Explainable Knowledge Distillation via Capsule Vision Transformers for Automated Kidney Disease Categorization

Abstract:   Kidney disease, characterized by the progressive decline in the renal system's ability to filter metabolic waste and excess fluids, poses a significant global health risk. When these physiological impairments persist beyond a three-month threshold, the condition is classified as Chronic Kidney Disease (CKD). Current diagnostic frameworks often struggle with high computational overhead, suboptimal precision, and a lack of architectural efficiency for deployment on lightweight devices. To resolve these challenges, this study introduces CapViT-DKD , an advanced kidney detection and classification framework utilizing a Capsule Vision Transformer integrated with Self-Supervised Diverse Knowledge Distillation. The methodology utilizes the CT-kidney dataset, initiated by a robust preprocessing pipeline comprising image resizing, normalization, and strategic data augmentation. The core architecture employs a teacher-student paradigm: a high-capacity Perceptive Capsule Transformer Network (PCapTN) serves as the teacher, transferring complex feature representations to a Lightweight Capsule Transformer Network (LCapTN) student model. This diverse knowledge distillation (DKD) approach significantly boosts the student model’s performance while maintaining a small computational footprint. To address the "black box" nature of deep learning, we incorporate Principal Component of Gradient-Class Activation Mapping (PCG-CAM), which provides visual explanations by highlighting the specific anatomical regions influencing the diagnostic output. Empirical results demonstrate that the proposed system achieves superior performance metrics, including an accuracy of 99.75%, precision of 99.50%, Recall of 99.15 % and an F1-score of 99.10%, validating its efficacy for clinical decision support