BrainMRDiff: A Diffusion Model for Anatomically Consistent Brain MRI Synthesis

TL;DR

BrainMRDiff addresses the challenge of synthesizing anatomically consistent brain MRI sequences when some acquisitions are missing or degraded. It introduces two modules, TSA for joint structure-tumor conditioning and TGAP for topology-guided preservation of tumor regions, integrated into a diffusion framework. The method leverages multiple anatomical masks, persistent-homology-based topology losses, and end-to-end training with a TSA-conditioned objective to achieve high fidelity in both brain structures and tumor morphology. Experimental results on BraTS-AG and BraTS-Met show improved image quality, segmentation performance, and clinically relevant task performance (MGMT prediction and survival analysis), suggesting strong potential for real-world neuro-oncology workflows.

Abstract

Accurate brain tumor diagnosis relies on the assessment of multiple Magnetic Resonance Imaging (MRI) sequences. However, in clinical practice, the acquisition of certain sequences may be affected by factors like motion artifacts or contrast agent contraindications, leading to suboptimal outcome, such as poor image quality. This can then affect image interpretation by radiologists. Synthesizing high quality MRI sequences has thus become a critical research focus. Though recent advancements in controllable generative AI have facilitated the synthesis of diagnostic quality MRI, ensuring anatomical accuracy remains a significant challenge. Preserving critical structural relationships between different anatomical regions is essential, as even minor structural or topological inconsistencies can compromise diagnostic validity. In this work, we propose BrainMRDiff, a novel topology-preserving, anatomy-guided diffusion model for synthesizing brain MRI, leveraging brain and tumor anatomies as conditioning inputs. To achieve this, we introduce two key modules: Tumor+Structure Aggregation (TSA) and Topology-Guided Anatomy Preservation (TGAP). TSA integrates diverse anatomical structures with tumor information, forming a comprehensive conditioning mechanism for the diffusion process. TGAP enforces topological consistency during reverse denoising diffusion process; both these modules ensure that the generated image respects anatomical integrity. Experimental results demonstrate that BrainMRDiff surpasses existing baselines, achieving performance improvements of 23.33% on the BraTS-AG dataset and 33.33% on the BraTS-Met dataset. Code will be made publicly available soon.

Figures (11)

• Figure 1: Overview of our proposed work. Baseline methods exhibit limitations in generating MR images with faithful anatomical representations. In contrast, our proposed BrainMRDiff framework integrates anatomical constraints—specifically WMT, CGM, LV, and tumor masks as control inputs—to produce MR images that accurately reflect anatomical structures
• Figure 2: BrainMRDiff architecture. Our proposed method consists of two components: a) Tumor+Structure Aggregation (TSA) module which aggregates the different anatomical structures and tumor segmentation masks as a unified control to the diffusion model, b) Topology-Guided Anatomy Preservation (TGAP) module which enforces topological constraints to ensure high fidelity tumor region generation.
• Figure 3: Tumor and Anatomy Structures. The tumor mask and the different anatomical structures namely whole Brain Mask (BM), White Matter Tracts (WMT), Cortical Gray Matter (CGM), Lateral Ventricles (LV) are shown overlaid on top of a FLAIR scan.
• Figure 4: Tumor+Structure Aggregation (TSA) module. The brain structure masks—WMT, CGM, LV, and BM—are fused with the tumor mask to create a unified representation, which serves as a conditional control for the diffusion model.
• Figure 5: Tumor-Guided Anatomy Preserving (TGAP) module. Predicted noise from the diffusion model is first deblurred, followed by masking of the tumor region. The PD is then computed from the mask, followed by loss calculation