radMLBench: A dataset collection for benchmarking in radiomics

Area of Science:

• Radiomics
• Machine Learning
• Data Science

Background:

• Machine learning in radiomics often lacks robust validation due to single-dataset testing.
• A need exists for larger, accessible datasets to benchmark new radiomics methods.
• This study curates a collection of radiomics datasets with binary outcomes.

Purpose of the Study:

• To create a comprehensive, accessible radiomics dataset collection for benchmarking.
• To evaluate the impact of feature decorrelation on predictive performance in radiomics pipelines.
• To enhance the reliability and robustness of machine learning applications in radiomics.

Main Methods:

• A systematic search identified tabular radiomics datasets with binary outcomes.
• Collected datasets were compiled into a homogeneous, Python-accessible collection.
• The dataset collection was used to test feature decorrelation's effect on predictive performance.

Main Results:

• Fifty radiomic datasets were curated, varying in sample size and feature count.
• Feature decorrelation prior to feature selection showed no significant average performance improvement.
• The findings suggest feature decorrelation may not be essential in radiomics pipelines.

Conclusions:

• A valuable, accessible radiomics dataset collection is now available for method benchmarking.
• Feature decorrelation does not consistently enhance predictive performance in radiomics.
• Omitting feature decorrelation can improve pipeline robustness and reliability.