<![CDATA[Introduction: Kidney diseases remain a major global health challenge, and computed tomography (CT) provides detailed renal imaging that supports accurate diagnosis. Although data augmentation is commonly used to improve deep learning performance on limited medical datasets, the effects of different augmentation strategies on image characteristics and model learning behavior remain insufficiently understood. Method: This study evaluated spatial, intensity, and hybrid augmentation for kidney CT image classification using a custom CNN, MobileNetV2, and EfficientNet-B0. A public dataset containing 12,446 CT images across Normal, Cyst, Tumor, and Stone classes was partitioned using stratified sampling into training, validation, and test sets. Four scenarios—baseline, spatial, intensity, and hybrid augmentation—were evaluated across three independent random seeds using accuracy, precision, recall, F1-score, and AUC. Results and Discussion: The baseline achieved mean accuracies of 99.96%, 98.32%, and 94.06% for CNN, MobileNetV2, and EfficientNet-B0, respectively. Intensity augmentation slightly improved CNN accuracy to 99.99% and consistently produced smaller performance degradation and more stable convergence than spatial and hybrid augmentation. Spatial and hybrid transformations generally reduced classification performance, indicating that excessive geometric changes may disrupt diagnostically relevant anatomical features. Conclusion: Baseline training provided the best overall performance, while intensity augmentation was the most effective augmentation strategy, demonstrating that preserving anatomically meaningful image characteristics is more important than indiscriminately increasing data diversity.]]>
Copyrights © 2026
