A Computer-Aided Diagnosis System Using Deep Learning for Multiclass Skin Lesion Classification.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin cancer, particularly melanoma, is a significant public health concern with a low survival rate.
• Early detection of skin cancer is vital for reducing mortality.
• Manual dermatoscopic inspection is time-consuming and costly, necessitating automated solutions.

Purpose of the Study:

• To propose a novel automated framework for multiclass skin lesion classification using deep learning.
• To enhance the accuracy and efficiency of skin cancer diagnosis through an AI-driven approach.

Main Methods:

• Data augmentation techniques including rotation and flipping were applied to the HAM10000 dataset.
• Deep learning models (ResNet-50, ResNet-101) were fine-tuned and trained using transfer learning on augmented data.
• Feature extraction, fusion via a modified serial approach, and selection using skewness-controlled SVR were performed before classification.

Main Results:

• The proposed automated framework achieved an accuracy of 91.7% on the augmented HAM10000 dataset.
• The performance of the augmented dataset significantly outperformed the original imbalanced dataset.
• The developed method demonstrated superior performance compared to recent studies in the field.

Conclusions:

• The automated deep learning framework offers a promising and effective solution for multiclass skin lesion classification.
• Data augmentation and advanced feature selection techniques improve diagnostic accuracy in skin cancer detection.
• This AI-based approach has the potential to aid clinicians in early and accurate skin cancer diagnosis.