CellApop: A knowledge-guided decoupled distillation framework for label-efficient apoptotic cell segmentation and...

Conventional apoptosis detection methods primarily depend on fluorescence staining, which is labor-intensive, potentially cytotoxic, and unsuitable for real-time monitoring. To overcome these limitations, this study presents a segmentation-based deep learning (DL) framework for label-free, dynamic detection of apoptotic cells in bright-field microscopy images.

A comprehensive training dataset comprising 16,472 bright-field cell images was curated from four sources-three public datasets (BF-C2DL-MuSC, DICC2DHHeLa, and LiveCell) and one proprietary apoptosis dataset. To achieve label-efficient learning, a Knowledge-guided Decoupled Distillation (KDD) framework was developed, wherein multiple expert models collectively guide the training of a lightweight student network, CellApop. The student model incorporates re-parameterization, depthwise separable convolutions, and an edge-aware module to improve segmentation accuracy under challenging conditions such as dense cellular overlap and indistinct boundaries. Performance was evaluated using the Dice similarity coefficient, Hausdorff Distance (HD), Intersection over Union (IoU), sensitivity, and specificity. Furthermore, CellApop was tested in an observer study for automated apoptosis-rate quantification across drug-treatment conditions, with its outputs compared against assessments by biological experts of varying experience levels.

CellApop achieved Dice scores of 0.843 for general cells and 0.754 for apoptotic cells, while markedly reducing model complexity and inference latency. The KDD strategy decreased manual labeling requirements by approximately 80 % on the proprietary dataset. In the observer study, model-derived apoptosis rates demonstrated high concordance with ground truth and were comparable to a senior expert's performance, surpassing those of junior and intermediate experts-particularly at early time points when apoptotic morphology was subtle.

The proposed CellApop framework delivers accurate, efficient, and label-free segmentation of apoptotic cells in bright-field microscopy, eliminating the need for fluorescent staining. Its robustness and scalability make it a promising tool for automated apoptosis quantification and drug-response assessment in routine experimental workflows.