Abstract:
Kidney disease continues to pose significant global health challenges, where early and precise diagnosis remains crucial for effective treatment planning. Recent advancements in deep learning, particularly Convolutional Neural Networks, have opened new possibilities for automated medical image analysis. This study investigates the potential of CNN architectures for kidney disease classification through systematic comparison of a custom-designed three-layer Sequential model against established pre-trained networks including Xception, ResNet50V2, and InceptionV3. The research utilizes a carefully curated dataset comprising 12,446 renal images divided into training (9,963) and testing (2,483) sets across four diagnostic categories: Cyst, Normal, Stone, and Tumor. A robust preprocessing pipeline was implemented, featuring standardized image resizing to 256×256 pixels, intensity normalization, and comprehensive data augmentation techniques including controlled rotation and flipping to enhance model generalization. The proposed CNN architecture was optimized using the Adam algorithm with carefully tuned parameters (learning rate=0.0001, β1=0.9, β2=0.999) and categorical cross-entropy loss function. Experimental results demonstrate the custom CNN's exceptional performance, achieving 99.84% classification accuracy and surpassing all benchmarked models - Xception (98.87%), ResNet50V2 (99.59%), and InceptionV3 (96.12%). The model maintains this high accuracy while exhibiting superior computational efficiency, requiring substantially fewer parameters than its pre-trained counterparts. Additional evaluation of class-specific metrics confirms consistent diagnostic reliability across all pathology types, with particularly strong performance in identifying tumor cases. These findings make important contributions to the field of medical AI by establishing that purpose-built CNN architectures can outperform complex pre-trained models for specialized diagnostic tasks. The study provides a validated framework for renal disease classification that successfully balances clinical-grade accuracy with practical implementation requirements. Future research directions include exploring hybrid architectures and attention mechanisms to further improve detection of challenging cases, as well as investigating multimodal integration for comprehensive diagnostic assessment.
