11 (14.9) 6 (19.4) 4 (12.9) >6cm 28 (37.8) 8 (10.8) 12 (38.7) 9 (29.0) Tumour centre location 0.350 1.000 0.980 Left hemisphere 25 (33.8) 11 (14.9) 9 (29.0) 6 (19.4) Right hemisphere 30 (40.5) 8 (10.8) 9 (29.0) 7 (22.6) Frontal lobe 0.674 0.013 0.902 Yes 32 (43.2) 10 (13.5) 14 (45.2) 4 (12.9) No 23 (31.1) 9 (12.2) 4 (12.9) 9 (29.0) Occipital lobe 0.702 0.497 0.742 Yes 5 (6.8) 3 (4.1) 2 (6.5) 0 (0.0) No 50 (67.6) 16 (21.6) 16 (51.6) 13 (41.9) Parietal lobe 0.264 0.023 0.116 Yes 13 (17.6) 7 (9.5) 0 (0.0) 4 (12.9) No 42 (56.8) 12 (16.2) 18 (58.1) 9 (29.0) Temporal lobe 0.706 0.710 0.728 Yes 35 (47.3) 6 (8.1) 6 (19.4) 6 (19.4) No 20 (27.0) 13 (17.6) 12 (38.7) 7 (22.6) Insular lobe 0.140 0.625 0.817 Yes 9 (12.2) 0 (0.0) 2 (6.5) 3 (9.7) No 46 (62.2) 19 (25.7) 16 (51.6) 10 (32.3) Involving cortex matter 0.635 1.000 0.742 Yes 48 (64.9) 18 (24.3) 17 (54.8) 12 (38.7) No 7 (9.5) 1 (1.4) 1 (3.2) 1 (3.2) Involving deep white matter 0.910 0.497 0.238 Yes 46 (62.2) 15 (20.3) 16 (51.6) 13 (41.9) No 9 (12.2) 4 (5.4) 2 (6.5) 0 (0.0) Involving pial matter 1.000 1.000 1.000 Yes 48 (64.9) 16 (21.6) 16 (51.6) 11 (35.5) No 7 (9.5) 3 (4.1) 2 (6.5) 2 (6.5) Involving ependymal membrane 0.555 1.000 0.600 Yes 19 (25.7) 8 (10.8) 8 (25.8) 5 (16.1) No 36 (48.6) 11 (14.9) 10 (32.3) 8 (25.8) Tumour cross midline 0.949 1.000 0.735 Yes 11 (14.9) 3 (4.1) 3 (9.7) 2 (6.5) No 44 (59.5) 16 (21.6) 15 (48.4) 11 (35.5) 880 Eur Radiol (2019) 29:877–888 Table 1 (continued) Characteristic Training cohort Validation cohort p (Inter) N=74 N=31 MGMT (+) MGMT (-) p (Intra) MGMT (+) MGMT (-) p (Intra) Oedema cross midline 1.000 1.000 0.416 Yes 12 (16.2) 4 (5.4) 5 (16.1) 4 (12.9) No 43 (58.1) 15 (20.3) 13 (41.9) 9 (29.0) Border 0.171 0.099 0.455 Well-defined 14 (55.4) 8 (10.8) 2 (6.5) 5 (16.1) Ill-defined 41 (18.9) 11 (14.9) 16 (51.6) 8 (25.8) Haemorrhage 0.254 1.000 0.134 Yes 11 (14.9) 1 (1.4) 5 (16.1) 4 (12.9) No 44 (59.5) 18 (24.3) 13 (41.9) 9 (29.0) Cystic and necrosis 0.751 1 0.053 No 15 (20.3) 5 (6.8) 7 (22.6) 5 (16.1) ≤25% 21 (28.4) 5 (6.8) 4 (12.9) 3 (9.7) 25%-50% 11 (14.9) 5 (6.8) 1 (3.2) 1 (3.2) ≥50% 8 (10.8) 4 (5.4) 6 (19.4) 4 (12.9) Oedema degree 0.011 0.275 0.563 ≤1.6 33 (44.6) 5 (6.8) 10 (32.3) 4 (12.9) >1.6 22 (29.7) 14 (18.9) 8 (25.8) 9 (29.0) * * Enhancement style 0.010 0.034 0369 No 15 (20.3) 3 (4.1) 4 (19.4) 0 (0.0) Ring enhancement 20 (27.0) 15 (20.3) 8 (25.8) 8 (25.8) Nodular enhancement 11 (14.9) 0 (0.0) 6 (19.4) 2 (6.5) Irregular reinforcement 9 (12.2) 1 (1.4) 0 (0.0) 3 (9.7) Enhancement degree 0.287 0.211 0.401 No 15 (20.3) 3 (4.1) 4 (12.9) 0 (0.0) Slight 5 (6.8) 0 (0.0) 2 (6.5) 1 (3.2) Obvious 35 (47.3) 16 (21.6) 12 (38.7) 12 (38.7) Signal characteristics 1.000 0.497 0.904 Homogeneous 5 (6.8) 2 (2.7) 2 (6.5) 0 (0.0) Heterogeneous 50 (67.6) 17 (23.0) 16 (51.6) 13 (41.9) MGMT(+) patients with oxygen 6-methylguanine-DNA methyltransferase (MGMT) methylation, MGMT(-) patients without MGMT methylation, p (Intra) the result of uni-variable analyses between methylated and unmethylated groups, p(Inter) significant difference between training and validation cohorts Unless otherwise specified, data are numbers of patients, with percentages in parentheses p < 0.05 MGMT methylation testing DNA modification kit (EpiGentek, Farmingdale, NY, USA). The PCR amplification and conditions are given in Online The methylation status of the MGMT promoter was assessed Supplemental Appendix E2. using pyrosequencing analysis as described elsewhere [19]. Briefly, each tumour specimen was histologically investigated by macro-dissection to guarantee a tumour cell content of at MRI acquisition least 80%. DNAwas extracted using the Simlex OUP ® FFPE DNA extraction kit (TIB, China) and quantified by spectro- Preoperative MRI was performed with a 3.0-T scanner (Signa photometry using a NanoDrop 2000 (Thermo Fisher HDxt, GE Healthcare, USA) using an 8-channel array coil. Scientific, Loughborough, UK). Bisulphite modification of The acquisition protocols for CE-T1-WI, T2-FLAIR and the extracted DNA was performed using the BisulFlash™ DWI are in given in Online Supplemental Appendix E3. Eur Radiol (2019) 29:877–888 881 Table 2 Clinical characteristics and MGMT predicted outcome for Process of radiomics analysis patients with TMZ chemotherapy The radiomics analysis was structured into four parts: ROI Characteristic Patients N=22 segmentation, feature extraction, feature selection and model construction (Fig. 1). In brief, we performed a manual ROI Gender with overlapped area by two radiologists (10 and 15 years of Male 13 (59.1) experience, respectively) on tumour and oedema habitats from Female 9 (40.9) T1-WI, T2-FLAIR and ADC maps, and the two radiologists Age, y were blinded to the final diagnosis and the MGMT methyla- ≤49 12 (54.5) tion status. Examples of six typical segmentation cases ac- >49 10 (45.5) cording to grade and MGMT methylation status are shown Grade in Fig. 2. The median ROI of each habitat and sequence is II 6 (27.3) shown in Online Supplemental Table 3. A set of 3,051 III 11 (50.0) imaging features was extracted including textural and IV 5 (22.7) non-textural features. Feature stability and reproducibil- MGMT ity was estimated by Intra-class correlation coefficients + 17 (77.2) and concordance correlation coefficients. Further feature – 5 (22.8) selection was performed based on minimum redundancy Predicted MGMT and maximum relevance algorithm. Single radiomics MGMT(+) 14 (82.3) signatures from each sequence and habitat, and a fusion MGMT(-) 3 (17.7) radiomics signature were finally constructed using logis- tic regression modelling. A detailed description is given MGMT(+) patients with oxygen 6-methylguanine-DNA methyltransfer- in Online Supplemental Appendix E4. ase (MGMT) methylation, MGMT(-) patients without MGMT methylation Clinical and radiological factors for MGMT prediction Unless otherwise specified, data are numbers of patients, with percent- ages in parentheses The average age was 49 years, thus we divided patients into an age ≤ 49 Clinical factor analysis and ADC parameter calculation Auni- years group and an age > 49 years group variate analysis was initially applied to select useful clinical factors (p < 0.1). Then, a forward selection (likelihood ratio) multi-variable analysis was performed to select clinical Demographic and clinical characteristic analysis factors with p < 0.05. Additionally, we introduced ADC parameters (mean tumour ADC values and mean Differences between the training and validation cohorts and peritumoral oedema ADC values) associated with between the intra- MGMT methylated and unmethylated MGMT promoter methylation as reported in the litera- groups in terms of demographic and clinical factors were ture [22]. Tumour and oedema ADC values were addi- assessed with Pearson’s chi-square tests or Fisher’s exact tests tionally integrated as a clinical model by logistic regres- for categorical variables and Student’s t-tests or Mann- sion based on the training cohort. Whitney U tests for continuous variables. Combined model with radiomics signature, clinical factor, and ADC values To achieve a holistic information-gathered Sample size and power calculation network, we generated a comprehensive model including the fusion radiomics signature, the selected clinical factor (oede- According to the thumb rule, the sample size needed to cover ma degree), and two ADC values (the tumour and oedema 10–15 observations per predictor variable to yield a stable areas). Considering a correlation between the radiomics sig- estimate [20]. In our study, the maximum number of included nature and other factors as well as the model complexity, we features for radiomics signature construction was 5 (T2- adopted the Akaike information criterion (AIC) to select op- FLAIR sequence on tumour area). Thus, the training dataset timal incorporated factors and used logistic regression model- needed to include at least 50 patients. For the validation ling to perform model construction. dataset power calculation, a sample of > 11 patients was re- quired to provide 80% power and a type I error rate of 5% Performance evaluation [21]. Our study cohort included 105 patients with 74 in the training dataset and 31 in the validation dataset, which met the Receiver operating characteristic (ROC) curves were plotted sample size requirement. and area under the curve (AUC), specificity and sensitivity 882 Eur Radiol (2019) 29:877–888 Fig. 1 Radiomics workflow. The radiomics process included four parts: and reproducibility analysis before using maximum relevance and region of interest (ROI) segmentation on each habitat and sequence, minimum redundancy algorithm. The radiomics signature was feature extraction, feature selection and model construction. ROI was generated by logistic regression with Bayesian information criteria as delineated on both the tumour and the peritumoral habitats on contrast- the stopping rule. Further performance evaluation was explored enhanced T1-weighted images and T2-FLAIR images. On each ROI, a including receiver operating characteristics, decision curve analysis and set of 3,051 features were extracted. We performed fundamental stability survival stratification were calculated for the fusion radiomics signature, selected version 3.4.1 (www.R-project.org). The threshold for clinical factor and ADC parameter. We further performed statistical significance was a two-sided p <0.05. stratification analysis for the fusion radiomics signature by grouping the cohorts according to age, gender and grade. We chose decision curve analysis (DCA) to estimate the clinical Results usefulness of the developed fusion radiomics signature and used the Delong test to explore whether the fusion radiomics Patient demographic data, clinical characteristics signature performed better than the traditional clinical model and molecular subtypes and ADC parameter. The baseline characteristics of the patients are shown in Table 1. There were no differences between the training and Prognostic value analysis validation cohorts in terms of demographic or clinical charac- teristics (p =0.053–1.000). A Kaplan-Meier curve was plotted based on the fusion In total, we included 73 (69.5%) patients with MGMT radiomics signature in order to stratify the OS in patients treat- promoter methylation and 32 (30.5%) patients without ed with adjuvant TMZ chemotherapy. The log-rank test was MGMT promoter methylation. No significant difference was used to determine whether there were statistical differences shown for the MGMT methylation status distribution in the between the two survival groups. training and validation cohorts (p =0.156). Statistical analysis Feature stability and reproducibility estimation We performed the statistical analysis with PASW Statistics, The statistical results of feature numbers after stability and repro- version 18.0 (SPSS Inc., Chicago, IL, USA) and R software, ducibility analysis are shown in Online Supplemental Fig. 2. Eur Radiol (2019) 29:877–888 883 Fig. 2 Tumour and oedema area segmentations are shown by the red and green lines, respectively. Oedema degree was obvious in the MGMT unmethylated group compared to the MGMT methylated group, especially for higher-grade astrocytomas (III and IV) Features extracted from tumour habitats exhibited statistically are shown in Online Supplemental Table 1 and Appendix E5, better performance than those extracted from peritumoral oede- respectively. A detailed explanation of each selected feature is ma habitats (reproducibility, p < 0.001; stability, p < 0.001). shown in Online Supplemental Table 2. Single radiomics sig- natures from T1-WI-tumour, T1-WI-oedema, T2-FLAIR-tu- Single radiomics signatures formula and evaluation mour, and T2-FLAIR-oedema were verified as eligible radiomics signatures with AUCs > 0.7 in both the training and validation cohorts (Table 3). These four single radiomics The selected features and integration formulas for single radiomics signatures derived from each sequence and habitat signatures all showed significant differences (p < 0.05) in the Table 3 Diagnostic performance of single radiomics signatures, fusion radiomics signature, clinical factors and ADC values Models Training cohort Validation cohort N=74 N=31 Sensitivity Specificity Accuracy AUC (95% CI) Sensitivity Specificity Accuracy AUC (95% CI) Tumour T1 0.789 0.691 0.716 0.706 (0.567–0.845) 0.615 0.667 0.645 0.739 (0.558–0.921) Tumour T2 0.947 0.782 0.824 0.916 (0.852–0.979) 0.538 0.889 0.742 0.701 (0.494–0.908) Tumour ADC 0.895 0.655 0.716 0.815 (0.714–0.917) 0.692 0.389 0.516 0.590 (0.381–0.799) Oedema T1 0.632 0.782 0.743 0.738 (0.601–0.875) 0.615 0.667 0.645 0.709 (0.519–0.900) Oedema T2 0.684 0.727 0.716 0.778 (0.654–0.902) 0.615 0.778 0.710 0.752 (0.567–0.937) Oedema ADC 0.632 0.782 0.743 0.678 (0.534–0.822) 0.615 0.889 0.774 0.816 (0.667–0.965) Fusion radiomics 0.872 0.842 0.865 0.925 (0.861–0.989) 0.944 0.539 0.774 0.902 (0.785–1.000) Clinical factors 0.737 0.600 0.6351 0.668 (0.548–0.789) 0.692 0.556 0.613 0.624 (0.448–0.800) ADC values 0.632 0.691 0.676 0.649 (0.511–0.787) 0.615 0.722 0.677 0.603 (0.382–0.823) 95% CI 95% confidence interval, AUC area under curve, T1 contrast-enhanced T1-weighted sequence, T2 T2-FLAIR sequence 884 Eur Radiol (2019) 29:877–888 MGMT methylated and unmethylated groups in both the (accuracy, sensitivity and specificity) of the fusion radiomics training and the validation cohorts. Boxplots describing the signature are shown in Table 3. distribution of the four radiomics signatures in the MGMT The fusion radiomics signature also had an outstanding per- methylated and unmethylated groups are shown in Online formance in the stratification analysis, considering age, gender Supplemental Fig. 3. However, neither ADC-tumour nor and grade (Table 4). It is noteworthy that the proposed ADC-oedema performed with satisfactory results for radiomics signature does not require a priori knowledge of MGMT identification. ADC-tumour did not perform well in grading information because it behaved well for distinguishing the validation cohort (AUC = 0.590) while ADC-oedema did MGMT methylation status not only in a grade II-IV cohort, but not perform well in the training cohort (AUC = 0.678). also in grade II, III and IV astrocytomas, separately. Detailed predictive indicators (AUC, accuracy, sensitivity and specificity) of each single radiomics signature are shown Clinical model and ADC parameter evaluation in Table 3. Only oedema degree was significantly different between the MGMT methylated and unmethylated groups in the Fusion radiomics signature formula and evaluation training cohort (p = 0.0015). The AUCs of this clinical factor were 0.668 and 0.624 in the training and valida- The fusion radiomics signature combining the four single tion cohorts, respectively. The ADC parameter achieved radiomics signatures was constructed with a Rad-score calcu- AUC values of 0.649 and 0.603 in the training and lated as follows: validation cohorts, respectively. Detailed predictive indi- cators (sensitivity and specificity) of the clinical model Rad‐scoreðÞ fusion¼ −6:785−1:026*Signature Oedema−T1 and ADC parameter are shown in Table 3. þ 3:950*Signature Oedema−T2 þ 2:907*Signature Performance comparison Tumour −T1 þ 5:427 Singature Tumour−T2 The fusion radiomics signature achieved the highest AUC The optimum cut-off value of the fusion radiomics was among the three models. The Delong test showed a significant 1.077 as per the Youden index. Patients were divided into difference between the fusion radiomics signature and the predicted MGMT methylated (Rad-score ≥ 1.077) and clinical model (p = 0.008 and 0.011), and between the fusion unmethylated groups (Rad-score < 1.077) based on fusion radiomics signature and ADC parameter (p =0.003 and Rad-scores. 0.027) in the training and validation cohorts, respectively. Barplots depicting the classification performance of the fusion signature in the training and validation cohorts are Combined model construction and evaluation shown in Fig. 3. The fusion radiomics signature achieved optimal AUC values of 0.925 and 0.902 in the training and During the combined model construction process, the AIC validation cohorts, respectively. Detailed predictive indicators value was minimum when only taking the fusion radiomics Fig. 3 Barplots depicting the classification performance of the fusion radiomics signature. The red bar with a prediction value > 0 indicates that the signature successfully classifies the MGMT methylation patients; the red bar with a prediction value < 0 indicates that the signature fails to classify the MGMT methylation patients. For the green bar, the contrary applies Eur Radiol (2019) 29:877–888 885 signature into account. The AIC values were 48.79, 49.79, 51.25 and 53.11 when subsequently adding oedema degree, ADC of tumour area and ADC of oedema area to the fusion radiomics signature. Notwithstanding, we calculated the AUC of the combined model integrating all factors and the result was concordant with the AIC value; the combined model achieved an AUC of 0.921 in the training cohort and 0.868 in the validation cohort, which were slightly lower than those for the fusion radiomics signature alone. Prognostic value of the fusion radiomics signature The fusion radiomics signature successfully divided patients treated with adjuvant TMZ chemotherapy into high-risk and low-risk OS groups with p = 0.03 (Fig. 4a). Moreover, the DCA showed that the fusion radiomics signature performed with higher net benefit (net benefit = 0.441) compared to simple stratification assuming that no patients had MGMT methylation or all patients had MGMT methylation (Fig. 4b). Discussion In this study, a preoperative and low-cost radiomics analysis was used to integrate imaging features from tumour and peritumoral oedema habitats on CE-T1-WI and T2-FLAIR images to predict MGMT promoter methylation in patients with grade II-IVastrocytoma. Moreover, we verified the prog- nostic value of the fusion radiomics signature for patients who underwent resection followed by adjuvant TMZ chemotherapy. Compared to clinical and conventional radiological factors [23, 24], our proposed radiomics signature exhibited excellent prediction performance. Additionally, the fusion radiomics signature integrating synergistic information outperformed each single radiomics signature from a simple habitat or se- quence. Potential reasons for this observation are as follows: first, the fusion signature included more comprehensive infor- mation reflecting granular textural differences in the microen- vironments and took in important archetypal imaging charac- teristics associated with MGMT methylation. Previous litera- ture supports the value of the multi-habitat radiomics for predicting survival in patients with GBM [25–27]. Second, most of the effective features extracted in our study were tex- tural features from Gabor transformation images, which con- ducted noise removal and filtration. Thus, these transformed images more effectively captured key tumour heterogeneity [25]. These findings agree with the radiomics hypothesis that gene phenotypic information of the tumour is reflected in ra- diological images [28, 29]. In previous work, Xi et al. showed that a radiomics signa- ture derived from T1-WI, T2-WI and enhanced T1-WI was a potential imaging marker for the prediction of MGMT Table 4 Stratification analysis of the fusion radiomics signature on training and validation cohorts Subgroups Fusion radiomics signature, median (IQR) Training cohort Validation cohort MGMT (+) MGMT (-) AUC (95% CI) p MGMT (+) MGMT (-) AUC (95% CI) p * * Age, y ≤52 3.024 (2.215–3.706) 0.846 (-0.692–1.769) 0.918 (0.8149–1) 0.0022 2.859 (2.681–3.135) 0.381 (-0.791–1.885) 0.9135 (0.775–1) 0.0012 >52 2.798 (1.500–3.136) -2.088 (-2.777–0.461) 0.9447 (0.8649–1) < 0.001 2.626 (2.108–3.364) 1.177 (-0.675–1.950) 0.92 (0.7361–1) 0.052 Gender Male 2.98 (2.219–3.371) -1.899 (-2.752–0.218) 0.9469 (0.8835–1) <0.001 2.554 (1.922–3.025) 1.177 (-0.373–2.28) 0.7625 (0.5136–1) 0.067 * * Female 2.741 (1.93–3.281) -0.53 (-2.088–0.751) 0.9802 (0.7849–1) <0.001 2.938 (2.722–3.316) -0.957 (-1.071–0.236) 1 (1–1) 0.0069 * * Grade II 2.991 (2.215–3.555) 0.444 (-0.46–0.942) 0.9417 (0.8473–1) <0.001 3.135 (2.747–3.309) 0.905 (0.046–1.657) 0.9643 (0.8653–1) 0.012 * * III 2.689 (1.992–3.265) -2.383 (-3.119–-0.692) 0.9091 (0.7708–1) <0.001 2.859 (2.626–3.017) 1.500 (-0.625–2.1) 0.8444 (0.6224–1) 0.042 IV 2.798 (0.83–3.072) –1.578 (-2.739–-0.264) 0.9444 (0.8227–1) 0.0028 2.554 (2.477–2.632) -0.165 (-1.227–1.257) 0.875 (0.5285–1) 0.27 * * —— 2.807 (2.026–3.309) -1.174 (-2.727–0.598) 0.9254 (0.8611–0.9896) <0.001 2.853 (2.631–3.162) 0.855 (-0.958–2.100) 0.9017 (0.7853–1) <0.001 MGMT(+)patients with oxygen 6-methylguanine-DNA methyltransferase (MGMT) methylation, MGMT(-) patients without MGMT methylation, IQR interquartile range p-value < 0.05 indicates significant difference in the median radiomics score between the MGMT(+) and MGMT(-) groups p <0.05 886 Eur Radiol (2019) 29:877–888 Fig. 4 (a) Kaplan-Meier curve verifying the prognostic value of the y-axis represents the net benefit and the x-axis represents the threshold fusion radiomics signature. Patients were successfully divided into high- probability. The threshold probability of the decision curve is 26% and risk (red line) and low-risk (green line) groups (p = 0.0308). (b) Decision the corresponding net benefit is 0.441. * p <0.05 curve analysis for the fusion radiomics signature on the overall cohort. The promoter methylation in GBMs, with prediction accuracy of increasing the prediction accuracy. The oedema degree and 86.59% in the training cohort and 80% in the validation cohort ADC values, as kinds of semi-quantitative clinical and quan- [30]. However, MGMT methylated patients not only behaved titative radiological factors, partially depended on the sub- well in GBM, but also presented with prolonged survival in jective judgement of radiologists (sometimes with strong lower-grade astrocytomas [4, 31]. Our study used an expand- reservations), and their prediction performance were poorer ed cohort that included grade II-IV astrocytomas, and per- compared with the radiomics signature. Thus when incorpo- formed with superior AUCs of 0.926 and 0.902 in the training rating these factors into the radiomics signature, there was and validation datasets, respectively. Notably, our proposed no additional positive effect on the improvement of the pre- fusion radiomics signature has the power to distinguish diction performance, but rather an increased complexity of MGMT methylation in separate grade II, III and IV (GBM) the prediction model. Additionally, even though it was re- cohorts, as well as in a grade II-IV cohort. It can predict ported that ADC values were correlated with MGMT pro- MGMT methylation status directly without the need for a moter methylation and prognosis in GBM [22, 32, 33], our pathological grading prerequisite. Considering that the most results indicated that radiomic features extracted from T1- significant advantage of radiomics is its non-invasive charac- CE and T2-FLAIR sequences performed better than those teristics, pre-knowledge of grading that requires biopsy criti- extracted from the ADC sequence. A potential reason for cally limits the clinical application of radiomics, while our this observation is the relatively poor imaging resolution of results showed great advances on the previous study with ADC, which limited the stability and robustness of the de- improved high accuracy. This finding will strongly promote rived radiomics features. radiomics application in clinical practice. MGMT promoter methylation has been shown to be asso- Furthermore, we also investigated whether a combined ciated with longer OS, [34]. In our study, MGMT promoter model integrating clinical factors, radiological factors and methylation status successfully stratified astrocytoma patients fusion radiomics signature would outperform the signature treated with adjuvant TMZ chemotherapy into two groups alone. However, adding oedema degree and ADC values to with significant prognostic differences, consistent with previ- the fusion radiomics signature caused minor deterioration ous research [6]. We also validated the proposed fusion rather than improvement in prediction performance. This radiomics signature for assessing TMZ chemotherapy effect. indicated that adding clinical and radiological factors to the Using the cut-off value of the fusion Rad-score, patients with radiomics signature increased the complexity without positive radiomics scores after TMZ chemotherapy had Eur Radiol (2019) 29:877–888 887 Methodology significantly longer OS than patients with negative scores • retrospective (p = 0.03), revealing another possible clinical application of • diagnostic or prognostic study this genetic prediction tool. • performed at one institution The present study had several limitations. First, our model Open Access This article is distributed under the terms of the Creative was trained and validated using retrospective data collected Commons Attribution 4.0 International License (http:// from a single institution. A large-scale prospective and creativecommons.org/licenses/by/4.0/), which permits unrestricted use, multicentre validation cohort collection is currently underway. distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link Second, our radiomics analysis only predicted MGMT pro- to the Creative Commons license, and indicate if changes were made. moter methylation prediction from T1-CE, T2-FLAIR and ADC map images, which are the most common structural References MR images. Additional scanning sequences such as dynamic susceptibility contrast, susceptibility-weighted imaging and 1. Wen PY, Kesari S (2008) Malignant gliomas in adults. N Engl J diffusional kurtosis imaging will be included in future studies Med 359(5):492–507 to further improve predictive performance. Third, the relation- 2. Walker DG, Kaye AH (2001) Diagnosis and management of astro- ship between imaging features and critical molecular markers cytomas, oligodendrogliomas and mixed gliomas: a review. Australas Radiol 45(4):472–482 such as IDH and 1p19q should also be studied in future re- 3. DeAngelis LM (2001) Brain tumors. N Engl J Med 344(2):114– search. Finally, the manual segmentation method used to de- lineate ROIs in this study (tumour and oedema) was quite time 4. Everhard S, Kaloshi G, Crinière E et al (2006) MGMT methylation: consuming. Semi-automatic or deep learning-based automatic a marker of response to temozolomide in low-grade gliomas. Ann Neurol 60(6):740–743 segmentation methods may enhance the objectivity of our 5. Stupp R, Mason WP, van den Bent MJ et al (2005) Radiotherapy method and promote the seamless integration of this technol- plus concomitant and adjuvant temozolomide for glioblastoma. N ogy into clinical application. Engl J Med 352(10):987–996 In conclusion, habitat-based MRI radiomics could provide 6. Hegi ME, Diserens AC, Gorlia T et al (2005) MGMT gene silenc- a non-invasive imaging strategy for the preoperative predic- ing and benefit from temozolomide in glioblastoma. N Engl J Med 352(10):997–1003 tion of MGMT promoter methylation in patients with grade 7. Cairncross G, Wang M, Shaw E et al (2012) Phase III trial of II–IV astrocytomas, and has prognostic implications for TMZ chemoradiotherapy for anaplastic oligodendroglioma: long-term re- chemotherapy, which may serve as a tool for guiding sults of RTOG 9402. J Clin Oncol 31(3):337–343 individualised treatment decision making. 8. Sanai N, Martino J, Berger MS (2012) Morbidity profile following aggressive resection of parietal lobe gliomas. J Neurosurg 116(6): 1182–1186 Funding: This study has received funding by the National Natural 9. Tate MC, Kim CY, Chang EF et al (2011) Assessment of morbidity Science Foundation of China (81471652, 81771824, 11705112, following resection of cingulate gyrus gliomas. J Neurosurg 114(3): 81701681, 81227901, 61231004, 81501616, 81671854 and 81771924); 640–647 the Precision Medicine Key Innovation Team Project (YT1601); the 10. Henson JW, Gaviani P, Gonzalez RG (2005) MRI in treatment of Social Development Projects of Key R&D programme in Shanxi adult gliomas. Lancet Oncol 6(3):167–175 Province (201703D321016); and the Natural Science Foundation of 11. Eoli M, Menghi F, Bruzzone MG et al (2007) Methylation of O6- Shanxi Province (201601D021162); the China Postdoctoral Science methylguanine DNA methyltransferase and loss of heterozygosity Foundation (2017M621108); the Natural Science Foundation of Shanxi on 19q and/or 17p are overlapping features of secondary glioblas- Province (201601D021162); National Key R&D Program of China tomas with prolonged survival. Clin Cancer Res 13(9):2606–2613 (2017YFA0205200, 2017YFC1308700, 2017YFC1308701 and 12. Drabycz S, Roldán G, de Robles P et al (2010) An analysis of image 2017YFC1309100), the Science and Technology Service Network texture, tumor location, and MGMT promoter methylation in glio- Initiative of the Chinese Academy of Sciences (KFJ-SW-STS-160). blastoma using magnetic resonance imaging. Neuroimage 49(2): 1398–1405 Compliance with ethical standards 13. Kumar V, Gu Y, Basu S et al (2012) Radiomics: the process and the challenges. Magn Reson Imaging 30(9):1234–1248 Guarantor The scientific guarantor of this publication is Jie Tian. 14. Lambin P, Rios-Velazquez E, Leijenaar R et al (2012) Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 48(4):441–446 Conflict of interest The authors of this article declare no relationships 15. Gillies RJ, Kinahan PE, Hricak H (2016) Radiomics: Images are with any companies whose products or services may be related to the more than pictures, they are data. Radiology 278(2):563–577 subject matter of the article. 16. Brown R, Zlatescu M, Sijben A et al (2008) The use of magnetic resonance imaging to non-invasively detect genetic signatures in Statistics and biometry One of the authors has significant statistical oligodendroglioma. Clin Cancer Res 14(8):2357–2362 expertise. 17. Yu J, Shi Z, Lian Y et al (2017) Noninvasive IDH1 mutation esti- mation based on a quantitative radiomics approach for grade II Informed consent Written informed consent was waived by the glioma. Eur Radiol 27(8):3509–3522 Institutional Review Board. 18. Zhou M, Hall L, Goldgof D et al (2014) Radiologically defined ecological dynamics and clinical outcomes in glioblastoma Ethical approval Institutional Review Board approval was obtained. multiforme: preliminary results. Transl Oncol 7(1):5–13 888 Eur Radiol (2019) 29:877–888 19. Dunn J, Baborie A, Alam F et al (2009) Extent of MGMT promoter 27. Huang Y, Liu Z, He L et al (2016) Radiomics signature: A potential biomarker for the prediction of disease-free survival in early-stage methylation correlates with outcome in glioblastomas given temo- zolomide and radiotherapy. Br J Cancer 101(1):124–131 (I or II) non-small cell lung cancer. Radiology 281(3):947–957 28. Rutman AM, Kuo MD (2009) Radiogenomics: creating a link be- 20. Helmreich JE (2016) Regression modeling strategies with applica- tween molecular diagnostics and diagnostic imaging. Eur J Radiol tions to linear models, logistic and ordinal regression and survival 70(2):232–241 analysis (2nd edition). J Stat Softw. 70 (Book Review 2) 29. Kuo MD, Gollub J, Sirlin CB et al (2007) Radiogenomic analysis to 21. Bock J (2001) Power and sample size calculations. Springer, New identify imaging phenotypes associated with drug response gene York 11(4):309-333 expression programs in hepatocellular carcinoma. J Vasc Interv 22. Romano A, Calabria LF, Tavanti F et al (2013) Apparent diffusion Radiol 18(7):821–831 coefficient obtained by magnetic resonance imaging as a prognostic 30. Xi YB, Guo F, Xu ZL et al (2017) Radiomics signature: A potential marker in glioblastomas: correlation with MGMT promoter meth- biomarker for the prediction of MGMT promoter methylation in ylation status. Eur Radiol 23(2):513–520 glioblastoma. J Magn Reson Imaging. https://doi.org/10.1002/ 23. Wick W, Weller M, van den Bent M et al (2014) MGMT testing- jmri.25860 The challenges for biomarker based glioma treatment. Nat Rev 31. Minniti G, Scaringi C, Arcella A et al (2014) IDH1 mutation and Neurol 10(7):372–385 MGMT methylation status predict survival in patients with anaplas- 24. Ellingson BM, Cloughesy TF, Pope WB et al (2012) Anatomic tic astrocytoma treated with temozolomide-based chemoradiother- localization of O6-methylguanine DNA methyltransferase apy. J Neurooncol 118(2):377–383 (MGMT) promoter methylated and unmethylated tumors: a radio- 32. Rundle-Thiele D, Day B, Stringer B et al (2015) Using the apparent graphic study in 358 de novo human glioblastomas. Neuroimage diffusion coefficient to identifying MGMT promoter methylation 59(2):908–916 status early in glioblastoma: importance of analytical method. J 25. Kickingereder P, Burth S, Wick A et al (2016) Radiomic profiling Med Radiat Sci 62(2):92–98 of glioblastoma: identifying an imaging predictor of patient survival 33. Yoon RG, Kim HS, Paik W et al (2016) Different diagnostic values with improved performance over established clinical and radiologic of imaging parameters to predict pseudoprogression in glioblasto- risk models. Radiology 280(3):880–889 ma subgroups stratified by MGMT promoter methylation. Eur 26. Kickingereder P, Neuberger U, Bonekamp D et al (2017) Radiomic Radiol 27(1):255–266 subtyping improves disease stratification beyond key molecular, 34. Wick W, Platten M, Meisner C et al (2012) Chemotherapy alone clinical and standard imaging characteristics in patients with glio- versus radiotherapy alone for malignant astrocytoma in the elderly: blastoma. Neuro Oncol. https://doi.org/10.1093/neuonc/nox188 the NOA-08 randomised, phase 3 trial. Lancet Oncol 13(7):707– http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png European Radiology Springer Journals

A multi-sequence and habitat-based MRI radiomics signature for preoperative prediction of MGMT promoter methylation in astrocytomas with prognostic implication

Download PDF
12 pages

Loading next page...
 
/lp/springer-journals/a-multi-sequence-and-habitat-based-mri-radiomics-signature-for-rgjURRvNPJ

References (39)

Publisher
Springer Journals
Copyright
Copyright © 2018 by The Author(s)
Subject
Medicine & Public Health; Imaging / Radiology; Diagnostic Radiology; Interventional Radiology; Neuroradiology; Ultrasound; Internal Medicine
ISSN
0938-7994
eISSN
1432-1084
DOI
10.1007/s00330-018-5575-z
Publisher site
See Article on Publisher Site

Abstract

Objectives Oxygen 6-methylguanine-DNA methyltransferase (MGMT) promoter methylation is a significant prognostic biomarker in astrocytomas, especially for temozolomide (TMZ) chemotherapy. This study aimed to preoperatively predict MGMT methylation status based on magnetic resonance imaging (MRI) radiomics and validate its value for evaluation of TMZ chemotherapy effect. Methods We retrospectively reviewed a cohort of 105 patients with grade II-IV astrocytomas. Radiomic features were extracted from the tumour and peritumoral oedema habitats on contrast-enhanced T1-weighted images, T2-weighted fluid-attenuated inver- sion recovery images and apparent diffusion coefficient (ADC) maps. The following radiomics analysis was structured in three phases: feature reduction, signature construction and discrimination statistics. A fusion radiomics signature was finally developed using logistic regression modelling. Predictive performance was compared between the radiomics signature, previously reported clinical factors and ADC parameters. Validation was additionally performed on a time-independent cohort (n = 31). The prognostic value of the signature on overall survival for TMZ chemotherapy was explored using Kaplan Meier estimation. Results The fusion radiomics signature exhibited supreme power for predicting MGMT promoter methylation, with area under the curve values of 0.925 in the training cohort and 0.902 in the validation cohort. Performance of the radiomics signature surpassed that of clinical factors and ADC parameters. Moreover, the radiomics approach successfully divided patients into high- risk and low-risk groups for overall survival after TMZ chemotherapy (p =0.03). Conclusions The proposed radiomics signature accurately predicted MGMT promoter methylation in patients with astrocytomas, and achieved survival stratification for TMZ chemotherapy, thus providing a preoperative basis for individualised treatment planning. Key Points � Radiomics using magnetic resonance imaging can preoperatively perform satisfactory prediction of MGMT methylation in grade II-IV astrocytomas. � Habitat-based radiomics can improve efficacy in predicting MGMT methylation status. � Multi-sequence radiomics signature has the power to evaluate TMZ chemotherapy effect. . . . . Keywords Astrocytoma Methylation Prognosis Diagnostic imaging ROC curve Jingwei Wei and Guoqiang Yang contributed equally to this work. Electronic supplementary material The online version of this article (https://doi.org/10.1007/s00330-018-5575-z) contains supplementary material, which is available to authorized users. * Hui Zhang Beijing Key Laboratory of Molecular Imaging, Beijing 100190, [email protected] China * Jie Tian University of Chinese Academy of Sciences, Beijing 100049, China [email protected] Department of Radiology, First Clinical Medical College, Shanxi Medical University, Taiyuan 030001, Shanxi Province, China Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China 878 Eur Radiol (2019) 29:877–888 Abbreviations feature data to targeted clinical outcomes [13, 14]. ADC Apparent diffusion coefficient Regions of interest (ROIs) are delineated on the tumour AUC Area under the curve and sub-regions of the lesion known as habitats. Thus, DCA Decision curve analysis radiomics not only offers holistic imaging information, GBM Glioblastoma but also explores the microenvironment of the tumour IDH Isocitrate dehydrogenase by analysing explicit sub-regional features that describe MGMT Oxygen 6-methylguanine- genetic heterogeneity more granularly [15]. For gliomas, DNA methyltransferase radiomics studies on molecular subtype prediction have MRI Magnetic resonance imaging demonstrated sufficient predictive accuracy for isocitrate OS Overall survival dehydrogenase (IDH) and 1p19q codeletion [16, 17]. ROC Receiver operating characteristic Habitat-based radiomics have also been shown to have ROI Region of interest the capability to identify survival stratification in glioblas- T1-WI T1-weighted imaging tomas (GBMs) [18]. These studies suggest that habitat- T2-FLAIR T2-weighted fluid-attenuated inversion recovery based radiomics may be similarly useful for the preoper- images ative prediction of MGMT promoter methylation in pa- TMZ Temozolomide tients with astrocytomas. In this study, we investigated the utility of a multi-sequence and multi-habitat MR radiomics signature as a preoperative Introduction and non-invasive biomarker of MGMT methylation predic- tion in patients with grade II–IV astrocytomas, and discuss Astrocytoma is the most common type of glioma, and carries a the prognostic implications for survival stratification on poor prognosis [1, 2]. The average survival time ranges from 17 TMZ chemotherapy response. weeks to 3 years [2, 3]. Fortunately, a subgroup of grade II-IV astrocytoma patients with oxygen 6-methylguanine-DNA methyltransferase (MGMT) promoter methylation show good responses to temozolomide (TMZ) chemotherapy and im- Materials and methods proved survival after treatment, which underscored the role of MGMT as a judicious molecular biomarker with a prognostic Patients implication [4–7]. Preoperative identification of MGMT pro- moter methylation would be of great clinical significance in This retrospective study was approved by the institutional selecting potential patients benefiting from TMZ chemothera- review board. We reviewed 105 patients who were newly py, thus assisting with planning the therapy regime. However, diagnosed with grade II–IV astrocytoma from October 2011 the standard approach for MGMT status confirmation requires to March 2017. The inclusion and exclusion criteria are given a large tissue sample that is usually obtained through surgery. in the Online Supplemental Appendix E1. The patient recruit- For patients with unresectable tumours, biopsy runs the risk of ment pathway is shown in Online Supplemental Fig. 1. neurological deficits and can accordingly jeopardise the quality A total of 105 patients met the study criteria and were divid- of life of the patient [8, 9]. Thus, there is an urgent need in ed into a training dataset (31 December 2011 to 2 November clinical practice for preoperative and non-invasive prediction 2015, n = 74) and a time-independent validation dataset (16 of MGMT promoter methylation in grade II-IV astrocytomas. November 2015 to 21 March 2017, n = 31). Demographic Magnetic resonance imaging (MRI), as a powerful non- and clinical data were acquired from medical reports, including invasive diagnostic imaging tool for astrocytoma manage- sex, age, grade and radiological characteristics (Table 1). ment [10], opens up the possibility of having this preop- For evaluation of the TMZ chemotherapy effect, the erative prediction. Previous studies have verified that cer- inclusion criteria were further expanded in the 105 pa- tain radiological characteristics on MR images such as tients: (1) patients receiving adjuvant treatment following tumour necrosis, enhancement patterns and tumour loca- surgery consisting of either (a) concomitant radiation plus tion are associated with MGMT promoter methylation TMZ followed by adjuvant TMZ or (b) TMZ; (2) overall [11, 12]. However, subjective judgements by radiologists survival (OS) was categorised according to death or sur- are vulnerable to inter-observer variability and generally vival exceeding the median survival of patients with each lack power and accuracy. In contrast, a newly emerging tumour grade (605 days for grade II, 398 days for grade technology termed radiomics could resolve this problem III and 169 days for grade IV). Twenty-two patients met by quantitative imaging analysis. Radiomics converts the inclusion criteria for the TMZ survival analysis. encrypted medical images into usable data by extracting Demographic and clinical data of the 22 patients are high-throughput imaging features and relating imaging shown in Table 2. Eur Radiol (2019) 29:877–888 879 Table 1 Clinical characteristics in the training and validation cohorts Characteristic Training cohort Validation cohort p (Inter) N=74 N=31 MGMT (+) MGMT (-) p (Intra) MGMT (+) MGMT (-) p (Intra) Gender 0.674 0.139 0.902 Male 32 (43.2) 10 (13.5) 8 (25.8) 10 (32.3) Female 23 (31.1) 9 (12.2) 10 (32.3) 3 (9.7) Age, y 0.225 0.701 0.195 ≤52 32 (43.2) 8 (10.8) 13 (41.9) 8 (25.8) >52 23 (31.1) 11 (14.9) 5 (16.1) 5 (16.1) Grade 0.258 0.509 0.905 II 24 (32.4) 5 (6.8) 7 (22.6) 4 (12.9) III 22 (29.7) 8 (10.8) 9 (29.0) 5 (16.1) IV 9 (12.2) 6 (8.1) 2 (6.5) 4 (12.9) Tumour size 0.508 1.000 0.073 ≤6cm 27 (36.5) 11 (14.9) 6 (19.4) 4 (12.9) >6cm 28 (37.8) 8 (10.8) 12 (38.7) 9 (29.0) Tumour centre location 0.350 1.000 0.980 Left hemisphere 25 (33.8) 11 (14.9) 9 (29.0) 6 (19.4) Right hemisphere 30 (40.5) 8 (10.8) 9 (29.0) 7 (22.6) Frontal lobe 0.674 0.013 0.902 Yes 32 (43.2) 10 (13.5) 14 (45.2) 4 (12.9) No 23 (31.1) 9 (12.2) 4 (12.9) 9 (29.0) Occipital lobe 0.702 0.497 0.742 Yes 5 (6.8) 3 (4.1) 2 (6.5) 0 (0.0) No 50 (67.6) 16 (21.6) 16 (51.6) 13 (41.9) Parietal lobe 0.264 0.023 0.116 Yes 13 (17.6) 7 (9.5) 0 (0.0) 4 (12.9) No 42 (56.8) 12 (16.2) 18 (58.1) 9 (29.0) Temporal lobe 0.706 0.710 0.728 Yes 35 (47.3) 6 (8.1) 6 (19.4) 6 (19.4) No 20 (27.0) 13 (17.6) 12 (38.7) 7 (22.6) Insular lobe 0.140 0.625 0.817 Yes 9 (12.2) 0 (0.0) 2 (6.5) 3 (9.7) No 46 (62.2) 19 (25.7) 16 (51.6) 10 (32.3) Involving cortex matter 0.635 1.000 0.742 Yes 48 (64.9) 18 (24.3) 17 (54.8) 12 (38.7) No 7 (9.5) 1 (1.4) 1 (3.2) 1 (3.2) Involving deep white matter 0.910 0.497 0.238 Yes 46 (62.2) 15 (20.3) 16 (51.6) 13 (41.9) No 9 (12.2) 4 (5.4) 2 (6.5) 0 (0.0) Involving pial matter 1.000 1.000 1.000 Yes 48 (64.9) 16 (21.6) 16 (51.6) 11 (35.5) No 7 (9.5) 3 (4.1) 2 (6.5) 2 (6.5) Involving ependymal membrane 0.555 1.000 0.600 Yes 19 (25.7) 8 (10.8) 8 (25.8) 5 (16.1) No 36 (48.6) 11 (14.9) 10 (32.3) 8 (25.8) Tumour cross midline 0.949 1.000 0.735 Yes 11 (14.9) 3 (4.1) 3 (9.7) 2 (6.5) No 44 (59.5) 16 (21.6) 15 (48.4) 11 (35.5) 880 Eur Radiol (2019) 29:877–888 Table 1 (continued) Characteristic Training cohort Validation cohort p (Inter) N=74 N=31 MGMT (+) MGMT (-) p (Intra) MGMT (+) MGMT (-) p (Intra) Oedema cross midline 1.000 1.000 0.416 Yes 12 (16.2) 4 (5.4) 5 (16.1) 4 (12.9) No 43 (58.1) 15 (20.3) 13 (41.9) 9 (29.0) Border 0.171 0.099 0.455 Well-defined 14 (55.4) 8 (10.8) 2 (6.5) 5 (16.1) Ill-defined 41 (18.9) 11 (14.9) 16 (51.6) 8 (25.8) Haemorrhage 0.254 1.000 0.134 Yes 11 (14.9) 1 (1.4) 5 (16.1) 4 (12.9) No 44 (59.5) 18 (24.3) 13 (41.9) 9 (29.0) Cystic and necrosis 0.751 1 0.053 No 15 (20.3) 5 (6.8) 7 (22.6) 5 (16.1) ≤25% 21 (28.4) 5 (6.8) 4 (12.9) 3 (9.7) 25%-50% 11 (14.9) 5 (6.8) 1 (3.2) 1 (3.2) ≥50% 8 (10.8) 4 (5.4) 6 (19.4) 4 (12.9) Oedema degree 0.011 0.275 0.563 ≤1.6 33 (44.6) 5 (6.8) 10 (32.3) 4 (12.9) >1.6 22 (29.7) 14 (18.9) 8 (25.8) 9 (29.0) * * Enhancement style 0.010 0.034 0369 No 15 (20.3) 3 (4.1) 4 (19.4) 0 (0.0) Ring enhancement 20 (27.0) 15 (20.3) 8 (25.8) 8 (25.8) Nodular enhancement 11 (14.9) 0 (0.0) 6 (19.4) 2 (6.5) Irregular reinforcement 9 (12.2) 1 (1.4) 0 (0.0) 3 (9.7) Enhancement degree 0.287 0.211 0.401 No 15 (20.3) 3 (4.1) 4 (12.9) 0 (0.0) Slight 5 (6.8) 0 (0.0) 2 (6.5) 1 (3.2) Obvious 35 (47.3) 16 (21.6) 12 (38.7) 12 (38.7) Signal characteristics 1.000 0.497 0.904 Homogeneous 5 (6.8) 2 (2.7) 2 (6.5) 0 (0.0) Heterogeneous 50 (67.6) 17 (23.0) 16 (51.6) 13 (41.9) MGMT(+) patients with oxygen 6-methylguanine-DNA methyltransferase (MGMT) methylation, MGMT(-) patients without MGMT methylation, p (Intra) the result of uni-variable analyses between methylated and unmethylated groups, p(Inter) significant difference between training and validation cohorts Unless otherwise specified, data are numbers of patients, with percentages in parentheses p < 0.05 MGMT methylation testing DNA modification kit (EpiGentek, Farmingdale, NY, USA). The PCR amplification and conditions are given in Online The methylation status of the MGMT promoter was assessed Supplemental Appendix E2. using pyrosequencing analysis as described elsewhere [19]. Briefly, each tumour specimen was histologically investigated by macro-dissection to guarantee a tumour cell content of at MRI acquisition least 80%. DNAwas extracted using the Simlex OUP ® FFPE DNA extraction kit (TIB, China) and quantified by spectro- Preoperative MRI was performed with a 3.0-T scanner (Signa photometry using a NanoDrop 2000 (Thermo Fisher HDxt, GE Healthcare, USA) using an 8-channel array coil. Scientific, Loughborough, UK). Bisulphite modification of The acquisition protocols for CE-T1-WI, T2-FLAIR and the extracted DNA was performed using the BisulFlash™ DWI are in given in Online Supplemental Appendix E3. Eur Radiol (2019) 29:877–888 881 Table 2 Clinical characteristics and MGMT predicted outcome for Process of radiomics analysis patients with TMZ chemotherapy The radiomics analysis was structured into four parts: ROI Characteristic Patients N=22 segmentation, feature extraction, feature selection and model construction (Fig. 1). In brief, we performed a manual ROI Gender with overlapped area by two radiologists (10 and 15 years of Male 13 (59.1) experience, respectively) on tumour and oedema habitats from Female 9 (40.9) T1-WI, T2-FLAIR and ADC maps, and the two radiologists Age, y were blinded to the final diagnosis and the MGMT methyla- ≤49 12 (54.5) tion status. Examples of six typical segmentation cases ac- >49 10 (45.5) cording to grade and MGMT methylation status are shown Grade in Fig. 2. The median ROI of each habitat and sequence is II 6 (27.3) shown in Online Supplemental Table 3. A set of 3,051 III 11 (50.0) imaging features was extracted including textural and IV 5 (22.7) non-textural features. Feature stability and reproducibil- MGMT ity was estimated by Intra-class correlation coefficients + 17 (77.2) and concordance correlation coefficients. Further feature – 5 (22.8) selection was performed based on minimum redundancy Predicted MGMT and maximum relevance algorithm. Single radiomics MGMT(+) 14 (82.3) signatures from each sequence and habitat, and a fusion MGMT(-) 3 (17.7) radiomics signature were finally constructed using logis- tic regression modelling. A detailed description is given MGMT(+) patients with oxygen 6-methylguanine-DNA methyltransfer- in Online Supplemental Appendix E4. ase (MGMT) methylation, MGMT(-) patients without MGMT methylation Clinical and radiological factors for MGMT prediction Unless otherwise specified, data are numbers of patients, with percent- ages in parentheses The average age was 49 years, thus we divided patients into an age ≤ 49 Clinical factor analysis and ADC parameter calculation Auni- years group and an age > 49 years group variate analysis was initially applied to select useful clinical factors (p < 0.1). Then, a forward selection (likelihood ratio) multi-variable analysis was performed to select clinical Demographic and clinical characteristic analysis factors with p < 0.05. Additionally, we introduced ADC parameters (mean tumour ADC values and mean Differences between the training and validation cohorts and peritumoral oedema ADC values) associated with between the intra- MGMT methylated and unmethylated MGMT promoter methylation as reported in the litera- groups in terms of demographic and clinical factors were ture [22]. Tumour and oedema ADC values were addi- assessed with Pearson’s chi-square tests or Fisher’s exact tests tionally integrated as a clinical model by logistic regres- for categorical variables and Student’s t-tests or Mann- sion based on the training cohort. Whitney U tests for continuous variables. Combined model with radiomics signature, clinical factor, and ADC values To achieve a holistic information-gathered Sample size and power calculation network, we generated a comprehensive model including the fusion radiomics signature, the selected clinical factor (oede- According to the thumb rule, the sample size needed to cover ma degree), and two ADC values (the tumour and oedema 10–15 observations per predictor variable to yield a stable areas). Considering a correlation between the radiomics sig- estimate [20]. In our study, the maximum number of included nature and other factors as well as the model complexity, we features for radiomics signature construction was 5 (T2- adopted the Akaike information criterion (AIC) to select op- FLAIR sequence on tumour area). Thus, the training dataset timal incorporated factors and used logistic regression model- needed to include at least 50 patients. For the validation ling to perform model construction. dataset power calculation, a sample of > 11 patients was re- quired to provide 80% power and a type I error rate of 5% Performance evaluation [21]. Our study cohort included 105 patients with 74 in the training dataset and 31 in the validation dataset, which met the Receiver operating characteristic (ROC) curves were plotted sample size requirement. and area under the curve (AUC), specificity and sensitivity 882 Eur Radiol (2019) 29:877–888 Fig. 1 Radiomics workflow. The radiomics process included four parts: and reproducibility analysis before using maximum relevance and region of interest (ROI) segmentation on each habitat and sequence, minimum redundancy algorithm. The radiomics signature was feature extraction, feature selection and model construction. ROI was generated by logistic regression with Bayesian information criteria as delineated on both the tumour and the peritumoral habitats on contrast- the stopping rule. Further performance evaluation was explored enhanced T1-weighted images and T2-FLAIR images. On each ROI, a including receiver operating characteristics, decision curve analysis and set of 3,051 features were extracted. We performed fundamental stability survival stratification were calculated for the fusion radiomics signature, selected version 3.4.1 (www.R-project.org). The threshold for clinical factor and ADC parameter. We further performed statistical significance was a two-sided p <0.05. stratification analysis for the fusion radiomics signature by grouping the cohorts according to age, gender and grade. We chose decision curve analysis (DCA) to estimate the clinical Results usefulness of the developed fusion radiomics signature and used the Delong test to explore whether the fusion radiomics Patient demographic data, clinical characteristics signature performed better than the traditional clinical model and molecular subtypes and ADC parameter. The baseline characteristics of the patients are shown in Table 1. There were no differences between the training and Prognostic value analysis validation cohorts in terms of demographic or clinical charac- teristics (p =0.053–1.000). A Kaplan-Meier curve was plotted based on the fusion In total, we included 73 (69.5%) patients with MGMT radiomics signature in order to stratify the OS in patients treat- promoter methylation and 32 (30.5%) patients without ed with adjuvant TMZ chemotherapy. The log-rank test was MGMT promoter methylation. No significant difference was used to determine whether there were statistical differences shown for the MGMT methylation status distribution in the between the two survival groups. training and validation cohorts (p =0.156). Statistical analysis Feature stability and reproducibility estimation We performed the statistical analysis with PASW Statistics, The statistical results of feature numbers after stability and repro- version 18.0 (SPSS Inc., Chicago, IL, USA) and R software, ducibility analysis are shown in Online Supplemental Fig. 2. Eur Radiol (2019) 29:877–888 883 Fig. 2 Tumour and oedema area segmentations are shown by the red and green lines, respectively. Oedema degree was obvious in the MGMT unmethylated group compared to the MGMT methylated group, especially for higher-grade astrocytomas (III and IV) Features extracted from tumour habitats exhibited statistically are shown in Online Supplemental Table 1 and Appendix E5, better performance than those extracted from peritumoral oede- respectively. A detailed explanation of each selected feature is ma habitats (reproducibility, p < 0.001; stability, p < 0.001). shown in Online Supplemental Table 2. Single radiomics sig- natures from T1-WI-tumour, T1-WI-oedema, T2-FLAIR-tu- Single radiomics signatures formula and evaluation mour, and T2-FLAIR-oedema were verified as eligible radiomics signatures with AUCs > 0.7 in both the training and validation cohorts (Table 3). These four single radiomics The selected features and integration formulas for single radiomics signatures derived from each sequence and habitat signatures all showed significant differences (p < 0.05) in the Table 3 Diagnostic performance of single radiomics signatures, fusion radiomics signature, clinical factors and ADC values Models Training cohort Validation cohort N=74 N=31 Sensitivity Specificity Accuracy AUC (95% CI) Sensitivity Specificity Accuracy AUC (95% CI) Tumour T1 0.789 0.691 0.716 0.706 (0.567–0.845) 0.615 0.667 0.645 0.739 (0.558–0.921) Tumour T2 0.947 0.782 0.824 0.916 (0.852–0.979) 0.538 0.889 0.742 0.701 (0.494–0.908) Tumour ADC 0.895 0.655 0.716 0.815 (0.714–0.917) 0.692 0.389 0.516 0.590 (0.381–0.799) Oedema T1 0.632 0.782 0.743 0.738 (0.601–0.875) 0.615 0.667 0.645 0.709 (0.519–0.900) Oedema T2 0.684 0.727 0.716 0.778 (0.654–0.902) 0.615 0.778 0.710 0.752 (0.567–0.937) Oedema ADC 0.632 0.782 0.743 0.678 (0.534–0.822) 0.615 0.889 0.774 0.816 (0.667–0.965) Fusion radiomics 0.872 0.842 0.865 0.925 (0.861–0.989) 0.944 0.539 0.774 0.902 (0.785–1.000) Clinical factors 0.737 0.600 0.6351 0.668 (0.548–0.789) 0.692 0.556 0.613 0.624 (0.448–0.800) ADC values 0.632 0.691 0.676 0.649 (0.511–0.787) 0.615 0.722 0.677 0.603 (0.382–0.823) 95% CI 95% confidence interval, AUC area under curve, T1 contrast-enhanced T1-weighted sequence, T2 T2-FLAIR sequence 884 Eur Radiol (2019) 29:877–888 MGMT methylated and unmethylated groups in both the (accuracy, sensitivity and specificity) of the fusion radiomics training and the validation cohorts. Boxplots describing the signature are shown in Table 3. distribution of the four radiomics signatures in the MGMT The fusion radiomics signature also had an outstanding per- methylated and unmethylated groups are shown in Online formance in the stratification analysis, considering age, gender Supplemental Fig. 3. However, neither ADC-tumour nor and grade (Table 4). It is noteworthy that the proposed ADC-oedema performed with satisfactory results for radiomics signature does not require a priori knowledge of MGMT identification. ADC-tumour did not perform well in grading information because it behaved well for distinguishing the validation cohort (AUC = 0.590) while ADC-oedema did MGMT methylation status not only in a grade II-IV cohort, but not perform well in the training cohort (AUC = 0.678). also in grade II, III and IV astrocytomas, separately. Detailed predictive indicators (AUC, accuracy, sensitivity and specificity) of each single radiomics signature are shown Clinical model and ADC parameter evaluation in Table 3. Only oedema degree was significantly different between the MGMT methylated and unmethylated groups in the Fusion radiomics signature formula and evaluation training cohort (p = 0.0015). The AUCs of this clinical factor were 0.668 and 0.624 in the training and valida- The fusion radiomics signature combining the four single tion cohorts, respectively. The ADC parameter achieved radiomics signatures was constructed with a Rad-score calcu- AUC values of 0.649 and 0.603 in the training and lated as follows: validation cohorts, respectively. Detailed predictive indi- cators (sensitivity and specificity) of the clinical model Rad‐scoreðÞ fusion¼ −6:785−1:026*Signature Oedema−T1 and ADC parameter are shown in Table 3. þ 3:950*Signature Oedema−T2 þ 2:907*Signature Performance comparison Tumour −T1 þ 5:427 Singature Tumour−T2 The fusion radiomics signature achieved the highest AUC The optimum cut-off value of the fusion radiomics was among the three models. The Delong test showed a significant 1.077 as per the Youden index. Patients were divided into difference between the fusion radiomics signature and the predicted MGMT methylated (Rad-score ≥ 1.077) and clinical model (p = 0.008 and 0.011), and between the fusion unmethylated groups (Rad-score < 1.077) based on fusion radiomics signature and ADC parameter (p =0.003 and Rad-scores. 0.027) in the training and validation cohorts, respectively. Barplots depicting the classification performance of the fusion signature in the training and validation cohorts are Combined model construction and evaluation shown in Fig. 3. The fusion radiomics signature achieved optimal AUC values of 0.925 and 0.902 in the training and During the combined model construction process, the AIC validation cohorts, respectively. Detailed predictive indicators value was minimum when only taking the fusion radiomics Fig. 3 Barplots depicting the classification performance of the fusion radiomics signature. The red bar with a prediction value > 0 indicates that the signature successfully classifies the MGMT methylation patients; the red bar with a prediction value < 0 indicates that the signature fails to classify the MGMT methylation patients. For the green bar, the contrary applies Eur Radiol (2019) 29:877–888 885 signature into account. The AIC values were 48.79, 49.79, 51.25 and 53.11 when subsequently adding oedema degree, ADC of tumour area and ADC of oedema area to the fusion radiomics signature. Notwithstanding, we calculated the AUC of the combined model integrating all factors and the result was concordant with the AIC value; the combined model achieved an AUC of 0.921 in the training cohort and 0.868 in the validation cohort, which were slightly lower than those for the fusion radiomics signature alone. Prognostic value of the fusion radiomics signature The fusion radiomics signature successfully divided patients treated with adjuvant TMZ chemotherapy into high-risk and low-risk OS groups with p = 0.03 (Fig. 4a). Moreover, the DCA showed that the fusion radiomics signature performed with higher net benefit (net benefit = 0.441) compared to simple stratification assuming that no patients had MGMT methylation or all patients had MGMT methylation (Fig. 4b). Discussion In this study, a preoperative and low-cost radiomics analysis was used to integrate imaging features from tumour and peritumoral oedema habitats on CE-T1-WI and T2-FLAIR images to predict MGMT promoter methylation in patients with grade II-IVastrocytoma. Moreover, we verified the prog- nostic value of the fusion radiomics signature for patients who underwent resection followed by adjuvant TMZ chemotherapy. Compared to clinical and conventional radiological factors [23, 24], our proposed radiomics signature exhibited excellent prediction performance. Additionally, the fusion radiomics signature integrating synergistic information outperformed each single radiomics signature from a simple habitat or se- quence. Potential reasons for this observation are as follows: first, the fusion signature included more comprehensive infor- mation reflecting granular textural differences in the microen- vironments and took in important archetypal imaging charac- teristics associated with MGMT methylation. Previous litera- ture supports the value of the multi-habitat radiomics for predicting survival in patients with GBM [25–27]. Second, most of the effective features extracted in our study were tex- tural features from Gabor transformation images, which con- ducted noise removal and filtration. Thus, these transformed images more effectively captured key tumour heterogeneity [25]. These findings agree with the radiomics hypothesis that gene phenotypic information of the tumour is reflected in ra- diological images [28, 29]. In previous work, Xi et al. showed that a radiomics signa- ture derived from T1-WI, T2-WI and enhanced T1-WI was a potential imaging marker for the prediction of MGMT Table 4 Stratification analysis of the fusion radiomics signature on training and validation cohorts Subgroups Fusion radiomics signature, median (IQR) Training cohort Validation cohort MGMT (+) MGMT (-) AUC (95% CI) p MGMT (+) MGMT (-) AUC (95% CI) p * * Age, y ≤52 3.024 (2.215–3.706) 0.846 (-0.692–1.769) 0.918 (0.8149–1) 0.0022 2.859 (2.681–3.135) 0.381 (-0.791–1.885) 0.9135 (0.775–1) 0.0012 >52 2.798 (1.500–3.136) -2.088 (-2.777–0.461) 0.9447 (0.8649–1) < 0.001 2.626 (2.108–3.364) 1.177 (-0.675–1.950) 0.92 (0.7361–1) 0.052 Gender Male 2.98 (2.219–3.371) -1.899 (-2.752–0.218) 0.9469 (0.8835–1) <0.001 2.554 (1.922–3.025) 1.177 (-0.373–2.28) 0.7625 (0.5136–1) 0.067 * * Female 2.741 (1.93–3.281) -0.53 (-2.088–0.751) 0.9802 (0.7849–1) <0.001 2.938 (2.722–3.316) -0.957 (-1.071–0.236) 1 (1–1) 0.0069 * * Grade II 2.991 (2.215–3.555) 0.444 (-0.46–0.942) 0.9417 (0.8473–1) <0.001 3.135 (2.747–3.309) 0.905 (0.046–1.657) 0.9643 (0.8653–1) 0.012 * * III 2.689 (1.992–3.265) -2.383 (-3.119–-0.692) 0.9091 (0.7708–1) <0.001 2.859 (2.626–3.017) 1.500 (-0.625–2.1) 0.8444 (0.6224–1) 0.042 IV 2.798 (0.83–3.072) –1.578 (-2.739–-0.264) 0.9444 (0.8227–1) 0.0028 2.554 (2.477–2.632) -0.165 (-1.227–1.257) 0.875 (0.5285–1) 0.27 * * —— 2.807 (2.026–3.309) -1.174 (-2.727–0.598) 0.9254 (0.8611–0.9896) <0.001 2.853 (2.631–3.162) 0.855 (-0.958–2.100) 0.9017 (0.7853–1) <0.001 MGMT(+)patients with oxygen 6-methylguanine-DNA methyltransferase (MGMT) methylation, MGMT(-) patients without MGMT methylation, IQR interquartile range p-value < 0.05 indicates significant difference in the median radiomics score between the MGMT(+) and MGMT(-) groups p <0.05 886 Eur Radiol (2019) 29:877–888 Fig. 4 (a) Kaplan-Meier curve verifying the prognostic value of the y-axis represents the net benefit and the x-axis represents the threshold fusion radiomics signature. Patients were successfully divided into high- probability. The threshold probability of the decision curve is 26% and risk (red line) and low-risk (green line) groups (p = 0.0308). (b) Decision the corresponding net benefit is 0.441. * p <0.05 curve analysis for the fusion radiomics signature on the overall cohort. The promoter methylation in GBMs, with prediction accuracy of increasing the prediction accuracy. The oedema degree and 86.59% in the training cohort and 80% in the validation cohort ADC values, as kinds of semi-quantitative clinical and quan- [30]. However, MGMT methylated patients not only behaved titative radiological factors, partially depended on the sub- well in GBM, but also presented with prolonged survival in jective judgement of radiologists (sometimes with strong lower-grade astrocytomas [4, 31]. Our study used an expand- reservations), and their prediction performance were poorer ed cohort that included grade II-IV astrocytomas, and per- compared with the radiomics signature. Thus when incorpo- formed with superior AUCs of 0.926 and 0.902 in the training rating these factors into the radiomics signature, there was and validation datasets, respectively. Notably, our proposed no additional positive effect on the improvement of the pre- fusion radiomics signature has the power to distinguish diction performance, but rather an increased complexity of MGMT methylation in separate grade II, III and IV (GBM) the prediction model. Additionally, even though it was re- cohorts, as well as in a grade II-IV cohort. It can predict ported that ADC values were correlated with MGMT pro- MGMT methylation status directly without the need for a moter methylation and prognosis in GBM [22, 32, 33], our pathological grading prerequisite. Considering that the most results indicated that radiomic features extracted from T1- significant advantage of radiomics is its non-invasive charac- CE and T2-FLAIR sequences performed better than those teristics, pre-knowledge of grading that requires biopsy criti- extracted from the ADC sequence. A potential reason for cally limits the clinical application of radiomics, while our this observation is the relatively poor imaging resolution of results showed great advances on the previous study with ADC, which limited the stability and robustness of the de- improved high accuracy. This finding will strongly promote rived radiomics features. radiomics application in clinical practice. MGMT promoter methylation has been shown to be asso- Furthermore, we also investigated whether a combined ciated with longer OS, [34]. In our study, MGMT promoter model integrating clinical factors, radiological factors and methylation status successfully stratified astrocytoma patients fusion radiomics signature would outperform the signature treated with adjuvant TMZ chemotherapy into two groups alone. However, adding oedema degree and ADC values to with significant prognostic differences, consistent with previ- the fusion radiomics signature caused minor deterioration ous research [6]. We also validated the proposed fusion rather than improvement in prediction performance. This radiomics signature for assessing TMZ chemotherapy effect. indicated that adding clinical and radiological factors to the Using the cut-off value of the fusion Rad-score, patients with radiomics signature increased the complexity without positive radiomics scores after TMZ chemotherapy had Eur Radiol (2019) 29:877–888 887 Methodology significantly longer OS than patients with negative scores • retrospective (p = 0.03), revealing another possible clinical application of • diagnostic or prognostic study this genetic prediction tool. • performed at one institution The present study had several limitations. First, our model Open Access This article is distributed under the terms of the Creative was trained and validated using retrospective data collected Commons Attribution 4.0 International License (http:// from a single institution. A large-scale prospective and creativecommons.org/licenses/by/4.0/), which permits unrestricted use, multicentre validation cohort collection is currently underway. distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link Second, our radiomics analysis only predicted MGMT pro- to the Creative Commons license, and indicate if changes were made. moter methylation prediction from T1-CE, T2-FLAIR and ADC map images, which are the most common structural References MR images. Additional scanning sequences such as dynamic susceptibility contrast, susceptibility-weighted imaging and 1. Wen PY, Kesari S (2008) Malignant gliomas in adults. N Engl J diffusional kurtosis imaging will be included in future studies Med 359(5):492–507 to further improve predictive performance. Third, the relation- 2. Walker DG, Kaye AH (2001) Diagnosis and management of astro- ship between imaging features and critical molecular markers cytomas, oligodendrogliomas and mixed gliomas: a review. Australas Radiol 45(4):472–482 such as IDH and 1p19q should also be studied in future re- 3. DeAngelis LM (2001) Brain tumors. N Engl J Med 344(2):114– search. Finally, the manual segmentation method used to de- lineate ROIs in this study (tumour and oedema) was quite time 4. Everhard S, Kaloshi G, Crinière E et al (2006) MGMT methylation: consuming. Semi-automatic or deep learning-based automatic a marker of response to temozolomide in low-grade gliomas. Ann Neurol 60(6):740–743 segmentation methods may enhance the objectivity of our 5. Stupp R, Mason WP, van den Bent MJ et al (2005) Radiotherapy method and promote the seamless integration of this technol- plus concomitant and adjuvant temozolomide for glioblastoma. N ogy into clinical application. Engl J Med 352(10):987–996 In conclusion, habitat-based MRI radiomics could provide 6. Hegi ME, Diserens AC, Gorlia T et al (2005) MGMT gene silenc- a non-invasive imaging strategy for the preoperative predic- ing and benefit from temozolomide in glioblastoma. N Engl J Med 352(10):997–1003 tion of MGMT promoter methylation in patients with grade 7. Cairncross G, Wang M, Shaw E et al (2012) Phase III trial of II–IV astrocytomas, and has prognostic implications for TMZ chemoradiotherapy for anaplastic oligodendroglioma: long-term re- chemotherapy, which may serve as a tool for guiding sults of RTOG 9402. J Clin Oncol 31(3):337–343 individualised treatment decision making. 8. Sanai N, Martino J, Berger MS (2012) Morbidity profile following aggressive resection of parietal lobe gliomas. J Neurosurg 116(6): 1182–1186 Funding: This study has received funding by the National Natural 9. Tate MC, Kim CY, Chang EF et al (2011) Assessment of morbidity Science Foundation of China (81471652, 81771824, 11705112, following resection of cingulate gyrus gliomas. J Neurosurg 114(3): 81701681, 81227901, 61231004, 81501616, 81671854 and 81771924); 640–647 the Precision Medicine Key Innovation Team Project (YT1601); the 10. Henson JW, Gaviani P, Gonzalez RG (2005) MRI in treatment of Social Development Projects of Key R&D programme in Shanxi adult gliomas. Lancet Oncol 6(3):167–175 Province (201703D321016); and the Natural Science Foundation of 11. Eoli M, Menghi F, Bruzzone MG et al (2007) Methylation of O6- Shanxi Province (201601D021162); the China Postdoctoral Science methylguanine DNA methyltransferase and loss of heterozygosity Foundation (2017M621108); the Natural Science Foundation of Shanxi on 19q and/or 17p are overlapping features of secondary glioblas- Province (201601D021162); National Key R&D Program of China tomas with prolonged survival. Clin Cancer Res 13(9):2606–2613 (2017YFA0205200, 2017YFC1308700, 2017YFC1308701 and 12. Drabycz S, Roldán G, de Robles P et al (2010) An analysis of image 2017YFC1309100), the Science and Technology Service Network texture, tumor location, and MGMT promoter methylation in glio- Initiative of the Chinese Academy of Sciences (KFJ-SW-STS-160). blastoma using magnetic resonance imaging. Neuroimage 49(2): 1398–1405 Compliance with ethical standards 13. Kumar V, Gu Y, Basu S et al (2012) Radiomics: the process and the challenges. Magn Reson Imaging 30(9):1234–1248 Guarantor The scientific guarantor of this publication is Jie Tian. 14. Lambin P, Rios-Velazquez E, Leijenaar R et al (2012) Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 48(4):441–446 Conflict of interest The authors of this article declare no relationships 15. Gillies RJ, Kinahan PE, Hricak H (2016) Radiomics: Images are with any companies whose products or services may be related to the more than pictures, they are data. Radiology 278(2):563–577 subject matter of the article. 16. Brown R, Zlatescu M, Sijben A et al (2008) The use of magnetic resonance imaging to non-invasively detect genetic signatures in Statistics and biometry One of the authors has significant statistical oligodendroglioma. Clin Cancer Res 14(8):2357–2362 expertise. 17. Yu J, Shi Z, Lian Y et al (2017) Noninvasive IDH1 mutation esti- mation based on a quantitative radiomics approach for grade II Informed consent Written informed consent was waived by the glioma. Eur Radiol 27(8):3509–3522 Institutional Review Board. 18. Zhou M, Hall L, Goldgof D et al (2014) Radiologically defined ecological dynamics and clinical outcomes in glioblastoma Ethical approval Institutional Review Board approval was obtained. multiforme: preliminary results. Transl Oncol 7(1):5–13 888 Eur Radiol (2019) 29:877–888 19. Dunn J, Baborie A, Alam F et al (2009) Extent of MGMT promoter 27. Huang Y, Liu Z, He L et al (2016) Radiomics signature: A potential biomarker for the prediction of disease-free survival in early-stage methylation correlates with outcome in glioblastomas given temo- zolomide and radiotherapy. Br J Cancer 101(1):124–131 (I or II) non-small cell lung cancer. Radiology 281(3):947–957 28. Rutman AM, Kuo MD (2009) Radiogenomics: creating a link be- 20. Helmreich JE (2016) Regression modeling strategies with applica- tween molecular diagnostics and diagnostic imaging. Eur J Radiol tions to linear models, logistic and ordinal regression and survival 70(2):232–241 analysis (2nd edition). J Stat Softw. 70 (Book Review 2) 29. Kuo MD, Gollub J, Sirlin CB et al (2007) Radiogenomic analysis to 21. Bock J (2001) Power and sample size calculations. Springer, New identify imaging phenotypes associated with drug response gene York 11(4):309-333 expression programs in hepatocellular carcinoma. J Vasc Interv 22. Romano A, Calabria LF, Tavanti F et al (2013) Apparent diffusion Radiol 18(7):821–831 coefficient obtained by magnetic resonance imaging as a prognostic 30. Xi YB, Guo F, Xu ZL et al (2017) Radiomics signature: A potential marker in glioblastomas: correlation with MGMT promoter meth- biomarker for the prediction of MGMT promoter methylation in ylation status. Eur Radiol 23(2):513–520 glioblastoma. J Magn Reson Imaging. https://doi.org/10.1002/ 23. Wick W, Weller M, van den Bent M et al (2014) MGMT testing- jmri.25860 The challenges for biomarker based glioma treatment. Nat Rev 31. Minniti G, Scaringi C, Arcella A et al (2014) IDH1 mutation and Neurol 10(7):372–385 MGMT methylation status predict survival in patients with anaplas- 24. Ellingson BM, Cloughesy TF, Pope WB et al (2012) Anatomic tic astrocytoma treated with temozolomide-based chemoradiother- localization of O6-methylguanine DNA methyltransferase apy. J Neurooncol 118(2):377–383 (MGMT) promoter methylated and unmethylated tumors: a radio- 32. Rundle-Thiele D, Day B, Stringer B et al (2015) Using the apparent graphic study in 358 de novo human glioblastomas. Neuroimage diffusion coefficient to identifying MGMT promoter methylation 59(2):908–916 status early in glioblastoma: importance of analytical method. J 25. Kickingereder P, Burth S, Wick A et al (2016) Radiomic profiling Med Radiat Sci 62(2):92–98 of glioblastoma: identifying an imaging predictor of patient survival 33. Yoon RG, Kim HS, Paik W et al (2016) Different diagnostic values with improved performance over established clinical and radiologic of imaging parameters to predict pseudoprogression in glioblasto- risk models. Radiology 280(3):880–889 ma subgroups stratified by MGMT promoter methylation. Eur 26. Kickingereder P, Neuberger U, Bonekamp D et al (2017) Radiomic Radiol 27(1):255–266 subtyping improves disease stratification beyond key molecular, 34. Wick W, Platten M, Meisner C et al (2012) Chemotherapy alone clinical and standard imaging characteristics in patients with glio- versus radiotherapy alone for malignant astrocytoma in the elderly: blastoma. Neuro Oncol. https://doi.org/10.1093/neuonc/nox188 the NOA-08 randomised, phase 3 trial. Lancet Oncol 13(7):707–

Journal

European Radiology – Springer Journals

Published: Jul 23, 2018

There are no references for this article.