A Hybrid Stacked Restricted Boltzmann Machine with Sobel Directional Patterns for Melanoma Prediction in Colored Skin...

Area of Science:

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

Background:

• Melanoma, a dangerous skin cancer, is characterized by uncontrolled cell growth and poses diagnostic challenges due to varied tissue involvement and atypical borders.
• Accurate melanoma detection is crucial for effective clinical management, yet identifying it from color images remains difficult despite various proposed methods.
• Existing approaches often struggle with the complexity and variability of skin lesion appearances.

Purpose of the Study:

• To develop and evaluate a comprehensive system for efficient and precise classification of skin lesions, specifically targeting melanoma detection.
• To introduce an innovative feature extraction method, the enhanced Sobel Directional Pattern (SDP), for improved skin image analysis.
• To validate the efficacy of a stacked Restricted Boltzmann Machine (RBM) classifier in accurately categorizing skin melanoma.

Main Methods:

• A multi-stage framework comprising preprocessing (DullRazor for hair artifact removal), semantic segmentation (Fully Connected Neural Network - FCNN) to identify Regions of Interest (ROIs), and feature extraction using the enhanced Sobel Directional Pattern (SDP).
• The enhanced SDP was employed for extracting relevant features from ROIs, demonstrating superiority over the ABCD method for skin image analysis.
• Classification of skin ROIs was performed using a stacked Restricted Boltzmann Machine (RBM), known for its effectiveness in pattern recognition and classification tasks.

Main Results:

• The proposed system achieved high classification accuracies across five diverse datasets: 99.8% (PH2), 96.5% (ISIC 2016), 95.5% (ISIC 2017), 87.9% (Dermnet), and 97.6% (DermIS).
• The enhanced Sobel Directional Pattern (SDP) proved more effective for skin image analysis compared to the ABCD method.
• Stacked Restricted Boltzmann Machines (RBMs) demonstrated superior performance in categorizing skin melanoma, highlighting their potential in dermatological diagnostics.

Conclusions:

• The developed comprehensive system offers an efficient and precise method for skin lesion classification, particularly for melanoma detection.
• The enhanced Sobel Directional Pattern (SDP) is a robust feature extraction technique for dermatological images.
• Stacked Restricted Boltzmann Machines (RBMs) are highly effective for accurate skin cancer classification, showing significant promise for clinical application.