Automated identification and quantification of myeloblasts among myeloid lineage cells on bone marrow biopsy using deep learning aligns with hematopathologist readings in patients with Acute Myeloid Leukemia and high-risk myelodysplastic syndrome

Abstract Background: Assessing myeloblast counts in bone marrow biopsy (BMBx) and aspiration is essential for diagnosis and disease status evaluation in patients with acute myeloid leukemia (AML) and higher-risk myelodysplastic syndromes (MDS). Accurate blast cell quantification has so far relied on time-consuming, labor-intensive manual reviews which can be subjective. We investigated the use of artificial intelligence (AI) for fully automated myeloblast detection and quantification from BMBx whole slide images (WSIs) compared with quantification by hematopathologists' assessments.Methods: This study utilized a multi-step computational pipeline for the detection and quantification of blast cells in bone marrow biopsy whole-slide images (WSIs). Annotated training and validation data were collected from 263 adult patients with AML (other than acute promyelocytic leukemia) or higher-risk MDS; 698 image patches containing blast cells from University Hospitals Cleveland Medical Center and 260 image patches with all WBCs manually labeled from 274 randomly selected WSIs from Duke University Medical Center. The pipeline consisted of three primary steps: (1) Myeloid Lineage Cell (MLC) detection – A YOLOv12 model was trained to detect GLCs using 260 manually annotated 256×256-pixel image patches. (2) Blast Classification – A total of 3,825 MLC patches (including 1,036 labeled as myeloblasts) were used to train and evaluate four classifiers using 5-fold cross-validation. Multi-scale image features were extracted using pretrained YOLOv12 and ResNet-18 backbones. These features were used to train four classifiers (Support Vector Machine [SVM], Random Forest, Linear Discriminant Analysis [LDA], and MLP[Multi-Layer Perception]) to distinguish myeloblast from other MLCs. (3) WSI Inference and Quantification – The pipeline was applied to 274 WSIs from 137 patients pre- and post- allogeneic hematopoietic cell transplantation. The number of myeloblasts and total MLCs per slide were counted, and the blast ratio was calculated and compared with hematopathologist interpretation.Results: The MLC detection model (YOLOv12) demonstrated strong performance, achieving a mean precision of 0.938 on the held-out validation set. For blast classification, the best average Area Under the ROC Curve (AUC) scores achieved across folds were Support Vector Machine (SVM): 94.82% We applied our pipeline to 274 WSIs, and detected and classified millions of WBCs successfully. Out of 274 WSIs, 28 WSIs (10.2%) showed blast ratios above the 5% clinical threshold. In 98% of cases, myeloblast quantification was within <5% difference from hematopathologists' interpretation. Six discordant slides revealing either borderline blast morphology or annotation inconsistencies.Conclusion: The overall agreement between AI-derived and hematopathologists' blast detection and quantification 98% cases demonstrates this pipeline's capability to provide efficient, robust, high-throughput histological quantification and flag borderline or difficult-to-interpret cases for further pathological review. Our automated pipeline for blast quantification can be further developed to serve as a support tool for hematopathologists in BMBx interpretation for myeloid disease cases.