Unsupervised Anomaly Detection in Brain MRI via Disentangled Anatomy Learning

TL;DR

This work tackles unsupervised brain MRI anomaly detection across multi-modality and multi-center data by introducing two novelty modules: a Disentangled Representation Module that separates anatomy from imaging information and a Edge-to-Image Restoration Module that reconstructs high-quality pseudo-healthy images using edge-derived anatomical codes. The approach enforces anatomy-focused reconstruction via brain priors and a differentiable one-hot binarization, enabling robust generalization, while edge-guided restoration suppresses abnormal residuals and preserves personalized structural details. A hybrid anomaly score combining pixel differences with SSIM, plus post-processing, yields strong voxel-level detection across nine public datasets and 4,443 abnormal MRIs, outperforming 17 SOTA methods. The framework demonstrates cross-sequence generalization (e.g., T1 and T2) and cross-center robustness, highlighting practical potential for scalable, multi-center clinical deployment of unsupervised brain anomaly detection.

Abstract

Detection of various lesions in brain MRI is clinically critical, but challenging due to the diversity of lesions and variability in imaging conditions. Current unsupervised learning methods detect anomalies mainly through reconstructing abnormal images into pseudo-healthy images (PHIs) by normal samples learning and then analyzing differences between images. However, these unsupervised models face two significant limitations: restricted generalizability to multi-modality and multi-center MRIs due to their reliance on the specific imaging information in normal training data, and constrained performance due to abnormal residuals propagated from input images to reconstructed PHIs. To address these limitations, two novel modules are proposed, forming a new PHI reconstruction framework. Firstly, the disentangled representation module is proposed to improve generalizability by decoupling brain MRI into imaging information and essential imaging-invariant anatomical images, ensuring that the reconstruction focuses on the anatomy. Specifically, brain anatomical priors and a differentiable one-hot encoding operator are introduced to constrain the disentanglement results and enhance the disentanglement stability. Secondly, the edge-to-image restoration module is designed to reconstruct high-quality PHIs by restoring the anatomical representation from the high-frequency edge information of anatomical images, and then recoupling the disentangled imaging information. This module not only suppresses abnormal residuals in PHI by reducing abnormal pixels input through edge-only input, but also effectively reconstructs normal regions using the preserved structural details in the edges. Evaluated on nine public datasets (4,443 patients' MRIs from multiple centers), our method outperforms 17 SOTA methods, achieving absolute improvements of +18.32% in AP and +13.64% in DSC.

Figures (12)

• Figure 1: The architecture diagram of the training and detection process of the proposed two modules: (a) disentangled representation module and (b) edge-to-image restoration module. These two modules are jointly trained on paired normal 3D brain MRIs and reconstruct abnormal images into pseudo-healthy images. The proposed hybrid anomaly score locates abnormal regions by measuring the pixel-level reconstruction differences.
• Figure 2: Comparison of anomaly maps generated by our method (three orthogonal views in \ref{['Anomaly map visualization']}) and 17 SOTA methods. All anomaly maps are scaled to [0, 1] using Max-Min normalization for clearer display. Our method shows superior detection of various brain anomalies in T1 and T2 scans.
• Figure 3: Comparison of pseudo-healthy images reconstructed by our method and other reconstruction-based methods. Our method repairs anomalies more effectively while preserving normal brain details with high fidelity.
• Figure 4: The pseudo-healthy image reconstruction pipeline and a visual analysis of the restored anatomical structure.
• Figure 5: Visualization of intermediate outputs of our method in different ablation setting. (a) Image restoration based on anatomical edges and modality representation (ours). (b) Image reconstruction without edge-to-image restoration. (c) Image restoration without disentangling representation.