MASAN: Multi-Attention Based Sliced Adversarial Network for Improved B-Cell Acute Lymphoblastic Leukemia Classification

B-cell Acute lymphoblastic leukemia (B-ALL) is a prevalent malignant tumor disease, often diagnosed using peripheral blood smears (PBS). Deep learning technologies have emerged as powerful tools to assist pathologists in analyzing white blood cell counts and types via microscopy for B-ALL subtype determination. However, training these pathologists requires a significant number of expert-annotated pathological slice images, which is time-consuming and costly. In addition, the limitation in the number of available images for training hinders the enhancement of the model's diagnostic precision. In this paper, we propose MASAN, a GAN-constructed model to generate PBS images specific to B-ALL subtypes. Built upon an efficient backbone, MASAN incorporates an efficient multi-scale attention(EMA) module, leading to high-quality synthetic image generation. Our experimental results demonstrate that MASAN can generate highly realistic B-ALL subtype PBS images, thus improving the performance on subsequent downstream classification tasks. The overall FID value of the generated images is 50.03. Moreover, the accuracy of two distinct classifiers improved by approximately 8% when using these generated images.