Classification of melanomas in situ using knowledge discovery with explained case-based reasoning.

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Early melanoma diagnosis relies on the ABCD rule (Asymmetry, Border irregularity, Color variegation, Diameter > 5mm).
• Dermoscopy improves lesion evaluation, reducing unnecessary excisions of benign lesions.
• Accurate dermatoscopic classification is crucial for effective melanoma prevention.

Purpose of the Study:

• To introduce LazyCL, a novel procedure designed to assist dermatologists in classifying skin lesions.
• To generate a domain theory for classifying in situ melanomas using LazyCL.
• To leverage artificial intelligence for improved diagnostic support in dermatology.

Main Methods:

• LazyCL combines case-based reasoning and clustering for domain theory generation.
• It employs lazy induction of descriptions (LID) with leave-one-out cross-validation.
• The process involves clustering, explanation collection, and refinement to create a preliminary domain theory.

Main Results:

• LazyCL produced results comparable to the Self-Organizing Maps (SOM) clustering method on a melanoma dataset.
• Experiments on standard machine learning datasets confirmed the correctness of LazyCL's generated clusters.
• The domain theory generated by LazyCL was found to be easily understandable by domain experts.

Conclusions:

• LazyCL is a feasible method for knowledge discovery and constructing domain theories in medical contexts.
• The system's explanations are interpretable by experts, aligning with their existing knowledge attributes.
• LazyCL demonstrates effectiveness as a clustering method, producing accurate classifications.