S2294
Physics - Machine learning and AI algorithms
ESTRO 2026
Tissue Complication Probability model to predict late xerostomia in head and neck cancer patients,” Int. J. Radiat. Oncol. Biol. Phys., vol. 0, no. 0, Aug. 2024. Keywords: NTCP, deep learning, uncertainty quantification
Digital Poster 2259
Rectified flow-based post-treatment brain MRI generation for patients with glioma from pre-RT priors Selena Huisman 1 , Vera Keil 2 , Joost Verhoeff 1 , Szabolcs David 1 1 Radiotherapy, Amsterdam UMC, Amsterdam, Netherlands. 2 Radiology and Nuclear Medicine, Amsterdam UMC, Amsterdam, Netherlands Purpose/Objective: Intracranial tumors such as gliomas represent a major clinical challenge due to their heterogeneity, limited therapeutic options, and profound impact on quality- of-life (QoL), neurological function, and survival. Standard therapies including surgery, radiotherapy, and systemic treatments induce complex, non-linear structural changes in the brain [1-3] that are monitored through standard MRI such as T1, T1 with contrast enhancement (T1Gd) and T2-FLAIR. Interpreting these post-treatment changes is difficult, as they can reflect a mixture of tumor (pseudo- )progression, treatment effect, or healthy-appearing tissue response. Recent advances in artificial intelligence (AI), allow image generation from multimodal clinical data. In this study, we investigate AI-driven generation of follow-up MRI in patients with intracranial tumors through conditional image generation, incorporating pre-radiotherapy MRI, time, radiotherapy dose distributions, and clinical information. The approach enables realistic and temporally resolved modeling of post-radiotherapy changes. Material/Methods: A public longitudinal dataset of 26 patients [2] was used to create a 2D rectified flow (RF) [3] model spatially conditioned on axial slices of pre-treatment MRI (T1, T1Gd, T2-FLAIR) and dose maps. Cross- attention conditioning was used to incorporate temporal (days after treatment) and chemotherapy data, as shown in figure 1. The resulting images were validated with image quality metrics such as Structural-similarity-index-measure (SSIM), Peak- signal-to-noise ratio (PSNR), Mean-squared-error (MSE) and non-linear image registration. Additionally, white-, grey-matter and CSF regions were segmented and Dice-scores were calculated between the real- and predicted images.
Results: The resulting model generates realistic follow-up MRI for any timepoint, while integrating treatment information. Comparing predicted- vs real MRI, SSIM is 0.88 and PSNR is 22.83. Segmenting tissue and CSF from predicted- vs real MRI results in a mean Dice of 0.87. The RF model allows up to 250x faster inference compared to DDPM, generating slices in 0.3 seconds. Fig. 2 shows a predicted T1 MRI at day 224, in addition to the real baseline MRI, dosimetry map and follow-up MRI at day 224. Jacobian determinants indicate a smaller difference between the predicted MRI and real follow-up MRI than between baseline MRI and real follow-up MRI.
Conclusion: 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 (fractionation, dose, distribution), producing predicted morphological changes. This capability has potential to support adaptive treatment dose planning and improve personalized outcome prediction and follow-up frequency for patients with intracranial tumors. References: [1] Nagtegaal, S. et al. Clinical and Translational Radiation Oncology 31, 14–20 (Nov. 2021) doi:
Made with FlippingBook - Share PDF online