Counterfactual MRI Data Augmentation using Conditional Denoising Diffusion Models:training AI for different clinical scenarios

Motivation: Medical deep learning (DL) models struggle with generalizability due to MR image acquisition parameters (IAP) variations, limiting their clinical utility. Goal(s): This study introduces a novel data augmentation method to enhance DL segmentation generalizability using conditional denoising diffusion models (cDDMs) to generate counterfactual MR images with altered IAPs, while preserving anatomy. Approach: We trained a cDDM to produce IAP-diverse counterfactual breast MR images. Segmentation models and IAP prediction models were trained to validate the cDDM. Results: Counterfactual data augmentation improved segmentation accuracy, particularly in out-of-distribution settings, demonstrating the potential of cDDMs to mitigate domain shifts in medical imaging. Impact: This method can improve the performance of DL models in clinical settings by enabling them to generalize across different acquisition settings. This could lead to more reliable and robust diagnoses, particularly in scenarios with limited access to diverse training data.