A HYBRID METAHEURISTIC AND TRANSFER LEARNING APPROACH FOR AUTOMATED SKIN CANCER CLASSIFICATION

Abstract

Skin cancer develops when skin cells begin to grow uncontrollably, often as a result of prolonged exposure to ultraviolet (UV) radiation from the sun. However, the disease is not limited to sun-exposed areas and can also occur on parts of the body that receive little or no sunlight. The three most common forms of skin cancer are melanoma, basal cell carcinoma, and squamous cell carcinoma. Among these, melanoma is considered the most dangerous because of its strong tendency to spread to other organs if not diagnosed and treated at an early stage. It may develop from an existing mole or emerge as a new dark-colored lesion that differs from the surrounding skin. In comparison, basal cell carcinoma and squamous cell carcinoma generally progress more slowly and are associated with a lower risk of metastasis. The outcome of skin cancer treatment largely depends on the cancer type, the patient's overall health, and, most importantly, the stage at which the disease is detected. Recent advances in artificial intelligence (AI) have significantly improved the accuracy and efficiency of medical image analysis, making AI-based techniques valuable tools for the early diagnosis of skin cancer. Deep learning methods, particularly Convolutional Neural Networks (CNNs), have demonstrated outstanding performance in automatically learning complex visual patterns from dermoscopic images. These models reduce the reliance on manual feature engineering and assist clinicians by providing fast and reliable diagnostic support, which can contribute to earlier treatment and improved patient outcomes. Several studies have explored the integration of deep learning with optimization algorithms to enhance skin cancer classification. In one approach, a CNN model was employed to extract discriminative features from dermoscopic images. Feature selection was then performed using metaheuristic optimization techniques, including Particle Swarm Optimization (PSO) and the Genetic Algorithm (GA), to identify the most informative feature subsets. The optimized features were subsequently classified using a Support Vector Machine (SVM), resulting in a classification accuracy of 89.17%, demonstrating the effectiveness of combining deep feature extraction with optimization-based feature selection. Another study proposed a segmentation framework based on the U-Net++ architecture with a DenseNet201 backbone. The model achieved strong performance across multiple evaluation metrics, reporting an accuracy of 94.16%, an F1-score of 91.39%, an AUC of 99.3%, an Intersection over Union (IoU) of 96.8%, a Dice coefficient of 77.19%, and a segmentation score of 75.47%. These results highlight the capability of advanced deep learning architectures to accurately identify and segment skin lesions in dermoscopic images. Skin cancer continues to represent a major public health concern, with approximately 3.5 million cases diagnosed annually in the United States. As the disease progresses, treatment becomes increasingly challenging and survival rates decline, emphasizing the importance of early and accurate diagnosis. However, conventional diagnostic procedures can be both time-consuming and expensive. To address these limitations, researchers have developed automated computer-aided diagnostic systems for lesion detection, segmentation, and classification. In one such study, a threshold-based framework was introduced for skin lesion analysis, while the hyper parameters of eight well-known CNN architectures—including VGG16, VGG19, MobileNet, and NASNet—were optimized using the SPASA optimization algorithm. This optimization strategy enhanced the predictive performance of the deep learning models, demonstrating the value of metaheuristic optimization in improving automatedskin cancer diagnosis.