MelanomaNet: RL-Guided Hyperparameter Optimization in Deep Meta-Ensembles for Melanoma Image Classification

Abstract

Skin cancers, particularly melanoma, constitute a public health issue as they are highly aggressive and deadly if not caught early. This research study proposes an energetically inspired, conscientious deep learning framework for dermoscopic image classification of skin lesions (benign/malignant). Our method relies on an ensemble of state-of-the-art convolutional neural networks combined with expressive morphological and color-based shape descriptors, which are further ensemble learning using ensemble learning, test-time augmentation (TTA), and reinforcement learning (RL) based hyperparameter search. Hybrid deep architecture of EfficientNetV2, NFNet, and ResNet were utilized here, with both focal loss and ArcFace margin product fused at training to balance class imbalance and improve feature discriminability. Shape-aware dual-branch model structure was utilized for fusing visual features and dermatologically important handcrafted features. Validation of the model was carried out with 5-fold cross-validation and large test set validation. The system performed 95.35% classification accuracy, AUC of 0.992, and PR- AUC of 0.992 on the test set with superior diagnostic performance and generalizability. Our findings illustrate the potential of state-of-the-art ensemble AI systems to aid dermatological diagnostics with high reliability and explainability.