Automated Bone Marrow Cellularity Estimation from Whole-Slide Images Using Deep Learning: Validation in an Indian Clinical Cohort

<title>Abstract</title> <p>Bone marrow evaluation is central to diagnosing leukemia, lymphoma, unexplained cytopenias and other hematologic disorders, yet it depends on specialised expertise and remains time-consuming, subjective and prone to interobserver variability. We present the Artificial Intelligence Slide Analysis Platform (AiSAP), a deep learning framework for automated bone marrow cellularity estimation directly from raw whole-slide images (WSIs), trained and validated on an Indian clinical trephine biopsy cohort that is underrepresented in existing automated pipelines. Unlike approaches restricted to cropped or pre-selected patches, AiSAP performs end-to-end WSI tiling, automated marrow region-of-interest extraction and adipocyte segmentation using a dual Mask R-CNN architecture, before computing a case-level cellularity ratio. Across five backbone configurations, average precision ranged from 45.3% to 51.9%, while case-level concordance reached up to a Pearson correlation coefficient of 0.95 and Lin's concordance correlation coefficient of 0.88 (depending on the backbone). These results show that AiSAP achieves expert-level concordance for automated cellularity estimation and represents a promising, generalizable step toward clinical decision support in bone marrow pathology, warranting further multi-institutional validation.</p>