Enhanced VGG16 Model for Kidney Tumor Classification

Background/Objectives: The timely identification of renal malignancies, particularly the accurate categorization of neoplastic subtypes, poses significant diagnostic challenges. Traditional techniques such as manual diagnosis and histopathological analysis are re-source-intensive, prone to inter-observer variability, and often lack scalability. Although recent deep learning (DL) approaches have demonstrated promising performance in kidney tumor classification, they still face critical limitations including poor interpreta-bility (“black-box” behavior), reduced generalization across heterogeneous cohorts, and computational inefficiency that hinders real-time clinical deployment. In response to these persistent challenges, we propose the M16+ model—an enhanced, VGG16-based deep learning architecture optimized to improve diagnostic accuracy, interpretability, and deployment feasibility in renal oncology workflows. Methods: The M16+ model consists of a pre-trained VGG16 convolutional backbone for robust feature extraction, followed by a custom classifier comprising batch-normalized dense layers and dropout regularization to reduce overfitting. The model was trained on a class-balanced cohort of 4,200 contrast-enhanced axial CT slices (2,100 benign, 2,100 malignant) obtained from 120 patients. The dataset was partitioned into 2,688 training (64%), 672 validation (16%), and 840 independent test slices (20%). A stratified 5-fold cross-validation was applied within the training set for hyperparameter tuning. The architectural design incorporates dual convolutional blocks to stabilize learning and promote discriminative representation across layers. Results: The M16+ model achieved a test-set accuracy of 98.0% (n = 840) with an AUC of 0.96, outperforming the baseline CNN-4 model by 2.6 percentage points. The incorporation of Gradient-weighted Class Activation Mapping (Grad-CAM++) ena-bled visual interpretation of the most influential regions contributing to each prediction, thereby enhancing clinician trust in model outputs. Conclusions: The proposed M16+ framework offers a robust, interpretable, and efficient computational solution for renal tumor classification. Its high diagnostic accuracy, coupled with model transparency and generalization, underscores its potential for clinical integration. Future efforts will focus on validating the model across external datasets and adapting it to heterogeneous im-aging protocols to assess real-world performance.