Feature Reduction in Graph Analysis.

Area of Science:

• Medical Image Analysis
• Machine Learning in Healthcare
• Computational Biology

Background:

• Improving medical image classification often involves adding features, increasing computational costs and not always enhancing performance.
• Existing feature selection algorithms struggle to effectively remove redundant features in medical image analysis.
• Over 50 features are commonly used in image feature classification literature.

Purpose of the Study:

• To develop a new feature selection algorithm for medical image classification that reduces computational complexity.
• To eliminate redundant features that offer similar contributions to classifier decisions.
• To maintain or improve cancer detection accuracy with a reduced feature set.

Main Methods:

• A novel feature selection algorithm based on graph analysis of feature interactions and feature-to-classifier decision relationships.
• Modification of path analysis incorporating regression analysis, multiple logistic, and posterior Bayesian inference.
• Experiments conducted on a database of 113 mammograms from the Mammographic Image Analysis Society.

Main Results:

• A 13-feature set selected by the new algorithm achieved comparable cancer detection accuracy (true positive and false-positive rates) to a 26-feature set selected by Sequential Floating Selection (SFS) and using all 50 features.
• The proposed algorithm significantly reduced the computational resources required for preprocessing and classifier training.
• Tested classifiers included Artificial Neural Network (ANN) and logistic regression.

Conclusions:

• The developed graph-based feature selection algorithm effectively identifies a minimal set of discriminative features for medical image classification.
• This approach offers a computationally efficient alternative to existing methods, reducing processing time and resource demands.
• The findings suggest that a reduced feature set can yield robust cancer detection performance in mammography.