Guided reconstruction with conditioned diffusion models for unsupervised anomaly detection in brain MRIs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised models for clinical screening require extensive annotated data.
• Unsupervised Anomaly Detection (UAD) identifies outliers from healthy data distributions.
• Generative models in UAD reconstruct healthy anatomy, with errors indicating anomalies.

Purpose of the Study:

• To address challenges in balancing healthy anatomy reconstruction and avoiding abnormal structure replication in UAD.
• To improve the accuracy and robustness of diffusion models for UAD in brain MRI.
• To enhance segmentation performance and domain adaptation for clinical screening tasks.

Main Methods:

• Proposed conditioning the denoising process of diffusion models with latent image representations.
• Utilized generative models for reconstructing healthy brain anatomy.
• Compared the novel approach against state-of-the-art methods on multiple datasets (BraTS, ATLAS, MSLUB, WMH).

Main Results:

• Achieved substantial improvements in segmentation performance, with Dice scores increasing by 11.9% (BraTS), 20.0% (ATLAS), and 44.6% (MSLUB).
• Demonstrated effective domain adaptation across different MRI acquisitions and simulated contrasts.
• Maintained competitive performance on the WMH dataset.

Conclusions:

• Conditioned diffusion models offer accurate and localized adaptation for UAD, preventing replication of unhealthy structures.
• The proposed method enhances segmentation accuracy and generalizability for anomaly detection in brain MRI.
• This approach shows promise for improving clinical screening by overcoming limitations of existing UAD techniques.