On the Automatic Detection and Classification of Skin Cancer Using Deep Transfer Learning

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer, including melanoma and non-melanoma, is a prevalent and often fatal disease.
• Early diagnosis of skin cancer significantly improves patient outcomes and can reduce treatment intensity and healthcare costs.
• Dermoscopy is a key diagnostic tool, and advancements in AI, particularly deep learning, are revolutionizing image-based diagnosis.

Purpose of the Study:

• To evaluate the effectiveness of raw deep transfer learning models for classifying skin lesions into seven categories.
• To assess the performance of 13 different deep transfer learning models using the HAM1000 dataset.
• To identify the advantages and limitations of applying deep transfer learning directly to dermoscopy images without explicit feature engineering.

Main Methods:

• Utilized the HAM1000 dataset comprising dermoscopy images of skin lesions.
• Developed a system employing 13 distinct deep transfer learning models.
• Input raw dermoscopy images directly into the models, bypassing traditional feature extraction and preprocessing steps.

Main Results:

• Some skin cancer types were classified with high accuracy by specific models.
• The overall best accuracy achieved was 82.9%.
• Dataset limitations, including class imbalance and a limited number of images in certain categories, impacted performance.

Conclusions:

• Raw deep transfer learning is a viable approach for skin lesion classification, demonstrating potential in AI-assisted dermatology.
• Dataset characteristics and the complexity of multi-class classification present challenges for achieving higher diagnostic accuracy.
• Further research and dataset refinement are necessary to enhance the reliability and performance of AI models for skin cancer detection.