ROI-BMC-DNNet: An efficient automatic analysis model of whole-slide scanned bone marrow aspirate images for the diagnosis of hematological disorders

Abstract

Cell differential counting in the scanned whole-slide image (WSI) is essential but still challenging in the intelligent diagnosis of hematological disorders based on bone marrow aspirate smear images. The huge size and the complex content of WSI pose multiple obstacles to training an efficient and accurate analysis model. This paper proposes a novel analysis framework termed ROI-BMC-DNNet. To capture patch images in the cell well-spread areas, we provide a pyramid segmentation network based on prior pathological knowledge to automatically locate regions of interest (ROI) that pathologists usually manually focus on to count nucleated cells. Specifically, we describe a simple but effective alignment method for multiscale ROI, aiming to maintain the accuracy of the semantic and the spatial information of the pyramid network. Furthermore, we design a patch sampling algorithm and a patch quality evaluating network to ensure the effectiveness of the sampled patches. Additionally, we provide a cell detection model to realize the automatic differential counting of the nucleated cells in the sampled patches. Experiments are conducted on the datasets of scanned WSIs from 120 patients. Our method achieves 92.27% retrieval precision for the high-quality patches (patches with nucleated cells with better morphology and without particles). When the missed nucleated cells are evaluated, the detection Recall and Precision of the bone marrow nucleated cells from the WSIs are 87.13% and 87.90% respectively. Notably, the recognition Recall and Precision of the blast cell are 91.37% and 90.47%, respectively. The results confirm that the proposed method is promising to work as an auxiliary diagnostic tool for hematological disorders.