Precision in Dermatology: Developing an Optimal Feature Selection Framework for Skin Lesion Classification

Area of Science:

• Dermatology
• Artificial Intelligence
• Machine Learning

Background:

• Melanoma is a dangerous skin cancer with increasing incidence.
• Early detection significantly improves patient survival rates.
• Current computer-based diagnostic methods still have limitations and a margin of error.

Purpose of the Study:

• To develop an advanced computer-based method for early melanoma detection.
• To maximize feature information by combining deep learning models.
• To reduce noise and redundancy using an evolutionary feature selection technique.

Main Methods:

• Utilized deep models (Darknet53, DenseNet201, InceptionV3, InceptionResNetV2) for feature extraction.
• Applied transfer learning to improve model performance.
• Integrated features from multiple models and employed a novel entropy-controlled gray wolf optimization (ECGWO) algorithm for feature selection.
• Validated the approach on PH2, ISIC-MSK, and ISIC-UDA dermoscopic datasets.

Main Results:

• The proposed method effectively maximized input feature information.
• The ECGWO algorithm successfully reduced noisy and redundant features.
• The integrated fusion and selection techniques generated highly discriminant feature information.
• The approach demonstrated effectiveness on benchmark datasets, outperforming established techniques.

Conclusions:

• The novel approach of combining deep models and evolutionary feature selection shows significant promise for accurate melanoma detection.
• This method enhances the discriminative power of features, leading to improved diagnostic accuracy.
• The study addresses a key research challenge in machine learning for medical image analysis.