<![CDATA[Kidney diseases, including cysts, stones, and tumors, are common among older adults and often progress asymptomatically, leading to delayed diagnoses. Manual interpretation of CT images by clinicians is labor-intensive and can vary significantly between observers, especially in high-volume settings. This study aims to develop and evaluate an artificial intelligence–based decision support system for multiclass kidney disease classification with an emphasis on robustness, computational efficiency, and clinical feasibility in elderly healthcare environments. The study proposes a medical informatics evaluation framework that integrates standard performance metrics with learning dynamics, overfitting analysis, and error distribution assessments to ensure reliable model selection. Three architectures were evaluated: a conventional CNN, MobileNet-V2, and EfficientNet-B0. Experiments were conducted on a publicly available dataset containing 12,446 CT images across four classes (Normal, Cyst, Stone, and Tumor). Models were trained under varying epoch settings and evaluated using weighted accuracy, precision, recall, F1-score, AUC, learning curve analysis, and confusion matrix assessment. The results indicate that the conventional CNN achieved perfect numerical performance but exhibited rapid convergence and early metric saturation, limiting the interpretability of generalization under the current dataset configuration. EfficientNet-B0 showed stable yet conservative performance, whereas MobileNet-V2 achieved near-optimal accuracy with gradual convergence, minimal overfitting, and superior computational efficiency. At the optimal configuration (epoch 50), MobileNet-V2 achieved an accuracy of 1.00, precision of 1.00, recall of 1.00, F1-score of 1.00, and an AUC of 0.9997. These findings suggest that lightweight architectures, particularly MobileNet-V2, offer a practical solution for CT-based kidney disease decision support, while acknowledging the need for patient-level and multi-institutional validation.]]>
