MSwinALL-Net: a multi-stage Swin Transformer network for acute lymphoblastic leukemia classification

Acute lymphoblastic leukemia (ALL) is a form of blood cancer that requires timely and precise diagnosis for effective treatment. The conventional method relies on manual examination of blood smear images which is a slow and inconsistent process due to cellular diversity and the limited availability of trained professionals. A Swin Transformer based architecture named MSwinALL-Net is proposed to overcome these challenges by incorporating a multi-stage shifted-window mechanism to capture both fine-grained local and global contextual features for identifying subtle abnormalities in leukemia cells. GAN based augmentation was used to synthesize high quality blood smear images. The proposed model outperforms transfer learning approaches by achieving an accuracy of 98.74%, precision of 1.0, recall of 0.99, F1-score of 0.99 and an AUC of 1.0 showing its potential for reliable, automated ALL diagnosis. The model's robustness and generalization are further validated through 5-fold cross-validation on benchmark leukemia datasets where the model outperformed recent transformer based models like ConvNeXt-Tiny and EfficientNetV2 which exhibited unstable and near random performance. These results highlighted MSwinALL-Net's effectiveness and reliability for automated ALL diagnosis in clinical applications. Future enhancement could incorporate more advanced generative frameworks such as StyleGAN and CycleGAN to further enhance synthetic image quality along with explainable AI techniques for providing interpretability and clinical transparency in real-world deployment.