AN INTELLIGENT DEEP LEARNING FRAMEWORK FOR SKIN CANCER DETECTION USING TRANSFER LEARNING AND PARTICLE SWARM OPTIMIZATION

Abstract

When the skin is exposed to sunlight for extended periods, it can trigger abnormal cell growth, leading to a condition commonly known as skin cancer. Interestingly, this disease can also develop in areas of the body that receive little or no sun exposure. The three primary types of skin cancer are melanoma, squamous cell carcinoma, and basal cell carcinoma.

Although skin cancer can be serious and even life threatening, the prognosis largely depends on factors such as the cancer type, the individual’s general health, and how early the disease is diagnosed. Melanoma is considered the most dangerous form since it has a high potential to spread to other parts of the body. It may arise from an existing mole or appear suddenly as an unusual dark spot that differs from the surrounding skin. In contrast, basal cell carcinoma and squamous cell carcinoma tend to progress more slowly and are less likely to be fatal if treated promptly.

Recent advances in artificial intelligence (AI) have revolutionized disease detection, allowing faster and more accurate diagnoses that lead to timely treatment. Among the most powerful AI methods are convolutional neural networks (CNNs), which excel at analyzing medical images and identifying subtle visual patterns with remarkable precision.

One study demonstrated the use of an efficient CNN model to analyze dermatological images of skin cancer. The researchers extracted essential visual features from these images and optimized them using two popular algorithms— Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) —to select the most informative features. The final classification was performed using a Support Vector Machine (SVM) , achieving an accuracy rate of  89.17% , indicating promising diagnostic capability.

In another investigation, a U-Net++ model combined with DenseNet201 as its backbone architecture showed even stronger performance across several evaluation metrics. It achieved an  accuracy of 94.16% , an  F1-score of 91.39% , an  AUC of 99.3% , an  IoU of 77.19% , and a  Dice coefficient of 75.47% , demonstrating superior precision and reliability in segmentation tasks.

Globally, skin cancer remains one of the most common malignancies, with approximately  3.5 million new cases reported annually in the United States alone . The survival rate decreases sharply as the disease advances, highlighting the importance of early detection. However, early diagnosis can be both challenging and costly. To mitigate this, another study introduced an automated, threshold-based system  for detecting, categorizing, and segmenting skin cancer lesions. The researchers employed a  smart optimization technique known as SPASA  to fine-tune the parameters of eight popular CNN architectures—including  VGG16, VGG19, MobileNet,  and  NASNet —to maximize diagnostic performance.