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

Learn More →

SINETRA: a Versatile Framework for Evaluating Single Neuron Tracking in Behaving Animals

SINETRA: a Versatile Framework for Evaluating Single Neuron Tracking in Behaving Animals This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. SINETRA: A VERSATILE FRAMEWORK FOR EVALUATING SINGLE NEURON TRACKING IN BEHAVING ANIMALS ⋆† † † Raphael Reme Alasdair Newson Elsa Angelini ⋆ ⋆ Jean-Christophe Olivo-Marin Thibault Lagache Institut Pasteur, Universite ´ de Paris-Cite, ´ CNRS UMR 3691, BioImage Analysis Unit F-75015 Paris, France LTCI, Telecom Paris, Institut Polytechnique de Paris, France Corresponding author: [email protected] (a) (b) t ABSTRACT 90 140 Accurately tracking neuronal activity in behaving animals presents significant challenges due to complex motions and background noise. The lack of annotated datasets limits the evaluation and improvement of such tracking algorithms. To address this, we de- veloped SINETRA, a versatile simulator that generates synthetic tracking data for particles on a deformable background, closely mimicking live animal recordings. This simulator produces anno- (c) (d) tated 2D and 3D videos that reflect the intricate movements seen in behaving animals like Hydra Vulgaris. We evaluated four state- of-the-art tracking algorithms highlighting the current limitations of these methods in challenging scenarios and paving the way for improved cell tracking techniques in dynamic biological systems. Index Terms— Single neuron tracking, Simulation, Mechanical Fig. 1. Synthetic image simulator (Hydra flow) (a) A synthetic im- forces, Fluorescence noise, Behaving animals age using the default imaging parameters α = 0.2 and ∆ = 50 . (b) 1. INTRODUCTION Synthetic tracks on 50 frames centered around a contraction (frames 90 to 140, with a contraction at t = 100). (c) 100x100 pixels Numerous algorithms have been developed to track neurons in fluo- from fluorescence videos of Hydra Vulgaris’ neurons [13, 14]. (d) rescence live imaging of behaving animals, particularly in the worm 100x100 pixels from synthetic videos using different imaging pa- Caenorhabditis (C.) Elegans, Zebrafish or the freshwater cnidarian rameters. Left: ∆ = 15 , α = 0.5. Right: ∆ = 200 , α = 0.2. Hydra Vulgaris. These animals undergo significant body deforma- tions during behaviors such as worm crawling or Hydra’s contrac- fluorescence imaging of behaving animals. This simulator models tion and complex somersaulting [1]. A first class of tracking meth- particles (neurons) as spots within a structured background that ac- ods employs analytical approaches, such as Bayesian filtering and counts for the auto-fluorescence of the embedding tissue. The tissue global distance minimization, to track neuronal activity [2, 3]. These deformations related to animal behaviors are either modeled with a methods can be complemented with tracklet stitching to assemble system of damped harmonic oscillators subjected to random contrac- fragmented cell trajectories into continuous tracks [4, 5]. These tion or elongation forces, or are directly extracted from optical flow Bayesian frameworks can also integrate optical flow (OF) estima- estimates of live animal deformations. tion of cell velocity allowing for more precise updates of Kalman Using this simulation framework, we benchmarked four track- filters during tracking [6]. A second class of methods relies on deep ing algorithms on our synthetic data: eMHT [3] and u-track [2] that learning to perform image registration and establish neuron corre- are standard Bayesian ltering fi and distance minimization methods, spondence across different poses of the animal over time. [7, 8, 9]. KOFT [6] that uses optical flow to accurately estimate cell velocity To objectively compare cell tracking algorithms, several soft- in a Kalman filter and ZephIR [9] which propagates semi-annotated ware tools[10, 11] and annotated datasets[12] have been developed. tracks using image registration. This benchmark underscores the ex- However, these mainly address classical cell tracking challenges, fo- isting limitations of these approaches when dealing with complex cusing on randomly moving and dividing cells in vitro, which do not movements, opening the door for enhanced cell tracking methods in reflect the conditions of neuron tracking in behaving animals. Neu- dynamic biological systems. rons do not divide and their motion is driven by the animal’s body deformations. As a result, evaluating neuron tracking algorithms 2. METHOD typically requires labor-intensive manual annotations of time-lapse 2.1. Modeling fluorescent particles and background sequences, which is impractical for large datasets, especially in three dimensions[13, 9, 4]. 2.1.1. Particles To facilitate the objective comparison and improvement of SIn- gle NEurons TRAcking algorithms, we developed SINETRA, a ver- We first modeled realistic fluorescent particles (neurons). A parti- satile framework for simulating two- and three-dimensional time- cle is characterized by its position, intensity and shape, all of which lapse sequences reproducing the complex motions of neurons in live change over time. Formally, each particle 1 ≤ i ≤ N in frame arXiv:2411.09462v2 [cs.CV] 15 Nov 2024 This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. (a) (c) 1 ≤ t ≤ T is modeled with a d ∈ {2, 3} dimensional Gaussian pro- file, with x ∈ R its position, w ∈ [0, 1] its intensity weight, t,i t,i ++ ++ and Σ ∈ S its elliptic shape, where S is the ensemble of t,i d d symmetric and positive matrices in dimension d. On the first frame, we defined the animal’s body as the pixels B ⊂ Ω , where Ω = {z ∈ R , z ≥ 0 and z < S} is the simula- (b) tion domain and S ∈ N the dimensions of the generated images. This pixels B are either obtained from experimental videos using coarse segmentation of the tissue based on intensity thresholding or randomly sampled as an ellipse that covers around 30% of the sur- face/volume of the simulation domain Ω . The initial position of par- ticles x are uniformly distributed within B with the sparsity of t=0,i Fig. 2. Synthetic image simulator (Springs 2D/3D). (a): Visual ap- the distribution tuned via a minimal distance between particles (neu- pearance of the initial 2D frame with springs between control points rons). Particle intensity weights w are initially set to 1 and the t=0,i (in blue). (b): Motion induced by random forces and springs con- axes of elliptic shapes Σ are initially sampled with a size from t=0,i straints over time. The motion of each particle (bright spots) is deter- 1 to 3 pixels and randomly rotated (see section 2.1.4). mined by interpolating control points. (c): 3D projection of springs Let Θ = (w , x , Σ ) be the parameters of a particle i at t,i t,i t,i t,i (in blue) and noisy particles (in white). frame t. The noise-free image of fluorescent particles I is defined at pixel location z ∈ Ω as: 2.1.3. Intensity variations N N X X − 1 1 T p − (x − z) Σ (x − z) Our simulation framework allows to model the dynamics of particle t,i t,i 2 t,i I (z) = q(z; Θ ) = w e (1) t,i t,i w intensity. This is particularly useful to account for changes in t,i i=1 i=1 particle fluorescence due to the biophysics of the dye such as calcium indicators in fluorescence imaging of neural activity [15, 16]. For the with q(z; Θ ) the weighted Gaussian probability density function t,i sake of simplicity, we have fixed the intensity weights to a constant evaluated at pixel z. (∀i, j, t, w = w = 1) in all our experiments. t,i t,j Since the intensity weight w are constrained in [0, 1] and the t,i particles are separated by a minimal distance, the brightest pixels in p p I will not exceed the maximum gray-level value (∀z, I (z) < 1). 2.1.4. Shape evolution t t For each Gaussian profile (particles and background), the covariance 2.1.2. Background & Noise matrix of the profile can be separated into size and rotation compo- nents: We model the background fluorescence as multiple Gaussian pro- T 2 Σ = R Diag(σ ) R (4) b b b b t,i t,i θ θ t,i t,i files of parameters (Θ ) b = (w , x , Σ ) b . Posi- t,i t,i t,i t,i 1≤ i≤ N 1≤ i≤ N tions of the background Gaussian profiles are also distributed homo- where σ ∈ R are the sizes along each axes of the ellipse, and t,i geneously in B. The sizes of background profiles are larger than R is the rotation matrix of angle θ . In the 2D case, θ ∈ R is θ t,i t,i t,i tracked fluorescent particles, ranging from 20 to 60 pixels. The 3 a single angle of rotation around the z-axis. In 3D, θ ∈ R are the t,i noise-free background image I (z) is defined at pixel location z as: rotation angles around the x, y and z axes. We model the smooth random evolution of profiles over time around an equilibrium position with damped harmonic oscillators b b I (z) = q(z; Θ ) (2) t t,i with random forces (see Section 2.2.3). The equilibrium value of the i=1 rotation angle of a particle i is θ = θ ∼ U ([0, π ]). Then, to eq,i t=0,i simulate rotation noise, we suppose that random forces are applied The background profiles overlap extensively, leading to sums b to the particles, modeled as i.i.d. Gaussian noise. We choose the of intensities greater than the weights w . Without scaling, the t,i magnitude of the force so that θ ∼ N θ , . We emphasize b t,i eq,i brightest pixels of I would exceed the maximum gray-level value. that (θ ) are correlated in time and are computed by solving t,i 0≤ t≤ T Therefore, we normalize the background images by a fixed constant, the damped harmonic oscillators equations (not sampled from this which is defined as the maximum value of the first background im- Gaussian distribution). b b age: G = max I (z). The initial particles sizes σ are sampled from a Uni- t=0,i Particle and background signals are linearly mixed with a pro- form distribution (1 to 3 pixels for particles, 20 to 60 pixels for portion α . We model the fluctuation of the number of photons de- background profiles). Intuitively, larger particles should have tected at each pixel z ∈ Ω using a Poisson Shot Noise process with larger size variations. We therefore define the normalized size ∆ the integration time. The noisy image at frame t is generated as: σ t,i d as s = ∈ R . Random forces are applied to increase t,i t=0,i or decrease the normalized size around its equilibrium value of 1. (1 − α ) p b I (z) = αI (z) + I (z) t They are modeled as i.i.d. Gaussian noise and chosen so that the (3) s ∼ N (1, 0.05). t,i I (z) ∼ P ∆ I (z) t t 2.2. Motion modeling This allows to parametrize the simulator to produce faithful im- To model the motion of neurons embedded in a deformable tissue, ages: α controls the visibility of particles over the background, and we propose two elastic motion models. At each time-step t, we up- ∆ the magnitude of the Poisson shot noise (see Figure 1). By default, date the position of both particles x (neurons) and background pro- we use α = 0.20 and ∆ = 50 (Figure 1.a). files x using the computed tissue deformation. t=0 t=20 t=40 This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. 2.2.1. Optical flow motion Let x(t) ∈ R be the quantity of interest at time t, we have the following system: We estimate the animal deformation by computing optical flows be- tween consecutive frames of an experimental video. The optical x¨ (t) = f(t) − λ x˙(t) − k (x(t) − x ) (8) eq flows are computed with Farneback algorithm [17], an analytical al- where f(t) is a random force, λ the dampening coefficient, k the gorithm that proved to be fast and robust in many applications. The stiffness coefficient and x the equilibrium state of the system. The eq positions of neurons and background profiles are initially sampled mass of the system is canceled out for simplicity. within B that corresponds to the thresholded mask of the behaving The system is discretized with a time interval dt: x = x(ndt). animals’ body, then they are iteratively moved from the computed We decided to use the semi-implicit Euler method which is a sym- flow at each time-step (see Figure 1). plectic integration method suited for our Hamiltonian system. Start- ing from initial conditions x , x˙ , it computes the next state follow- 0 0 2.2.2. Springs motion ing: To simulate localized, unpredictable deformations of the animal’s x¨ = f − λ x˙ − k(x − x ) n n n n eq body, we employ a system of damped springs subjected to random x˙ = x˙ + dtx¨ (9) n+1 n n contraction and elongation forces. We position a grid of n control x = x + dtx˙ n+1 n n+1 points of equal mass attached by springs within the initial domain B For the processes modeled here, oscillations are not realistic. (see Figure 2). We then apply random forces to the control points and Therefore we decided to critically damp the system so that the tran- solve the n-body damped harmonic oscillator equations to estimate sient solution (without random forces) decays to the equilibrium the dynamics of control points along time (see Section 2.2.3). state. This implies to set λ = 2 k. Please note that the control points are not the particles and are Finally, we introduce the critical time τ = which corresponds not visible on the generated images. The positions of the particles to the exponential decay of the transient solution. We decided to use and background Gaussian profiles are computed with an elastic in- a reasonable value of τ = 10 frames: without forces, the system terpolation (Thin Plate Spline) between the n control points. returns to the equilibrium in around 10 frames. In all our springs, we More formally, let (p (t) ∈ R ) be the coordinates of i 1≤ i≤ n 2 1 have λ = and k = . the n control points of our system at time t. For each point p , we 2 τ τ define its dampening coefficient λ ∈ R . For each pair (p , p ), a i + i j 3. RESULTS AND DISCUSSIONS spring is created with a stiffness k ∈ R and an equilibrium length ij + eq 3.1. Evaluation l ∈ R . We set k = 0 (no spring) for all pairs except the 8 (26 in + ij ij 3D) closest neighbors. Let l (t) be the Euclidean distance between Using our simulator, we generated a synthetic dataset of fluorescent ij p and p at time t. The n-body damped harmonic oscillators can neurons in a behaving animal. We simulated 800 neurons over 200 i j be written: frames, using ∆ = 50 and α = 0.2. We focused on three different scenarios: springs rand p¨ (t) = f (t) − λ p˙ (t) + f (t) (5) i i i i i 1. Hydra Flow: we generated 1024x1024 images using optical- flow based motion. The flows are extracted from 200 selected p (t) − p (t) i j springs eq f (t) = − k l (t) − l (6) ij ij i ij frames of an experimental video where the Hydra Vulgaris is l (t) ij contracting [14]. rand where f (t) is a random contraction or elongation force applied on 2. Springs 2D: we used springs-based motion (with the amplitude control point i at time t. To compute this force, we start by randomly of random forces a = 4 pixels) to generate 1024x1024 im- max selecting a subset S(t) ⊂ J1, nK of control points, where |S(t)| ∼ ages. U (J2, mK) and m = 10 is the fixed maximum number of control 3. Springs 3D: we simulated 200x200x200 volumes with springs- points involved in the force. Then, we apply a contraction (resp. based motion (with a = 3 pixels). max elongation) force toward (resp. from) the barycenter of the selected For a set of ground-truth and computed tracks, we measure the ac- control points. Formally, letp¯ (t) = p (t) be the curacy of the tracking algorithm with the HOTA score [18] that esti- i∈S(t) |S(t)| mates both localization, detection and association performance. We barycenter of the selected control points. Then: used a tolerance distance η = 2 pixels (i.e. a predicted particle is never associated with a ground-truth particle distant by more than η 0 if i ∈/ S(t) pixels). rand f (t) = (7) The simulator and the tracking experiments are available p (t) − p¯ (t) i i d(t)a (t) otherwise alongside these synthetic datasets at https://github.com/ ||p (t) − p¯ (t)|| raphaelreme/SINETRA. where d(t) ∈ {− 1, 1} controls the random direction of the motion 3.2. Evaluated algorithms (contraction or elongation) and a (t) ∼ U( a , a ) is the ran- i max max Using this synthetic dataset, we compared four state-of-the-art dom amplitude of the motion. tracking algorithms that are based on distinct methodological frame- works: (1) u-track [2] from trackmate/Fiji software [19, 20] is 2.2.3. Damped harmonic oscillators a global distance minimization algorithm in two steps (frame-to- In our simulator, we rely on damped harmonic oscillators to model frame linking, tracklet stitching). We used the advanced version that the temporal evolution of positions, angles and sizes of particles. We model motion with Kalman filters. (2) eMHT [3] from Icy software detail here the calculations underlying these equations in the one- [21] is a probabilistic algorithm based on multiple motion models, dimensional case with a single spring, which benefits from theoreti- with a probabilistic handling of tracks and a multiple hypothesis cal background. For our multiple interacting springs (Section 2.2.2), association method. (3) KOFT [6] implemented in ByoTrack [13] we have extrapolated the results from this simpler case. is a probabilistic algorithm that exploits optical flow within Kalman This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Scenario Hydra Flow Springs 2D Springs 3D We also compared the computational efficiencies of the differ- u-track 67.7 ± 5.7% 80.3 ± 3.7% 70.8 ± 6.2% ent tracking methods. u-track is the fastest method, running at 25 eMHT 73.4 ± 8.1% 82.4 ± 2.8% 76.5 ± 6.4% FPS on a Lenovo Legion 5 laptop on our 2D/3D sequences. The KOFT 93.5 ± 2.1% 95.8 ± 0.6% 80.2 ± 5.8% multiple hypothesis association in eMHT is more expensive, and the ZephIR@3 76.4 ± 6.0% 74.6 ± 7.4% 51.5 ± 2.5% algorithm runs at 2 FPS on 2D/3D sequences. KOFT is slower in ZephIR@10 91.9 ± 3.2% 91.3 ± 1.7% 75.3 ± 2.7% 3D than in 2D because of the optical flow computations, running at 10 FPS in 2D and 2 FPS in 3D. Finally ZephIR is the slowest option Table 1. HOTA at 2 pixels [18] of the different tracking algorithms, because it needs to optimize its loss at each frame. It runs at 0.3 FPS on three synthetic scenarios. We report the mean and std on 5 dif- on CPU but is much faster on GPU, running at 3 FPS on a nvidia ferent random simulations sharing the same parameters (except the RTX3070. random seed). 4. CONCLUSION t t t 90 140 90 140 90 140 In this paper, we presented SINETRA, a versatile framework to gen- erate synthetic annotated single-particle-tracking datasets with a re- alistic imaging noise and complex motions such as those experienced by neurons in behaving animals like Hydra Vulgaris or C. elegans. ZephIR@3 u-track eMHT To model animal’s deformation and neurons’ motion, we either use t t t 90 140 90 140 90 140 optical flow estimates from experimental datasets, or model elastic deformations with a system of damped springs that experience ran- dom contraction and elongation forces. Through various simulation scenarios in both 2D or 3D, we as- sessed and highlighted the limitations of four different tracking algo- ZephIR@10 KOFT Ground-Truth rithms. We believe that developing realistic simulation frameworks is crucial for creating more robust tracking algorithms for in vivo Fig. 3. Temporal projection of tracks (Hydra flow scenario) on 50 monitoring of neuronal activity in behaving animals. frames centered around a contraction (frames 90 to 140, contraction at t = 100). 5. COMPLIANCE WITH ETHICAL STANDARDS This is a numerical simulation study for which no ethical approval filters. (4) ZephIR [9] is a frame by frame registration algorithm that was required. propagates partially annotated tracks. We tested ZephIR with three (ZephIR@3) and ten (ZephIR@10) annotated frames. 6. ACKNOWLEDGMENTS Each algorithm requires a pre-detection of the particles to track. We chose to rely on the wavelet thresholding method described in This work is supported by the Institut Pasteur and France-BioImaging [22] to detect particles for which we measured a detection f1 score Infrastructure (ANR-10-INBS-04). R.R and T.L. are supported by of around 80% across our synthetic dataset. the ANR (ANR-21-CE45-0020-01 REBIRTH). None of the authors declare to have a financial conflict of interest 3.3. Results in the results of this study. Table 1 summarizes the performance of the different tracking algo- 7. REFERENCES rithms on the three simulation scenarios. u-track [2, 19, 20] and eMHT [3, 21] model a near-constant velocity of tracked particles, [1] W. Yamamoto and R. Yuste, “Peptide-driven control of som- and are less robust to large and sudden contractions that occurs in ersaulting in hydra vulgaris,” Current Biology, vol. 33, no. 10, behaving animals compared to KOFT and ZephIR (see Figure 3). pp. 1893–1905, 2023. u-track reaches an averaged of 67.7% HOTA on the Hydra flow sce- [2] K. Jaqaman, D. Loerke, M. Mettlen, et al., “Robust single- nario, whereas eMHT is slighlty more robust thanks to its multiple particle tracking in live-cell time-lapse sequences,” Nature hypothesis association and reaches 73.4% HOTA. Methods, vol. 5, no. 8, pp. 695–702, 2008. ZephIR registration method is more resilient to the animal’s large motions thanks to well chosen annotated frames [9]. Using 3 an- [3] N. Chenouard, I. Bloch, and J.-C. Olivo-Marin, “Multiple hypothesis tracking for cluttered biological image sequences,” notated frames in elongated and contracted animals, it outperforms u-track and eMHT on Hydra Flow scenario, reaching 80.9% HOTA. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 35, no. 11, pp. 2736–3750, 2013. Yet, on springs-based videos where contractions are more frequent, ZephIR requires more annotated frames to outperforms u-track and [4] T. Lagache, A. Hanson, J. Perez-Orte ´ ga, et al., “Tracking cal- eMHT. We emphasize that frame annotating is tedious in experimen- cium dynamics from individual neurons in behaving animals,” tal datasets, limiting the capabilities of ZephIR in highly deforming PLoS Computational Biology, vol. 17, pp. e1009432, 10 2021. organisms like Hydra. Finally, KOFT [6, 13] outperformed the other [5] R. Reme, V. Piriou, A. Hanson, et al., “Tracking intermittent methods across the different simulations, due to its precise estima- particles with self-learned visual features,” in IEEE Interna- tion of neuronal flow motion and the use of Kalman filters, both tional Symposium on Biomedical Imaging (ISBI). IEEE, 2023, of which significantly improve particle tracking. (see Figure 3). It pp. 1–5. reaches above 90% HOTA on 2D scenarios. However, we noticed that it had difficulty tracking neurons during the animal’s fastest [6] R. Reme, A. Newson, E. Angelini, et al., “Particle tracking movements when neurons were densely packed, or in 3D where its in biological images with optical-flow enhanced kalman filter- optical flow estimates are less precise (80.2 % on springs 3D sce- ing,” in 2024 IEEE International Symposium on Biomedical nario). Imaging (ISBI), 2024, pp. 1–5. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. [7] X. Yu, M. S. Creamer, F. Randi, et al., “Fast deep neural cor- respondence for tracking and identifying neurons in c. elegans using semi-synthetic training,” Elife, vol. 10, pp. e66410, 2021. [8] C. F. Park, M. Barzegar-Keshteli, K. Korchagina, et al., “Auto- mated neuron tracking inside moving and deforming c. elegans using deep learning and targeted augmentation,” Nature Meth- ods, vol. 21, no. 1, pp. 142–149, 2024. [9] J. Ryu, A. Nejatbakhsh, M. Torkashvand, et al., “Versatile mul- tiple object tracking in sparse 2d/3d videos via deformable im- age registration,” PLoS Computational Biology, vol. 20, no. 5, pp. e1012075, 2024. [10] D. Svoboda and V. Ulman, “Mitogen: a framework for gen- erating 3d synthetic time-lapse sequences of cell populations in fluorescence microscopy,” IEEE Transactions on Medical Imaging, vol. 36, no. 1, pp. 310–321, 2016. [11] A. Matyjaszkiewicz, G. Fiore, F. Annunziata, et al., “Bsim 2.0: an advanced agent-based cell simulator,” ACS Synthetic Biology, vol. 6, no. 10, pp. 1969–1972, 2017. [12] V. Ulman, M. Maska, ˇ K. E. Magnusson, et al., “An objective comparison of cell-tracking algorithms,” Nature Methods, vol. 14, no. 12, pp. 1141–1152, 2017. [13] A. Hanson, R. Reme, N. Telerman, et al., “Automatic moni- toring of neural activity with single-cell resolution in behaving hydra,” Scientific Reports , vol. 14, no. 1, pp. 5083, 2024. [14] C. Dupre and R. Yuste, “Non-overlapping neural networks in hydra vulgaris,” Current Biology, vol. 27, no. 8, pp. 1085– 1097, 2017. [15] J. Nakai, M. Ohkura, and K. Imoto, “A high signal-to-noise ca2+ probe composed of a single green fluorescent protein,” Nature Biotechnology, vol. 19, no. 2, pp. 137–141, 2001. [16] R. Yuste, “Fluorescence microscopy today,” Nature Methods, vol. 2, no. 12, pp. 902–904, 2005. [17] G. Farneback, ¨ “Two-frame motion estimation based on poly- nomial expansion,” in Image Analysis: 13th Scandinavian Conference, SCIA. Springer, 2003, pp. 363–370. [18] J. Luiten, A. Osep, P. Dendorfer, et al., “Hota: A higher or- der metric for evaluating multi-object tracking,” International Journal of Computer Vision, vol. 129, pp. 548–578, 2021. [19] J. Schindelin, I. Arganda-Carreras, E. Frise, et al., “Fiji: an open-source platform for biological-image analysis,” Nature Methods, vol. 9, no. 7, pp. 676–682, 2012. [20] J.-Y. Tinevez, N. Perry, J. Schindelin, et al., “Trackmate: An open and extensible platform for single-particle tracking,” Methods, vol. 115, pp. 80–90, 2017. [21] F. De Chaumont, S. Dallongeville, N. Chenouard, et al., “Icy: an open bioimage informatics platform for extended repro- ducible research,” Nature Methods, vol. 9, no. 7, pp. 690–696, [22] J.-C. Olivo-Marin, “Extraction of spots in biological images using multiscale products,” Pattern Recognition, vol. 35, no. 9, pp. 1989–1996, 2002. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Computing Research Repository arXiv (Cornell University)

SINETRA: a Versatile Framework for Evaluating Single Neuron Tracking in Behaving Animals

Download PDF
5 pages

Loading next page...
 
/lp/arxiv-cornell-university/sinetra-a-versatile-framework-for-evaluating-single-neuron-tracking-in-rbsavR7SNt

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

eISSN
ARCH-3344
DOI
10.48550/arxiv.2411.09462
Publisher site
See Article on Publisher Site

Abstract

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. SINETRA: A VERSATILE FRAMEWORK FOR EVALUATING SINGLE NEURON TRACKING IN BEHAVING ANIMALS ⋆† † † Raphael Reme Alasdair Newson Elsa Angelini ⋆ ⋆ Jean-Christophe Olivo-Marin Thibault Lagache Institut Pasteur, Universite ´ de Paris-Cite, ´ CNRS UMR 3691, BioImage Analysis Unit F-75015 Paris, France LTCI, Telecom Paris, Institut Polytechnique de Paris, France Corresponding author: [email protected] (a) (b) t ABSTRACT 90 140 Accurately tracking neuronal activity in behaving animals presents significant challenges due to complex motions and background noise. The lack of annotated datasets limits the evaluation and improvement of such tracking algorithms. To address this, we de- veloped SINETRA, a versatile simulator that generates synthetic tracking data for particles on a deformable background, closely mimicking live animal recordings. This simulator produces anno- (c) (d) tated 2D and 3D videos that reflect the intricate movements seen in behaving animals like Hydra Vulgaris. We evaluated four state- of-the-art tracking algorithms highlighting the current limitations of these methods in challenging scenarios and paving the way for improved cell tracking techniques in dynamic biological systems. Index Terms— Single neuron tracking, Simulation, Mechanical Fig. 1. Synthetic image simulator (Hydra flow) (a) A synthetic im- forces, Fluorescence noise, Behaving animals age using the default imaging parameters α = 0.2 and ∆ = 50 . (b) 1. INTRODUCTION Synthetic tracks on 50 frames centered around a contraction (frames 90 to 140, with a contraction at t = 100). (c) 100x100 pixels Numerous algorithms have been developed to track neurons in fluo- from fluorescence videos of Hydra Vulgaris’ neurons [13, 14]. (d) rescence live imaging of behaving animals, particularly in the worm 100x100 pixels from synthetic videos using different imaging pa- Caenorhabditis (C.) Elegans, Zebrafish or the freshwater cnidarian rameters. Left: ∆ = 15 , α = 0.5. Right: ∆ = 200 , α = 0.2. Hydra Vulgaris. These animals undergo significant body deforma- tions during behaviors such as worm crawling or Hydra’s contrac- fluorescence imaging of behaving animals. This simulator models tion and complex somersaulting [1]. A first class of tracking meth- particles (neurons) as spots within a structured background that ac- ods employs analytical approaches, such as Bayesian filtering and counts for the auto-fluorescence of the embedding tissue. The tissue global distance minimization, to track neuronal activity [2, 3]. These deformations related to animal behaviors are either modeled with a methods can be complemented with tracklet stitching to assemble system of damped harmonic oscillators subjected to random contrac- fragmented cell trajectories into continuous tracks [4, 5]. These tion or elongation forces, or are directly extracted from optical flow Bayesian frameworks can also integrate optical flow (OF) estima- estimates of live animal deformations. tion of cell velocity allowing for more precise updates of Kalman Using this simulation framework, we benchmarked four track- filters during tracking [6]. A second class of methods relies on deep ing algorithms on our synthetic data: eMHT [3] and u-track [2] that learning to perform image registration and establish neuron corre- are standard Bayesian ltering fi and distance minimization methods, spondence across different poses of the animal over time. [7, 8, 9]. KOFT [6] that uses optical flow to accurately estimate cell velocity To objectively compare cell tracking algorithms, several soft- in a Kalman filter and ZephIR [9] which propagates semi-annotated ware tools[10, 11] and annotated datasets[12] have been developed. tracks using image registration. This benchmark underscores the ex- However, these mainly address classical cell tracking challenges, fo- isting limitations of these approaches when dealing with complex cusing on randomly moving and dividing cells in vitro, which do not movements, opening the door for enhanced cell tracking methods in reflect the conditions of neuron tracking in behaving animals. Neu- dynamic biological systems. rons do not divide and their motion is driven by the animal’s body deformations. As a result, evaluating neuron tracking algorithms 2. METHOD typically requires labor-intensive manual annotations of time-lapse 2.1. Modeling fluorescent particles and background sequences, which is impractical for large datasets, especially in three dimensions[13, 9, 4]. 2.1.1. Particles To facilitate the objective comparison and improvement of SIn- gle NEurons TRAcking algorithms, we developed SINETRA, a ver- We first modeled realistic fluorescent particles (neurons). A parti- satile framework for simulating two- and three-dimensional time- cle is characterized by its position, intensity and shape, all of which lapse sequences reproducing the complex motions of neurons in live change over time. Formally, each particle 1 ≤ i ≤ N in frame arXiv:2411.09462v2 [cs.CV] 15 Nov 2024 This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. (a) (c) 1 ≤ t ≤ T is modeled with a d ∈ {2, 3} dimensional Gaussian pro- file, with x ∈ R its position, w ∈ [0, 1] its intensity weight, t,i t,i ++ ++ and Σ ∈ S its elliptic shape, where S is the ensemble of t,i d d symmetric and positive matrices in dimension d. On the first frame, we defined the animal’s body as the pixels B ⊂ Ω , where Ω = {z ∈ R , z ≥ 0 and z < S} is the simula- (b) tion domain and S ∈ N the dimensions of the generated images. This pixels B are either obtained from experimental videos using coarse segmentation of the tissue based on intensity thresholding or randomly sampled as an ellipse that covers around 30% of the sur- face/volume of the simulation domain Ω . The initial position of par- ticles x are uniformly distributed within B with the sparsity of t=0,i Fig. 2. Synthetic image simulator (Springs 2D/3D). (a): Visual ap- the distribution tuned via a minimal distance between particles (neu- pearance of the initial 2D frame with springs between control points rons). Particle intensity weights w are initially set to 1 and the t=0,i (in blue). (b): Motion induced by random forces and springs con- axes of elliptic shapes Σ are initially sampled with a size from t=0,i straints over time. The motion of each particle (bright spots) is deter- 1 to 3 pixels and randomly rotated (see section 2.1.4). mined by interpolating control points. (c): 3D projection of springs Let Θ = (w , x , Σ ) be the parameters of a particle i at t,i t,i t,i t,i (in blue) and noisy particles (in white). frame t. The noise-free image of fluorescent particles I is defined at pixel location z ∈ Ω as: 2.1.3. Intensity variations N N X X − 1 1 T p − (x − z) Σ (x − z) Our simulation framework allows to model the dynamics of particle t,i t,i 2 t,i I (z) = q(z; Θ ) = w e (1) t,i t,i w intensity. This is particularly useful to account for changes in t,i i=1 i=1 particle fluorescence due to the biophysics of the dye such as calcium indicators in fluorescence imaging of neural activity [15, 16]. For the with q(z; Θ ) the weighted Gaussian probability density function t,i sake of simplicity, we have fixed the intensity weights to a constant evaluated at pixel z. (∀i, j, t, w = w = 1) in all our experiments. t,i t,j Since the intensity weight w are constrained in [0, 1] and the t,i particles are separated by a minimal distance, the brightest pixels in p p I will not exceed the maximum gray-level value (∀z, I (z) < 1). 2.1.4. Shape evolution t t For each Gaussian profile (particles and background), the covariance 2.1.2. Background & Noise matrix of the profile can be separated into size and rotation compo- nents: We model the background fluorescence as multiple Gaussian pro- T 2 Σ = R Diag(σ ) R (4) b b b b t,i t,i θ θ t,i t,i files of parameters (Θ ) b = (w , x , Σ ) b . Posi- t,i t,i t,i t,i 1≤ i≤ N 1≤ i≤ N tions of the background Gaussian profiles are also distributed homo- where σ ∈ R are the sizes along each axes of the ellipse, and t,i geneously in B. The sizes of background profiles are larger than R is the rotation matrix of angle θ . In the 2D case, θ ∈ R is θ t,i t,i t,i tracked fluorescent particles, ranging from 20 to 60 pixels. The 3 a single angle of rotation around the z-axis. In 3D, θ ∈ R are the t,i noise-free background image I (z) is defined at pixel location z as: rotation angles around the x, y and z axes. We model the smooth random evolution of profiles over time around an equilibrium position with damped harmonic oscillators b b I (z) = q(z; Θ ) (2) t t,i with random forces (see Section 2.2.3). The equilibrium value of the i=1 rotation angle of a particle i is θ = θ ∼ U ([0, π ]). Then, to eq,i t=0,i simulate rotation noise, we suppose that random forces are applied The background profiles overlap extensively, leading to sums b to the particles, modeled as i.i.d. Gaussian noise. We choose the of intensities greater than the weights w . Without scaling, the t,i magnitude of the force so that θ ∼ N θ , . We emphasize b t,i eq,i brightest pixels of I would exceed the maximum gray-level value. that (θ ) are correlated in time and are computed by solving t,i 0≤ t≤ T Therefore, we normalize the background images by a fixed constant, the damped harmonic oscillators equations (not sampled from this which is defined as the maximum value of the first background im- Gaussian distribution). b b age: G = max I (z). The initial particles sizes σ are sampled from a Uni- t=0,i Particle and background signals are linearly mixed with a pro- form distribution (1 to 3 pixels for particles, 20 to 60 pixels for portion α . We model the fluctuation of the number of photons de- background profiles). Intuitively, larger particles should have tected at each pixel z ∈ Ω using a Poisson Shot Noise process with larger size variations. We therefore define the normalized size ∆ the integration time. The noisy image at frame t is generated as: σ t,i d as s = ∈ R . Random forces are applied to increase t,i t=0,i or decrease the normalized size around its equilibrium value of 1. (1 − α ) p b I (z) = αI (z) + I (z) t They are modeled as i.i.d. Gaussian noise and chosen so that the (3) s ∼ N (1, 0.05). t,i I (z) ∼ P ∆ I (z) t t 2.2. Motion modeling This allows to parametrize the simulator to produce faithful im- To model the motion of neurons embedded in a deformable tissue, ages: α controls the visibility of particles over the background, and we propose two elastic motion models. At each time-step t, we up- ∆ the magnitude of the Poisson shot noise (see Figure 1). By default, date the position of both particles x (neurons) and background pro- we use α = 0.20 and ∆ = 50 (Figure 1.a). files x using the computed tissue deformation. t=0 t=20 t=40 This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. 2.2.1. Optical flow motion Let x(t) ∈ R be the quantity of interest at time t, we have the following system: We estimate the animal deformation by computing optical flows be- tween consecutive frames of an experimental video. The optical x¨ (t) = f(t) − λ x˙(t) − k (x(t) − x ) (8) eq flows are computed with Farneback algorithm [17], an analytical al- where f(t) is a random force, λ the dampening coefficient, k the gorithm that proved to be fast and robust in many applications. The stiffness coefficient and x the equilibrium state of the system. The eq positions of neurons and background profiles are initially sampled mass of the system is canceled out for simplicity. within B that corresponds to the thresholded mask of the behaving The system is discretized with a time interval dt: x = x(ndt). animals’ body, then they are iteratively moved from the computed We decided to use the semi-implicit Euler method which is a sym- flow at each time-step (see Figure 1). plectic integration method suited for our Hamiltonian system. Start- ing from initial conditions x , x˙ , it computes the next state follow- 0 0 2.2.2. Springs motion ing: To simulate localized, unpredictable deformations of the animal’s x¨ = f − λ x˙ − k(x − x ) n n n n eq body, we employ a system of damped springs subjected to random x˙ = x˙ + dtx¨ (9) n+1 n n contraction and elongation forces. We position a grid of n control x = x + dtx˙ n+1 n n+1 points of equal mass attached by springs within the initial domain B For the processes modeled here, oscillations are not realistic. (see Figure 2). We then apply random forces to the control points and Therefore we decided to critically damp the system so that the tran- solve the n-body damped harmonic oscillator equations to estimate sient solution (without random forces) decays to the equilibrium the dynamics of control points along time (see Section 2.2.3). state. This implies to set λ = 2 k. Please note that the control points are not the particles and are Finally, we introduce the critical time τ = which corresponds not visible on the generated images. The positions of the particles to the exponential decay of the transient solution. We decided to use and background Gaussian profiles are computed with an elastic in- a reasonable value of τ = 10 frames: without forces, the system terpolation (Thin Plate Spline) between the n control points. returns to the equilibrium in around 10 frames. In all our springs, we More formally, let (p (t) ∈ R ) be the coordinates of i 1≤ i≤ n 2 1 have λ = and k = . the n control points of our system at time t. For each point p , we 2 τ τ define its dampening coefficient λ ∈ R . For each pair (p , p ), a i + i j 3. RESULTS AND DISCUSSIONS spring is created with a stiffness k ∈ R and an equilibrium length ij + eq 3.1. Evaluation l ∈ R . We set k = 0 (no spring) for all pairs except the 8 (26 in + ij ij 3D) closest neighbors. Let l (t) be the Euclidean distance between Using our simulator, we generated a synthetic dataset of fluorescent ij p and p at time t. The n-body damped harmonic oscillators can neurons in a behaving animal. We simulated 800 neurons over 200 i j be written: frames, using ∆ = 50 and α = 0.2. We focused on three different scenarios: springs rand p¨ (t) = f (t) − λ p˙ (t) + f (t) (5) i i i i i 1. Hydra Flow: we generated 1024x1024 images using optical- flow based motion. The flows are extracted from 200 selected p (t) − p (t) i j springs eq f (t) = − k l (t) − l (6) ij ij i ij frames of an experimental video where the Hydra Vulgaris is l (t) ij contracting [14]. rand where f (t) is a random contraction or elongation force applied on 2. Springs 2D: we used springs-based motion (with the amplitude control point i at time t. To compute this force, we start by randomly of random forces a = 4 pixels) to generate 1024x1024 im- max selecting a subset S(t) ⊂ J1, nK of control points, where |S(t)| ∼ ages. U (J2, mK) and m = 10 is the fixed maximum number of control 3. Springs 3D: we simulated 200x200x200 volumes with springs- points involved in the force. Then, we apply a contraction (resp. based motion (with a = 3 pixels). max elongation) force toward (resp. from) the barycenter of the selected For a set of ground-truth and computed tracks, we measure the ac- control points. Formally, letp¯ (t) = p (t) be the curacy of the tracking algorithm with the HOTA score [18] that esti- i∈S(t) |S(t)| mates both localization, detection and association performance. We barycenter of the selected control points. Then: used a tolerance distance η = 2 pixels (i.e. a predicted particle is never associated with a ground-truth particle distant by more than η 0 if i ∈/ S(t) pixels). rand f (t) = (7) The simulator and the tracking experiments are available p (t) − p¯ (t) i i d(t)a (t) otherwise alongside these synthetic datasets at https://github.com/ ||p (t) − p¯ (t)|| raphaelreme/SINETRA. where d(t) ∈ {− 1, 1} controls the random direction of the motion 3.2. Evaluated algorithms (contraction or elongation) and a (t) ∼ U( a , a ) is the ran- i max max Using this synthetic dataset, we compared four state-of-the-art dom amplitude of the motion. tracking algorithms that are based on distinct methodological frame- works: (1) u-track [2] from trackmate/Fiji software [19, 20] is 2.2.3. Damped harmonic oscillators a global distance minimization algorithm in two steps (frame-to- In our simulator, we rely on damped harmonic oscillators to model frame linking, tracklet stitching). We used the advanced version that the temporal evolution of positions, angles and sizes of particles. We model motion with Kalman filters. (2) eMHT [3] from Icy software detail here the calculations underlying these equations in the one- [21] is a probabilistic algorithm based on multiple motion models, dimensional case with a single spring, which benefits from theoreti- with a probabilistic handling of tracks and a multiple hypothesis cal background. For our multiple interacting springs (Section 2.2.2), association method. (3) KOFT [6] implemented in ByoTrack [13] we have extrapolated the results from this simpler case. is a probabilistic algorithm that exploits optical flow within Kalman This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Scenario Hydra Flow Springs 2D Springs 3D We also compared the computational efficiencies of the differ- u-track 67.7 ± 5.7% 80.3 ± 3.7% 70.8 ± 6.2% ent tracking methods. u-track is the fastest method, running at 25 eMHT 73.4 ± 8.1% 82.4 ± 2.8% 76.5 ± 6.4% FPS on a Lenovo Legion 5 laptop on our 2D/3D sequences. The KOFT 93.5 ± 2.1% 95.8 ± 0.6% 80.2 ± 5.8% multiple hypothesis association in eMHT is more expensive, and the ZephIR@3 76.4 ± 6.0% 74.6 ± 7.4% 51.5 ± 2.5% algorithm runs at 2 FPS on 2D/3D sequences. KOFT is slower in ZephIR@10 91.9 ± 3.2% 91.3 ± 1.7% 75.3 ± 2.7% 3D than in 2D because of the optical flow computations, running at 10 FPS in 2D and 2 FPS in 3D. Finally ZephIR is the slowest option Table 1. HOTA at 2 pixels [18] of the different tracking algorithms, because it needs to optimize its loss at each frame. It runs at 0.3 FPS on three synthetic scenarios. We report the mean and std on 5 dif- on CPU but is much faster on GPU, running at 3 FPS on a nvidia ferent random simulations sharing the same parameters (except the RTX3070. random seed). 4. CONCLUSION t t t 90 140 90 140 90 140 In this paper, we presented SINETRA, a versatile framework to gen- erate synthetic annotated single-particle-tracking datasets with a re- alistic imaging noise and complex motions such as those experienced by neurons in behaving animals like Hydra Vulgaris or C. elegans. ZephIR@3 u-track eMHT To model animal’s deformation and neurons’ motion, we either use t t t 90 140 90 140 90 140 optical flow estimates from experimental datasets, or model elastic deformations with a system of damped springs that experience ran- dom contraction and elongation forces. Through various simulation scenarios in both 2D or 3D, we as- sessed and highlighted the limitations of four different tracking algo- ZephIR@10 KOFT Ground-Truth rithms. We believe that developing realistic simulation frameworks is crucial for creating more robust tracking algorithms for in vivo Fig. 3. Temporal projection of tracks (Hydra flow scenario) on 50 monitoring of neuronal activity in behaving animals. frames centered around a contraction (frames 90 to 140, contraction at t = 100). 5. COMPLIANCE WITH ETHICAL STANDARDS This is a numerical simulation study for which no ethical approval filters. (4) ZephIR [9] is a frame by frame registration algorithm that was required. propagates partially annotated tracks. We tested ZephIR with three (ZephIR@3) and ten (ZephIR@10) annotated frames. 6. ACKNOWLEDGMENTS Each algorithm requires a pre-detection of the particles to track. We chose to rely on the wavelet thresholding method described in This work is supported by the Institut Pasteur and France-BioImaging [22] to detect particles for which we measured a detection f1 score Infrastructure (ANR-10-INBS-04). R.R and T.L. are supported by of around 80% across our synthetic dataset. the ANR (ANR-21-CE45-0020-01 REBIRTH). None of the authors declare to have a financial conflict of interest 3.3. Results in the results of this study. Table 1 summarizes the performance of the different tracking algo- 7. REFERENCES rithms on the three simulation scenarios. u-track [2, 19, 20] and eMHT [3, 21] model a near-constant velocity of tracked particles, [1] W. Yamamoto and R. Yuste, “Peptide-driven control of som- and are less robust to large and sudden contractions that occurs in ersaulting in hydra vulgaris,” Current Biology, vol. 33, no. 10, behaving animals compared to KOFT and ZephIR (see Figure 3). pp. 1893–1905, 2023. u-track reaches an averaged of 67.7% HOTA on the Hydra flow sce- [2] K. Jaqaman, D. Loerke, M. Mettlen, et al., “Robust single- nario, whereas eMHT is slighlty more robust thanks to its multiple particle tracking in live-cell time-lapse sequences,” Nature hypothesis association and reaches 73.4% HOTA. Methods, vol. 5, no. 8, pp. 695–702, 2008. ZephIR registration method is more resilient to the animal’s large motions thanks to well chosen annotated frames [9]. Using 3 an- [3] N. Chenouard, I. Bloch, and J.-C. Olivo-Marin, “Multiple hypothesis tracking for cluttered biological image sequences,” notated frames in elongated and contracted animals, it outperforms u-track and eMHT on Hydra Flow scenario, reaching 80.9% HOTA. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 35, no. 11, pp. 2736–3750, 2013. Yet, on springs-based videos where contractions are more frequent, ZephIR requires more annotated frames to outperforms u-track and [4] T. Lagache, A. Hanson, J. Perez-Orte ´ ga, et al., “Tracking cal- eMHT. We emphasize that frame annotating is tedious in experimen- cium dynamics from individual neurons in behaving animals,” tal datasets, limiting the capabilities of ZephIR in highly deforming PLoS Computational Biology, vol. 17, pp. e1009432, 10 2021. organisms like Hydra. Finally, KOFT [6, 13] outperformed the other [5] R. Reme, V. Piriou, A. Hanson, et al., “Tracking intermittent methods across the different simulations, due to its precise estima- particles with self-learned visual features,” in IEEE Interna- tion of neuronal flow motion and the use of Kalman filters, both tional Symposium on Biomedical Imaging (ISBI). IEEE, 2023, of which significantly improve particle tracking. (see Figure 3). It pp. 1–5. reaches above 90% HOTA on 2D scenarios. However, we noticed that it had difficulty tracking neurons during the animal’s fastest [6] R. Reme, A. Newson, E. Angelini, et al., “Particle tracking movements when neurons were densely packed, or in 3D where its in biological images with optical-flow enhanced kalman filter- optical flow estimates are less precise (80.2 % on springs 3D sce- ing,” in 2024 IEEE International Symposium on Biomedical nario). Imaging (ISBI), 2024, pp. 1–5. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. [7] X. Yu, M. S. Creamer, F. Randi, et al., “Fast deep neural cor- respondence for tracking and identifying neurons in c. elegans using semi-synthetic training,” Elife, vol. 10, pp. e66410, 2021. [8] C. F. Park, M. Barzegar-Keshteli, K. Korchagina, et al., “Auto- mated neuron tracking inside moving and deforming c. elegans using deep learning and targeted augmentation,” Nature Meth- ods, vol. 21, no. 1, pp. 142–149, 2024. [9] J. Ryu, A. Nejatbakhsh, M. Torkashvand, et al., “Versatile mul- tiple object tracking in sparse 2d/3d videos via deformable im- age registration,” PLoS Computational Biology, vol. 20, no. 5, pp. e1012075, 2024. [10] D. Svoboda and V. Ulman, “Mitogen: a framework for gen- erating 3d synthetic time-lapse sequences of cell populations in fluorescence microscopy,” IEEE Transactions on Medical Imaging, vol. 36, no. 1, pp. 310–321, 2016. [11] A. Matyjaszkiewicz, G. Fiore, F. Annunziata, et al., “Bsim 2.0: an advanced agent-based cell simulator,” ACS Synthetic Biology, vol. 6, no. 10, pp. 1969–1972, 2017. [12] V. Ulman, M. Maska, ˇ K. E. Magnusson, et al., “An objective comparison of cell-tracking algorithms,” Nature Methods, vol. 14, no. 12, pp. 1141–1152, 2017. [13] A. Hanson, R. Reme, N. Telerman, et al., “Automatic moni- toring of neural activity with single-cell resolution in behaving hydra,” Scientific Reports , vol. 14, no. 1, pp. 5083, 2024. [14] C. Dupre and R. Yuste, “Non-overlapping neural networks in hydra vulgaris,” Current Biology, vol. 27, no. 8, pp. 1085– 1097, 2017. [15] J. Nakai, M. Ohkura, and K. Imoto, “A high signal-to-noise ca2+ probe composed of a single green fluorescent protein,” Nature Biotechnology, vol. 19, no. 2, pp. 137–141, 2001. [16] R. Yuste, “Fluorescence microscopy today,” Nature Methods, vol. 2, no. 12, pp. 902–904, 2005. [17] G. Farneback, ¨ “Two-frame motion estimation based on poly- nomial expansion,” in Image Analysis: 13th Scandinavian Conference, SCIA. Springer, 2003, pp. 363–370. [18] J. Luiten, A. Osep, P. Dendorfer, et al., “Hota: A higher or- der metric for evaluating multi-object tracking,” International Journal of Computer Vision, vol. 129, pp. 548–578, 2021. [19] J. Schindelin, I. Arganda-Carreras, E. Frise, et al., “Fiji: an open-source platform for biological-image analysis,” Nature Methods, vol. 9, no. 7, pp. 676–682, 2012. [20] J.-Y. Tinevez, N. Perry, J. Schindelin, et al., “Trackmate: An open and extensible platform for single-particle tracking,” Methods, vol. 115, pp. 80–90, 2017. [21] F. De Chaumont, S. Dallongeville, N. Chenouard, et al., “Icy: an open bioimage informatics platform for extended repro- ducible research,” Nature Methods, vol. 9, no. 7, pp. 690–696, [22] J.-C. Olivo-Marin, “Extraction of spots in biological images using multiscale products,” Pattern Recognition, vol. 35, no. 9, pp. 1989–1996, 2002.

Journal

Computing Research Repository – arXiv (Cornell University)

Published: Nov 15, 2024

There are no references for this article.