Ensuring Reproducibility and Deploying Models with the Image2Radiomics Framework: An Evaluation of Image Processing...

Area of Science:

• Radiomics
• Medical Imaging
• Artificial Intelligence in Oncology

Background:

• Image processing variability challenges the clinical deployment of radiomics models.
• Reproducibility is essential for reliable artificial intelligence (AI) tools in medical diagnostics.

Purpose of the Study:

• To assess the impact of image processing on a validated pancreatic neuroendocrine tumor (PanNET) detection model.
• To introduce Image2Radiomics, a novel framework for reproducible radiomics image processing.

Main Methods:

• Re-extracted radiomics features using Image2Radiomics from CT images.
• Evaluated the effect of nine image processing pipeline alterations on model performance.
• Quantified prediction discrepancies using probability differences and classification disagreement.

Main Results:

• Image2Radiomics successfully replicated the reference PanNET model (Cohen's kappa = 1).
• Image processing alterations significantly reduced model performance (AUC dropped from 0.87 to 0.71).
• Disagreements in predictions occurred in up to 45% of patients, with minor changes causing up to 21% disagreement.

Conclusions:

• Standardizing image processing pipelines is critical for radiomics reproducibility and clinical adoption.
• The Image2Radiomics framework offers a standardized approach to define and share processing pipelines.
• This framework facilitates the deployment of robust radiomics models in clinical and multicenter settings.