Abstract:
Early and accurate cancer detection remains one of the most pressing challenges in clinical oncology, and deep convolutional neural networks (CNNs) have shown strong promise in automating diagnosis from medical images. However, the diagnostic performance of CNN architectures such as ResNet-50, InceptionV3, and VGG16 is highly sensitive to hyperparameter choices — including learning rate, batch size, and weight initialization — and manual or grid-based tuning is often computationally expensive and prone to local minima. This talk presents a unified research programme that addresses this challenge by integrating nature-inspired metaheuristic optimization algorithms with pretrained CNN backbones across four distinct cancer imaging domains: skin lesions, renal malignancy, oral cancer, and breast cancer.
In each study, a bio-inspired optimizer — the Whale Optimization Algorithm, the Walrus Optimization Algorithm, the Aquila Optimizer, and the Osprey Optimization Algorithm respectively — was used to automatically fine-tune network hyperparameters, replacing conventional manual and gradient-based tuning strategies. The optimized models were benchmarked against standard pretrained architectures (AlexNet, GoogLeNet, VGG16, ResNet-50, InceptionV3, Xception, MobileNet) using accuracy, precision, recall, F1-score, specificity, AUC-ROC, Matthews Correlation Coefficient, log loss, and inference time.
Across all four applications, the metaheuristic-optimized models consistently and substantially outperformed their conventional counterparts, achieving classification accuracies of 98.29% for skin lesion detection, 94.53% for renal malignancy prediction, 97.80% for oral cancer diagnosis, and 97.70% for breast cancer detection, alongside markedly reduced false-negative rates and improved AUC-ROC and MCC values. Notably, the optimized models also achieved favorable inference times, supporting their feasibility for real-time and resource-constrained clinical deployment. Interpretability techniques such as occlusion sensitivity analysis further validated that model attention aligned with clinically relevant regions of interest.
Collectively, these findings establish that bio-inspired metaheuristic hyperparameter optimization offers a generalizable, architecture-agnostic strategy for enhancing the accuracy, robustness, and computational efficiency of deep learning-based diagnostic systems across diverse cancer types. This presentation will discuss the shared methodological framework, comparative results, current limitations, and future directions, including multi-class classification and multimodal data integration for next-generation AI-assisted cancer diagnostics.
