Classification of skin lesions in dermoscopic images via deep transfer learning and quaternion Hahn moment invariant-based feature extraction

Skin lesions are areas of skin that differ in appearance, texture, or color from the surrounding skin. They can result from a variety of causes, ranging from harmless conditions like minor injuries to serious cases such as skin cancer. Distinguishing between benign and malignant lesions typically requires the expertise of a dermatologist. Using tools like dermoscopy, dermatologists can achieve a diagnostic accuracy rate of 75 % to 90 %. We propose a novel classification framework that integrates Quaternion Hahn Moment Invariants, optimized via the Firefly Algorithm, into a customized ResNeXt50 transfer learning architecture, combining the geometric invariance and inter-channel color encoding of quaternion moments with the hierarchical feature learning of deep pretrained networks. Moment invariance enables the model to detect lesions regardless of their scale, rotation, or position, reducing the reliance on extensive geometric data augmentation. The architecture features a dedicated input processing head that adapts moment-based representations to the backbone, selective fine-tuning of the deeper ResNeXt50 layers while preserving pretrained low- and mid-level features, and a custom classification head consisting of residual dense blocks with dropout regularization for robust seven-class prediction. Using the HAM10000 dataset, our proposed model demonstrated 99.24 % AUC, 99.63 % balanced accuracy, 0.9753 precision, 0.9931 recall, and 0.9841 F1-score.