A Hematologist-Level Deep Learning Algorithm (BMSNet) for Assessing the Morphologies of Single Nuclear Balls in Bone Marrow Smears: Algorithm Development

• DOI: 10.2196/15963
• Publication Date: 2020-04-08T12:45:16Z
• Authors: Yi-Ying Wu; Tzu-Chuan Huang; Ren-Hua Ye; Wen-Hui Fang; Shiue-Wei Lai; Ping-Ying Chang; Wei-Nung Liu; Tai-Yu Kuo; Cho-Hao Lee; Wen-Chiuan Tsai; Chin Lin
• Keywords: Gold standard (test)

Bone marrow aspiration and biopsy remain the gold standard for diagnosis of hematological diseases despite development flow cytometry (FCM) molecular gene analyses. However, interpretation results is laborious operator dependent. Furthermore, obtained exhibit inter- intravariations among specialists. Therefore, it important to develop a more objective automated analysis system. Several deep learning models have been developed applied in medical image but not field histology, especially bone smear applications.The aim this study was model (BMSNet) assisting hematologists smears faster disease monitoring.From January 1, 2016, December 31, 2018, 122 were photographed divided into cohort (N=42), validation (N=70), competition (N=10). The included 17,319 annotated cells from 291 high-resolution photos. In total, 20 photos taken each patient cohort. This eight annotation categories: erythroid, blasts, myeloid, lymphoid, plasma cells, monocyte, megakaryocyte, unable identify. BMSNet convolutional neural network with YOLO v3 architecture, which detects classifies single model. Six visiting staff members participated human-machine competition, FCM regarded as ground truth.In cohort, according 6-fold cross-validation, average precision bounding box prediction without consideration classification 67.4%. After removing error, recall similar those most categories. detecting than 5% blasts area under curve (AUC) (0.948) higher AUC (0.929) lower pathologists (0.985). 20% AUCs (0.981) (0.980) (0.942). Further showed that performance difference could be attributed myelodysplastic syndrome cases. mean value correlations between 0.960, values ranged 0.952 0.990.Our can assist interpreting by facilitating accelerating detection hematopoietic cells. detailed morphological still requires trained hematologists.