(648f) An Objective Method Screening Approach for Optimizing Cell Tracking and Identity Annotation in Dense Fluorescent Microscopic Images

C. elegans is an extremely useful model organism in understanding neural basis of behaviors because of several advantageous features. The nematode has a small and fully mapped nervous system, displays complex behaviors, and is amenable to fluorescence imaging. Recent advances in microscopy have enabled simultaneous recording of activities of the entire brain of C. elegans thus producing a mine of rich datasets. However to extract useful information from these datasets, several challenges must be overcome. First, neurons must be detected and tracked in the 4-dimensional (3D + t) videos as the animal non-rigidly deforms across the video frames. This is required for extraction of meaningful calcium signals (neuron activity data) from videos. Second, biological identities of cells tracked in the video must be determined. This enables comparing neuron activities across animals and experimental conditions, and incorporating pre-existing knowledge in literature. Automated methods are needed for both tasks to handle large datasets and accelerate knowledge discovery.

While several machine learning and computer vision based methods have been proposed for the two tasks – tracking and annotation, there is no general consensus among which methods work, why the methods work, when they work, which methods are robust to noises in data etc. due to a lack of systematic comparison. To address these questions we performed a method screen comparing 22 different methods (combination of previously unexplored methods and previously proposed methods) across 6 different accuracy metrics on synthetic datasets with realistic properties across a range of noise levels (3 types of noise simulated). The methods compared included 8 registration-based methods (with linear objective functions) and 14 graph-matching based methods (with quadratic objective functions). For tracking task, methods were also compared for 3 different track linking strategies – sequential tracking, all frames matched to one frame from video, all frames matched to an atlas.

Systematic comparison reveals several important conclusions. 1) All methods perform well on data with low amounts of noise; 2) as expected linear methods are orders of magnitude faster than quadratic methods; 3) quadratic methods – CRF[1], IPFP-MAP[2] and L2QP-MAP[3] are robust to noise maintaining high accuracy at high noise levels, in comparison accuracy for registration based methods (linear methods) fall sharply with noise in data; 4) for some but not all quadratic methods, accuracy is heavily dependent on pairwise edge features used in the method thus highlighting the need for appropriate feature selection; 5) among tracking strategies sequential tracking performs worst, strategy where all frames are matched to an atlas frame performs best, better than when frames are matched to a randomly selected video frame.

This work addresses important methodological needs in quantitative biology, particularly in neuroscience. It points to unbiased evaluation and choice of the best method for cell tracking and annotation tasks in microscopic images, based on data properties. Insights generated by our comparison based approach will guide development of cell tracking and identity annotation methods in future. Further we make our code freely available thus providing a model zoo for future researchers to compare and optimize their methods.

1. Chaudhary, S., Lee, S. A., Li, Y., Patel, D. S., & Lu, H. (2021). Graphical-model framework for automated annotation of cell identities in dense cellular images. ELife, 10, e60321. https://doi.org/10.7554/eLife.60321
2. Leordeanu, M., Hebert, M., & Sukthankar, R. (2009). An Integer Projected Fixed Point Method for Graph Matching and MAP Inference. In Y. Bengio, D. Schuurmans, J. Lafferty, C. Williams, & A. Culotta (Eds.), Advances in Neural Information Processing Systems (Vol. 22). Curran Associates, Inc. https://proceedings.neurips.cc/paper/2009/file/fc2c7c47b918d0c2d792a719dfb602ef-Paper.pdf
3. Leordeanu, M., & Hebert, M. (2006). Efficient MAP Approximation for Dense Energy Functions. Proceedings of the 23rd International Conference on Machine Learning, 545–552. https://doi.org/10.1145/1143844.1143913