Unsupervised sub-segmentation for pigmented skin lesions.

Area of Science:

• Medical image analysis
• Computational dermatology
• Artificial intelligence in healthcare

Background:

• Early detection of malignant melanoma is crucial for effective skin cancer treatment.
• Computer-aided diagnosis systems aid in improving diagnostic accuracy for skin lesions.
• Automated segmentation of skin lesions is a critical first step, complicated by darker areas within lesions that hold diagnostic value.

Purpose of the Study:

• To present an unsupervised segmentation scheme for simultaneously isolating normal skin, pigmented skin lesions, and darker areas within lesions.
• To develop a method that enhances the diagnostic information extracted from skin lesion images.

Main Methods:

• An adaptive mean-shift algorithm was applied in a 5D spatial color-texture feature space to create homogenous regions.
• Sub-segmentation maps were generated using maximal similarity-based region merging and kernel k-means clustering.
• Cluster validity measurement was employed to determine the optimal number of segments.

Main Results:

• The proposed method demonstrated competitive performance in automatic segmentation on both digital photographs and dermoscopy images.
• Isolated dark areas within lesions proved valuable in discriminating malignant melanomas from benign nevi.
• Incorporating asymmetry descriptors of isolated dark areas increased malignant melanoma identification accuracy from 65.38% to 80.77%.

Conclusions:

• The segmentation scheme achieved lesion boundaries comparable to those delineated by experienced dermatologists.
• Analysis of darker area distributions within lesions shows significant promise for melanoma characterization.