Ensemble-ASNet: An Optimized Weighted CNN Ensemble for Multi-Class Kidney Lesion Classification Using CT Images

Preventing the progression of chronic kidney disease and minimizing diagnostic delays requires accurate detection of renal abnormalities such as cysts, nephrolithiasis, tumors, and normal tissue. Although deep learning-based medical imaging has made significant progress, many current models still suffer from issues such as overfitting, uneven class performance, poor interpretability, and difficulty combining extra feature representations effectively. To overcome these difficulties, this work proposes Ensemble-ASNet. This optimized weighted soft-voting ensemble combines three highly effective pre-trained CNN backbones: InceptionV3, VGG16, and MobileNetV2, to enhance robustness and discriminative power in four-class renal CT classification. Optimal model weights were determined through a constrained optimization strategy applied to validation predictions. Ensemble-ASNet outperformed all individual models and numerous cutting-edge CNN, transformer, and hybrid ensemble techniques documented in the literature, achieving 99.42% accuracy, precision, recall, and F1-score. Grad-CAM++ visualizations demonstrate that the model concentrates on clinically significant anatomical regions, which improves interpretability and decision-making. Future research will examine multi-institutional datasets and lightweight ensemble deployment; limitations include the absence of Hounsfield Unit-calibrated imaging and patient-level validation. Overall, the results demonstrate that Ensemble-ASNet has significant potential for integration into clinical diagnostic workflows and offers a dependable and efficient framework for automated renal CT analysis.