Enhancing Skin Cancer Diagnosis Through Fine-Tuning of Pretrained Models: A Two-Phase Transfer Learning Approach.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer is a global health concern requiring early detection for better outcomes.
• Traditional diagnostic methods for skin cancer can be subjective and time-consuming.
• Machine learning offers potential for automated and accurate skin lesion classification.

Purpose of the Study:

• To evaluate the efficacy of transfer learning and fine-tuning techniques in classifying skin lesions using deep learning models.
• To compare the performance of nine different pretrained convolutional neural network (CNN) models on the HAM10000 dataset.

Main Methods:

• A two-phase deep learning approach was implemented: transfer learning with frozen layers, followed by fine-tuning all layers.
• Nine pretrained models (VGG16, VGG19, InceptionV3, Xception, DenseNet121, etc.) were evaluated.
• Performance was assessed using accuracy, precision, recall, and F1-score metrics on the HAM10000 dataset.

Main Results:

• The fine-tuned VGG16 model achieved the highest test set accuracy of 99.3%.
• Other models also demonstrated high performance, indicating the effectiveness of the proposed methodology.
• The study highlights the potential of deep learning for precise skin cancer classification.

Conclusions:

• Advanced machine learning models, particularly VGG16, show high efficacy in skin cancer classification.
• These findings support the integration of AI in dermatology for enhanced diagnostic accuracy.
• The study offers valuable insights for improving clinical decision-making in skin cancer detection.