Papers

Accepted Papers

Following double-blind peer review, 27 papers were accepted to MIART 2026. All accepted papers were presented as posters at the workshop, and 9 of them were additionally selected for oral presentation across three thematic sessions. Slides for each oral presentation are linked from its entry below; see the programme for timings and the keynote slides.

The MIART 2026 Best Paper Award went to Blanca Rodriguez-Gonzalez et al. for Can We Trust Synthetic CT Algorithms? Uncertainty-Aware Evaluation of CBCT to CT Synthesis for Adaptive Proton Therapy.

All accepted papers will be published in the MICCAI 2026 Springer proceedings. Papers are grouped below by theme and listed alphabetically by title within each theme.

Image Synthesis & Cross-Modality Translation (7 papers)

Oral & Poster 🏆 Best Paper Award

Can We Trust Synthetic CT Algorithms? Uncertainty-Aware Evaluation of CBCT to CT Synthesis for Adaptive Proton Therapy

Blanca Rodriguez-Gonzalez, Ivan Escobar Corominas, Angel Torrado-Carvajal, Borja Rodriguez-Vila, Alejandro Mazal, Norberto Malpica

Presented in Oral Session 2: Cross-Modality Image Synthesis. Winner of the MIART 2026 Best Paper Award.

OpenReview → Slides (PDF) →
Abstract

Synthetic CT (sCT) generation from cone beam CT (CBCT) has emerged as a key component of adaptive proton therapy (APT), enabling daily dose recalculation. However, image quality metrics alone provide limited information regarding prediction reliability. This work studies whether uncertainty estimates can identify regions where synthesis errors may have dosimetric consequences. A 2.5D Attention U-Net was trained using three uncertainty quantification strategies (Monte Carlo Dropout, Deep Ensembles and Evidential Deep Learning (EDL)). Beyond conventional image-quality evaluation, uncertainty was assessed through its spatial association with Hounsfield Unit (HU) errors and dose discrepancies. EDL achieved the best image-domain performance and the strongest uncertainty-error agreement. While global dosimetric differences between methods were modest, uncertainty maps localized regions of increased HU and dose error. These findings suggest that uncertainty provides complementary reliability information and should be considered as part of quality assurance for sCT-based APT.

Oral & Poster

Cross-Modal Spatial Gating of Planning-MRI Priors for Longitudinal CBCT-to-CT Synthesis

Vincent Jaouen, Vincent Bourbonne, Julien Bert, Ulrike Schick, Pierre-Henri Conze, Dimitris Visvikis

Presented in Oral Session 2: Cross-Modality Image Synthesis.

OpenReview → Slides (PDF) →
Abstract

Synthetic computed tomography (sCT) from cone-beam CT (CBCT) is a major objective in online adaptive radiotherapy. Planning magnetic resonance image (MRI) can supply strong soft-tissue priors to improve sCT synthesis in combination with CBCT inputs. Prior multimodal approaches, however, combine daily CBCT and planning MRI through simple channel-wise concatenation, and it remains unclear whether this fusion mechanism itself limits the benefit obtained from the MRI prior. We systematically compare three fusion strategies within a common Residual-in-Residual Dense Block (RRDB) conditional GAN backbone: input concatenation with channel attention, a large separable kernel attention (LSKA) spatial gate, and a proposed cross-modal spatial gating (CMSG) that couples modality-specific encoders with joint channel and spatial recalibration. On a cohort of 190 brain-metastasis patients (152 train, 38 hold-out), with three sequential CBCTs and a planning CT reference, we evaluate every fusion strategy across four input configurations and assess significance with patient-level paired Wilcoxon signed-rank tests in native Hounsfield-Unit (HU) space. The proposed CMSG achieves the best fidelity (PSNR 30.98 dB, MAE 55.87 HU) and reveals a fusion × modality interaction: the planning MRI yields a significant MAE reduction only when fused with modality-specific encoders (median -6.61 HU within CMSG, p<10^-5; a net -3.06 HU against the best CBCT-only baseline, p=2.2 × 10^-3), while under input concatenation with a spatial gate its benefit is not significant (p=0.20). These results demonstrate that the design of the fusion mechanism is key for a better exploitation of the planning MRI prior. Code is available at github.com/vhxjaouen/CrossModalSpatialGating.

Poster

EPC-3D-Diff: Equivariant Physics Consistent Conditional 3D Latent Diffusion for CBCT to CT Synthesis

Alzahra Altalib, Chunhui Li, Haytham Ahmad Alewaidat, Khaled Z. Alawneh, Ahmad Awad Qandeel, Alessandro Perelli

OpenReview →
Abstract

Cone-beam CT (CBCT) is routinely acquired during radiotherapy for patient setup, but its quantitative reliability is degraded by scatter, noise, and reconstruction artifacts, limiting Hounsfield Unit (HU) accuracy. We propose EPC-3D-Diff, a novel conditional 3D latent diffusion framework for volumetric CBCT to CT synthesis that introduces a projection domain equivariance loss derived from acquisition physics. Unlike common image domain equivariance, we exploit the fact that an in plane rotation of the volume corresponds to an angular shift in its projections. During training, we enforce this relationship by forward projecting rotated synthesized CT volumes and matching them to appropriately angle shifted projections of the paired target CT, yielding a physics consistent equivariance constraint integrated into the diffusion objective. To capture full 3D context efficiently, conditional diffusion is performed in a compact latent space learnt by a lightweight 3D autoencoder, preserving axial depth while downsampling in plane resolution for stable training. We validate on a paired head CBCT/CT phantom dataset, including repeat scans, and paired clinical data using patient wise splits, and perform single and mixed domain training, ablations, and comparisons with diffusion and CycleGAN. EPC-3D-Diff generalizes well and achieved substantial improvements, +7.4 dB (phantom) and +1.8 dB (clinical data) in PSNR compared to state of the art methods, alongside improved SSIM and HU accuracy, within tissue boundaries. Overall, EPC-3D-Diff improves robustness and physics consistency, supporting HU aware synthesis for downstream radiotherapy workflows. The open source code for EPC-3D-Diff is available at https://github.com/ALZAHRAALTALIB/EPC-3D-Diff

Poster

Mamba-Driven MRI-to-CT Synthesis for MRI-Only Radiotherapy Planning

Konstantinos Barmpounakis, Theodoros P. Vagenas, Maria Vakalopoulou, George K. Matsopoulos

OpenReview →
Abstract

Radiotherapy workflows for oncological patients increasingly rely on multi-modal medical imaging, commonly involving both Magnetic Resonance Imaging (MRI) and Computed Tomography (CT). MRI-only treatment planning has emerged as an attractive alternative, as it reduces patient exposure to ionizing radiation and avoids errors introduced by inter-modality registration. While nnU-Net-based frameworks are predominantly used for MRI-to-CT synthesis, we explore Mamba-based architectures for this task to investigate the applicability of state-space modeling for cross-modality medical image translation and assess its performance relative to established convolutional architectures. Specifically, we adapt the SegMamba architecture, originally proposed for segmentation, to perform image-to-image generation. Our 3D Mamba architecture effectively captures complex volumetric features and long-range dependencies, thus allowing accurate CT synthesis while maintaining a relatively low parameter count. Experiments were conducted on a subset of SynthRAD2025 dataset, comprising registered single-channel MRI–CT volume pairs across three anatomical regions. Quantitative evaluation is performed via a combination of image similarity metrics computed in Hounsfield Units (HU) and segmentation-based metrics obtained from TotalSegmentator to ensure geometric consistency is preserved. The findings pave the way for the integration of state-space models into radiotherapy workflows.

Poster

MRI-to-Synthetic CT for Intracranial Stereotactic Radiotherapy: A Comparative Study of Deep Learning Architectures with Focus on Bone Fidelity

Danilo Paglialunga, Alfonso Belardo, Maria Giulia Ubeira Gabellini, Simone Raggio, Sara Broggi, Gabriele Palazzo, Bianca Bordigoni, Gianluca Gragnaniello, Antonella del Vecchio, Lorenzo Placidi, Claudio Fiorino

OpenReview →
Abstract

Intracranial stereotactic radiation therapy (SRS/SRT) with robotic systems demands sub-millimetric geometric accuracy. As image guidance relies on Digitally Reconstructed Radiographs (DRRs) derived from the planning CT for skull tracking along the bone-tissue gradient, accurate reconstruction of this interface drives treatment precision. MRI-only workflows would remove the MRI-CT registration error, but they shift onto the synthesis model the full responsibility of reproducing correct spatial information and Hounsfield Unit (HU) values. We present a modular pipeline for MRI-to-synthetic-CT (sCT) generation, built around a 3D Attention U-Net trained on the SynthRAD2023 Brain dataset (180 brain MRI/CT pairs). The pipeline combines additive attention gates on skip connections with a bone-weighted composite loss, a cranial-specific HU normalization window ([-1000, 2000], bone defined above 300 HU), bone-aware patch sampling, Exponential Moving Average (EMA) of the weights, and Hann-weighted sliding window inference. We compared four architectures under one shared protocol: a conditional GAN, a denoising diffusion probabilistic model (DDPM), a standard 3D U-Net, and the proposed 3D Attention U-Net. Attention U-Net provided the best HU accuracy, with a global MAE of ~106 HU and a bone-region MAE of ~195 HU. A per-patient reliability assessment module then combines bone geometry, DRR consistency, and HU accuracy into a single confidence score, providing a structured retrospective framework for assessing case-wise suitability in stereotactic workflows.

Oral & Poster

SynthRCT: Scalable Conditional Deformation Synthesis for Synthetic Repeat CT Generation

Tomas Guija-Valiente, Norberto Malpica, Blanca Rodriguez-Gonzalez

Presented in Oral Session 2: Cross-Modality Image Synthesis.

OpenReview → Slides (PDF) →
Abstract

In proton therapy, plans are typically optimized on a single planning CT, making robustness evaluation essential under anatomical changes. However, current scenarios often rely on simplified perturbations that poorly capture complex, patient-specific variability. We propose SynthRCT, a scalable conditional generative framework for 3D anatomical deformation synthesis. Based on a conditional variational autoencoder, SynthRCT learns a latent deformation space and decodes sampled latent codes into local stationary velocity fields conditioned on an input anatomy. Local fields are assembled into coherent full-volume transformations, enabling memory-scalable generation for large field-of-view CT data. We validate the approach on respiratory 4DCT data with multiple breathing-phase anatomies per subject. SynthRCT enables patient-specific sampling of plausible anatomical transformations beyond predefined robustness scenarios. Code available at: https://github.com/TomasGuija/RCTSynth.

Poster

WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis

Siyuan Mei, Yan Xia, Yipeng Sun, Siming Bayer, Zirong Li, Chengze Ye, Daiqi Liu, Fuxin Fan, Yixing Huang, Andreas Maier

OpenReview →
Abstract

Generating CT volumes from MRI and CBCT can improve treatment planning in adaptive radiotherapy while avoiding additional radiation exposure. However, direct regression of CT intensities is challenged by the inherently high dynamic range and long-tailed distributions, thereby averaging out sparse yet clinically important structures. To alleviate this issue, we reformulate the regression target into multiple windowed representations, leveraging the inductive prior that CT intensities are structure-deterministic and window-separable. These windowed views exhibit smoother distributions and admit structured fusion back to the full-range CT. Building on this reformulation, we introduce WING, a WINdow-prior-based Generative network comprising: 1) a new Gated Inception Generator to produce multi-window predictions, enabling multi-shape kernel interactions to capture cross-modality correspondence; 2) a Fuse-and-Refine Transformer to aggregate the windowed outputs and learn residuals for detail refinement; and 3) a joint adversarial training objective to enhance window-conditioned realism. Extensive experiments demonstrate that our compact WING achieves state-of-the-art performance on the MRI-to-CT and CBCT-to-CT benchmarks, while supporting multi-anatomy synthesis with a single model.

Segmentation & Contour Quality Assurance (7 papers)

Poster

An Uncertainty-Guided Multi-Scale SwinUNETR Framework for 3D Larynx Segmentation in CT

Olga Menegaki, Konstantinos Georgas, George K. Matsopoulos

OpenReview →
Abstract

Accurate segmentation of the larynx is a fundamental preprocessing step in radiation therapy planning for head and neck (H&N) cancer. However, reliable laryngeal delineation from computed tomography (CT) remains challenging due to low soft-tissue contrast, morphological variation, and significant class imbalance. To address these limitations, this study presents an uncertainty-aware segmentation framework combining a SwinUNETR backbone with an S²Net-inspired multi-scale refinement head. By employing a memory-efficient Multi-Scale Feature Fusion (MSFF) module, the model enhances 3D encoder features to produce a coarse prediction. Crucially, a top-down Localization Calibration (LC) mechanism leverages voxel-wise uncertainty to refine this prediction, ensuring precise boundary definition in ambiguous regions without compromising anatomical integrity. The proposed method is validated on a private CT cohort comprising both healthy subjects and patients with laryngeal cancer, achieving Dice scores of 88.5% and 78.00%, respectively. Additionally, the framework is evaluated on the larynx subset of the publicly available SegRap2023 dataset, attaining a Dice score of 92.54%. Across both cohorts the framework performs on par with strong state-of-the-art baselines, with a statistically significant improvement on the cancer class, where boundary definition is most challenging. In clinically realistic radiation planning conditions, these results indicate consistent generalization performance.

Oral & Poster

CBCT Segmentation in Head and Neck ART: A Longitudinal Deep Learning Approach

Lucía Cubero, Elena Navarro Alfaro, Irina Rakotoarisedy, Anaïs Barateau, Joël Castelli, Renaud de Crevoisier, Oscar Acosta, Javier Pascau

Presented in Oral Session 3: Adaptive Radiotherapy (Longitudinal Imaging, Motion & Dose).

OpenReview → Slides (PDF) →
Abstract

Adaptive radiotherapy (ART) for head and neck cancer (HNC) relies on repeated cone-beam CT (CBCT) imaging to monitor anatomical changes throughout treatment course. Accurate organ-at-risk (OAR) segmentation on these images is essential for dose monitoring, yet most existing CBCT segmentation methods process each fraction independently, ignoring temporal anatomical changes. We propose a longitudinal ConvLSTM-UNet for multi-organ segmentation of CBCT-derived synthetic CT (sCT) images. The framework was evaluated under two clinically relevant settings: fully automatic autoregressive inference and replanning-guided inference that updates temporal memory using contours from replanning CTs. Compared with a nnU-Net baseline, replanning-guided inference achieved the best overall performance (mean DSC 77.4%, mean ASD 1.67 mm), significantly improving 12 of 17 OARs, while autoregressive inference remained competitive without manual updates. We further performed a longitudinal dosimetric analysis that revealed significant dose increases in the parotid and submandibular glands. 46% of patients exceeded the 26 Gy contralateral parotid mean-dose constraint during treatment despite remaining below it on the planning CT based on rigid dose propagation. These results demonstrate that longitudinal deep learning improves multi-organ CBCT segmentation while enabling automated dosimetric monitoring, supporting its potential for ART workflows.

Oral & Poster

DoMa-Seg: Dose Map Guidance for Medical Image Segmentation in Head and Neck Radiotherapy

Jihe Li, Amith Kamath, Daniel Schanne, Olgun Elicin, Mauricio Reyes

Presented in Oral Session 1: Segmentation, Contour Quality & Calibration.

OpenReview → Slides (PDF) →
Abstract

Deep learning-based segmentation models for radiotherapy treatment planning are typically trained with geometric losses (e.g., Dice). However, clinical studies report a weak correlation between geometric contour accuracy and dosimetric outcomes, particularly in complex anatomical regions like the Head and Neck (H&N). We propose Dose Map Guided Segmentation (DoMa-Seg), a novel training paradigm that directly integrates dosimetric awareness into the segmentation loss using readily available dose distributions, without requiring an auxiliary network, as in prior work. To overcome the non-differentiability of standard dosimetric endpoints derived from Dose-Volume Histograms (DVHs), we introduce a differentiable training loss based on the generalized Equivalent Uniform Dose (gEUD), a radiobiologically grounded metric that naturally extends from binary masks to soft probability maps. Experiments on 3D H&N segmentation show that DoMa-Seg achieves a 12.0% lower Dosimetric Impact Approximation (DIA) error for target volumes, and 1.8 × faster training compared to a state-of-the-art approach incorporating dosimetric information during model training, highlighting its clinical potential for radiotherapy workflows. Our code is available at https://github.com/Jihe-Li/DoMa-Seg.

Oral & Poster

Parameter-Efficient Pretrained-CT-to-MRI Transfer for Rectal Cancer Segmentation: Performance-Calibration Trade-Offs

Aneesh Rangnekar, Jorge Tapias Gomez, Joseph O. Deasy, Harini Veeraraghavan

Presented in Oral Session 1: Segmentation, Contour Quality & Calibration.

OpenReview → Slides (PDF) →
Abstract

Accurate rectal cancer segmentation from magnetic resonance imaging (MRI) is essential for adaptive radiotherapy and tumor response assessment, but deployment also requires computational efficiency and informative, calibrated uncertainty estimates. We therefore introduce SWIFT, a SWin pretrained model with parameter-eFficient and Tumor-aware fine-tuning for rectal cancer segmentation. A Swin V2 encoder pretrained on 10,444 public 3D CT volumes using a DINOv2-style objective was adapted to T2-weighted MRI through four cumulative configurations: full fine-tuning (SWIFT), decoder compression (SWIFTe), low-rank adaptation (SWIFTe-LoRA), and a four-member LoRA-decoder ensemble (SWIFTe-LDE4). Geometric accuracy, tumor detection, radiomic agreement, and probability calibration were evaluated on a held-out 247-case test set from a single-institution cohort acquired using 1.5 or 3 Tesla GE scanners. Compared with SWIFT, SWIFTe reduced total parameters by 70.1% (from 72.8M to 21.8M) and increased tumor detection rate from 89.9% to 93.9%, while achieving a slightly lower median surface DSC (0.61 versus 0.62) and improved radiomic agreement. In a separate SWIFTe ablation, removing tumor-aware augmentation reduced detection from 93.9% to 89.9% but increased surface DSC from 0.61 to 0.64, demonstrating a detection-boundary-agreement trade-off. SWIFTe-LoRA used 14.6% of SWIFTe's trainable parameters while retaining similar segmentation performance. SWIFTe-LDE4 achieved the lowest calibration errors among the four configurations after temperature scaling (expected calibration error, 0.217; Brier score, 0.222), although the absolute expected calibration error indicates residual miscalibration. Similar efficiency-calibration patterns were observed using the public VoCo checkpoint, supporting robustness across pretrained initializations rather than external clinical generalizability. Code and selected model checkpoints are available in our https://github.com/The-Veeraraghavan-Lab/RectalTxSegblueGitHub repository.

Oral & Poster

Recognizing Dosimetrically Irrelevant Edits of Auto-Segmentations in Head and Neck Organs of Interest

Mar Fernandez Salamanca, Rita Simões, Joëlle E. van Aalst, Charlotte L. Brouwer, Tomas M. Janssen

Presented in Oral Session 1: Segmentation, Contour Quality & Calibration.

OpenReview → Slides (PDF) →
Abstract

Manual editing of auto-segmentations of organs of interest (OOI) contours remains a major bottleneck in radiotherapy (RT) workflows. Without clear criteria, clinicians may over- or under-edit, without knowing whether their changes provide a clinically relevant dosimetric benefit. This study assesses the relationship between manual editing patterns and dosimetric impact, and translates the known relationship between OOI–planning target volume (PTV) distance and dosimetric impact into practical contour editing recommendations. A retrospective cohort of 766 head-and-neck RT patients was analyzed, where ATLAS based-auto-segmentation was clinically used for all but 166 patients where deep learning (DL)-based-auto-segmentation model was used. Dosimetric impact was quantified using the dose impact approximation (DIA), by comparing the dose metrics of both segmentations using the clinical dose distribution under a non-replanned framework. Despite substantial variability in manual editing across time, DIA remained stable, indicating limited overall dosimetric impact. In the DL cohort, DIA was strongly associated with OOI–PTV distance, with larger deviations observed for structures overlapping the PTV. Hybrid segmentations were used to simulate spatially constrained editing strategies. Non-overlapping OOIs (42%) exhibited minimal residual differences (<0.5 Gy) even without editing. For overlapping OOIs (58%), restricting edits to within 2 cm of the PTV reduced residual differences to below 0.1 Gy for the submandibular, with limited additional benefit beyond this distance. Based on these findings, we propose a distance-based editing strategy consisting of (1) no routine editing for non-overlapping OOIs and (2) restricting edits of overlapping submandibulars OOIs to regions within 2 cm of the PTV.

Poster

Supporting Clinical Trial Quality Assurance of Organ-at-Risk Contours with Image-Conditioned Diffusion Models

Clea Dronne, Catharine H. Clark, Xavier Loizeau, Elizabeth Miles, Peter Hoskin, Jamie R. McClelland

OpenReview →
Abstract

Clinical trial quality assurance of organs-at-risk contours is currently performed manually and can be time-consuming and subjective. We evaluate whether reconstruction-based normative modelling can support this process by identifying contours that deviate from a learned distribution of clinically acceptable segmentations. A previously developed VAE model and an image-conditioned segmentation diffusion model were trained on brainstem and spinal cord contours from the RADCURE dataset and evaluated on real rejected and corrected cases from the ARTDECO clinical trial. Quality assurance was performed by comparing input contours under review with model reconstructions using global and regional Distance-to-Agreement maps. The diffusion model more consistently assigned lower DTA values to corrected contours and better localised clinically corrected regions than the VAE. These results suggest that image-conditioned diffusion reconstruction is a promising model-agnostic tool for supporting organ-at-risk contour review in radiotherapy clinical trials. Code is available here.

Poster

Which Edits Matter? Simulating Realistic Local Corrections to Organ-of-Interest DL Segmentation and Predicting Dosimetric Impact

Joëlle van Aalst, Tomas Janssen, Federica Maruccio, Rita Simões, Mar Fernandez Salamanca, Peter van Ooijen, Charlotte Brouwer

OpenReview →
Abstract

Clinical application of deep learning (DL)-based organ-of-interest (OOI) segmentation in radiotherapy requires case-level manual verification and adjustment. In current clinical practice, the full structure is typically verified and adjusted, limiting efficiency as not every local adjustment translates into a meaningful change in the treatment plan. Identifying such negligible adjustments could allow their omission, however no framework currently predicts this at the level of local corrections. In order to evaluate which edits matter, we introduce a framework that clusters the total discrepancy between a pair of DL and clinically used contours into spatially and directionally coherent local correction regions. We next apply each correction region individually to the original DL contour to simulate a realistic, isolated local correction. We quantify its dosimetric impact as the resulting change in mean dose under the original treatment plan compared to the original DL contour. We apply this framework to a set of 15 head and neck OOIs across 61 patients. Using this set, we train a gradient boosting classifier to distinguish negligible ( 0.01 Gy dose difference) from non-negligible local corrections (>0.01 Gy) using the following features: OOI-type, target location, original dose-level, distance to target, and correction volume. The classifier achieved an ROC-AUC of 0.93 and average precision of 0.89. This framework could be applied to inform the development of standardised, population-level review guidelines that move from full structure correction towards regional, prioritised editing. Furthermore, when applied directly to a given patient, the same approach could provide case-specific guidance during segmentation correction, for example by decomposing a full uncertainty map into candidate editing regions. This could be particularly valuable in time-critical settings such as online adaptive radiotherapy.

Dose Calculation, Dose Prediction & Plan Optimisation (5 papers)

Poster

4D Pencil Beam Treatment Plan with Conditional Weight Predictions

Nair N. von Mühlenen, Florentin Bieder, Philippe C. Cattin

OpenReview →
Abstract

Objective. This work aims to accelerate our four-dimensional (4D) proton pencil beam delivery strategy, which incorporates respiratory motion into a dynamic treatment plan, to improve dose conformity and treatment efficiency for gantry-less and magnet-free scanner designs. Approach. To accelerate the generation and adaptation of 4D treatment plans, we propose a conditional hybrid ResNet18-Transformer model for predicting pencil beam weights. The model predicts either the full set of pencil beam weights or a subset thereof. Main Results. The conditional model can successfully predict the remaining beam weights of a treatment, and the results suggest that extending this approach to full treatment prediction is feasible but requires further investigation. Significance. The acceleration of 4D treatment generation via pencil beam weight prediction takes us one step closer to the feasibility of treating mobile targets with simplified, gantry-less, and magnet-free scanner designs. Reducing system complexity while preserving dosimetric conformity may offer a pathway toward more accessible and cost-effective proton beam therapy for motion-affected tumours. Our code is available on github.com/NairVonMuehlenen/Pencil-Beam-Weight-Prediction.

Poster

A Differentiable Gaussian Mixture Model-Based Scorecard for Cohort-Aware Radiotherapy Plan Evaluation and Optimization

Ho-Hsin Chang, Rex Cardan, Richard A. Popple, Dennis N. Stanley, John B. Fiveash, Sandeep Bodduluri, Joseph Harms, Carlos E. Cardenas

OpenReview →
Abstract

Review of radiotherapy (RT) treatment plans often relies on subjective analysis by physicists and physicians. Qualitative step-wise metrics are non-differentiable as optimization objectives, posing hurdles to automated RT planning. While continuous plan quality scorecards are used for retrospective evaluation, their use as optimization objectives remains limited. This paper proposes a data-driven, differentiable scorecard based on a Gaussian Mixture Model cumulative distribution function (GMM-CDF) for cohort-aware radiotherapy plan evaluation. Historical clinical prostate plans were categorized by prescription and target complexity, including 70 Gy hypofractionation and 40 Gy SBRT cohorts. A 3-component GMM mapped dosimetric parameters to continuous percentile-like scores. The naturally differentiable GMM-CDF yields probability density functions as exact derivatives, enabling fast gradient-based optimization. The GMM-CDF scorecard was evaluated via four experiments: an 80/20 train/test split, a target-complexity cohort assessment, a cross-prescription mismatch test, and scalar dose-scaling optimization using an L-BFGS-B solver. Results demonstrated similar grade distributions between train and testing set. Cohort-specific scorecards produced distinct organ-at-risk (OAR) score distributions versus generalized scorecards, particularly across single-, dual-, and tri-target groups. In cross-prescription testing, OAR scores shifted substantially with mismatched prescription scorecards, whereas prescription-normalized target scores remained relatively stable. In optimization testing, average plan scores improved by 22.8 points. The proposed GMM-CDF scorecard provides a continuous, differentiable approach of selected DVH-based plan quality metrics, supporting differentiable statistical scorecards as optimization-aware plan quality tools.

Poster

Generalizing Monte Carlo Dose Reconstruction Beyond the Training Anatomy Using Conditional Diffusion Models

Hong-Phuong Dang, Richard Gozan

OpenReview →
Abstract

Accurate Monte Carlo (MC) dose calculation is essential in radiotherapy, but obtaining high-sampling (HS) dose maps requires a very large number of particle histories and remains computationally expensive. Deep learning has recently been investigated as an asymptotic variance reduction strategy to reconstruct HS dose maps from low-sampling (LS) simulations. However, most existing approaches formulate this task as deterministic image-to-image regression and may be sensitive to anatomical distribution shifts. This limitation is particularly important when HS references are scarce for a new anatomical site. In this work, we investigate conditional denoising diffusion probabilistic models (DDPMs) for MC dose reconstruction beyond the training anatomy. The proposed model learns the conditional distribution of the HS dose map given the corresponding LS simulation and planning computed tomography (CT) image. It reconstructs the HS dose map through an iterative stochastic denoising process rather than a single deterministic prediction. The model is trained on abdominal data and evaluated on abdominal test data and five unseen pelvic and head-and-neck slices. On the abdominal test set, it performs close to a deterministic U-Net. On the limited cross-anatomy set, the DDPM produces visually more faithful reconstructions in the examples considered. Geometric alignment further improves its results. Although preliminary, these findings motivate larger-scale evaluation of conditional diffusion models for MC dose reconstruction.

Poster

Heterogeneous IMRT and VMAT Dose Calculation Using Deep Learning with Beam Path Reconstruction and PTV Information

Loes Vandenbroucke, Emma Van Riet, Dylan Callens, Patrick Berkovic, Maarten Lambrecht, Sandra Nuyts, Wouter Crijns, Frederik Maes

OpenReview →
Abstract

Deep learning (DL) enables fast 3D dose calculation for radiotherapy, but existing approaches are restricted to a single treatment technique or tumor site. Furthermore, representing delivery parameters in the patient domain remains challenging. We propose a generalizable DL framework for full-plan dose calculation that reconstructs the full beam path by projecting the fluence maps into the patient domain and that incorporates PTV information to guide target dose prediction. A heterogeneous model trained jointly on lung IMRT and head-and-neck (H&N) VMAT was compared with modality-specific models to evaluate generalization across treatment techniques and tumor sites. Compared with homogeneous training, heterogeneous training increased the average gamma passing rate (98.42% vs. 98.26% for H&N VMAT; 98.60% vs. 98.54% for lung IMRT) and reduced mean dose difference for the planning target volume (2.07% vs. 3.34% for H&N VMAT; 2.16% vs. 2.82% for lung IMRT). Organs at risk mean dose differences remained below 3% for both approaches. Ablation experiments demonstrated that beam-path reconstruction and PTV information improve prediction accuracy.

Oral & Poster

Uncertainty-Aware Out-of-Field Dose Prediction in External Beam Radiotherapy

Mohammed El Aichi, Nathan Benzazon, Meïssane M'Hamdi, Rodrigue Allodji, Florent de Vathaire, Eric Deutsch, Ibrahima Diallo, Maria Vakalopoulou, Charlotte Robert

Presented in Oral Session 3: Adaptive Radiotherapy (Longitudinal Imaging, Motion & Dose).

OpenReview → Slides (PDF) →
Abstract

Low and very low radiation doses delivered outside the treatment field in radiotherapy can affect the immune system and contribute to radiation-induced lymphopenia, which is associated with poorer clinical outcomes. This is particularly relevant in radio-immunotherapy settings, where preservation of immune function may directly impact therapeutic efficacy. In this work, we propose an energy-conditioned deep learning framework for voxel-wise out-of-field (OOF) dose prediction from the planned in-field dose and whole-body mask. Beam energy is incorporated through additive conditioning, and the deterministic model is further extended to a heteroscedastic mean variance formulation for voxel-wise uncertainty estimation. Our framework is developed and evaluated on 1,519 reconstructed whole-body dose distributions from retrospective photon-based LINAC treatment records. The best-performing model, an energy-conditioned U-Net, achieved an MAE of 9.76 ± 9.28 cGy and an RMSD of 15.61 ± 12.50 cGy. Energy conditioning significantly improved the U-Net baseline, while mean variance modeling provided uncertainty estimates positively correlated with prediction errors. These results demonstrate the potential of energy-conditioned deep learning for rapid and uncertainty-aware whole-body OOF dose estimation. Code is available at https://github.com/maichi98/DoseFieldNet.

Outcome, Toxicity & Treatment-Response Modelling (4 papers)

Poster

Dynamic Modeling of Brain Metastasis Response with Multitask Temporal Learning

Vincent Andrearczyk, Lorenz Achim Kuhn, Jonas Richiardi, Andreas F. Hottinger, Luis Schiappacasse, Vincent Dunet, Adrien Depeursinge

OpenReview →
Abstract

Patients with brain metastases (BMs) treated using stereotactic radiosurgery (SRS) undergo longitudinal MRI to monitor response, yet outcomes vary substantially between lesions across and within patients, reflecting biological heterogeneity and differences in radiosensitivity. Modeling response is challenging due to heterogeneous temporal dynamics and irregular follow-up schedules, with varying numbers of scans per patient. Conventional machine learning relying on fixed, regularly sampled inputs is therefore limited to predict BM response to SRS. We present a large retrospective post-SRS dataset of 4,766 lesions from 568 patients, monitored with contrast-enhanced T1-weighted MPRAGE and T2-weighted MRI over a mean/median duration of 316/214 days. We formulate lesion-level response prediction as (i) future complete response (CR) classification at flexible temporal horizons and (ii) survival analysis with time to complete response (TCR) as the event (53% event rate). We propose a time-aware framework for irregular longitudinal data based on gated recurrent units (GRU). Absolute time since SRS is explicitly encoded and integrated with dynamic lesion-level imaging features and patient-level static covariates. The model processes variable-length input sequences and supports flexible prediction horizons within a unified architecture. Our approach outperforms MLP and XGBoost baselines that require fixed input representations, achieving a C-index of up to 0.77 for TCR prediction and AUCs between 0.74 and 0.90 for CR classification. This work establishes a scalable framework for personalized longitudinal response modeling in SRS-treated BMs. Code will be released upon acceptance.

Poster

Mammo-LIFE: Longitudinal Mammographic Imaging and Clinical Feature Enrichment for Post-Radiotherapy Outcome Prediction

Farnoush Bayatmakou, Maryam Hosseini, Reza Taleei, Arash Mohammadi

OpenReview →
Abstract

Recent advances in Artificial Intelligence (AI)-powered Computer-Aided Diagnosis (CAD) systems have substantially improved breast cancer screening, diagnosis, and prognosis. Comparatively, post-radiotherapy outcome prediction using paired longitudinal mammograms has received considerably less attention. This is largely due to the limited availability of well-annotated longitudinal datasets. Longitudinal mammograms, coupled with paired pre- and post-treatment information, provide a unique opportunity to characterize treatment-induced breast tissue changes following radiotherapy. The resulting learned representations can serve as a valuable asset for advancing personalized radiotherapy planning and post-treatment management. In this context, we propose Mammo-LIFE, a patient-level multimodal framework for post-radiotherapy outcome prediction that combines longitudinal mammographic features with patient-level clinical variables. The imaging branch processes paired pre- and post-treatment mammograms acquired from the four standard views using a mammography-specific encoder adapted via Low-Rank Adaptation (LoRA). Within each view, pre- and post-treatment representations are explicitly compared through a longitudinal comparison module to capture treatment-related changes. The resulting view-level embeddings are then aggregated using learned view-attention pooling to form a unified patient-level mammographic representation. Selected clinical variables are subsequently combined with the image-derived prediction probability through a late-fusion strategy. To evaluate the effectiveness of combining paired longitudinal mammograms with clinical information, experiments were conducted on an in-house clinical cohort using patient-level stratified five-fold cross-validation. The proposed Mammo-LIFE demonstrated strong cross-validated performance across the evaluated configurations. The best setting achieved an AUC of 0.86 ± 0.14, accuracy of 0.79 ± 0.15, and F1-score of 0.83 ± 0.11.

Poster

More Comprehensive and Reliable Deep Learning Normal Tissue Complication Probability Models for Head and Neck Cancer Patients

Daniel C. MacRae, Luuk van der Hoek, Suzanne P.M. de Vette, Hendrike Neh, Joëlle E. van Aalst, Matias Valdenegro-Toro, Nanna M. Sijtsema, Peter M.A. van Ooijen, Lisanne V. van Dijk

OpenReview →
Abstract

Conventional and existing deep learning (DL) normal tissue complication probability (NTCP) models for head and neck cancer (HNC) typically predict a single radiation-induced toxicity at a time and provide only deterministic risk estimates, despite evidence that toxicities share biological mechanisms, and despite a clinical need for information about the reliability of these risk estimates. Thus, this study proposes a multi-toxicity (MT) deep learning model that jointly predicts five toxicity endpoints (aspiration, dysphagia, sticky saliva, taste alteration, and xerostomia) and evaluates two uncertainty quantification (UQ) methods, Monte Carlo dropout (MCD) and deep ensembles (DE). Using a cohort of 1,373 HNC patients, the MT model achieved comparable performance to single-toxicity (ST) models and higher discriminative performance and calibration for sticky saliva and taste alteration. For UQ, both DE and MCD produced reliable uncertainty estimates when combined with binary entropy, with DE consistently outperforming MCD. These findings demonstrate that MT modelling can provide a more comprehensive and better-calibrated toxicity risk profile than ST models, and that DE combined with binary entropy offers the most reliable basis for uncertainty-aware, trustworthy DL-based NTCP modelling.

Poster

PR3DICTR: A 3D Image-Based Deep Learning Prediction Modelling Framework and Its Use in Radiotherapy

Luuk van der Hoek, Daniel C. MacRae, Robert van der Wal, Suzanne P.M. de Vette, Hendrike Neh, Baoqiang Ma, Nanna M. Sijtsema, Peter M.A. van Ooijen, Lisanne V. van Dijk

OpenReview →
Abstract

Imaging data is becoming more ubiquitous in the medical domain, particularly in radiotherapy. Deep learning (DL) models have been shown to be effective in utilizing medical image data in prediction tasks, however they remain challenging to develop and are yet to be widely adopted in clinical practice. To aid standardized development of 3D image-based DL prediction models we introduce PR3DICTR: Platform for Research in 3D Image Classification and sTandardised tRaining. PR3DICTR uses a 6-step process for effective use: 1-3 prerequisites (image/tabular data preprocessing and organization); 4-5 model construction and user input (including training and hyperparameter tuning); and 6 monitoring and evaluation. To highlight its use in radiotherapy, PR3DICTR was used for two radiotherapy related prediction tasks: 1) dysphagia 6 months after head and neck cancer (HNC) treatment; and 2) overall survival after HNC treatment. For both tasks, the PR3DICTR model achieves comparable or slightly better performance than the published counterparts (task 1: AUC: 0.86 vs 0.84, task 2: C-index: 0.78 vs 0.71). PR3DICTR was conceived to be a continuously evolving platform, and we invite any researcher to see if it is applicable for their research efforts.

Motion Management, Longitudinal Adaptation & Registration (3 papers)

Poster

Is Deformable Image Registration Ready for Brain Metastasis Reirradiation Dose Accumulation? A Longitudinal MRI Benchmark of Registration Accuracy

Hengjie Liu, Manju Sharma, Xinyi Fu, Di Xu, Ke Sheng

OpenReview →
Abstract

Dose accumulation is increasingly important in adaptive radiation therapy and reirradiation, but its clinical validity depends on the performance of deformable image registration (DIR). Reirradiation of brain metastases (BMs) with stereotactic radiosurgery (SRS) provides a controlled but clinically meaningful DIR test case: intra-subject brain deformation is usually limited after rigid alignment, yet recurrent lesions can undergo substantial local shape and volume changes that rigid registration cannot capture and can affect dose accumulation. We benchmarked a wide range of learning-based and optimization-based DIR methods on 87 manually screened longitudinal contrast-enhanced T1-weighted MRI lesion pairs from an institutional BM SRS retreatment cohort. Learning-based methods pretrained on healthy-brain MRI were evaluated zero-shot and after instance-specific optimization (ISO) or tumor-proximity target-specific optimization (TSO). Registration was assessed using lesion overlap (Dice), surface distance metrics (HD95 and sASD), target-volume recovery, and runtime and memory. Pretrained learning-based methods showed variable zero-shot performance, while ISO/TSO improved all tested learning-based families. However, optimization-based methods remained the best-performing approach while maintaining reasonable runtime. These findings suggest that even state-of-the-art DIR methods do not yet provide sufficiently accurate and consistent registration for unmonitored use in brain metastasis reirradiation dose accumulation. Because accurate registration is a prerequisite for deformable dose accumulation, clinical application will require case-level quality control and direct assessment of how registration uncertainty affects downstream dose metrics.

Oral & Poster

Motion-Consistent Memory for Foundation Model-Based Tumor Tracking in MR-Guided Radiotherapy

Pauline Ornela Megne Choudja, Marcel Nachbar, Aya Ghoul, Cihan Gani, Daniela Thorwarth, Thomas Kuestner

Presented in Oral Session 3: Adaptive Radiotherapy (Longitudinal Imaging, Motion & Dose).

OpenReview → Slides (PDF) →
Abstract

Accurate tumor tracking in MR-guided radiotherapy is critical for real-time beam gating and adaptive radiotherapy delivery. Prolonged tracking failures risk geometric misses of the target volume. Dynamic MR imaging is usually performed to capture the underlying motion with subsequent segmentation and tracking of the lesion. While foundation models such as SAM2 have shown strong performance in video object segmentation, their memory retrieval mechanisms rely predominantly on appearance-based representations, making them susceptible to object drifts under non-rigid deformable motion. We propose a motion-guided memory modulation framework that augments SAM2 with explicit motion consistency cues without retraining the base model. An optical flow model is used to estimate motion and propagate tumor masks between consecutive frames. Agreement between the propagated mask and the current SAM2 prediction is quantified using an IoU-based consistency metric. The metric acts as a reliability weight to modulate the memory representations prior to retrieval, suppressing unreliable memories before they corrupt the tracking state. Evaluated on the TrackRAD2025 Challenge dataset, our framework markedly improves over both SAM2 and MedSAM2 baselines. Against MedSAM2, we reduce the miss rate from 6.1% to 1.3% (109 additional recovered frames), improve J&F (mean Jaccard and contour F-score) by 3.3 percentage points, and reduce maximum failure duration from 90 to 8 frames, corresponding to approximately 22s versus 2s at clinical cine MRI acquisition rates, while preserving segmentation quality (median Dice 0.884). All variants operate within real-time clinical constraints, processing each frame in under 200ms. These results demonstrate that motion-guided memory modulation is a lightweight, plug-and-play, and clinically motivated strategy for robust tumor tracking in MRI-guided adaptive radiotherapy.

Poster

Patient-Informed XCAT Modelling of Free Breathing and Deep-Inspiration Breath Hold Anatomy: A Feasibility Study

Maria Grazia Carnevale, Stevan Vrbaski, Anna Cavallo, Carlotta Giandini, Alessandro Cicchetti, Maria Carmen De Santis

OpenReview →
Abstract

Deep-inspiration breath-hold (DIBH) reduces cardiac exposure in left-breast radiotherapy, but its anatomical benefit varies substantially between patients. We evaluated whether patient-informed 4D Extended Cardiac-Torso (XCAT) phantoms could accurately reproduce clinical free-breathing (FB) and DIBH organ-volume distributions. For sixteen patients with paired FB/DIBH CT scans and valid respiratory traces, patient-specific XCAT phantoms were generated by matching clinical breast, lung, and heart volumes. Respiratory traces were then used to parameterise thoracic expansion and generate the corresponding DIBH configuration. Agreement was assessed using patient-level errors and bootstrap resamples. Mean absolute percentage errors (MAPE) were 0.17%, 0.72%, and 0.11% for FB breast, lung, and heart volumes, and 0.22%, 3.61%, and 1.74% for DIBH scans. Mean clinical and XCAT left-lung expansions were 828.6 and 852.6 cm³, respectively. These findings support the feasibility of patient-informed XCAT modelling for controlled FB/DIBH population studies, while spatial and dosimetric validation remain necessary.

Agentic Large Language Models for Image Analysis (1 paper)

Poster

Agentic Large Language Models for Training-Free Neuro-Radiological Image Analysis

Ayhan Can Erdur, Daniel Scholz, Jiazhen Pan, Benedikt Wiestler, Daniel Rueckert, Jan C. Peeken

OpenReview →
Abstract

State-of-the-art large language models (LLMs) show high performance in general visual question answering. However, a fundamental limitation remains: current architectures lack the native 3D spatial reasoning required to directly analyze volumetric medical imaging, such as CT or MRI. Emerging agentic AI offers a new solution, eliminating the need for intrinsic 3D processing by enabling LLMs to orchestrate and leverage specialized external tools. Yet, the feasibility of such agentic frameworks in complex, multi-step radiological workflows remains underexplored. In this work, we present a training-free agentic pipeline for automated brain MRI analysis. Validating our methodology on several LLMs (GPT-5.4, Gemini 3.1 Pro, Claude Sonnet 4.6) with off-the-shelf domain-specific tools, our system autonomously executes complex end-to-end workflows, including preprocessing (skull stripping, registration), pathology segmentation (glioma, meningioma, metastases), and volumetric reasoning. We evaluate our framework across increasingly complex radiological tasks, from single-scan tumor and anatomy analysis to longitudinal response assessment requiring multi-timepoint comparisons. We analyze the impact of architectural design by comparing single-agent models against multi-agent "domain-expert" collaborations. To support rigorous evaluation of future agentic systems, we release a benchmark dataset of image-prompt-answer tuples derived from public data. Our results demonstrate that agentic AI can solve highly neuro-radiological image analysis tasks through tool use without the need for training or fine-tuning.

Call for Papers

The call for papers is closed. Submissions closed on July 17, 2026, and the review process is complete. The accepted contributions are listed above and in the final programme. This section is retained as a record of the call.

Submissions were invited across the full breadth of the therapeutic AI workflow, including but not limited to:

Submission and Review Process

Important Dates

← Back to Home Committee & Contact →