Automatically Diagnosing Skin Cancers From Multimodality Images Using Two-Stage Genetic Programming.

Area of Science:

• Computer-aided diagnosis
• Machine learning
• Dermatology

Background:

• Skin malignancy detection from images is a growing research area.
• Genetic programming (GP) offers flexible model evolution and automatic feature selection/construction.
• Existing GP methods may use redundant features, impacting classification performance.

Purpose of the Study:

• To develop a novel two-stage GP approach for improved skin malignancy classification.
• To enhance classification performance by optimizing feature selection and construction.
• To create interpretable models for dermatologists.

Main Methods:

• A two-stage genetic programming (GP) method was developed.
• Stage one: Prominent feature selection.
• Stage two: New feature construction from selected features and operators (e.g., multiplication). Features extracted using local binary patterns and pyramid-structured wavelet decomposition.

Main Results:

• The proposed two-stage GP method significantly improved classification performance.
• The approach effectively selected relevant features and constructed informative new ones.
• Performance was validated on two real-world skin image datasets and compared favorably against other methods.

Conclusions:

• The novel two-stage GP method enhances machine learning classification for skin malignancy detection.
• The evolved GP models are interpretable, aiding dermatologists in identifying key diagnostic features.
• This approach offers a promising tool for computer-aided diagnosis in dermatology.