Abstract
Brain tumors are among the most clinically significant neurological diseases and remain a major cause of morbidity and mortality due to their aggressive growth and structural heterogeneity. As tumors expand, they induce substantial anatomical deformation that disrupts both local tissue organization and global brain architecture, complicating diagnosis, treatment planning, and surgical navigation. Yet a subject-specific reference of how the brain would appear without tumor-induced changes is fundamentally unobtainable in clinical practice. We present BrainNormalizer, an anatomy-informed diffusion framework that reconstructs pseudo-healthy MRIs directly from tumorous scans by conditioning the generative process on boundary cues extracted from the subject's own anatomy. This boundary-guided conditioning enables anatomically plausible pseudo-healthy reconstruction without requiring paired non-tumorous and tumorous scans. BrainNormalizer employs a two-stage training strategy. The pretrained diffusion model is first adapted through inpainting-based fine-tuning on tumorous and non-tumorous scans. Next, an edge-map-guided ControlNet branch is trained to inject fine-grained anatomical contours into the frozen decoder while preserving learned priors. During inference, a deliberate misalignment strategy pairs tumorous inputs with non-tumorous prompts and mirrored contralateral edge maps, leveraging hemispheric correspondence to guide reconstruction. On the BraTS2020 dataset, BrainNormalizer achieves strong quantitative performance and qualitatively produces anatomically plausible reconstructions in tumor-affected regions while retaining overall structural coherence. BrainNormalizer provides clinically reliable anatomical references for treatment planning and supports new research directions in counterfactual modeling and tumor-induced deformation analysis.
Introduction
When a brain tumor grows, it displaces tissue, creates edema, and deforms the brain’s natural architecture. For neurosurgeons and radiologists, this makes it difficult to determine exactly where the healthy brain ends and the tumor begins. The ideal solution would be to compare the patient's current scan to a "pre-tumor" scan of their own brain, but such references are rarely available. While researchers have attempted "pseudo-healthy reconstruction" to erase tumors, previous methods often produced blurry results or anatomically incorrect structures.
Methodology
BrainNormalizer treats pseudo-healthy reconstruction as an anatomy-informed image inpainting task using a two-stage training process.
Stage 1: Domain Adaptation
The framework begins with a pretrained Stable Diffusion (v1.5) model. Since this model was originally trained on natural images, it is fine-tuned on brain MRIs using an inpainting objective. This process teaches the model the general appearance and texture of brain tissue.
Stage 2: Structural Guidance with ControlNet
To ensure the model respects the patient's specific anatomical layout, a ControlNet branch is added. This branch is trained on Canny edge maps—anatomical contours of the brain’s internal boundaries, such as cortical folds and ventricles. By freezing the main diffusion model and training only the ControlNet branch, the system learns to inject fine-grained anatomical constraints without losing the learned tissue priors.
Inference: The Hemispheric Mirroring Strategy
During inference, BrainNormalizer employs a "deliberate misalignment" strategy. It takes the tumorous scan, masks the tumor, and provides the model with a text prompt describing a healthy individual. Crucially, it uses an edge map mirrored from the patient's own healthy hemisphere. By leveraging the brain's natural symmetry, the model uses the healthy side as a blueprint to fill in the tumor-affected region.
Experimental Results
The study utilized the BraTS2020 dataset, focusing on T1-weighted contrast-enhanced (T1CE) scans. BrainNormalizer significantly outperformed existing state-of-the-art methods:
- FID (Fréchet Inception Distance): Achieved 27.1 (vs. 43.1 for the baseline), indicating generated images are much closer to real healthy brain distributions.
- SSIM (Structural Similarity Index): Achieved 0.76 (vs. 0.37 for the baseline), proving high structural coherence with the healthy hemisphere.
- Clinical Plausibility: Only 7.1% of reconstructed images were flagged by a separate tumor-detection AI as still containing a tumor, compared to 15.7% for previous methods.
Discussion and Limitations
BrainNormalizer demonstrates that symmetry is a powerful prior for subject-specific guidance. However, the mirroring strategy relies on having at least one healthy hemisphere; it may struggle with large bilateral tumors. Additionally, the current 2D slice-based approach may introduce inconsistencies when reconstructed into 3D volumes. Future work will explore native 3D diffusion frameworks and multi-modal integration (T1, T2, FLAIR) to provide a more comprehensive clinical reference.
Summary updated