Background: Recently, CT manufacturers have developed deep‐learning‐based reconstruction (DLR) algorithms for conventional and spectral CT images. On dual‐energy CT platforms, DLR algorithms have been developed in particular to reduce image noise in virtual monoenergetic images (VMIs), and to improve the accuracy of iodine concentrations on iodine maps.Purpose: To assess the impact of a prototype, dedicated Spectral AI Reconstruction (SAIR) algorithm (Spectral Precise Image; Philips Healthcare) compared with a hybrid iterative reconstruction (HIR) algorithm (“Spectral”, Philips Healthcare) on the quality of low‐keV VMIs and iodine maps in abdomen‐pelvis CT conditions.Methods: A spectral phantom was scanned on the dual‐layer spectral CT system at 120 kVp and a CTDIvol of 12 mGy. VMIs at 40/50/60/70 keV were generated using Level 4 of HIR algorithm and the five levels of SAIR algorithm named Smoother/Smooth/Standard/Sharp/Sharper. Noise power spectrum (NPS) was computed to assess noise magnitude and noise texture (fav). Task‐based transfer function (TTF) was calculated to assess spatial resolution (f50) on blood‐mimicking material plus iodine inserts at 2 and 4 mg/mL. A detectability index (d’) was computed to model the detection task of two contrast‐enhanced lesion according to keV. The accuracy of the iodine concentrations measured from six iodine inserts was calculated on iodine maps for HIR and all SAIR levels.Results: For all keV, noise magnitude values were higher than with HIR using the Sharp (11.9 ± 5.6%; p > 0.05) and Sharper (39.3 ± 7.3%; p = 0.004) levels and lower than HIR for the other SAIR levels, decreasing from Standard (−15.3 ± 3.9%; p > 0.05) to Smooth (−42.3 ± 2.2%; p = 0.004) and Smoother levels (−70.5 ± 1.1%; p < 0.001). For all energy levels, fav values were higher with all SAIR levels than with HIR, except when using the Smoother level. For both inserts, differences in f50 values between the five SAIR levels and HIR did not exceed ± 0.05 mm−1. For all energy levels and both simulated lesions, d’ values were only higher than with HIR when using the Standard (32.4 ± 7.5%; p > 0.05), Smooth (95.6 ± 10.1%; p = 0.008) and Smoother (303.4 ± 16.6%; p < 0.001) levels. The accuracy of measured iodine values, calculated for each type of reconstruction, was similar (p > 0.05) and equal to 0.25 ± 0.03 mg/mL.Conclusion: Compared to the HIR algorithm used in clinical practice for low‐keV VMIs, Standard, Smooth and Smoother levels of SAIR reduced image noise and improved the detectability of contrast‐enhanced lesions. Better noise texture was found only with the Standard and Smooth levels compared to HIR.

The COMS (Collaborative Ocular Melanoma Study) eye plaques are based on the application of low-energy photon sources such as iodine-125 (125I) and palladium-103 (103Pd). Compared to 125I, 103Pd is characterized by different radiobiological and physical parameters, which include the half-life, relative biological effectiveness (RBE), and average photon energy. Variation of these parameters may increase or decrease the integrated cell-killing effect of eye plaques; therefore, quantitative evaluation of biological effects may be used to optimize the dose prescription and improve the therapeutic ratio. The goal of this article is to show that the higher RBE of lower-energy photons from 103Pd sources counteracts the steeper dose fall-off in heterogeneous eye plaque geometry and the effect of radiation protraction to produce the higher cell killing effect compared to 125I sources. Biological effectiveness of COMS eye plaques with 125I and 103Pd sources is evaluated using the equivalent uniform RBE-weighted dose (EUDRBE). The EUDRBE is defined as the uniform dose distribution with RBE=1 that produces the cell survival equal to a nonuniform dose distribution with variable RBE. The EUDRBE was proposed for comparison of cell survival in radiotherapy with a nonuniform dose and the RBE effect, and such a concept can be applied to both external beam fractionated radiotherapy and continuous irradiation in brachytherapy. The EUDRBE is computed in a 1D model of COMS eye plaques using the dose distributions in a water-equivalent medium with heterogeneity correction from MC calculations. The EUDRBE is computed in five dose fractions using the linear quadratic (LQ) cell survival model corrected for the effects of protracted irradiation and RBE. The EUDRBE for eye plaques is compared to the reference dose of 50 Gy(RBE) that is the lowest total dose (delivered in five fractions) used in hypofractionated proton therapy for ocular melanoma. In the simulations with a hypothetical RBE=1, the EUDRBE for eye plaques with both 125I and 103Pd sources is lower than the reference dose of 50 Gy(RBE) for most of the simulated implant durations and tumor heights. Also, the EUDRBE for eye plaques with 103Pd sources is lower compared to 125I sources. If the measured RBE of 1.4 and 1.9 is applied for 125I and 103Pd, respectively, the EUDRBE for eye plaques becomes larger than the reference dose of 50 Gy(RBE). The range of EUDRBE for 125I is within 50-60 Gy(RBE) depending on the implant duration and tumor height; therefore, the lowest EUDRBE is comparable to the dose used in hypofractionated proton therapy. The range of EUDRBE for 103Pd is within 57-67 Gy(RBE), which is higher compared to 125I. The effects of RBE for COMS eye plaques with 125I and 103Pd sources should be considered to explain the tumor control probability (TCP)>80% observed in clinical practice. The EUDRBE for 103Pd sources is larger than the EUDRBE for both the 125I sources and hypofractionated proton therapy; therefore, the dose de-escalation relative to 85 Gy may be considered.

Increasing evidence of the great normal tissue sparing effect in ultra-high dose-rate (UHDR) irradiation points to the potential for improving the therapeutic ratio. However, limited accumulation of data still prevents further understanding of the normal tissue sparing effect mechanism and its subsequent clinical translation. This study aimed to be the first to provide in vivo data for intestinal toxicity in UHDR carbon-ion irradiation. The abdominal part of eight week old C3H/HeSlc female mice was locally irradiated at an entrance region of 400-MeV/u monoenergetic carbon-ion beams either with a conventional dose rate (CONV) or UHDR. The CONV and UHDR irradiations had mean dose rates (MDRs) of ∼0.3Gy/s and ∼96Gy/s, respectively. Multiple dose levels (12-18Gy) were used under the dedicated dosimetry system for UHDR irradiation. Acute intestinal toxicity was evaluated by monitoring the survival of 96 irradiated mice for 30 days post-irradiation. To investigate the influence of dose rate on intestinal crypts, CONV, and UHDR irradiations were conducted for an additional six mice for each. These mice were autopsied 6 or 84h after irradiation along with two non-irradiated mice and their immunohistochemically stained intestines were used to count the apoptotic cells in crypts or regenerating crypts. Survival fractions of mice were comparable for both dose rates. At the sub-lethal dose, 14Gy, normal tissue complication probabilities (NTCPs) for the endpoint of survival were 0.78 and 0.89 for CONV and UHDR irradiations, respectively. The 13Gy dose resulted in NTCP of 0.11 for both dose rates. Doses leading to 50% NTCP were 13.6 and 13.5Gy for CONV and UHDR irradiations, respectively, which produced a dose modifying factor of 1.008. The percentages of apoptotic crypts at 6h post-irradiation were 36% for CONV and 34% for UHDR irradiations, showing a small difference (p=0.782). The percentages of remaining crypts at 84h post-irradiation were 7.9% for CONV and 5.8% for UHDR irradiations (p=0.214), indicating a slightly worse result for UHDR irradiation, albeit without statistical significance. The intestinal toxicity in UHDR carbon-ion irradiation was investigated for the first time. UHDR carbon-ion irradiation did not produce a substantial sparing effect for intestinal toxicity at least under the beam conditions tested here, i.e., a 400-MeV/u carbon-ion beam at an MDR of ∼96Gy/s. Additional experiments with different endpoints and beam conditions are mandatory to assess the potential of FLASH carbon-ion therapy.

BackgroundSecondary‐radiation background (BG) generated in kilovoltage X‐ray flat‐panel detectors (FPDs) under scanning proton beam irradiation has not been thoroughly analyzed.PurposeThis study aimed to determine and mathematically model the characteristics of secondary‐radiation BG in FPDs under scanning proton beam irradiation.MethodsUsing a synchrotron‐based proton system and two FPDs mounted on a gantry tilted at 135° to the proton beam axis, we acquired images of two phantoms (block and thorax). The FPD images were captured during three conditions: only X‐ray exposure, proton irradiation, and pauses between proton irradiations. Because the FPD readout rate (30 fps) was twice the X‐ray exposure rate (15 pps), the images were further categorized into two types: those corresponding to with and without X‐ray exposure. These FPD images were analyzed to determine the characteristics of the secondary‐radiation BG. In addition, a Monte Carlo simulation was conducted to complement the performed measurements.ResultsAnalysis of the FPD images without X‐ray exposure revealed that the pure secondary‐radiation BG appeared as sparse spike patterns without a noticeable spatial bias across the FPD area in most images. The BG affected only 1.25% of the FPD pixels and could be modeled as an exponentially decreasing function with increasing pixel intensity. When considering typical X‐ray image brightness for the block and thorax phantoms, the impact of the BG on mean‐pixel‐intensity variation was < 1%. Monte Carlo simulations suggested that prompt photons were the primary source reaching the FPD.ConclusionsThis study demonstrates that the secondary‐radiation BG characteristics observed in the X‐ray FPD images during scanning proton beam irradiation can be quantitatively modeled. The impact of the secondary‐radiation BG on mean‐pixel‐intensity of FPD images was minimal.

Breast tumor segmentation is crucial for the diagnosis of breast cancer, as it enables radiologists to rapidly identify areas of interest and facilitate subsequent analysis, diagnosis, and treatment. Present breast tumor segmentation methods are typically applied to high-resolution computed tomography images. However, fewer segmentation methods are utilized for positron emission tomography/computed tomography (PET/CT) imaging systems. Our goal is to develop a deep learning algorithm which combines functional and structural information for breast tumor segmentation in PET/CT images. This can enhance analytical accuracy and speed up the process of obtaining segmentation outcomes, thereby assisting physicians in subsequent patient diagnosis and treatment. In this study, we explore an automatic image segmentation model to segment breast tumors in PET images. The proposed CT-Guided Transformer modules utilize features of various scales from CT images to generate attention maps for PET features. During the fusion process, effective consensus information is extracted from the features of two different modalities using similarity-based contrastive learning, thus enhancing the segmentation performance. Five evaluation metrics (Jaccard coefficient, Dice score, precision, sensitivity, and Hausdorff distance) are utilized to evaluate segmentation performance. The proposed algorithm is compared to the single-modality method and other multimodal fusion strategies. Experiments are conducted using a collected clinical breast dataset alongside the public QIN-Breast benchmark, and the results show that the proposed algorithm accurately segments the outline of breast tumors, achieving superior performance (86.19% Dice, 75.73% Jaccard) on the primary dataset and outperforming standard cross-attention on the public dataset by 3.86% Dice and 3.64% Jaccard. The quantitative and visualization results confirm that our method outperformed single-modality input methods and fusion methods. Additionally, we evaluate the distribution of metrics collected by cases to further demonstrate the superiority of our approach. We present a deep-learning-based method for the joint segmentation of anatomical and functional PET/CT images. Compared to single-modality and dual-modality methods with various fusion strategies, our approach significantly improves the accuracy of breast tumor delineation, demonstrating great potential for breast tumor diagnosis.

The use of MRI-guided radiation therapy (MRIgRT) has shown considerable advantages. However, the acquisition of electron density information still relies on Computed tomography (CT) images. In the context of medical image synthesis, achieving high global and local accuracy, as well as efficient synthetic CT (sCT) generation, remains a key challenge. This study aims to improve the accuracy of sCT image generation from MRI by introducing a novel deep learning model that leverages global context and local detail, with a minimal increase in model complexity. In this study, we propose a novel generative adversarial network based on Mamba block and Residual Constraint (RC) strategy, namely RC-MambaGAN, which achieves a significant improvement in accuracy with only a minimal increase in computation time required for sCT generation. Specifically, we integrate Mamba block into the generator to enhance the model's ability to capture long-range contextual information, while preserving the precision of local features and maintaining the global structure of the image. The RC strategy improves the alignment between sCT and real CT by minimizing residuals. The model was validated on multi-center datasets that included both tumor and pelvic-region data. Furthermore, the dosimetric accuracy of the generated sCT images was evaluated to assess their clinical applicability. RC-MambaGAN achieved high performance with mean absolute error (MAE) of 37.094±8.761 HU, peak signal-to-noise ratio (PSNR) of 28.424±1.145dB, structural similarity index measure (SSIM) of 0.947±0.004, and Mutual Information (MI) of 1.426±0.012. External validation demonstrated that RC-MambaGAN outperformed existing state-of-the-art methods in both image quality and quantitative accuracy, even for structures visible in MRI but not in CT images. The proposed RC-MambaGAN markedly improves the image quality and anatomical accuracy of sCT images derived from MRI, thereby advancing the feasibility of MRI-only radiotherapy workflows.

The purpose of this study is to adapt a Bayesian dual-virtual non-contrast (VNC) method by integrating prior anatomical knowledge from AI-based multi-organ segmentation and to generalize it for spectral photon-counting CT (PCCT) with an arbitrary number of energychannels. A previously proposed Bayesian VNC method is reformulated for any number of energies and adapted for integration with AI segmentation. TotalSegmentator, an open-access whole-body AI segmentation model, is used to provide spatial priors. The method is applied to simulated contrast-enhanced dual-energy CT (DECT) and PCCT datasets from eight virtual patients, with and without AI segmentation. Key radiotherapy-relevant parameters such as electron density ( ) and proton stopping power ratio (SPR) are estimated and compared to ground truth values. Additional results are obtained for non-contrast scans by setting contrast agent uptake tozero. AI-based segmentation improved the accuracy of parameter estimation for both DECT and PCCT, with a more pronounced effect for PCCT. The combination of high spectral resolution and anatomical priors led to reduced RMS errors in SPR and . Mean absolute water-equivalent path length (WEPL) errors confirmed the superiority of segmentation-assisted PCCT over othermethods. This proof of concept demonstrates a flexible, AI-assisted Bayesian framework for extracting quantitative information from contrast-enhanced spectral CT. By integrating AI segmentation and generalizing to PCCT, the method shows improved tissue characterization, suggesting the value of AI in extracting quantitative information beyond DECT. Further validation on clinical datasets isneeded. Quantitative VNC methods offer the potential to extract radiotherapy-related parameters from contrast-enhanced spectral CT without the need for additional non-contrast imaging. However, the inherently ill-posed nature of tissue characterization from limited spectral data remains a major limitation, which requires advancedtechniques.

Generating three-dimensional (3D) ultrasound (US) data from conventional two-dimensional (2D) freehand acquisitions typically necessitates external tracking systems to ascertain probe positioning. However, these hardware-based solutions often present practical challenges, including substantial cost, increased setup complexity, and susceptibility to interference or line-of-sight issues, thereby limiting their widespread clinical integration. A tracker-free approach to 3D US reconstruction could dramatically improve the accessibility and applicability of volumetric ultrasound in standard medicalprocedures. This study presents a method for 3D US reconstruction from 2D B-mode image sequences that eliminates the need for external tracking. The goal is to improve reconstruction accuracy and usability for diagnostic, preoperative, and intraoperativeapplications. A hybrid Transformer-convolutional neural network (CNN) is trained end-to-end to regress inter-frame six degrees of freedom (6-DoF) poses. Global self-attention captures long-range probe motion, whereas convolutional layers refine local speckle patterns. Experiments used 53 forearm scans (Dataset 1) and 36 prostate scans (Dataset 2). Reconstruction performance was assessed with Dice Similarity Coefficient (DSC) and trajectory drift (Dataset 1) plus structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) (Dataset 2). In Dataset 1, the proposed model achieved a DSC of 0.71 0.19 and a drift of 13.25 9.17 , outperforming the 2D CNN (0.62 0.26, 18.92 11.20 ) and ConvLSTM (0.67 0.32, 18.36 5.44 ). In Dataset 2, it obtained an SSIM of 0.845 and a PSNR of 29.10 dB, exceeding the 2-D CNN (0.742, 26.40 dB) and ConvLSTM (0.812, 27.85 dB). All improvements were statistically significant as determined by paired t-tests (p 0.01). The tracker-free Transformer-CNN consistently improves volumetric overlap, trajectory stability, and image quality relative to established CNN- or recurrent neural network (RNN)-based schemes, demonstrating a practical, cost-effective route to high-fidelity 3D ultrasound across diverse clinical protocols.

Segmentation of cardiac structures is essential for cardiac function evaluation using cine magnetic resonance imaging (MRI). Deep learning can be used to segment cardiac structures in cine cardiac MRI with high accuracy, but this approach requires fully annotated datasets for training, which are difficult to obtain. Semi-supervised segmentation methods provide a way to alleviate the burden of manual labeling by using labeled and unlabeled data for training. However, these methods generally provide suboptimal segmentationaccuracies. To develop a semi-supervised method that utilizes relatively small training datasets and under-annotations for improved cine cardiac MRIsegmentation. The proposed approach consists of deformable registration, fully and weakly supervised segmentation, and a temporal attention perceiver (TAP). The registration module was trained to warp labeled frames to generate pseudo labels for unlabeled frames. The warped labeled images were used to train the fully supervised segmentation network. The unlabeled images and the pseudo label were used to train the weakly supervised segmentation model, and the segmentation prediction was compared with the input pseudo label as an auxiliary loss to the registration module. The TAP module was employed to generate optimized features for the warped labeled and the original unlabeled images both paired with the original labeled image. Consistency between the resulting features was enforced to refine cross-instance feature alignment to facilitate the registration. One hundred, twenty, and ten subjects from the Automatic Cardiac Diagnosis Challenge (ACDC) and seventy-five, thirty, and fifteen cases from the Multi-Vendor & Multi-Disease (M&Ms) Cardiac Image Segmentation Challenge were used for training, each with random end-systolic (ES)/end-diastolic (ED) frames labeled. The optimized models were used to segment the remaining 50 ACDC and 50 M&Ms subjects. The proposed approach was compared with several commonly used semi-supervised segmentation methods in terms of Dice-similarity-coefficients (DSC), average-symmetric-surface-distance (ASSD), and Hausdorff-distance (HD) for left (LV) and right (RV) ventricular cavity and myocardium (Myo). A Unet trained on the same subjects each with both frames labeled was used as an upper bound (Unet_UB). Using 100 ACDC training subjects, our approach yielded DSC = 0.910 0.063, ASSD = 1.37 0.63 mm, and HD = 6.38 2.99 mm for RV, DSC = 0.894 0.024, ASSD = 1.20 1.12 mm, and HD = 4.67 3.22 mm for Myo, and DSC = 0.934 0.056, ASSD = 1.25 1.63 mm, and HD = 3.97 5.76 mm for LV. A bidirectional copy-paste (BCP) method performed the best among the comparative methods and generated DSC = 0.902 0.060, ASSD = 1.45 0.60 mm, and HD = 7.50 3.20 mm for RV, DSC = 0.885 0.030, ASSD = 1.28 0.80 mm, and HD = 5.80 2.80 mm for Myo, and DSC = 0.920 0.068, ASSD = 1.15 0.40 mm, and HD = 4.20 3.30 mm for LV. For Unet_UB, these were 0.905 0.068, 1.48 0.61 mm, and 6.35 2.85 mm for RV, 0.895 0.030, 1.05 0.45 mm, and 4.40 3.09 mm for Myo, and 0.941 0.044, 1.02 0.34 mm, and 3.17 1.63 mm for LV. Similar trends were observed when using 75 M&Ms training subjects. For all the experiments, our approach outperformed BCP in general and yielded segmentation accuracies comparable to Unet_UB. The proposed approach outperformed several commonly used semi-supervised segmentation methods and yielded segmentation accuracies on par with fully supervised Unet using various relatively small datasets and under annotations fortraining.

BackgroundFLASH radiotherapy using ultra‐high dose rates (UHDR, > 40 Gy/s) demonstrates significant healthy‐tissue sparing while maintaining tumor‐control effectiveness. However, the underlying mechanisms remain unclear, with hypotheses suggesting that reductions in reactive oxygen species (ROS) yields could contribute to the FLASH effect. Direct experimental measurements of ROS dynamics under UHDR conditions are challenging, making Monte Carlo simulations valuable complementary tools for tracking individual water radiolysis species over time.PurposeTo enable a mechanistic investigation of ROS yield reductions under UHDR conditions, we developed a GPU‐based water radiolysis simulation platform capable of modeling key radiation chemistry processes.MethodsUsing our MPEXS2.1‐DNA framework with step‐by‐step molecular tracking, we simulated sequential 55 MeV proton irradiation in a 1 × 1 × 1 µm3 target volume within a 2 × 2 × 2 µm3 water phantom filled with neutral pH water at dose rates ranging from 0.02 Gy/s to 500 Gy/s, with a total absorbed dose of 10 Gy. Simulations were performed under oxygenated (pO2 = 25%, 239.4 µM) and deoxygenated (pO2 = 0%) conditions without additional scavengers. Radiation chemical yields (G values, species/100 eV) of hydroxyl radicals, hydrated electrons, and hydrogen peroxide were calculated at time points from 1 ms to 1000 s post‐irradiation. Statistical analysis was performed using 1000 independent sequential proton irradiation scenarios per dose rate condition, processed in parallel on a single GPU. Results were compared with published experimental data.ResultsThe calculated dose rate dependence of the G values of hydroxyl radicals showed agreement with experimental data, with relative G values decreasing monotonically from 1.0 at 0.02 Gy/s to approximately 0.12 at 500 Gy/s at 10 ms post‐irradiation. Our simulations revealed that intertrack chemical reactions between neighboring proton tracks occurred, leading to decreases in the G values of hydroxyl radicals through enhanced radical‐radical interactions. The G values of hydrated electrons remained constant (G ≈ 2.5 species/100 eV) across all dose rates under deoxygenated conditions, consistent with experimental observations. Oxygen consumption followed a depletion rate of 0.028%/Gy (0.27 µM/Gy), in agreement with experimental measurements, but was insufficient to cause significant depletion at typical UHDR doses (10–40 Gy). Additionally, the calculated G values of hydrogen peroxide increased by 20% with dose rate (contrary to measured decreases of 20%–40%), suggesting the presence of competitive reaction pathways not included in current models.ConclusionsWe developed a GPU‐based computational framework for water radiolysis simulations under UHDR conditions using MPEXS2.1‐DNA. Our step‐by‐step approach enabled spatially precise tracking of intertrack reactions among radiolysis species originating from different proton tracks, a capability not achievable with conventional simplified methods. Additionally, we demonstrated that large‐scale statistical analysis is computationally feasible under UHDR conditions through GPU acceleration. Our results successfully reproduced the experimental trends for the G values of hydroxyl radicals and hydrated electrons. Oxygen depletion rates also showed good agreement with experimental measurements. However, the discrepancy between simulated and measured G values of hydrogen peroxide indicates the need to incorporate additional competitive reactions, potentially including third‐order mechanisms, for future UHDR modeling studies.