Progressive BAT for Neural Tuning (PBNT): A Bio-Inspired Hyperparameter Optimization Framework for Skin Lesion Classification

Abstract

Advanced architectures and effective hyperparameter tuning are necessary to achieve high diagnostic accuracy, despite the fact that deep learning (DL) has become a critical instrument for automated skin cancer detection.  This research suggests an optimised deep learning framework for the classification of multi-class skin lesions. The framework integrates state-of-the-art CNNs with a novel bio-inspired optimiser, Progressive Bat for Neural Tuning (PBNT).  VGG16, ResNet50, EfficientNetV2, AlexNet, and DenseNet were assessed in experiments conducted on the ISIC 2024 (3D-TBP) and HAM10000 datasets.  Bayesian Optimisation, Bat Algorithm, Grey Wolf Optimiser, and Firefly Algorithm were benchmarked against PBNT, and the performance was evaluated using the F1-score, precision, recall, and accuracy.  PBNT consistently outperformed existing optimisers, with AlexNet-PBNT achieving a 99.0% F1-score, 99.1% recall, 98.9% precision, and 99.0% accuracy, surpassing all other model-optimizer combinations and recent benchmarks.  The automated diagnosis of skin cancer is considerably improved by the integratioSSn of CNN architectures with hyperparameter tuning driven by PBNT.  The AlexNet-PBNT model offers a clinically viable and extremely accurate solution for early detection.