An AI-enabled tool for quantifying overlapping red blood cell sickling dynamics in microfluidic assays

TL;DR

An automated deep learning framework that integrates AI-assisted annotation, segmentation, classification, and instance counting to quantify red blood cell (RBC) populations across varying density regimes in time-lapse microscopy data is presented.

Abstract

Understanding sickle cell dynamics requires accurate identification of morphological transitions under diverse biophysical conditions, particularly in densely packed and overlapping cell populations. Here, we present an automated deep learning framework that integrates AI-assisted annotation, segmentation, classification, and instance counting to quantify red blood cell (RBC) populations across varying density regimes in time-lapse microscopy data. Experimental images were annotated using the Roboflow platform to generate labeled dataset for training an nnU-Net segmentation model. The trained network enables prediction of the temporal evolution of the sickle cell fraction, while a watershed algorithm resolves overlapping cells to enhance quantification accuracy. Despite requiring only a limited amount of labeled data for training, the framework achieves high segmentation performance, effectively addressing challenges associated with scarce manual annotations and cell overlap. By quantitatively tracking dynamic changes in RBC morphology, this approach can more than double the experimental throughput via densely packed cell suspensions, capture drug-dependent sickling behavior, and reveal distinct mechanobiological signatures of cellular morphological evolution. Overall, this AI-driven framework establishes a scalable and reproducible computational platform for investigating cellular biomechanics and assessing therapeutic efficacy in microphysiological systems.

Figures (6)

• Figure 1: Schematic representation of the AI-enhanced segmentation framework for quantifying RBC sickling dynamics in dense and overlapping fields. A subset of images from microfluidic experiments is sampled to obtain representative frames at user-defined intervals. These frames are subsequently annotated in Roboflow using AI-assisted manual labeling to generate instance masks of red blood cells (RBCs) labeled as healthy or sickled. The annotated images are used to train an enhanced nnU-Net segmentation model that automatically optimizes preprocessing and network configurations. The optimized model weights are then applied during inference to produce predicted segmentation masks for unseen experimental datasets. A marker-controlled watershed post-processing step refines and separates overlapping cells, enabling accurate instance counting and classification. Finally, the temporal evolution of the sickled fraction is quantified from videos of sickling dynamics.
• Figure 2: Workflow for dataset acquisition and annotation using the Roboflow platform dwyer2024roboflow. Starting from microscopy frames (Data Collection), cell classes are defined (Healthy and Sickled) and used to guide annotation in Roboflow via AI-assisted manual labeling. The resulting labeled dataset (Dataset Output) provides class-specific masks for downstream model training and evaluation, with background in black, healthy cells in green, and sickled cells in red.
• Figure 3: Overview of the watershed pipeline for separating overlapping RBCs. The left panel shows the nnU-Net segmentation result, where overlapping same-class cells may appear as merged regions. The middle panel summarizes the marker-controlled watershed steps, and the right panel shows the instance-separated output after watershed. Green indicates healthy cells, red indicates sickle cells, and yellow outlines highlight regions that were successfully split.
• Figure 4: (A--C) Representative sickling dynamics at different cell suspension densities from the same patient, shown at $t = 0$ s (top row) and $t = 120$ s (bottom row). For each condition, the left column shows the original micrograph, the middle column displays the nnU-Net prediction with watershed-based instance separation (green: healthy; red: sickle; yellow: overlapping regions separated by watershed), and the right column compares the predicted sickle-cell fraction (solid blue) with manual counts (red dashed).
• Figure 5: Effect of hemoglobin modification on RBC sickling dynamics. (A,B) Time-resolved micrographs of RBCs under 2% O$_2$ with (A) 0% and (B) 100% hemoglobin modification by osivelotor at $t = 0$ s and $t = 120$ s (left panel). Comparison between predicted and manually counted sickling dynamics under different hemoglobin modification levels (right panel).