Guided Reconstruction with Conditioned Diffusion Models for Unsupervised Anomaly Detection in Brain MRIs

TL;DR

This work tackles unsupervised anomaly detection in brain MRI by addressing diffusion-based reconstruction challenges, notably preserving local intensity characteristics and handling domain shifts. It introduces conditioned DDPMs (cDDPMs) that fuse an input-derived latent context with the diffusion denoising process, guiding reconstructions toward aligned intensity distributions while avoiding copying unhealthy structures. Empirically, cDDPMs deliver significant improvements in Dice scores on BraTS, ATLAS, and MSLUB, demonstrate effective domain adaptation under simulated contrasts, and achieve competitive segmentation performance across multiple datasets. The approach offers a practical, annotation-efficient pathway for robust brain MRI anomaly detection, with open-source code for reproduction.

Abstract

The application of supervised models to clinical screening tasks is challenging due to the need for annotated data for each considered pathology. Unsupervised Anomaly Detection (UAD) is an alternative approach that aims to identify any anomaly as an outlier from a healthy training distribution. A prevalent strategy for UAD in brain MRI involves using generative models to learn the reconstruction of healthy brain anatomy for a given input image. As these models should fail to reconstruct unhealthy structures, the reconstruction errors indicate anomalies. However, a significant challenge is to balance the accurate reconstruction of healthy anatomy and the undesired replication of abnormal structures. While diffusion models have shown promising results with detailed and accurate reconstructions, they face challenges in preserving intensity characteristics, resulting in false positives. We propose conditioning the denoising process of diffusion models with additional information derived from a latent representation of the input image. We demonstrate that this conditioning allows for accurate and local adaptation to the general input intensity distribution while avoiding the replication of unhealthy structures. We compare the novel approach to different state-of-the-art methods and for different data sets. Our results show substantial improvements in the segmentation performance, with the Dice score improved by 11.9%, 20.0%, and 44.6%, for the BraTS, ATLAS and MSLUB data sets, respectively, while maintaining competitive performance on the WMH data set. Furthermore, our results indicate effective domain adaptation across different MRI acquisitions and simulated contrasts, an important attribute for general anomaly detection methods. The code for our work is available at https://github.com/FinnBehrendt/Conditioned-Diffusion-Models-UAD

Figures (8)

• Figure 1: Overview of our proposed approach. The encoder representations are learned along with the DDPM in the main training stage to condition the denoising process. The timestep embedding $c_t$ is concatenated with the input image's projected encoder representation $c$. The resulting conditioning vector $c'$ is used to scale and shift feature maps of the denoising Unet in the residual blocks. Each residual block consists of two convolution operators (conv1, conv2), group normalization (Norm) and Sigmoid Linear Units (SiLU). During evaluation, the residual map between unhealthy brain images and their healthy reconstructions is used for anomaly detection.
• Figure 2: Overview of the pre-training strategy. Random patches are erased from the input image. $F_{enc}$ is used to derive a latent representation and $F_{dec}$ is used to reconstruct the removed patches. After pre-training, the encoder $F_{enc}$ is then fine-tuned along with the DDPM in the main training stage to condition the denoising process.
• Figure 3: Simulating the conditioning effect of the cDDPM for 50 $\%$ of noise and 100$\%$ noise in the input image. In the first block, the input image that is fed to the DDPM or cDDPM is shown. In the second block, the reconstructions of cDDPM for different conditioning inputs are shown when a noise level of 50$\%$ is applied. In the third block, the reconstructions of cDDPMs and DDPMs are compared at a noise level of 100$\%$. From top to bottom, the contrast level of the conditioning and input image is increased, respectively, for all columns.
• Figure 4: Comparison of the histograms for input-reconstruction pairs of the healthy IXI (left) and the unhealthy BraTS21 (right) data set with original and augmented contrast. The top row shows the baseline DDPM without conditioning and the bottom row our proposed cDDPM with conditioning. The Kullback-Leibler divergence (KLD) for both histograms is indicated within each plot (lower values are preferable). Both models are evaluated by ensembling different values for $t_{test}=[250,500,750]$.
• Figure 5: Exemplary reconstructions and anomaly maps for DDPMs (second row) and cDDPMs (third row). The input and the corresponding ground truth annotation are provided in the first row. For each case, the reconstruction, the anomaly map and the histograms of intensity values in input and reconstruction are shown. Note that for histogram calculation, only healthy areas are considered. For visualization purposes, we provide segmentation maps next to the anomaly maps. We derive the binarization threshold by optimizing for the best possible dice score.