Melanoma and Nevus Skin Lesion Classification Using Handcraft and Deep Learning Feature Fusion via Mutual Information...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Early detection of dangerous skin lesions, particularly melanoma, is crucial for patient survival.
• Existing diagnostic methods can be subjective and benefit from objective, automated analysis.
• Computer-Aided Detection (CAD) systems offer potential for enhanced diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate a novel CAD system for detecting and classifying melanoma.
• To fuse traditional dermatoscopic features (ABCD rule) with deep learning features for improved performance.
• To propose a method for adjusting evaluation metrics for imbalanced skin lesion datasets.

Main Methods:

• A CAD system incorporating preprocessing, feature extraction, feature fusion, and classification was designed.
• Handcrafted features based on the ABCD rule (Asymmetry, Borders, Colors, Dermatoscopic Structures) were extracted.
• Deep learning features were extracted using a pre-trained Convolutional Neural Network (CNN).
• Mutual Information (MI) was used for feature fusion, and classification was performed using methods like Support Vector Machines (SVMs).
• The framework was validated on the ISIC 2018 public dataset.

Main Results:

• The proposed CAD system demonstrated improved accuracy, specificity, and sensitivity compared to state-of-the-art methods.
• The fusion of handcrafted and deep learning features enhanced diagnostic performance.
• A novel procedure for adjusting evaluation metrics for imbalanced datasets was proposed and justified.

Conclusions:

• The developed CAD system shows significant promise for accurate melanoma detection and classification.
• Feature fusion combining dermatoscopic rules and deep learning is effective.
• Addressing data imbalance is critical for reliable evaluation of skin lesion diagnostic tools.