Advances in dermatological imaging: enhancing skin melanoma classification for improved patient outcomes

Abstract

The study presents an enhanced AlexNet-based deep learning system for binary classification of melanoma skin cancer as either benign or malignant using two paired dermatoscopic and clinical image datasets. The study evaluates the resilience of the models across different image sets with common preprocessing and specific data augmentation, using a melanoma dataset containing 10,000 images and a benign versus malignant dataset with 3,600 images. The AlexNet refinement exceeded several standard machine learning (ML) classifiers and other deep architectures on the two datasets with practical training times, gaining 97.12% and 96.21% in balanced accuracy. The training proceeded with SGD as optimiser and cross-entropy as loss on 256×256 images. Benchmarking against support vector machine (SVM), k-nearest neighbour (KNN), and other convolutional neural networks (CNNs) designs shows that the selected architecture and hyperparameters achieved the highest performance on cost-effective computation for the routine melanoma triage. The report highlights the need for external validation, incorporation into dermatological workflows, and explainability to improve trust, diminish dataset bias, and support the safe clinical deployment in practice.