<![CDATA[Kidney diseases remain a significant global health challenge, and computed tomography (CT) plays an important role in supporting their diagnosis through detailed visualization of renal structures. In deep learning-based medical image classification, data augmentation is commonly employed to mitigate the limitations of small training datasets; nevertheless, the effects of distinct augmentation strategies on image characteristics and learning behavior across architectures remain insufficiently explored. This study investigates the impact of spatial, intensity, and hybrid augmentation on kidney CT image classification using CNN, MobileNetV2, and EfficientNet-B0 architectures. A stratified data split was adopted, and all experiments were repeated using three random seeds (1, 42, and 123), with results reported as mean ± standard deviation. Four training scenarios (baseline, spatial, intensity, and hybrid augmentation) were evaluated using accuracy, precision, recall, F1-score, and Area Under the Curve (AUC). The baseline configuration achieved the highest overall performance, reaching mean accuracies of 99.96 ± 0.04%, 98.32 ± 0.06%, and 94.06 ± 0.58% for CNN, MobileNetV2, and EfficientNet-B0, respectively. Among the augmentation strategies, intensity augmentation achieved the highest performance only for CNN while consistently exhibiting smaller performance degradation relative to the baseline and more stable convergence than the spatial and hybrid augmentation approaches. The consistently high baseline performance, particularly for CNN, underscores the importance of rigorous validation on public medical imaging datasets. These findings indicate that preserving anatomically relevant image characteristics is more beneficial than indiscriminately increasing data diversity. Therefore, while the baseline remained the best overall training strategy, intensity augmentation was the most effective augmentation approach]]>
Copyrights © 2026
