Sequential boundary tracking via reinforcement learning for overlapping cell resolution in sickle cell imaging

Sickle cell disease requires rapid, accurate diagnostic screening to facilitate early therapeutic actions with potential life-saving impact. Unfortunately, traditional manual microscopy screening is heavily affected by high rates of human error, long processing delays, and detection variability, which are especially pronounced when identifying truly challenging microscopic fields of view, such as those consisting of clusters of overlapping red blood cells. These diagnostic delays result in poor patient outcomes, particularly in resource-limited settings where specialized healthcare providers are scarce. Highly precise automated tracking solutions are thus essential to ensure diagnostic fairness and reliability. To address these challenges, we introduce a general end-to-end hybrid in which feature extraction is based on deep localized features and decision-making is performed by sequential RL. The method is based on a dual-stage pipeline: a bespoke U-Net architecture first separates individual cells within complex, overlapping cell bunches, and afterward, a Q-learning agent that finds optimized sequential tracking paths on the coordinate grid matrices maps the cells to final labels. Extensive experimental results on the public Kaggle dataset show that the segmentation module achieves a high-fidelity dice similarity coefficient of 95.51%, and the proposed integrated RL pathfinder improves the final classification results, achieving an overall accuracy of 98.00%. Our solution achieves a +16:60% absolute diagnostic improvement over a traditional single CNN classifier (81:40%), demonstrating that the proposed framework is robust, has novel methodological characteristics, and is clinically viable for fully automated hematological screening.