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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Renal malignancy detection is crucial for patient outcomes.
• Traditional deep learning models require optimization for complex medical image analysis.
• Interpretability and transparency are key challenges in AI-driven medical diagnosis.

Purpose of the Study:

• To optimize ResNet-50 hyperparameters using the Walrus Optimization Algorithm (WaOA) for improved renal malignancy detection.
• To compare the performance of the WaOA-optimized ResNet-50 against conventional deep learning models.
• To enhance model interpretability and transparency through Occlusion Sensitivity Analysis.

Main Methods:

• A dataset of 12,446 abdominal CT images was utilized, categorized into cyst, normal, stone, and tumor.
• ResNet-50, AlexNet, GoogLeNet, and Inception V3 models were trained and evaluated.
• The Walrus Optimization Algorithm (WaOA) was employed for hyperparameter tuning of ResNet-50.
• Occlusion Sensitivity Analysis was performed for model interpretability.

Main Results:

• The WaOA-optimized ResNet-50 achieved 94.53% accuracy, outperforming other models in precision, recall, F1-score, and AUC-ROC.
• The model demonstrated high reliability with an MCC of 0.9038 and low log loss of 0.1597.
• Occlusion Sensitivity Analysis provided insights into critical image regions influencing classification decisions.

Conclusions:

• Metaheuristic-based hyperparameter tuning is effective for deep learning in medical imaging.
• The WaOA-optimized ResNet-50 shows significant potential for accurate and reliable renal malignancy detection.
• Integrating Occlusion Sensitivity Analysis ensures transparency and trustworthiness in AI-assisted medical diagnosis.