[Paper Review] Patched Diffusion Models for Unsupervised Anomaly Detection in Brain MRI

Finn Behrendt, Debayan Bhattacharya+3 more|arXiv (Cornell University)|Mar 7, 2023

Advanced Neuroimaging Techniques and Applications10 citations

TL;DR

This paper presents patched diffusion models (pDDPMs) for unsupervised anomaly detection in brain MRI and shows improvements over baselines on BraTS21 and MSLUB datasets.

ABSTRACT

The use of supervised deep learning techniques to detect pathologies in brain MRI scans can be challenging due to the diversity of brain anatomy and the need for annotated data sets. An alternative approach is to use unsupervised anomaly detection, which only requires sample-level labels of healthy brains to create a reference representation. This reference representation can then be compared to unhealthy brain anatomy in a pixel-wise manner to identify abnormalities. To accomplish this, generative models are needed to create anatomically consistent MRI scans of healthy brains. While recent diffusion models have shown promise in this task, accurately generating the complex structure of the human brain remains a challenge. In this paper, we propose a method that reformulates the generation task of diffusion models as a patch-based estimation of healthy brain anatomy, using spatial context to guide and improve reconstruction. We evaluate our approach on data of tumors and multiple sclerosis lesions and demonstrate a relative improvement of 25.1% compared to existing baselines.

Motivation & Objective

• Address the challenge of detecting pathologies in brain MRI without pixel-level annotations by using unsupervised anomaly detection.
• Reformulate diffusion-based generation as patch-based estimation of healthy brain anatomy guided by spatial context.
• Incorporate patch-level denoising with surrounding context via a sliding window and patch stitching to reconstruct whole-brain anatomy.
• Evaluate pDDPM on BraTS21 and MSLUB datasets and compare against established UAD baselines to demonstrate performance gains.

Proposed method

• Apply a forward diffusion process to patches of the input image while keeping the rest of the image intact.
• Use a U-Net to learn the reverse diffusion step and estimate the original patch within the context of its surrounding tissue.
• Train with Lp-based loss on the patch region (L_p) or full reconstruction loss (L_rec), and use a fixed test-time step t_test for efficiency.
• During evaluation, perform sliding-window patch diffusion-denoising across the image, reconstruct all patches, and stitch them to form the full image; average overlapping regions.

Experimental results

Research questions

• RQ1 Can patch-based diffusion models leveraging surrounding spatial context improve unsupervised anomaly detection in brain MRI compared to non-patch DDPMs and other baselines?
• RQ2 How do patch size and test-time diffusion parameters (e.g., t_test) affect UAD performance on BraTS21 and MSLUB datasets?

Key findings

• pDDPM with fixed sampling and Lp loss achieves the best Dice and AUPRC on BraTS21 and MSLUB among all evaluated methods (BraTS21 Dice 49.00%, AUPRC 54.07%; MSLUB Dice 10.35? Note: see table for exact values).
• The pDDPM method outperforms all baselines on both data sets for DICE and AUPRC with statistical significance on BraTS21 (p<0.05).
• A patch size of 60x60 pixels yielded the best performance, and a test-time noise level corresponding to t_test around 400 provided a performance peak.
• Reconstruction quality on healthy IXI data showed that DAE achieved the lowest l1 error, while pDDPM achieved smaller errors than DDPM, suggesting better healthy-brain reconstructions while preserving anomaly detectability.