Automatic histologically-closer classification of skin lesions

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Melanoma is a prevalent and dangerous form of skin cancer, necessitating accurate and efficient diagnostic tools.
• Dermoscopy is crucial for evaluating skin lesions, but automated analysis methods are continuously being improved.
• Existing automated methods for melanoma classification have limitations in accuracy and adaptability.

Purpose of the Study:

• To develop and validate a novel automated melanoma classification approach using structural co-occurrence matrices (SCM) derived from dermoscopy image frequencies.
• To enhance the discriminative power of SCM by transforming it into an adaptive feature extractor.
• To compare the performance of the proposed SCM method against established techniques and recent works.

Main Methods:

• Extracted main frequencies from dermoscopy images to compute structural co-occurrence matrices (SCM).
• Developed an adaptive feature extraction technique based on SCM for improved melanoma classification.
• Validated the approach using the ISIC 2016, ISIC 2017, and PH2 dermoscopy image datasets.
• Evaluated performance using metrics including specificity, sensitivity, accuracy, and area under the curve (AUC).

Main Results:

• The proposed SCM method demonstrated high performance across all tested datasets.
• Achieved excellent accuracy (up to 99%) and sensitivity (up to 99.2%) in melanoma classification.
• Outperformed other methods like local binary patterns, gray-level co-occurrence matrix, and invariant moments of Hu.

Conclusions:

• The SCM approach offers an effective and automatic method for melanoma classification from dermoscopy images.
• This technique shows significant potential for improving the accuracy and efficiency of skin cancer diagnosis.
• The adaptive feature extraction capability of SCM enhances its utility in medical image analysis.