RLTNT: An explainable residual learning-based transformer model for kidney disease classification

Kidney diseases are among the most prevalent and important health issues affecting the world population today. It is crucial to diagnose the illness of the kidney in the early stages for effective treatment, reducing the number of deaths and side effects. Medical imaging significantly impacts early disease detection and diagnosis, with deep learning models like convolutional neural networks (CNNs), contributing significantly to advancements in this domain. However, the black-box behavior of CNNs elevates concerns regarding their reliability and interpretability, which are vital for clinical decision-making. In this paper, we developed a novel Residual Learning-based Transformer-in-Transformer (RLTNT) model, integrating ResNet-18 for feature extraction with a Transformer-in-Transformer (TNT) architecture for classification, which enhances classification accuracy and addresses the black box nature of deep learning techniques through explainable artificial intelligence in medical imaging. The proposed RLTNT model is applied to classify kidney CT images into four categories: normal, cyst, tumor, and stone for dataset 1 and into two categories: normal and stone for dataset 2. Highlights of the proposed architecture are the proficiency of the ResNet-18 to efficiently extract hierarchical features and the ability of the TNT to capture long-range dependencies and complex patterns. The RLTNT model achieved validation accuracies of 99.60% on Dataset 1 and 99.79% on Dataset 2, surpassing DenseNet121 at 98.30%, ViT at 98.60%, and FINDWELL at 99.58%, demonstrating superior performance in kidney CT image classification. These results demonstrate the ability of the model to enhance both the accuracy and interpretability of kidney disease diagnosis, supporting its prospective role in clinical decision support systems.