Skin lesion classification using features of 3D border lines.

Area of Science:

• Dermatology
• Medical Imaging
• Machine Learning

Background:

• Machine learning aids clinical diagnosis, particularly in classifying skin lesions from RGB images.
• Current 2D image features (shape, color, texture) have limitations in guaranteeing optimal classification results.

Purpose of the Study:

• To investigate the utility of 3D depth information for classifying pigmented skin lesions.
• To enhance diagnostic accuracy by exploiting lesion border line characteristics in a new dimension.

Main Methods:

• Extraction of features from the depth information of 3D images.
• Classification of lesions using a quadratic Support Vector Machine (SVM).
• Focus on utilizing depth data for melanoma detection.

Main Results:

• Achieved a geometric mean of 94.87% for melanoma detection.
• Demonstrated 100.00% sensitivity and 90.00% specificity.
• Outperformed other settings by providing more balanced sensitivity and specificity.

Conclusions:

• 3D depth information offers significant potential for improving skin lesion classification.
• Exploiting this overlooked dimension leads to more robust and balanced diagnostic performance.
• The proposed algorithm effectively addresses class imbalance issues in medical datasets.