Autonomous Region-of-Interest Detection, Cell Segmentation and Cell Classification in Bone Marrow Cytomorphology Using Deep Learning

Résumé fourni par la source

The differential nucleated cell count in bone marrow smears forms the diagnostic basis in hematological disorders. The current gold standard involves manually detecting evaluable regions of interest (ROI) and subjectively categorizing up to 500 cells to form a diagnosis. This process is time-consuming, labor-intensive, and suffers from high inter-observer variability. To address this issue, we propose an end-to-end deep learning pipeline for automating cytomorphology in bone marrow smears, covering ROI selection in whole slide images (WSI), cell detection and classification of 20 cell types. ROI detection was performed using a U-Net model to segment aspirate spicules in WSI. Automatically selected ROIs were ranked for image quality based on feature similarity to a reference pool of 4074 ROIs manually selected by hematologists. 47669 cells on 479 ROIs were manually labeled by hematologists and experienced medical technical assistants (n=10) using a majority vote system, requiring each cell to be labeled by at least three observers. Using these labeled cells, we trained a Faster R-CNN model with a ResNet-50 backbone for cell detection and a ResNet-50 for cell classification to provide automated differential counts for nucleated cells at the WSI level. Aspirate spicules were segmented with a Dice Score of 0.86, achieving Recall and Precision of 0.85 and 0.88 for the target class. Cell detection achieved a mean average precision of 0.84 at an intersection over union threshold of 50% (mAP50). Subsequently, the cell classification model achieved an overall accuracy of 0.87 across all 20 cell types in a test set of 1,200 cells. The specific area under the curve (AUC) with 95% confidence interval (CI) for each of the different cell classes is listed as follows: blast 0.991 [CI: 0.982 to 1.0], promyelocyte 0.980 [CI: 0.946 to 1.0], myelocyte 0.992 [CI: 0.980 to 1.0], metamyelocyte 0.976 [CI: 0.934 to 1.0], band neutrophil 0.991 [CI: 0.975 to 1.0], segmented neutrophil 0.990 [CI: 0.969 to 1.0], basophil 0.989 [CI: 0.846 to 1.0], eosinophil 1.0 [CI: 1.0 to 1.0], promonocyte 0.809 [CI: 0.444 to 1.0], immature monocyte 0.946 [CI: 0.829 to 1.0], monocyte 0.978 [CI: 0.950 to 1.0], proerythroblast 0.996 [CI: 0.964 to 1.0], erythroblast 0.999 [CI: 0.996 to 1.0], macrophage 0.968 [CI: 0.724 to 1.0], immature lymphocyte 0.899 [CI: 0.493 to 1.0], lymphocyte 0.993 [CI: 0.985 to 1.0], plasma cell 0.999 [CI: 0.992 to 1.0], megakaryocyte 0.994 [CI: 0.884 to 1.0], smudge cell 0.983 [CI: 0.959 to 1.0], artifact 0.980 [CI: 0.918 to 1.0]. The fully trained model's operating speed to provide a differential count for 500 cells was 146.5s on average. We propose a seamless, fully autonomous pipeline to provide ROI detection, cell segmentation, and differential counts for nucleated cells in bone marrow aspirates, achieving high accuracies across 20 cell types at a fraction of time of standard manual differential counts. By utilizing a majority voting approach for training labels, we reduce inter-observer variability, incorporate robust ground truth labels, and account for uncertainty in label assignment. Our models may therefore provide a highly accurate and time-efficient alternative to current standard practices based on manual labor.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.