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

Technical Summary: Automated Bone Marrow Cellularity Estimation from Whole-Slide Images Using Deep Learning

Problem Statement
Bone marrow cellularity—the proportion of hematopoietic cells relative to adipocytes—is a critical diagnostic parameter for hematologic disorders such as leukemia, myelodysplastic syndromes (MDS), and aplastic anemia. Currently, the assessment of cellularity relies on manual microscopic evaluation of Hematoxylin and Eosin (H&E) stained trephine biopsy sections by expert hematopathologists. This process is time-consuming, subjective, and prone to interobserver variability. While deep learning has advanced biomedical imaging, existing automated pipelines often rely on pre-selected or cropped image patches rather than processing raw Whole-Slide Images (WSIs) end-to-end. Furthermore, there is a lack of validated automated tools specifically trained on Indian clinical cohorts, which are underrepresented in existing datasets.

Methodology: The AiSAP Framework
The authors propose the Artificial Intelligence Slide Analysis Platform (AiSAP), a deep learning framework designed for end-to-end automated cellularity estimation directly from raw WSIs. The system utilizes a dual-stage instance segmentation approach based on the Mask R-CNN architecture implemented via the Detectron2 library.

1. Pipeline Architecture:

   • Stage 1 (Model-1): The raw WSI is tiled into smaller patches. Model-1 performs multi-class instance segmentation to extract the "cells" region of interest (ROI). It distinguishes between four classes: 'cells' (hematopoietic cells, adipocytes, and other cellular structures), 'bone' (trabeculae), 'discarded_cells' (hemorrhage, necrosis, compression artifacts), and 'foreign_label' (staining artifacts, glass cracks).
   • Stage 2 (Model-2): The regions identified as 'cells' by Model-1 are passed to Model-2. This model is dedicated to segmenting the 'fat' (adipocyte) class within the extracted cellular regions.
   • Cellularity Calculation: The final cellularity percentage is computed using the formula:
     $$\text{Cellularity (\%)} = \left(1 - \frac{A_{fat}}{A_{cells}}\right) \times 100$$
     Where $A_{fat}$ is the total adipocyte area and $A_{cells}$ is the total marrow area identified by Model-1.

2. Model Configurations:
   Five different backbone architectures were evaluated under identical hyperparameter conditions to determine the optimal configuration:

   • FPN-based with PointRend: ResNet-50 (R50_FPN), ResNet-101 (R101_FPN), and ResNeXt-101 (X101_FPN).
   • Single-scale with Standard Mask Head: ResNet-50 (R50_C4) and ResNet-101 (DC5).
   • All models were initialized with COCO-pretrained weights and fine-tuned on a custom dataset derived from the Post Graduate Institute of Medical Education and Research (PGIMER), Chandigarh.

3. Dataset:
   The study utilized a retrospective cohort of 220 whole-slide images. The data was split into training (70%), validation (15%), and testing (15%) subsets. Annotations were performed by a pathologist under medical supervision, covering approximately 45,000 annotations for Model-1 and 4,226 for Model-2. The final validation involved 170 clinical cases compared against the average assessment of two independent expert hematopathologists.

Key Results

• Instance Segmentation Performance:
  • For the multi-class bone marrow task (Model-1), the R101_DC5 backbone achieved the highest Average Precision (AP) of 50.0% and recall of 82.2% on the validation set.
  • For the single-class fat/adipocyte task (Model-2), the R50_C4 backbone achieved the highest AP (88.9%) and recall (94.5%).
  • The 'discarded_cells' class remained the most challenging to segment across all backbones (F1-score ~63–65%).
• Case-Level Concordance:
  • When comparing model-predicted cellularity against expert assessments across 170 cases, the X101_FPN backbone achieved the highest Lin's Concordance Correlation Coefficient (LCCC) of 0.882 with the smallest systematic bias (+6.9%).
  • The R50_C4 backbone achieved the highest Pearson correlation coefficient (0.948) and an LCCC of 0.857.
  • All backbones exhibited a consistent positive bias (model overestimation of 6.9% to 8.6% compared to experts), which the authors attribute to the model's ability to detect small cellular elements often missed in rapid visual inspection, though potential misclassification of fibrosis or clots as cellular area is also noted.
  • Bland–Altman analysis confirmed that the majority of cases fell within the 95% limits of agreement.

Significance and Claims
The paper claims that AiSAP represents a significant step toward clinical decision support in bone marrow pathology by addressing the gap in end-to-end WSI processing for Indian clinical cohorts. Unlike prior studies that rely on cropped fields or semantic segmentation alone, AiSAP integrates automated tiling, ROI extraction, and dual-stage instance segmentation to provide a quantitative cellularity ratio directly from raw slides.

The authors assert that the framework achieves expert-level concordance (LCCC up to 0.88, Pearson $r$ up to 0.95), demonstrating that transfer learning with Mask R-CNN architectures can yield reproducible assessments that reduce interobserver variability. While acknowledging limitations such as single-center data and systematic bias, the study positions AiSAP as a promising, generalizable tool warranting further multi-institutional validation to confirm robustness across varying staining protocols and clinical scenarios.