Identification and Visualization of Fusion Gene Subtypes in APL Using Spatial Attention Mechanisms in Vision Models

Peirou Yan¹Guo Pu²Ping Wu³Lijun Wen⁴Suning Chen⁴Song Xue^5*

• ¹Aerospace Clinical Medical College of Peking University, Beijing, Beijing Municipality, China
• ²Peking University, Beijing, Beijing Municipality, China
• ³Hebei Yanda Hospital, Langfang, Hebei, China
• ⁴The First Affiliated Hospital of Soochow University, Suzhou, Jiangsu Province, China
• ⁵Lu Daopei Hospital, Beijing, China

Acute promyelocytic leukemia (APL) features leukemic cell differentiation arrest at the promyelocytic stage, mainly due to the t (15;17), (q24; q21) translocation that forms the PML-RARA fusion protein. Variant RARα translocations, with distinct biological traits and all-trans retinoic acid (ATRA) responses, often cause misdiagnosis and lengthy genetic testing. To solve these problems, we propose a spatial attention mechanism-enhanced convolutional neural network integrating ResNet Blocks and a spatial attention module (CNN with spatial attention), which can achieve high-precision identification of APL fusion gene subtypes and pixel-level visualization of key areas. Data collected from two hospitals and Kaggle, including bone marrow smear images of PML-RARA, TTMV-RARA, NPM1-RARA, STAT5B-RARA, and NUP98-RARG subtypes, were preprocessed to form a five-class dataset. The model attains an overall accuracy of 98.04% in five-class classification, with good performance in each category. The attention maps also improve model interpretability. Such a novel and rapid diagnostic approach for APL subtypes holds significant clinical value.