Improving Skin Lesion Diagnosis: A Hybrid Approach Using Orthogonal Combination of Local Binary Pattern Features and...

Area of Science:

• Dermatology
• Medical Imaging
• Machine Learning

Background:

• Increasing global wealth and healthcare expectations drive demand for services.
• Healthcare systems face pressure, necessitating efficient diagnostic solutions.
• Early skin lesion diagnosis is crucial for effective treatment and patient outcomes.

Purpose of the Study:

• To develop a machine learning (ML) model for automated skin lesion diagnosis.
• To focus on early detection of skin lesions for improved patient care.
• To classify eight different types of skin lesions.

Main Methods:

• Utilized Gaussian Mixture Models (GMM) with geometric features for image artifact removal.
• Employed a color descriptor based on hybrid orthogonal local binary patterns for lesion characterization.
• Applied ReliefF feature selection to identify key diagnostic features.
• Evaluated multiple ML models including Decision Tree, Random Forest, k-NN, MLP, and Ensemble Extra Trees (ET).

Main Results:

• Ensemble Extra Trees (ET) demonstrated superior performance in classifying skin lesions.
• Achieved a high accuracy rate of 97.31% with the ET model.
• The selected features significantly contributed to accurate classification.

Conclusions:

• The developed ML approach enhances early skin lesion diagnosis.
• This advancement improves the quality of dermatological patient care.
• Automated diagnosis offers a potential solution to increasing healthcare demands.