Diffusion Models for Counterfactual Generation and Anomaly Detection in Brain Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Segmentation masks of pathological areas are crucial for managing conditions like brain tumors and strokes.
• Generating healthy counterfactuals of diseased images aids radiologist training and segmentation model interpretability.

Purpose of the Study:

• To develop a weakly supervised method for generating healthy versions of diseased images.
• To create pixel-wise anomaly maps from these generated healthy images.
• To improve brain lesion segmentation and model interpretability.

Main Methods:

• Utilized a saliency map (obtained with ACAT) to identify pathological areas.
• Employed a diffusion model trained on healthy samples, combining Denoising Diffusion Probabilistic Model (DDPM) and Denoising Diffusion Implicit Model (DDIM).
• DDPM modified lesion areas, while DDIM reconstructed normal anatomy, with seamless fusion at each step.

Main Results:

• The method accurately reconstructed healthy images when applied to healthy samples.
• Achieved the highest mean Dice and IoU scores in brain lesion segmentation compared to alternative weakly supervised methods.
• Demonstrated effective generation of anomaly maps for pathological areas.

Conclusions:

• The proposed method successfully generates healthy counterfactuals of diseased images.
• This approach enhances the performance of brain lesion segmentation and offers improved interpretability.
• The technique shows promise for medical training and diagnostic support systems.