Optimization Techniques for Multi-object Detection and Tracking on Live-cell Fluorescence Microscopy Images and Their Applications

Abstract

Fluorescence microscopy is a pivotal imaging technique to visualize biological processes and has been extensively utilized in live-cell morphology analysis. Despite its utility, related object detection and tracking tasks still face challenges due to large data scales, inferior data quality, and insufficient annotations, leading to reliance on adaptive thresholding. Current adaptive thresholding approaches have two significant limitations: Firstly, they cannot handle the heteroscedasticity of image data well and result in biased outputs. Secondly, they deal with frames of time-series imaging data independently and result in inconsistent detections over time. We introduce two novel optimization techniques to address these limitations and enhance detection and tracking results in live-cell imaging. The first one, ConvexVST, is a convex optimization approach to transform heteroscedastic data into homoscedastic data, making them more tractable for subsequent analysis. The second one, Joint Thresholding, is a graph-based approach to get the optimal adaptive thresholds while maintaining temporal consistency. Our methods demonstrate superior performance across various object detection and tracking tasks. Specifically, when applied to microglia imaging data, our techniques enable the acquisition of more complete cell morphology and more accurate detection of microglia tips. Furthermore, by integrating these techniques with existing frameworks, we propose an advanced pipeline for embryonic cell detection and tracking in light-sheet microscopy images, which is streets ahead of state-of-the-art peer methods and sets a new benchmark in the field.