Image-Based Bladder Cancer Classification Using Transformer Architectures

Abstract

This study presents a multiclass approach to distinguish low-grade cancer (LGC), high-grade cancer (HGC), cystitis, non-suspicious tissue (NST), and non-tumor lesions (NTL) in endoscopic images. Transformer models Vision Transformer, DeiT, CaiT, MedViT, and Swin Transformer were used for classification. The highest classification accuracy was 98.6% with the Swin Transformer. In addition to the evaluation of the obtained results, Grad-CAM was used to determine the accuracy of the Swin Transformer's decision-making process. The results demonstrated that the model identified meaningful tissue features rather than irrelevant background noise. © 2025 IEEE.