A LIGHTWEIGHT CUSTOM CNN FOR ROBUST AND CLINICALLY SAFE CT KIDNEY DISEASE CLASSIFICATION

ABSTRACT: Kidney diseases such as renal cysts, stone, and malignant tumors represent a significant and increasing health burden worldwide. Manual CT diagnosis is labour-intensive and has high inter-observer variability, which requires the support of automatic diagnosis. In this paper, a lightweighted custom Convolutional Neural Network (CNN) is proposed, which is used to automatically classify kidney computed tomography (CT) images into four classes: Cyst, Normal, Stone, and Tumor. It includes residual skip connections, a dual-pooling head (Global Average and Global Max Pooling), Spatial Dropout2D regularization, and a composite training objective that explicitly favors Tumor Recall. The CT-KIDNEY-DATASET was mad leakage-safe through MD5 exact deduplication and perceptual-hash grouping, so that all reported evaluations are free of data leakage. Across the five validation folds, KidneyCNN achieved a mean accuracy of 99.82 ± 0.06% and a mean Tumor Recall of 99.31 ± 0.30%. Separately, ensemble inference on the independent held-out test set of 1,904 images achieved 99.84% accuracy, with AUC = 1.000 for all classes and only three misclassifications. The Grad-CAM++ results show that the model attention as anatomically relevant, which further indicates that KidneyCNN is a clinically safe and deployable kidney disease screening tool.