A Framework for Automated Tracking of Morphologically Distinct Cell Populations in Time-Lapse Microscopy

Abstract

Identifying morphologically distinct cell populations in time-lapse fluorescence microscopy is central to understanding how complex tissues develop and remodel, yet remains labor-intensive and subject to observer bias despite advances in cell segmentation. We present a generalizable computational framework for automated detection and tracking of cell populations with characteristic morphological signatures, demonstrated through detection of ventral midline cells - a critical anatomical landmark in developing Drosophila embryos. Our modular pipeline integrates deep learning segmentation with morphological feature extraction using rotated bounding box analysis to characterize individual cells by elongation, orientation, and size. Multi-criteria filtering identifies candidates meeting population-specific morphological profiles, followed by spatial clustering to assemble spatially coherent structures. The clusters undergo validation to ensure geometric consistency through aspect ratio and orientation constraints, with temporal interpolation using circular statistics for missing frames. Systematic evaluation of five clustering methods on a stratified test set reveals performance tradeoffs between detection sensitivity and geometric accuracy. The modular architecture enables straightforward adaptation to diverse experimental conditions through adjustable morphological criteria, making it broadly applicable to any tissue analysis requiring identification of morphologically distinct cell groups across developmental and morphogenetic processes.