Image feature selection by a genetic algorithm: application to classification of mass and normal breast tissue

Area of Science:

• Medical Imaging
• Machine Learning
• Biomedical Engineering

Background:

• Accurate differentiation of masses from normal tissue in mammograms is crucial for early cancer detection.
• Traditional feature selection methods may not optimally capture complex patterns in mammographic data.

Purpose of the Study:

• To introduce and evaluate a novel genetic algorithm (GA) based approach for image feature selection in mammogram analysis.
• To compare the performance of GA-based feature selection against stepwise feature selection for classifying regions of interest (ROIs).

Main Methods:

• Utilized a genetic algorithm (GA) for feature selection, guided by the area under the receiver operating characteristic (ROC) curve (Az) as the fitness measure.
• Employed linear discriminant classifiers and backpropagation neural networks (BPNs) for classification tasks.
• Extracted 587 features (572 texture, 15 morphological) from 168 mass ROIs and 504 normal ROIs.

Main Results:

• The GA-based feature selection achieved an average test Az of 0.90 with a linear discriminant classifier, outperforming stepwise selection (0.89).
• Similar performance (0.90 Az) was observed with a BPN classifier and GA-based selection.
• The study explored the impact of various GA parameters on classification accuracy.

Conclusions:

• Genetic algorithms offer a versatile and effective method for feature selection in mammogram analysis.
• GA-based feature selection demonstrates comparable or superior performance to traditional methods, enhancing classifier design.
• This approach supports the development of robust linear and nonlinear classifiers for medical image analysis.