Using Deep Learning Models for Classifying Kidney CT Scans

Abstract

This thesis presents a well-structured and technically rigorous comparison of vision transformer architectures for kidney CT classification, demonstrating a strong understanding of both deep learning theory and medical imaging challenges. The experimental design is particularly commendable, as training all models from scratch allows for a fair evaluation of architectural differences without the influence of pretrained features. The results are clearly presented and supported by thorough analysis, including training dynamics, confusion matrices, and interpretability methods. A key strength is the critical discussion of limitations, especially regarding potential data leakage and its impact on near-perfect accuracy, which reflects strong scientific awareness. The inclusion of interpretability analysis further enhances the depth of the study, even though it reveals important concerns about model behavior. Overall, this is a thoughtful and well-executed thesis that contributes meaningful insights into transformer-based approaches for medical image classification.