Comparative Analysis of MobileNetV2, EfficientNetB0, and U-Net Models for Kidney Disease Classification from CT Scans

Abstract

Kidney disease includes affected cysts, tumors, and stones which is a significant worldwide health problem, where delayed or inaccurate diagnosis can negatively impact results. Computed tomography (CT) affords high anatomical context and is standardly utilized to assess renal parenchyma and identify abnormal areas. This study explores deep learning based automated classification of kidney abnormalities in CT. A publicly accessible multi-class renal CT dataset was developed and preprocessed through resizing, intensity normalization, and augmentation. Three convolutional architectures which are EfficientNetB0, MobileNetV2, and U-Net adapted to image-level prediction those were trained and compared in repeated cross validation using accuracy, precision, recall, and F1-score as primary metrics. EfficientNetB0 delivered the most consistent and highest accuracy performance in evaluations, with some evidence of clinical adequacy. MobileNetV2 achieved competitive, stable performance at lower computational expense, with a preference for deployment in resource-constrained settings. The U-Net baseline, although computationally efficient, demonstrated higher variability for direct classification. In total, the findings confirm the potential of lightweight CNNs for CT-based kidney disease diagnosis and the importance of multi-test evaluation for stability. Future directions include extension to larger, more heterogeneous cohorts, exploration of transformer-based multimodal fusion with clinical text, incorporation of explainable AI for transparent decision-making, and validation in actual radiology workflows.