Deep Learning-Based Transfer Learning for Classification of Skin Cancer

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer is a major global health concern, with diagnosis challenging due to similar lesion appearances.
• Early detection of skin lesions is critical for successful treatment and patient outcomes.
• Automated diagnostic tools, powered by artificial intelligence (AI), show promise in assisting dermatologists.

Purpose of the Study:

• To conduct a comparative analysis of six distinct transfer learning neural networks for multi-class skin cancer classification.
• To evaluate the performance of these AI models using the HAM10000 dataset.

Main Methods:

• Employed image replication for underrepresented classes to address dataset imbalance.
• Evaluated VGG19, InceptionV3, InceptionResNetV2, ResNet50, Xception, and MobileNet transfer learning networks.
• Utilized the HAM10000 dataset for multi-class skin cancer classification.

Main Results:

• Image replication proved effective in improving classification performance.
• All tested models achieved high classification accuracies and F-measures with reduced false negatives.
• The Xception Net demonstrated superior performance, achieving 90.48% accuracy, highest recall, precision, and F-Measure.

Conclusions:

• Transfer learning models, particularly Xception Net, show significant potential for accurate automated skin cancer diagnosis.
• AI-assisted tools can enhance dermatologists' capabilities in identifying malignant skin lesions.
• Image augmentation techniques like replication are valuable for improving AI model performance on imbalanced medical datasets.