Answers are grounded in this paper and cite their sources.
Detecting diverse lesions in brain MRI is clinically important. Unsupervised anomaly detection methods have attracted widespread attention because they learn healthy patterns to detect anomalies without lesion annotations. Among them, reconstruction-based methods generate a pseudo-healthy image (PHI) and localize lesions using reconstruction errors. However, existing methods struggle to suppress anomaly reconstruction while recovering normal details. To address this challenge, we propose a pseudo-healthy image reconstruction framework based on dual-encoder feature fusion. A structural reference encoder takes binary edge maps as input to extract structural features and limit the direct propagation of abnormal intensity information. An appearance detail encoder extracts appearance features from the original image to supplement normal tissue details. Discrepancy-guided adaptive fusion uses structural reference features as an anchor to regulate appearance contributions according to local feature discrepancies. At local positions with larger discrepancies, appearance weights are reduced to limit the introduction of abnormal information; at positions with smaller discrepancies, more appearance details are incorporated to improve normal tissue reconstruction. A shared decoder reconstructs a PHI from the fused features. Across eight brain MRI datasets (3055 volumes), our method outperforms 16 baselines, improving average precision and Dice similarity coefficient by 4.15 and 3.19 percentage points over the strongest baseline.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: