Multi-Scale Transformer Architecture for Accurate Medical Image Classification

This research proposes an AI-based skin lesion classification algorithm by utilizing an improved Transformer architecture to break the limitations regarding accuracy and robustness for medical image evaluation. The incorporation of a multi-scale feature fusion mechanism and optimization of the self-attentive process provides the model with the capability to better extract global and local features and improve its capability for the identification of lesions with fuzzy boundaries and complex structures. Evaluation of the improved Transformer on the ISIC 2017 dataset shows the improved Transformer outperforms existing top-performing AI architectures such as ResNet50, VGG19, ResNext, and Vision Transformer on the leading metrics of accuracy, AUC, F1-Score, and Precision. Grad-CAM visualization also highlights the interpretability of the model by validating high correspondence between the attention region selected by the algorithm and real lesion locations. This research further highlights the revolutionary prospects of advanced AI architectures for medical imaging and the way forward toward the development of accurate and trustworthy diagnostic aids. Future research will explore the scalability of the proposed solution toward general medical image processing and examine the incorporation of multimodal data toward the advancement of AI-based diagnostics to intelligent medicine.