Dynamic clustering detection through multi-valued descriptors of dermoscopic images

Area of Science:

• Dermatology and Medical Imaging
• Computational Biology and Bioinformatics

Background:

• Melanoma is a deadly skin cancer, and early diagnosis is crucial but challenging for clinicians.
• Current data analysis for skin lesion discrimination often relies on categorical or numerical features.
• Distinguishing between malignant melanoma and benign nevi in uncertain cases requires advanced diagnostic tools.

Purpose of the Study:

• To introduce a novel dynamic clustering methodology for analyzing multi-valued descriptors from dermoscopic images.
• To support medical diagnosis by classifying uncertain pigmented skin lesions.
• To improve the accuracy of differentiating malignant melanoma from benign nevi.

Main Methods:

• Utilized multi-valued descriptors (intervals, histograms) beyond scalar variables for image analysis.
• Developed a dynamic clustering method employing Wasserstein distance for comparing multi-valued data.
• Applied discriminant analysis to identify key features distinguishing benign and malignant lesions.
• Segmented dermoscopic images to extract multi-valued descriptors for classification.

Main Results:

• The proposed dynamic clustering method effectively characterized uncertain skin lesions.
• Specific descriptors associated with dermoscopic characteristics showed discriminatory power.
• The methodology provided a novel approach to aid dermatologists in diagnosing uncertain lesions.

Conclusions:

• The dynamic clustering methodology offers a promising tool for improving the diagnosis of uncertain pigmented skin lesions.
• The use of multi-valued descriptors in conjunction with Wasserstein distance clustering enhances lesion characterization.
• This approach can assist dermatologists in making more informed decisions regarding melanoma versus nevi classification.