Beyond Model Stochasticity: External Randomness for Diffusion-Based Unsupervised MRI Anomaly Detection

Authors: Oh, Changsuk; He, Qiudi; Zhao, Kai; Pang, Kaifeng; Bang, Chulmin; Jank, Erika; Raj, Vinith; Savjani, Ricky; Sung, Kyunghyun

Source journal: Journal of imaging informatics in medicine (2026-11-17)

DOI: 10.1007/s10278-026-02500-2

Abstract

Unsupervised anomaly detection (UAD) in MRI aims to identify abnormal regions through a model trained on normal MR images. Recently, reconstruction-based methods leveraging denoising diffusion probabilistic models (DDPMs) as base generative models (GMs) have gained attention. These approaches typically compute an anomaly score by measuring the difference between a query image and a single reconstructed counterpart (or low-diversity counterparts). However, these approaches typically struggle to differentiate erroneous reconstructions of normal regions from actual abnormalities. To address this problem, we propose a novel approach that enhances anomaly detection by leveraging statistical discrepancy from high-diversity reconstructions. We employ image inpainting with randomly sampled masks to generate diversified reconstructions. We can produce diverse reconstructions through image inpainting even when the input masks differ only slightly, depending on the information retained in the unmasked regions. By sampling multiple random input masks, we can obtain a diverse set of reconstructions, enabling more reliable distinction between normal and abnormal regions. In abnormal lesions, large discrepancies consistently appear across reconstructions, whereas normal regions exhibit irregular and less pronounced differences. Since existing anomaly definitions often fail to adequately capture statistical discrepancy in inpainted images, we introduce a new anomaly score. We define a pixel-wise anomaly score as the average of the normalized differences between the query image and the inpainted images generated from randomly sampled masks covering the pixel. Experimental results demonstrate that our method, with its reconstruction diversification strategy and novel anomaly score, substantially outperforms existing reconstruction-based approaches.

Read the full text

Related topics

Related open access articles

Abstract and bibliographic data from PubMed (U.S. National Library of Medicine). Copyright of the abstract belongs to the original publisher or authors.