Enhancing Kidney Stone Detection in Ultrasound Images with Vision Transformers and Hybrid Attention

Kidney stone disease is a significant health issue worldwide, requiring accurate diagnosis for effective management. Ultrasound imaging is a common diagnostic tool due to its safety and cost-effectiveness; however, its limitations include low resolution and high noise. This paper presents a novel approach for kidney stone classification using Vision Transformers (ViTs) augmented with hybrid attention mechanisms. The hybrid framework combines global self-attention from ViTs with local attention techniques to improve feature extraction and noise robustness. Additionally, the proposed method incorporates advanced preprocessing techniques to enhance image quality and applies data augmentation to mitigate class imbalance. The method demonstrates improved classification accuracy, interpretability, and generalizability compared to best performing models or architectures of convolutional neural networks (CNNs), offering a promising solution for automated kidney stone detection in clinical settings. This framework also includes visualization tools for model explainability, aiding clinicians in understanding the model's predictions.