Identification of Acute Lymphoblastic Leukemia Subtypes on a Microscopic Image of Multicellular Blood Using Object Detection Model with Swin Transformer

Detecting subtypes of acute lymphoblastic leukemia (ALL) in multicellular microscopic images is crucial for early diagnosis and treatment. In the previous study, ALL subtype detection has often employed conventional methods requiring multiple phases, including WBC segmentation, touch cell separation, feature extraction, and classification. We compare object detection algorithms that require only a single learning framework and no additional steps. The performance of the YOLO, Mask R-CNN, and Mask R-CNN with Swin Transformer models for detecting ALL subtypes are compared. The aim of model comparison is to evaluate the performance in detecting the subtype of ALL with the best mAP value. In the detection of ALL subtypes, the Mask R-CNN with Swin Transformer surpasses all other models. The Mask R-CNN model with Swin Transformer produced the best global test results for the L1, L2 and L3 detection process, with mAP(0.5) and mAP(0.95) values of 94.5% and 68%, respectively.