Abstract Radiation therapy (RT) is a crucial component of cancer management, with most patients receiving RT at some point, often in combination with surgery or systemic treatments. Cell sparing at ultrahigh dose rates was found in in vitro cell survival fraction experiments dating back to the 1960s. In the 1970s, studies revealed an unexpected dose-rate effect on radiation-induced mortality in mice and a dose-rate-dependent response in rat skin. This effect was assumed to be affecting healthy and tumour tissue equally and was not pursued further. Decades later, this modality of RT was rediscovered using ultra-high dose-rate irradiation (average dose rate >40 Gy/s) and it was demonstrated that this improves the therapeutic ratio, sparing healthy tissue while being isoeffective for tumours. This effect is referred to as the FLASH effect. Currently, the first human trials using electron and proton beams are ongoing, although significant challenges remain in fully understanding the radiobiological and chemical mechanisms behind this phenomenon. This roadmap provides an overview of the current and future challenges of FLASH radiation therapy and outlines research directions to optimally advance the field and design future clinical trials: radiobiology of the FLASH effect, opportunities and beam parameters for the future in FLASH with electrons, intraoperative FLASH radiotherapy with electrons, ultra-high dose-rate dosimetry, opportunities for the future in FLASH with protons, FLASH treatment planning and machine QA with protons, clinical translation of FLASH with electrons and first experiences and challenges of initial proton FLASH trials.
Abstract Objective: Quantitative assessment of renal cell carcinoma (RCC) ablation margins requires reliable registration of pre- and post-ablation images. However, ablation-related tissue changes weaken local anatomical correspondence, making deformation estimates potentially non-unique. We developed an uncertainty-aware registration framework for more reliable margin estimation. Approach: An ensemble of plausible registrations was generated by stochastically sampling key registration settings, including the registration mask, similarity metric, B-spline grid spacing, and initial alignment, with Latin hypercube sampling used to improve parameter-space coverage. Registration uncertainty was characterized from this ensemble and propagated into probabilistic margin estimation. Evaluation used 25 biomechanically simulated cases with known deformation fields and 47 clinical RCC cases. Main results: In simulation, 98 ± 2% of kidney voxels had a mean displacement error below 1 mm, and uncertainty was strongly associated with registration error (median case-wise Spearman correlation, 0.99). In the clinical study, our method achieved anatomical alignment consistent with the deterministic baselines, with a median kidney Dice similarity coefficient of 0.944 and a median local mean absolute surface distance of 0.5 mm. In general, our method maintained a sensitivity of 1.00 while achieving higher positive predictive value (PPV) than the deterministic methods. Under the global criterion, PPV was 0.70 versus 0.54–0.58, although the pairwise differences between our method and each deterministic method were not statistically significant. Under the focal criterion, PPV was 0.78 versus 0.29–0.50. The pairwise differences were significant between our method and the standard and penalty-constrained B-spline methods (both p<0.001), but not between our method and the diffeomorphic method (p=0.063). Significance: The framework extends registration uncertainty estimation to RCC ablation-margin assessment by converting variability in plausible correspondences into probabilistic margin maps. It highlights negative-margin regions consistently identified across registrations, reduces registration-sensitive false-positive local tumor progression predictions, and may support more reliable postoperative evaluation.
Abstract Objective. MRI is increasingly required to operate near electronic devices that emit dynamic electromagnetic interference (EMI). Existing reference-channel EMI removal methods estimate a transfer function between EMI sensors and imaging coils from the measured data alone, and do not use prior knowledge about the structure of MR images. We develop and validate an EMI removal method that uses the inherent spatial smoothness of MR images as an additional prior, demonstrated on a head-only 0.5T scanner. Approach. Smooth Total variation Regularization for Interference Detection and Elimination (STRIDE) operates in the image domain. For each image column, EMI sensor data are combined and subtracted from the imaging coil data such that the L2 norm of the total variation (TV) along the readout direction is minimized. This exploits the smoothness of MR images, a property well-known in compressed sensing but not previously used in EMI removal. STRIDE was evaluated on phantom and in-vivo brain datasets under EMI scenarios replicating sources encountered at our affiliated hospitals, and compared against two implementations of External Dynamic Interference Estimation and Removal (EDITER). Main results. STRIDE produced visually better EMI removal, higher temporal signal-to-noise ratio, and larger EMI removal percentage than both EDITER implementations. STRIDE also resulted in lower or equal root-mean-square error than both EDITER implementations in all 8 phantom EMI conditions tested. These results held across both phantom and in-vivo datasets. Significance. STRIDE shows that incorporating known properties of MR images into the EMI removal problem provides meaningful improvements over methods that rely on measured data alone. The approach is deterministic, training-free, sequence-independent, and compatible with existing EMI sensor hardware. STRIDE therefore extends reference-channel EMI removal to scenarios where existing methods fail, particularly point-of-care and interventional MRI outside fully shielded environments, providing a complementary path to machine-learning approaches.
Abstract Objective. To derive an exact analytical expression for the diffusion MRI signal arising from diffusion confined to a cylindrical surface under finite rectangular pulsed-gradient spin-echo (PGSE) gradients, and to develop computationally efficient strategies for repeated model evaluations. Approach. The Bloch–Torrey equation was solved using a spectral matrix formalism of the Laplace operator in the eigenbasis of the cylindrical surface. The resulting signal is expressed as a product of non-commuting matrix exponentials and is valid for arbitrary rectangular gradient durations and separations, without approximations to the diffusion propagator or spin phase distribution. A reduced real spectral basis was introduced to decrease the dimensionality of the problem. Accelerated implementations based on Strang splitting and Gauss–Legendre quadrature were developed for repeated signal evaluations and computation of the spherical mean. The analytical signal was validated against Monte Carlo diffusion simulations. Main results. The exact formulation showed excellent agreement with Monte Carlo simulations over a wide range of cylinder radii and acquisition parameters. The spectral expansion converged rapidly with mode order, and the reduced basis substantially decreased computational cost without loss of accuracy. The accelerated implementations provided large additional speed-ups while retaining controllable errors relative to the exact solution. Significance. The proposed framework extends exact finite-pulse solutions of the Bloch–Torrey equation to diffusion confined to cylindrical surfaces and provides practical implementations for directional and orientationally averaged diffusion MRI signals. It therefore offers an approximation-free reference for assessing faster signal approximations and for applications requiring large numbers of model evaluations.
Abstract Objective: This study investigates how domain-aligned self-supervised pretraining influences downstream PET/CT tumor segmentation, with a focus on disentangling the roles of domain proximity and pretraining dataset scale. Approach: We compared nnU-Net segmentation models initialized from random weights or from DINOv3 encoders pretrained on natural images, X-ray mammography, CT, or PET, using datasets of different sizes. To disentangle the effects of pretraining domain and dataset size, CT-DINOv3 encoders were pretrained using different numbers of 2D CT slices, including subsets matched in size to the pretraining datasets from other domains. Models were evaluated on two PET/CT benchmarks, autoPET (whole-body, FDG/PSMA) and HECKTOR (head-and-neck primary tumors and metastatic lymph nodes), using five-fold cross-validation with frozen encoders. Main results: CT-pretrained models achieved the highest Dice scores for lymph nodes, FDG lesions, and PSMA lesions. For primary tumors, PET pretraining produced the highest observed Dice score, followed by CT pretraining, suggesting the potential benefit of in-domain representations when PET data are sufficient. Notably, PET pretraining exceeded CT pretraining using equally sized datasets for PSMA but remained below models pretrained on larger CT datasets for both PSMA and FDG targets. In contrast, natural-image pretraining did not outperform from-scratch training despite its much larger pretraining scale, indicating that scale alone is insufficient when domain proximity is low. Across comparisons, strong transfer reflected the combination of domain proximity and sufficient pretraining scale, especially for CT-based models with broad data availability. Significance: These results demonstrate that both domain proximity and dataset scale jointly determine the effectiveness of self-supervised pretraining. While higher domain alignment improves performance at comparable scale, sufficient data is required to realize this benefit. CT-based pretraining provides a favorable balance between domain relevance and data availability, leading to robust performance across tasks. This study provides practical guidance for selecting pretraining strategies in PET/CT segmentation.
