Four-class classification of skin lesions with task decomposition strategy

Area of Science:

• Dermatology and Medical Imaging
• Computer-Aided Diagnosis
• Computational Pathology

Background:

• Computer-aided skin lesion classification aids in skin cancer detection.
• Existing methods often overlook nonmelanocytic skin lesions (NoMSLs) like basal cell carcinoma (BCC) and seborrheic keratosis (SK).
• Accurate classification of both melanocytic skin lesions (MSLs) and NoMSLs is crucial for dermatologists in training and non-specialists.

Purpose of the Study:

• To develop and evaluate a novel computer-aided method for classifying four types of skin lesions: melanomas, nevi, BCCs, and SKs.
• To compare the performance of a layered classification model against baseline flat models.
• To identify key features contributing to accurate skin lesion classification.

Main Methods:

• A computer-aided method was developed, extracting 828 features categorized into color, subregion, and texture.
• Two classification models were implemented: a layered model using task decomposition and flat models for baseline comparison.
• The models were trained and tested on 964 dermoscopy images comprising melanomas, nevi, BCCs, and SKs.

Main Results:

• The layered model achieved superior performance compared to flat models.
• Detection rates for the layered model were 90.48% for melanomas, 82.51% for nevi, 82.61% for BCCs, and 80.61% for SKs.
• Irregularity of color distribution was identified as an effective classification feature.

Conclusions:

• The proposed layered computer-aided method shows significant promise for improving the accuracy of skin lesion classification.
• The method's ability to classify both MSLs and NoMSLs enhances its clinical utility.
• Further development could lead to more robust diagnostic tools for skin cancer detection.