PhagoStat a scalable and interpretable end to end framework for efficient quantification of cell phagocytosis in neurodegenerative disease studies

TL;DR

PhagoStat addresses the challenge of quantifying phagocytosis in dynamic, unstained cells from time-lapse microscopy by delivering an end-to-end, scalable pipeline that combines open-source data handling, robust frame-level quality control, and interpretable DL-based cell segmentation. The approach integrates a data loading/normalization strategy, a CECC-based frame-registration module with blur detection, and joint aggregation–cell analysis to yield comprehensive phagocytosis metrics, supplemented by an explainable AI layer including heatmaps and a sensitivity-based smart annotation framework. In a microglia use case for frontotemporal dementia, PhagoStat reveals that FTD-mutant cells are larger and more phagocytic than wild-type cells, demonstrating the method’s ability to extract biologically meaningful phenotypes and to generate a large, public 2D+time dataset for reproducibility. The pipeline achieves competitive performance with reduced model size, lower hardware demands, and real-time processing potential on HPC resources, and it is openly released to accelerate translational research in neurodegenerative diseases.

Abstract

Quantifying the phagocytosis of dynamic, unstained cells is essential for evaluating neurodegenerative diseases. However, measuring rapid cell interactions and distinguishing cells from background make this task very challenging when processing time-lapse phase-contrast video microscopy. In this study, we introduce an end-to-end, scalable, and versatile real-time framework for quantifying and analyzing phagocytic activity. Our proposed pipeline is able to process large data-sets and includes a data quality verification module to counteract perturbations such as microscope movements and frame blurring. We also propose an explainable cell segmentation module to improve the interpretability of DL methods compared to black-box algorithms. This includes two interpretable DL capabilities: visual explanation and model simplification. We demonstrate that interpretability in DL is not the opposite of high performance, by additionally providing essential DL algorithm optimization insights and solutions. Besides, incorporating interpretable modules results in an efficient architecture design and optimized execution time. We apply our pipeline to analyze microglial cell phagocytosis in FTD and obtain statistically reliable results showing that FTD mutant cells are larger and more aggressive than control cells. The method has been tested and validated on public benchmarks by generating state-of-the art performances. To stimulate translational approaches and future studies, we release an open-source end-to-end pipeline and a unique microglial cells phagocytosis dataset for immune system characterization in neurodegenerative diseases research. This pipeline and the associated dataset will consistently crystallize future advances in this field, promoting the development of interpretable algorithms dedicated to the domain of neurodegenerative diseases' characterization. github.com/ounissimehdi/PhagoStat

Figures (16)

• Figure 1: PhagoStat: A comprehensive end-to-end pipeline for quantifying microglial cell phagocytosis in the context of frontotemporal dementia (FTD). The PhagoStat pipeline is a fully operational system comprised of the following stages: (i) efficient loading of raw data (Fig.\ref{['fig2']}.b), (ii) applying data quality checks and quantifying aggregates over time (Fig \ref{['fig3']}.c), and (iii) performing cell instance segmentation using an interpretable deep learning (IDL) approach (Fig.\ref{['fig5']}, which incorporates Fig.\ref{['fig4']}.c). This comprehensive pipeline streamlines the analysis process and facilitates accurate and reliable results for researchers working with microglial cell phagocytosis data.
• Figure 2: Efficient data loading and normalization pipeline. (a) Detailed steps of the data loading and normalization module where: the two channels (aggregates and cells) are extracted directly from the microscope raw data, then it applies the local and global normalization to standardize the data. (b) High performance computing (HPC) cluster compatible scheme that scales to big datasets. (c) Quantitative comparison of our single-CPU/multi-CPU method and the GPU-accelerated Carl Zeiss ZEN software when processing 76GB CZI file (raw data). To facilitate direct comparison, 'Frame input & output' times are the combination of "read and write" times for all systems. As an insight, our method's time allocation for SSDs (25% reading and 75% saving) and HDDs (76.6% reading and 23.3% saving).
• Figure 3: Detailed data quality workflow. (a) Detailed CECC registration approach. (b) Detailed data quality check modules: (i) CECC based scene shift correction module, (ii) blurry frames detection module, and (iii) saving registration information and the rejected blurry frames. (c) Overview of the aggregates quantification workflow: data quality check + segmentation and matching.
• Figure 4: Detailed DL and IDL architectures for cell instance segmentation. (a) Detailed architecture of the segmentation module during the training phase: custom loss functions (global and local) were used during the retro-propagation on the LSTM modules. (b) Detailed inference phase combining U-Net like architectures, LSTM modules and watershed for instance-level cell segmentation. (c) The details of the explainable segmentation module that contains: (i) light U-Net like models attached to a visualization module applied for each time point, (ii) time coherence module (TTCM) that extracts cell seeds, and (iii) watershed module that combines all signals for a full separation.
• Figure 5: Scene cell instance segmentation and tracking. The scene instance-level segmentation module can use the DL module (Fig.\ref{['fig4']}.b) or the IDL module (Fig.\ref{['fig4']}.c) for a scene cell instance segmentation. This module quantifies cell count, area and coordinates for each frame. Cell speed and total movement quantification loads the scene cell features (frame id, centroid, area). Cell centroids are corrected using scene shift correction module (Fig.\ref{['fig3']}.b). For cell speed and total movement quantification, any tracking algorithm (i.e., Bayesian Tracker) can be applied on the corrected cell features. Results of all complementary modules are saved in an open-source format (i.e., CSV).