Radiomics Applicability Domain Analysis Classification Framework (RADAN-CF): A method for evaluating prediction...

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Radiomics applications

Background:

• Radiomics models offer clinical decision support but lack reliable assessment, especially with data shifts.
• Existing uncertainty methods fail with differing test data structures.
• A new framework is needed for reliable radiomics prediction assessment.

Purpose of the Study:

• Introduce the Radiomics Applicability Domain ANalysis - Classification Framework (RADANCF).
• Enable transparent, per-prediction reliability assessment for radiomics classification.
• Address limitations in current uncertainty estimation methods under distributional shift.

Main Methods:

• RADANCF uses six criteria across data representativeness and model behavior.
• Reliability categories are summarized using a traffic-light scheme.
• Evaluated on six datasets with five classifiers, including external validation on 2689 prostate cancer patients.

Main Results:

• Prediction error associated with RADANCF categories, with intermediate categories showing high error.
• RADANCF criteria were complementary, with low correlations.
• RADANCF category significantly impacted predictions, controlling for dataset/model effects.
• Specific criteria combinations overrepresented in high-error predictions.
• External validation confirmed RADANCF's diagnostic and risk-oriented nature.

Conclusions:

• RADANCF offers transparent, per-prediction reliability assessment for radiomics.
• It accounts for data representativeness and model behavior.
• Complements traditional metrics, supporting cautious radiomics model deployment.