Abstract
Acute myeloid leukemia (AML) is a hematologic malignancy in which accurate subtype identification is important for treatment planning and prognosis. This study developed a deep learning approach for AML subtype classification using multiple single-cell images from each patient rather than conventional instance-based classification. An open-source dataset containing 81,214 single-cell images from 189 individuals was analyzed. The dataset included four genetically defined AML subtypes and a healthy control group. Two multi-view convolutional neural network (MV-CNN) models were evaluated: a 99-view model based on random sampling and a 500-view model in which image augmentation was used when fewer than 500 images were available for a patient. A single-view CNN served as the comparison model. Performance was assessed using tenfold cross-validation and accuracy, F1 score, sensitivity, specificity, precision, and Matthews correlation coefficient (MCC). The 500-view MV-CNN achieved the best performance, with an accuracy of 0.8783, F1 score of 0.8622, sensitivity of 0.8637, specificity of 0.8774, precision of 0.8616, and MCC of 0.8441. The 99-view model showed lower but comparable performance, while the single-view CNN yielded the lowest values. Patient-level multi-view analysis improved AML subtype classification compared with single-image analysis, supporting the potential of multi-view deep learning for hematologic image classification.