A multi-objective based radiomics feature selection method for response prediction following radiotherapy.

Area of Science:

• Radiomics and Medical Imaging Analysis
• Computational Biology and Bioinformatics
• Oncology and Personalized Medicine

Background:

• Radiomics offers rich data for predicting treatment response in personalized medicine.
• High-dimensional and redundant features in radiomics necessitate effective feature selection for predictive model development.

Purpose of the Study:

• To introduce a novel multi-objective radiomics feature selection method (MRMOPSO) for enhanced treatment response prediction.
• To jointly optimize the number of features, sensitivity, and specificity in radiomics feature selection.

Main Methods:

• The MRMOPSO method incorporates Fisher score for faster convergence and min-redundancy particle generation with truncation for feature reduction.
• Elitism strategies are employed to enhance the local search capabilities of the algorithm.
• The method was validated on a multi-institution oropharyngeal cancer dataset, including training, cross-validation, and independent evaluation sets.

Main Results:

• The MRMOPSO achieved Area Under the Curve (AUC) values of 0.82 for cross-validation and 0.84 for the independent dataset.
• MRMOPSO demonstrated significantly higher AUCs compared to Lasso, mRMR, F-score, and mutual information methods.
• The proposed method outperformed single-objective (GA, PSO) and other multi-objective (MOPSO, NSGA2) feature selection algorithms.

Conclusions:

• The developed MRMOPSO effectively reduces feature dimensions in radiomics analyses.
• This multi-objective approach significantly improves the performance of prediction models for radiotherapy response.
• The MRMOPSO method offers superior sensitivity and specificity for treatment response prediction in oncology.