eess.IVMar 9, 2026

Rectified flow-based prediction of post-treatment brain MRI from pre-radiotherapy priors for patients with glioma

Authors: Selena Huisman, Nordin Belkacemi, Vera C. Keil, Joost Verhoeff, Szabolcs David

Organizations: Amsterdam UMC, Department of Radiation Oncology, De Boelelaan 1117, 1081 HV Amsterdam, The Netherlands · Cancer Center Amsterdam, Imaging and Biomarkers, De Boelelaan 1117, 1081 HV Amsterdam, The Netherlands · Department of Radiology and Nuclear Medicine, De Boelelaan 1117, 1081 HV Amsterdam, The Netherlands · Amsterdam UMC, Vrije Universiteit Amsterdam, Amsterdam Neuroscience, Brain Imaging De Boelelaan 1117, 1081 HV Amsterdam, the Netherlands

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Brain tumors result in 20 years of lost life on average. Standard therapies induce complex structural changes in the brain that are monitored through MRI. Recent developments in artificial intelligence (AI) enable conditional multimodal image generation from clinical data. In this study, we investigate AI-driven generation of follow-up MRI in patients with intracranial tumors through conditional image generation. This approach enables realistic modeling of post-radiotherapy changes, allowing for treatment optimization. The public SAILOR dataset of 25 patients was used to create a 2D rectified flow model conditioned on axial slices of pre-treatment MRI and RT dose maps. Cross-attention conditioning was used to incorporate temporal and chemotherapy data. The resulting images were validated with structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), Dice scores and Jacobian determinants. The resulting model generates realistic follow-up MRI for any time point, while integrating treatment information. Comparing real versus predicted images, SSIM is 0.88, and PSNR is 22.82. Tissue segmentations from real versus predicted MRI result in a mean Dice-Sørensen coefficient (DSC) of 0.91. The rectified flow (RF) model enables up to 250x faster inference than Denoising Diffusion Probabilistic Models (DDPM). The proposed model generates realistic follow-up MRI in real-time, preserving both semantic and visual fidelity as confirmed by image quality metrics and tissue segmentations. Conditional generation allows counterfactual simulations by varying treatment parameters, producing predicted morphological changes. This capability has potential to support adaptive treatment dose planning and personalized outcome prediction for patients with intracranial tumors. Code will be available upon peer-reviewed publication at: https://github.com/SelenaIHuisman/RF-GlioPREDICT

Figures & tables

Figure 1: The architecture and training process of the rectified flow diffusion U-Net. The model is trained with pre-treatment information as conditioning, and post treatment information for the L1 loss function.
Figure 2: The architecture and inference process of the rectified flow diffusion U-Net. Follow-up MRIs are generated for any timepoint and treatment combination based on the input information.
Figure 3: An example slice from the middle of the brain, with the radiotherapy dose, real baseline MRI, real follow-up MRI and a predicted follow-up MRI. Zoomed-in sections of the ventricles are shown below the images to highlight local morphological changes predicted by the model. For the RT dose, warmer colors mean a higher dose. For each image, right is on the right.
Figure 4: The Jacobian determinants between baseline and real follow-up and baseline and predicted follow-up, with the baseline image underlaid. Red means that a pixel expands during non-linear registration, while blue means that the pixel contracts. Values between -0.1 and 0.1 are excluded. For each image, right is on the right.
Figure 5: Counterfactual simulations showing how different treatment combinations affect the brain. The baseline, real followup and the true treatment (1x RT dose and temozolomide (TMZ) are shown in the top row as reference. The bottom 3 x 3 MRI slices show the prediction outcomes at 224 days, each for a different treatment combination. The predicted images are all subtracted from the predicted image based on the true treatment (the middle one), resulting in an overlay that shows pixel-intensity differences. Red means positive values, while blue means negative values. The baseline, real follow-up and RT dose are exactly the same as figure 3 . For each image, right means right. Values between -10 and 10 are excluded for clarity.
Figure 6: Eight counterfactual generations for every 60 days, varying only in the temporal conditioning. To highlight tumor and tissue morphological evolution, each image is subtracted from the previous one, therefore a difference map for each interval is created. Blue indicates an intensity decrease from one image to the next, while red indicates an intensity increase. The same baseline MRI slice from Figure 3 is included, and right is on the right. Values between -10 and 10 are excluded for clarity.

Explore similar work

CardsList
Mar 8, 2026

Brain-WM: Brain Glioblastoma World Model

Precise prognostic modeling of glioblastoma (GBM) under varying treatment interventions is essential for optimizing clinical outcomes. While generative AI has shown promise in simulating GBM evolution, existing methods typically treat interventions as static conditional inputs rather than dynamic decision variables. Consequently, they fail to capture the complex, reciprocal interplay between tumor evolution and treatment response. To bridge this gap, we present Brain-WM, a pioneering brain GBM world model that unifies next-step treatment prediction and future MRI generation, thereby capturing the co-evolutionary dynamics between tumor and treatment. Specifically, Brain-WM encodes spatiotemporal dynamics into a shared latent space for joint autoregressive treatment prediction and flow-based future MRI generation. Then, instead of a conventional monolithic framework, Brain-WM adopts a novel Y-shaped Mixture-of-Transformers (MoT) architecture. This design structurally disentangles heterogeneous objectives, successfully leveraging cross-task synergies while preventing feature collapse. Finally, a synergistic multi-timepoint mask alignment objective explicitly anchors latent representations to anatomically grounded tumor structures and progression-aware semantics. Extensive validation on internal and external multi-institutional cohorts demonstrates the superiority of Brain-WM, achieving 91.5% accuracy in treatment planning and SSIMs of 0.8524, 0.8581, and 0.8404 for FLAIR, T1CE, and T2W sequences, respectively. Ultimately, Brain-WM offers a robust clinical sandbox for optimizing patient healthcare. The source code is made available at https://github.com/thibault-wch/Brain-GBM-world-model.

Chenhui Wang, Boyun Zheng, Liuxin Bao +4

Institute of Science and Technology for Brain-inspired Intelligence, Fudan University, Shanghai 200433, China · Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China · School of Automation, Hangzhou Dianzi University, Hangzhou 310018, China +1

Clinical PredictionOncologyWorld Model LearningMedical Image Synthesis
AskPDF
Oct 29, 2025

MedForj: An open, large-scale foundational generative prior for high-resolution 3D brain MRI

This work introduces MedForj, a suite of 3D foundational generative priors based on diffusion models. The MedForj models were trained on 72,65972{,}659 1~mm isotropic 3D T1T_1-weighted MRI human brain image volumes from 38,17438{,}174 subjects, drawn from a curated corpus of 80,67580{,}675 volumes from 42,50642{,}506 subjects spanning 3838 publicly available datasets. These training images were manually inspected to exclude those with poor quality and excessive pathology, and otherwise were minimally processed. The models include six different diffusion training strategies: rectified flow, latent diffusion rectified flow, flow matching, velocity prediction, clean prediction, and noise prediction. Image samples produced by each of these models were compared to each other and against real, ground truth data under downstream segmentation distributions, FID, five inverse problems, and blind human inspection in an observer study. Flow matching was the strongest strategy overall, achieving the best inverse problem solving results at 28.8028.80~dB PSNR and 0.8740.874 SSIM averaged over the five forward problems, the highest rate of reconstructions judged real by blind human raters at 72.6%72.6\%, and the closest per-structure match to real segmented anatomy in a permutation test. It was not best everywhere: rectified flow produced the most convincing unconditional samples in the observer study and the best FID, and the latent rectified-flow model achieved the smallest joint distributional distance to real anatomy. No other strategy, however, performed consistently well across all four evaluations. We therefore recommend flow matching as the default MedForj prior, while releasing every strategy so that the choice can be revisited per application. All model weights and corresponding code are publicly available at https://github.com/piksl-research/medforj.

Samuel W. Remedios, Aaron Carass, Jerry L. Prince +1

Image Analysis and Communications Laboratory, Johns Hopkins University · Department of Neurology, Johns Hopkins Hospital

Flow MatchingHealthcare3D Medical ImagingMedical Image SynthesisDiffusion-Based Inverse Problems3D Diffusion Models+1
AskPDF
Oct 6, 2026

Zero-Shot Brain MRI Inpainting with 2.5D Unconditional Flow Priors

Generative inpainting of brain MRI volumes is essential for synthesizing healthy tissue in pathological regions, improving the accuracy and reliability of automated downstream brain analysis applications such as image registration, brain extraction, and segmentation. However, standard 3D approaches are computationally prohibitive, while efficient 2D slice-wise methods suffer from severe inter-slice discontinuities. Furthermore, traditional models rely on conditional training, requiring task-specific learning of masked inputs. We propose a zero-shot brain MRI inpainting framework utilizing 2.5D unconditional flow priors to capture spatial context along the superior-inferior axis without the overhead of full 3D convolutions. During training, our flow matching model learns the joint distribution of adjacent axial slice triplets, modeling the manifold of healthy brain anatomy while explicitly excluding pathological regions from the loss function. At inference, the model processes the input triplets autoregressively along the depth axis. We employ the Restora-Flow solver to constrain the unconditional prior using the input mask, achieving accurate zero-shot inpainting. Evaluations show our 2.5D strategy resolves the structural discontinuities of 2D baselines, synthesizing plausible healthy tissue while maintaining volumetric consistency across the axial, sagittal, and coronal planes. As a final step, we generate and average an ensemble of multiple stochastic reconstructions to form the final prediction. Quantitative results benchmarked on the official BraTS 2026 Inpainting Challenge validation set demonstrate the effectiveness of our proposed approach, yielding an SSIM of 0.816 ±\pm 0.112, MSE of 0.007 ±\pm 0.005, and PSNR of 22.923 ±\pm 4.343. Code is available at https://github.com/imigraz/brats2026-inpainting.

Arnela Hadzic, Franz Thaler, Simon Johannes Joham +1

Institute for Medical Informatics, Statistics and Documentation, Medical University of Graz, Graz, Austria

Image InpaintingMedical Image SynthesisBrain MRI
AskPDF