Walrus Optimization-Enhanced ResNet-50 for AI-Driven Renal Malignancy Prediction with Occlusion Sensitivity-Based Interpretation

Objective: Main goal is to optimize the hyperparameters of ResNet-50 using the Walrus Optimization Algorithm (WaOA) to enhance classification performance for renal malignancy detection. The study aims to compare the WaOA-optimized ResNet-50 with conventional deep learning models, evaluate its effec-tiveness through various performance metrics, and integrate Occlusion Sensitivity Analysis to ensure model interpretability and transparency in AI-driven medical diagnosis. Methods: A total of 12,446 abdominal CT images were collected from multiple hospitals in Dhaka, Bangladesh, comprising four diagnostic categories: cyst (3,709 images), normal (5,077), stone (1,377), and tumor (2,283). Several deep learning models AlexNet, GoogLeNet, Inception V3, and ResNet-50 were trained and evaluated. The Walrus Optimization Algorithm (WaOA) was implemented to fine-tune hy-perparameters, including weight and bias learning rate for ResNet-50. The models were assessed using various performance metrices. Additionally, Occlusion Sensitivity Analysis was applied to visualize and interpret the model’s decision-making process by identifying critical regions in CT images that influence classification. Result: The WaOA-optimized ResNet-50 achieved superior performance with 94.53% accuracy, outper-forming other models in precision (93.28%), recall (91.32%), F1-score (92.16%), and AUC-ROC (99.33%), indicating enhanced diagnostic efficiency. The model demonstrated a strong MCC (0.9038) and lower log loss (0.1597), ensuring better reliability and confidence in predictions. Despite a slightly higher inference time (0.1133 sec), accuracy and computational efficiency was minimal. Conclusion: This article confirms the use of metaheuristic-based hyperparameter tuning for deep learning models in medical imaging. The WaOA-optimized ResNet-50 demonstrates significant improve-ments over conventional models, making it a promising tool for renal malignancy detection. The integra-tion of Occlusion Sensitivity Analysis ensures transparency and reliability in AI-assisted diagnosis. Future work can explore hybrid optimization techniques, multi-modal learning, and real-time clinical deployment to further enhance the model’s applicability in healthcare.