ENHANCED DEEP TRANSFER LEARNING FRAMEWORK USING PARTICLE SWARM OPTIMIZATION FOR AUTOMATED AND INTELLIGENT SKIN CANCER CLASSIFICATION

Abstract

Prolonged exposure to ultraviolet (UV) radiation from sunlight can damage skin cells and contribute to uncontrolled or abnormal cell growth, which may eventually lead to skin cancer. However, skin cancer is not restricted to sun-exposed regions and can also develop on parts of the body that receive little or no direct sunlight. The three major forms of skin cancer are melanoma, squamous cell carcinoma (SCC), and basal cell carcinoma (BCC). The severity and prognosis of the disease depend on several factors, including the type of cancer, the patient's general health condition, and the stage at which the disease is diagnosed. Among these types, melanoma is generally considered more aggressive because of its higher potential to invade surrounding tissues and spread to other parts of the body. It may originate from an existing mole or emerge as a new, unusually dark or irregular lesion. In comparison, basal cell carcinoma and squamous cell carcinoma generally have a lower risk of becoming life-threatening when detected and treated appropriately.Recent advances in artificial intelligence (AI) have created new possibilities for the early and automated diagnosis of various diseases. In particular, deep learning techniques have demonstrated considerable potential in medical image analysis because they can automatically learn meaningful patterns and visual features from large collections of images. Convolutional neural networks (CNNs) have become especially important in this area because of their ability to extract discriminative features from medical images and support accurate disease classification.

In one investigation, an efficient CNN-based architecture was employed to analyze skin cancer images and generate informative image features. The researchers subsequently investigated different feature combinations to identify the most discriminative subset. Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) techniques were applied to optimize the feature-selection process. The selected features were then provided to a Support Vector Machine (SVM) classifier for classification. The proposed approach achieved an accuracy of 89.17%, demonstrating the potential of combining deep feature extraction, evolutionary optimization, and machine learning for skin cancer diagnosis.

Another study investigated a U-Net++ segmentation architecture incorporating DenseNet201 as its backbone. The model demonstrated strong performance across several evaluation measures, including accuracy, F1-score, area under the curve (AUC), intersection over union (IoU), and Dice coefficient. The reported values were 94.16% accuracy, 91.39% F1-score, 99.30% AUC, 96.80% IoU, and 75.47% Dice coefficient, indicating the effectiveness of the architecture for automated skin lesion analysis and segmentation.

Skin cancer continues to represent a substantial public health challenge, with approximately 3.5 million cases reportedly diagnosed annually in the United States. As the disease advances, treatment becomes more difficult and the likelihood of favorable outcomes can decrease. Despite the importance of early diagnosis, conventional skin cancer assessment can be time-consuming, costly, and dependent on expert interpretation. To overcome some of these limitations, researchers have explored automated approaches for detecting, classifying, and segmenting skin lesions. One such study proposed a threshold-based automated methodology and incorporated an optimization technique known as SPASA to determine suitable parameter settings for eight established CNN architectures. These architectures included models such as VGG16, VGG19, MobileNet, and NASNet. The optimization process was intended to improve the performance of the CNN models and identify configurations capable of producing more reliable skin cancer analysis.

Keywords

Skin cancer detection, Transfer learning, Classical machine learning-based approach, Deep learning-based approach.