Comparative Study Of Gradient Optimizers For Enhanced Skin Cancer Image Analysis

Skin disease is one of the most prominent types of cancer affecting all genders. Mostly, it’s characterized by abnormal skin growths and strange spots on the skin surface caused by prolonged exposure to ultraviolet (UV) light from the sun or other external sources. Skin cancer/disease is generally classified into four main categories: basal cell carcinoma, Pigmented Benign Keratosis, Nevus, and melanoma. This study evaluates the performance of three gradient-based optimizers Adam, Stochastic Gradient Descent (SDG), and RMSprop on a DL model for skin disease classification. The model classifies various skin conditions as basal cell carcinoma, melanoma, nevus, and pigmented benign keratosis. The classes are evaluated by the following accurate Metrics; F1 score, sensitivity, specificity, and area under the curve (AUC) are used to compare the optimizers. Our findings indicate that gradient optimizers, Adam and RMSprop significantly outperform SDG across all disease classes. RMSprop demonstrates the highest overall accuracy, particularly for melanoma, with an AUC of 1.00, strong sensitivity of 0.91, and precision of 0.96. Adam also shows competitive performance with high F1 scores and generalization. In contrast, SDG underperforms due to its fixed learning rate, leading to lower precision and sensitivity in critical conditions. The findings suggest that adaptive optimizers, especially RMSprop, improve the accuracy and validity of deep learning models in skin cancer/disease detection. Thus, making them suitable for clinical applications in skin disease diagnosis.