Medical Physics

Background Safe implementation and translation of FLASH radiotherapy to the clinic requirehs development of beam monitoring devices capable of high temporal resolution with wide dynamic ranges. Ideal detectors should be able to monitor LINAC pulses, withstand high doses and dose rates, and provide information about the beam output, energy/range, and profile. Purpose Two novel detectors have been designed and tested for ultra‐high dose‐rate (UHDR) monitoring: a multilayer nano‐structured 3‐layer high‐energy‐current (HEC3) detector, and a segmented large area, 4‐section flat (S4) detector with the goal of exploring their properties for a future combined design. Methods A Novalis‐TX LINAC was converted to produce a 10 MeV electron‐FLASH beam. Pulses were monitored using both HEC3 and S4 detectors. The HEC3 detector structure consisted of three electrode layers separated by a nanoporous aerogel (Aero): Al(50 µm)–Aero(100 µm)–Ta(10 µm)–Aero(100 µm)–Al(50 µm). The S4 structure was comprised of three layers: Cu(100 nm)–air(1 mm)–Al(100 nm) with contact potential for charge collection. Both detectors are self‐powered as they do not require an external voltage bias for charge collection. The beam was also characterized using a photodiode, Gafchromic EBT‐XD Film, OSLDs, and an Advanced Markus Chamber. Results The electron‐FLASH beam displayed a Gaussian‐like profile with 15 cm FWHM at isocenter. Electron‐FLASH dose rates up to an average of 260 Gy/s were measured on the surface of a solid water phantom at isocenter with an instantaneous dose rate of 1.8 × 10⁵ Gy/s and a dose per pulse of up to 1 Gy/pulse. Both HEC3 and S4 detectors could record individual pulses for repetition rates of 360 Hz with a 4 µs pulse‐width. The HEC3 detector signal increased linearly with dose, MU, number of pulses, and dose rate up to 850 Gy/s with no loss of functionality at high doses or dose rates. The S4 detector showed linearity with MU and number of pulses at each of the four channels independently showing potential for spatial information and steering but lacked dose rate independence. Conclusions Two novel detectors, HEC3 and S4, successfully measured electron‐FLASH pulses and hence can be considered capable of electron‐FLASH beam monitoring in different capacities. HEC3 detector technology is suitable for monitoring high‐dose and UHDR beams with high temporal resolution required for pulse counting. We envision the combination of the HEC3 internal structure with the S4 piece‐wise design for real‐time monitoring of the temporal structure, spatial profiles, energy, and dosimetric properties of UHDR beams.

Background Transbronchial needle biopsy is crucial for diagnosing lung cancer, yet its efficacy depends on accurately localizing the target lesion and biopsy needle. Digital tomosynthesis (DTS) is considered a promising imaging modality for guiding bronchoscopy procedures due to its low radiation dose and small footprint relative to cone‐beam computed tomography (CBCT). However, the image quality of DTS is still not sufficient for an accurate guidance, mainly due to its limited‐angle acquisition. Purpose Preoperative computed tomography (CT) scans are often performed prior to bronchoscopy interventions for diagnosis or to plan the procedure. The CT images are of high quality and are characterized by a high spatial resolution compared to intraoperative DTS images. These patient‐specific prior CT images and the intraoperative DTS images share a fair amount of anatomical information. The main differences only stem from patient positioning and respiratory motion. When these differences are addressed properly, prior CT images augment intraoperative DTS image reconstruction with strong prior knowledge and potentially enhance DTS image quality. Methods We propose in this work a prior‐aided DTS image reconstruction technique leveraging prior CT images to improve DTS image quality. This technique is based on a recently published deformable CT‐to‐DTS image registration algorithm which is customized for bronchoscopy interventions. The main idea is to register a prior CT image to an intermediate DTS image reconstructed using the standard iterative algebraic reconstruction technique (ART), then to re‐reconstruct the DTS image using ART and the registered prior CT image as a first estimation. Results The proposed prior‐aided reconstruction method was tested on a physical phantom and six patient bronchoscopy datasets. Real DTS data acquired with a pseudo‐linear (PL) scan geometry and simulated DTS data generated according to a spherical ellipse (SE) scan geometry were considered. Results evaluated qualitatively by visual inspection and quantitatively by computing Pearson's correlation (PC) with respect to the reference CBCT images suggest significant improvements in image quality using the prior‐aided DTS reconstruction compared to the standard zero‐initialized ART reconstruction. PC coefficients of the six patient datasets were on average 0.64±0.13$0.64\pm 0.13$ and 0.55±0.13$0.55\pm 0.13$ using a zero‐initialized ART reconstruction with SE data and PL data, respectively, and 0.82±0.09$0.82\pm 0.09$ and 0.72±0.09$0.72\pm 0.09$ using the proposed prior‐aided reconstruction with SE data and PL data, respectively. Conclusions While the initial estimation in iterative reconstruction algorithms is often overlooked, we proved that initial estimation is of critical importance in DTS image reconstruction and we have demonstrated the profound advantages of integrating prior CT images in intraoperative DTS image reconstruction. CT‐augmented DTS offers a viable alternative to CBCT in guiding bronchoscopy interventions at a fraction of the radiation dose. Further clinical studies are needed to validate improved diagnostic yield.

Background Cone‐beam computed tomography (CBCT) is a crucial daily imaging modality in image‐guided and adaptive radiotherapy. However, the use of ionizing radiation in CBCT imaging increases the risk of secondary cancers, which is particularly concerning for pediatric patients. Deep learning super‐resolution has shown promising results in enhancing the resolution of photographic and medical images but has not yet been explored in the context of CBCT dose reduction. Purpose To facilitate CBCT imaging dose reduction, we propose using an enhanced super‐resolution generative adversarial network (ESRGAN) in both the projection and image domains to restore the image quality of low‐dose CBCT. Methods An extensive projection database, containing 2997 CBCT scans from head and neck cancer patients, was used to train two different ESRGAN models to generate super‐resolution CBCTs. One model operated in the projection domain, using pairs of simulated low‐resolution (low‐dose) and original high‐resolution (high‐dose) projections and yielded CBCTSRpro. The other model operated in the image domain, using pairs of axial slices from reconstructed low‐resolution and high‐resolution CBCTs (CBCTLR and CBCTHR) and resulted in CBCTSRimg. Super‐resolution CBCTs were evaluated in terms of image similarity (MAE, ME, PSNR, and SSIM), noise characteristics, spatial resolution, and registration accuracy, using the original CBCT as a reference. To test the perceptual difference between the original and super‐resolution CBCT, we performed a visual Turing test. Results Visually, both super‐resolution approaches in the projection and image domains improved the image quality of low‐dose CBCTs. This was confirmed by the visual Turing test, that showed low accuracy, sensitivity, and specificity, indicating almost no perceptual difference between CBCTHR and the super‐resolution CBCTs. CBCTSRimg (accuracy: 0.55, sensitivity: 0.59, specificity: 0.50) performed slightly better than CBCTSRpro (accuracy: 0.59, sensitivity: 0.61, specificity: 0.57). Image similarity metrics were affected by varying noise levels and did not reflect the visual improvements, with MAE/ME/PSNR/SSIM values of 110.4 HU/2.9 HU/40.4 dB/0.82 for CBCTLR, 136.6 HU/−0.4 HU/38.6 dB/0.77 for CBCTSRpro, and 128.2 HU/1.9 HU/39.0 dB/0.80 for CBCTSRimg. In terms of spatial resolution, quantified by calculating 10% levels of the task transfer function, both CBCTSRpro and CBCTSRimg outperformed CBCTLR and nearly matched the reference CBCTHR (CBCTLR: 0.66 lp/mm, CBCTSRpro: 0.88 lp/mm, CBCTSRimg: 0.95 lp/mm, CBCTHR: 1.01 lp/mm). Noise characteristics of CBCTSRimg and CBCTSRpro were comparable to the reference CBCTHR. Registration parameters showed negligible differences for all CBCTs (CBCTLR, CBCTSRpro, CBCTSRimg), with average absolute differences in registration parameters being below 0.4° for rotations and below 0.06 mm for translations (CBCTHR as reference). Conclusions This study demonstrates that deep learning can be a valuable tool for CBCT dose reduction in CBCT‐guided radiotherapy by acquiring low‐dose CBCTs and restoring the image quality using deep learning super‐resolution. The results suggest that higher quality images can be generated when super‐resolution is performed in the image domain compared to the projection domain.

Background Intensity‐modulated proton therapy (IMPT) planning often relies on physical dose constraints to balance tumor control and sparing of organs at risk (OARs). However, focusing solely on these dose objectives does not always minimize the normal‐tissue toxicity, which is quantified as normal tissue complication probability (NTCP). NTCP is also a quantitative criterion for patient selection between proton and photon treatments. Purpose This study introduces an NTCP‐based IMPT planning (NTCP‐IMPT) method designed to directly minimize normal‐tissue toxicity while maintaining tumor coverage. Methods NTCP‐IMPT simultaneously optimizes NTCP and dose‐volume histogram (DVH)‐based physical dose objectives while adhering to the minimum‐monitor‐unit (MMU) constraint for plan deliverability. The optimization problem is solved by the interior‐point method. To assess its efficacy in reducing normal‐tissue toxicity, NTCP‐IMPT is compared with standard IMPT (without NTCP optimization) for four head‐and‐neck (HN) cancer patients in terms of physical dose quality and NTCP of xerostomia and dysphagia. Results Across all four patients, NTCP‐IMPT plans met target dose criteria (D95% ≥ 100% and D2% ≤ 110%) while maintaining maximum doses to the spinal cord and brainstem comparable to standard IMPT. NTCP‐IMPT also reduced mean doses to parotid glands, submandibular glands, oral cavity, and pharyngeal constrictor muscles (PCMs). Compared to the standard IMPT, NTCP‐IMPT achieved average reductions in NTCP for xerostomia (grade ≥ 2: 3.67%; grade ≥3: 1.07%) and dysphagia (grade ≥ 2: 7.54%; grade ≥ 3: 3.72%). Conclusions NTCP‐IMPT effectively minimizes normal‐tissue toxicity and improves the sparing of OARs associated with side effects while maintaining comparable tumor coverage compared to standard IMPT.

Background In magnetic resonance image (MRI)‐guided radiotherapy (MRgRT), 2D rapid imaging is commonly used to track moving targets with high temporal frequency to minimize gating latency. However, anatomical motion is not constrained to 2D, and a portion of the target may be missed during treatment if 3D motion is not evaluated. While some MRgRT systems attempt to capture 3D motion by sequentially tracking motion in 2D orthogonal imaging planes, this approach assesses 3D motion via independent 2D measurements at alternating instances, lacking a simultaneous 3D motion assessment in both imaging planes. Purpose We hypothesized that a motion model could be derived from prior 2D orthogonal imaging to estimate 3D motion in both planes simultaneously. We present a manifold learning technique to estimate 3D motion from 2D orthogonal imaging. Methods Five healthy volunteers were scanned under an IRB‐approved protocol using a 3.0 T Siemens Skyra simulator. Images of the liver dome were acquired during free breathing (FB) with a 2.6 mm × 2.6 mm in‐plane resolution for approximately 10 min in alternating sagittal and coronal planes at ∼5 frames per second. The motion model was derived using a combined manifold learning and alignment approach based on locally linear embedding (LLE). The model utilized the spatially overlapping MRI signal shared by both imaging planes to group together images that had similar signals, enabling motion estimation in both planes simultaneously. The model's motion estimates were compared to the ground truth motion derived in each newly acquired image using deformable registration. A simulated target was defined on the dome of the liver and used to evaluate model performance. The Dice similarity coefficient and distance between the model‐tracked and image‐tracked contour centroids were evaluated. Motion modeling error was estimated in the orthogonal plane by back‐propagating the motion to the currently imaged plane and by interpolating the motion between image acquisitions where ground truth motion was available. Results The motion observed in the healthy volunteer studies ranged from 12.6 to 38.7 mm. On average, the model demonstrated sub‐millimeter precision and > 0.95 Dice coefficient compared to the ground truth motion observed in the currently imaged plane. The average Dice coefficient and centroid distance between the model‐tracked and ground truth target contours were 0.96 ± 0.03 and 0.26 mm ± 0.27 mm respectively across all volunteer studies. The out‐of‐plane centroid motion error was estimated to be 0.85 mm ± 1.07 mm and 1.26 mm ± 1.38 mm using the back‐propagation (BP) and interpolation error estimation methods. Conclusions The healthy volunteer studies indicate promising results using the proposed motion modeling technique. Out‐of‐plane modeling error was estimated to be higher but still demonstrated sub‐voxel motion accuracy.

Background Effective breast cancer treatment planning requires balancing tumor control while minimizing radiation exposure to healthy tissues. Choosing between intensity‐modulated radiation therapy (IMRT) and three‐dimensional conformal radiation therapy (3D‐CRT) remains pivotal, influenced by patient anatomy and dosimetric constraints. Purpose This study aims to develop a decision‐making framework utilizing deep learning to predict dose distributions, aiding in the selection of optimal treatment techniques. Methods A 2D U‐Net convolutional neural network (CNN) model was used to predict dose distribution maps and dose‐volume histogram (DVH) metrics for breast cancer patients undergoing IMRT and 3D‐CRT. The model was trained and fine‐tuned using retrospective datasets from two medical centers, accounting for variations in CT systems, dosimetric protocols, and clinical practices, over 346 patients. An additional 30 consecutive patients were selected for external validation, where both 3D‐CRT and IMRT plans were manually created. To show the potential of the approach, an independent medical physicist evaluated both dosimetric plans and selected the most appropriate one based on applicable clinical criteria. Confusion matrices were used to compare the decisions of the independent observer with the historical decision and the proposed decision‐making framework. Results Evaluation metrics, including dice similarity coefficients (DSC) and DVH analyses, demonstrated high concordance between predicted and clinical dose distribution for both IMRT and 3D‐CRT techniques, especially for organs at risk (OARs). The decision‐making framework demonstrated high accuracy (90%$\%$), recall (95.7%$\%$), and precision (91.7%$\%$) when compared to independent clinical evaluations, while the historical decision‐making had lower accuracy (50%$\%$), recall (47.8%$\%$), and precision (78.6%$\%$). Conclusions The proposed decision‐making model accurately predicts dose distributions for both 3D‐CRT and IMRT, ensuring reliable OAR dose estimation. This decision‐making framework significantly outperforms historical decision‐making, demonstrating higher accuracy, recall, and precision.

Background Computed tomography (CT) reconstruction problems are always framed as inverse problems, where the attenuation map of an imaged object is reconstructed from the sinogram measurement. In practice, these inverse problems are often ill‐posed, especially under few‐view and limited‐angle conditions, which makes accurate reconstruction challenging. Existing solutions use regularizations such as total variation to steer reconstruction algorithms to the most plausible result. However, most prevalent regularizations rely on the same priors, such as piecewise constant prior, hindering their ability to collaborate effectively and further boost reconstruction precision. Purpose This study aims to overcome the aforementioned challenge a prior previously limited to discrete tomography. This enables more accurate reconstructions when the proposed method is used in conjunction with most existing regularizations as they utilize different priors. The improvements will be demonstrated through experiments conducted under various conditions. Methods Inspired by the discrete algebraic reconstruction technique (DART) algorithm for discrete tomography, we find out that pixel grayscale values in CT images are not uniformly distributed and are actually highly clustered. Such discovery can be utilized as a powerful prior for CT reconstruction. In this paper, we leverage the collaborative filtering technique to enable the collaboration of the proposed prior and most existing regularizations, significantly enhancing the reconstruction accuracy. Results Our experiments show that the proposed method can work with most existing regularizations and significantly improve the reconstruction quality. Such improvement is most pronounced under limited‐angle and few‐view conditions. Furthermore, the proposed regularization also has the potential for further improvement and can be utilized in other image reconstruction areas. Conclusions We propose improving the performance of iterative CT reconstruction algorithms by applying the collaborative filtering technique along with a prior based on the densely clustered distribution of pixel grayscale values in CT images. Our experimental results indicate that the proposed methodology consistently enhances reconstruction accuracy when used in conjunction with most existing regularizations, particularly under few‐view and limited‐angle conditions.

Background Lung computed tomography (CT) scan image registration is being used for lung function analysis such as ventilation. Given the high sensitivity of functional analyses to image registration errors, an image registration error scoring tool that can measure submillimeter image registration errors is needed. Purpose To propose an image registration error scoring tool, termed λ, whose spatial sensitivity can be used to quantify image registration errors in steep image gradient regions under realistic noise conditions. Methods λ compares two images, termed reference and evaluated. The HU and distance scales of both images are normalized by user‐selected scaling criteria. For each voxel in the reference image, the 4D Euclidian distances between the reference voxel and the nearby evaluated voxels are calculated, and the minimum of these distances is λ$\lambda$. We tested λ$\lambda$ in simulated individual blood vessels comprised of 1, 3, and 5 mm diameter cylinders in 1 × 1 × 1 mm³ voxel images, which were blurred to simulate CT scanner intrinsic resolution and volume averaging. We placed the simulated vessels in a homogeneous background simulating parenchymal tissue density and injected 20, 40, and 60 HU standard deviation Gaussian noise. We used isotropic Gaussian filters with 0.5, 1.0, and 1.5 mm standard deviation kernels to smooth the simulated images. We assessed λ$\lambda$ using reference‐evaluated vessel shifts of −1.0 to 1.0 mm in 0.05 mm steps via rigid translational and rotational deformations. We examined whether λ$\lambda$ tracked the translation vector via its internal spatial component. We restricted λ$\lambda$ to voxels using the angle, termed θ$\theta$, between the λ$\lambda$ vector and the normalized spatial‐distance axes, terming the results the restricted‐λ$\lambda$, λR${{\lambda }_R}$, where θ$\theta$ was hypothesized to be a proxy for image gradient. We determined whether θ$\theta$ was coincident with the image gradient by examining if the voxels with |θ|≤30∘$| \theta | \le {{30}^ \circ }$ tracked the evaluated vessels. We used the 95th percentile of λR${{\lambda }_R}$, λR95$\lambda _R^{95}$, to determine spatial sensitivity, which we took as a conservative estimate of registration error, by fitting λR95$\lambda _R^{95}$ to a modified absolute‐value function for each tested rigid translation, noise level, smoothing kernel, and vessel radius combination. We demonstrated the use of λ$\lambda$ on a clinical example consisting of a set of 25 deformably registered free‐breathing thoracic CT scans. We visually compared the λ$\lambda$ and λR${{\lambda }_R}$ results against the HU differences between each clinical image pair. Results We found θ to be coincident with the image gradient. We found that λ$\lambda$’s spatial component tracked the vessel shifts. We determined the spatial sensitivity limit of λR95$\lambda _R^{95}$ to be < 0.2 mm. The noise level and smoothing kernel influenced λR95$\lambda _R^{95}$ sensitivity, worsening with increasing noise, and improving with increasing smoothing. For the clinical images, we observed λ$\lambda$ to qualitatively match the absolute difference of intensity in the image pairs and λR${{\lambda }_R}$ to restrict itself to high gradient regions or regions of visually apparent errors. Conclusion λR95$\lambda _R^{95}$ detected sub‐millimeter positioning errors between simulated vessels in the presence of typical CT noise. The noise magnitude and choice of noise smoothing kernel were inversely related to λR95$\lambda _R^{95}$ sensitivity, implying that study‐specific tuning of the pre‐smoothing kernel may be required. The demonstrated ability in geometric tests of λR95$\lambda _R^{95}$ to detect subvoxel DIR errors warrants further evaluation and testing.

Background Accurate musculoseletal ultrasound (MSKUS) image segmentation is crucial for diagnosis and treatment planning. Compared with traditional segmentation methods, deploying deep learning segmentation methods that balance segmentation efficiency, accuracy, and model size on edge devices has greater advantages. Purpose This paper aims to design a MSKUS image segmentation method that has fewer parameters, lower computation complexity and higher segmentation accuracy. Methods In this study, an attention mechanism‐based lightweight UNet (AML‐UNet) is designed to segment target muscle regions in MSKUS images. To suppress the transmission of redundant feature, Channel Reconstruction and Spatial Attention Module is designed in the encoding path. In addition, considering the inherent characteristic of MSKUS image, Multiscale Aggregation Module is developed to replace the skip connection architecture of U‐Net. Deep supervision is also introduced to the decoding path to refine predicted masks gradually. Our method is evaluated on two MSKUS 2D‐image segmentation datasets, including 3917 MSKUS and 1534 images respectively. In the experiments, a five‐fold cross‐validation method is adopted in ablation experiments and comparison experiments. In addition, Wilcoxon Signed‐Rank Test and Bonferroni correction are employed to validate the significance level. 0.01 was used as the statistical significance level in our paper. Results AML‐UNet yielded a mIoU of 84.17% and 90.14% on two datasets, representing a 3.38% (p<0.001$p<0.001$) and 3.48% (p<0.001$p<0.001$) over the Unext model. The number of parameters and FLOPs are only 0.21M and 0.96G, which are 1/34 and 1/29 of those in comparison with UNet. Conclusions Our proposed model achieved superior results with fewer parameters while maintaining segmentation efficiency and accuracy compared to other methods.

Background Electron therapy, vital for treating skin lesions and superficial tumors, demands precise electron cutouts to ensure accurate dose delivery. Traditional manual tracing methods introduce uncertainties and inefficiencies, necessitating innovative solutions for custom block creation. Purpose To introduce and validate the Optical Automatic Contour Tracing (O‐ACT) method, enhancing electron cutout generation's accuracy and efficiency through optical imaging and software automation. Methods Utilizing a 3D‐printed holder, a centrally‐mounted charge‐coupled device (CCD) camera on the electron applicator secured two distinct perspective images of the skin's contour at varying heights. Through image binarization and skeletonization, we identified the clinical target contour's region‐of‐interests (ROIs) in each setup image. Employing distances from each ROI's center, assessed at 5‐degree intervals from both images, we reconstructed the target contour on the skin. The magnification factor, set at a 95 cm source‐to‐point distance, determined the final cutout shape. We crafted an in‐house software in MATLAB for camera calibration and image processing and juxtaposed our results against the standard clinical cutout from the treatment planning system (TPS) using a correlation coefficient based on masked binary images' mutual information. Additionally, we performed dosimetric evaluations using abstract shapes to compare O‐ACT with other methods. Results Our methodology yielded cutout shapes exhibiting remarkable alignment with the TPS clinical cutout. O‐ACT demonstrated superior precision in generating cutout shapes that closely align with the contours in TPS, improving upon other methods in terms of adaptability to patient body shapes and contour accuracy. Dosimetric evaluations showed minimal differences between methods, with O‐ACT providing slightly more consistent results. Dose profile analyses in penumbra regions indicated O‐ACT's improved accuracy compared to conventional methods. Conclusions Pushing the boundaries of traditional practices, our O‐ACT offers a more accurate, efficient, and reproducible method for custom electron cutout creation from clinical setup images. This innovation promises not only to streamline clinical workflow but also to potentially uplift clinical outcomes in radiation oncology by providing more accurate patient‐specific treatment accessary.

Background Cardiac radioablation (CRA) is a new and promising treatment modality for patients with ventricular tachycardia refractory to standard‐of‐care treatment. Electroanatomic maps are used to define radiation target volumes; however, there is currently no native method to import electroanatomic maps into the treatment planning system (TPS). Purpose To develop Edico, a semi‐automated tool to enable electroanatomic map import into a TPS, by converting electroanatomic maps to a Digital Imaging and COmmunications in Medicine (DICOM) standard. The overall aim is to facilitate target volume delineation and improve workflow efficiency in treating patients. Methods Edico imports voltage and spatial data from electroanatomic maps and sorts these into voxels to be exported in a DICOM format, with each voxel containing the average voltage value of the data that falls within it. Three different rectangular electroanatomic maps were created and processed using Edico to ensure that expected features are maintained through processing. A sensitivity analysis of voxel size was completed using 19 different electroanatomic maps processed at five different sets of voxel dimensions, for a total of 95 resulting voxelized datasets. The coefficient of variation in each populated voxel in the datasets was analyzed to determine which voxel sizes are necessary to ensure that data loss is kept to a minimum throughout processing, despite averaging. Five electroanatomic maps were used to re‐contour clinical target volumes and planning target volumes for previously‐treated patients with their electroanatomic maps now directly registered to their planning computed tomography (CT) scans. Results All three rectangular test electroanatomic maps were processed as expected. All tested voxel sizes resulted in low coefficients of variation overall, with the exception of the largest voxel size of 1.8 × 1.8 × 8 mm. When using Edico, a user should choose voxel dimensions similar to or smaller than those of a planning CT. Of five pairs of clinical and planning target volumes from previously treated patients, adjustments were made to four (80%), retrospectively, using the electroanatomic maps generated using Edico, registered to the patients’ planning CTs. Conclusions Edico provides a reliable solution for electroanatomic map import into a TPS and facilitates clinical and planning target volume identification in CRA.

Background Image guided radiotherapy (IGRT) with cone‐beam computed tomography (CBCT) is limited by the sub‐optimal soft‐tissue contrast and spatial resolution of energy‐integrating flat panel detectors (FPDs) which produce quasi‐quantitative CT numbers. Spectral CT with high resolution photon‐counting detectors (PCDs) could improve tumor delineation by enhancing the soft‐tissue contrast, spatial resolution, dose‐efficiency, and CT number accuracy. Purpose This study presents the first linac‐mounted PCD. On the journey to developing spectral cone‐beam CT for IGRT, the planar image quality of a linac‐mounted PCD is first fundamentally characterized and compared to an FPD in terms of the 2D spatial resolution, noise, and contrast. Methods A Medipix3RX‐based PCD was mounted to the kV FPD of an x‐ray volume imaging (XVI) system on an Elekta linac and the PCD acquisition was synchronized with the pulsed kV source. The energy calibration of the Medipix3RX was determined with various radioisotope gamma emissions up to 60 keV. To compare the 2D spatial resolution and noise between the PCD and FPD, the pre‐sampling modulation transfer function (MTF) and normalized noise power spectrum (NPS) were measured using an RQA5 spectrum and a fluoroscopy phantom was imaged to determine the limiting resolution of line pairs. Spectral planar images of phantom inserts containing two different concentrations of calcium (60 and 240 mg/cc) and iodine (5 and 15 mg/cc) were optimally energy weighted to maximize the contrast using tube voltages of 60, 80, 100, and 120 kV. To account for drifts in the sensor temperature, the PCD was dynamically translated in and out of the insert shadow during acquisitions to obtain flat field corrections per frame. The raw contrast of the resultant planar images was compared to the energy‐integrating FPD. Results The energy calibration of the Medipix3RX was observed to be linear up to 60 keV. The limiting resolution observed on the fluoroscopy phantom was 2 lp/mm for the FPD and 5 lp/mm for the PCD. The pre‐sampling MTF was higher across all frequencies comparing the PCD to the FPD. The normalized NPS of the PCD did not vary with frequency, whereas the spectrum for the FPD decreased monotonically and was lower than the PCD noise power across most of the spatial frequency range studied due to optical light spreading. Optimal energy weights were applied to the dynamically acquired PCD images and the raw contrast of the 60 mg/cc calcium insert increased by factors of 1.12±0.09$1.12\pm 0.09$ and 1.52±0.22$1.52\pm 0.22$ at 60 and 120 kV respectively compared to the FPD. Conclusions A Medipix3RX‐based PCD was successfully integrated with the kilovoltage imaging system on an Elekta linac. The initial planar image quality characterization indicated improvements in the MTF and energy‐weighted contrast compared to the FPD. Future work will focus on obtaining linac‐mounted spectral CBCT images with a translate‐rotate geometry, however this initial study indicates that variations in the PCD sensor response during acquisitions must be addressed to realise the full potential of linac‐mounted spectral CBCT.

Background The multi‐leaf collimator (MLC) is an advanced device utilized for beam shaping and intensity modulation in radiotherapy. With the framework of the contemporary single‐layer MLC featuring a rounded leaf tip, the leaf tip transmission and leakage exert a considerable influence on radiotherapy. Purpose To scale down the leakage and transmission from the leaf gap when the opposite leaves are closed. Methods This study proposes a new design of MLC, named dynamic leaf machine (DLM), specifically engineered to diminish the transmission and leakage from the MLC leaf tip. The DLM incorporates an innovative leaf configuration that involves a combination of fixed and movable layers in a single MLC leaf named fixed and movable (FM) layers leaf, which is advantageous in dealing with transmission and leakage from the leaf tip by employing a staggered arrangement of the fixed and movable layer when the opposite leaves are closed. Subsequently, the DLM is assembled on the Elekta VERSA HD accelerator to assess its performance, focusing on aspects of mechanical characters, transmission, leakage, and penumbra. Results In comparison to conventional MLC leaves, the FM leaf design in the DLM has achieved a remarkable reduction in the leakage between opposed closed leaves, decreasing it from 48.41% to 0.41%. Additionally, the transmission between the adjacent leaves has been measured at 0.59%, and the penumbra in 16 mm × 32 mm field is 2.82 mm, aligning with the performance of conventional MLC. Conclusions The DLM with FM leaf performs a significant reduction in transmission from the opposite leaf tip in comparison to conventional MLC while maintaining minimal penumbra, effectively mitigating the transmission and leakage problem between opposing leaves, thereby enhancing the effect of radiotherapy.

Background An organ‐specific Positron emission tomography (PET) scanner can achieve the same sensitivity with much fewer detectors as compared to a whole‐body PET scanner, thereby substantially reducing the system cost. It can also achieve much better spatial resolution as compared to a whole‐body PET scanner since the photon noncollinearity effect is reduced by using smaller detector ring diameter. Consequently, the development of organ‐specific PET scanners with high spatial resolution, high sensitivity, and low cost has been a major focus of international research in PET instrument development for many years. Purpose The focus of this work is to develop high‐resolution depth encoding PET detectors with high timing resolution and a reduced number of signal processing electronic channels. Consequently, PET scanners tailored for specific organs can be developed with high spatial and timing resolutions, enhanced sensitivity, and affordable cost. Methods An 8 × 8 silicon photomultiplier (SiPM) array with a pixel size of 3 × 3 mm² and a multiplexed signal readout circuit is coupled to one end of the lutetium yttrium orthosilicate (LYSO) array with a glass light guide between them to achieve a good crystal identification of small crystals by using only four position‐encoding energy signals. A 4 × 4 SiPM array with a pixel size of 6 × 6 mm² and an individual readout circuit is coupled to the other end of the crystal array without a light guide to achieve a good coincidence timing resolution (CTR). The depth of interaction (DOI) of the detector is measured by ratio of the energies measured by the two SiPM arrays and can be used to correct the depth dependency of the timing. The flood histograms, energy resolutions (ERs), DOI resolutions, and CTRs of two detectors utilizing LYSO arrays with different crystal sizes were measured with each of the two SiPM arrays alternately placed at the front of the detectors. Results A better flood histogram was obtained by placing the 8 × 8 SiPM array in front of the detector. The depth dependency of timing was larger when the 4 × 4 SiPM array was placed at the front of the detector. A better CTR was obtained by placing the 4 × 4 SiPM array at the back of the detector as compared to placing it at the front of the detector if the depth‐dependent timing correction was not performed. If the depth‐dependent timing correction was performed, a better CTR can be obtained by placing the 4 × 4 SiPM array at the front of the detector. The first detector using a 12 × 12 LYSO crystal array with a crystal size of 1.95 × 1.95 × 20 mm³ provided a flood histogram with all crystals clearly resolved, an ER of 11.7%, a DOI resolution of 2.9 mm, and a CTR of 275 ps with the depth‐dependent timing correction. The second detector using a 23 × 23 LYSO crystal array with a crystal size of 0.95 × 0.95 × 20 mm³ provided a flood histogram with all but the edge crystals clearly resolved, an ER of 17.6%, a DOI resolution of 2.3 mm, and a CTR of 293 ps with the depth‐dependent timing correction. Conclusions PET detectors with small crystal cross‐sectional sizes, good DOI and timing resolutions and a reduced number of electronics channels were developed. The detectors can be used to develop high performance organ‐specific PET scanners.

Background Treatment planning in radiation therapy (RT) is performed on image sets acquired with commercial x‐ray computed tomography (CT) scanners. Considering an increased frequency of verification scans for adaptive RT and the advent of alternatives to x‐ray CTs, there is a need to review the requirements for image sets used in RT planning. Purpose This study aims to derive the required image quality (IQ) for the computation of the dose distribution in proton therapy (PT) regarding spatial resolution and the combination of spatial resolution and noise. The knowledge gained is used to explore the potential for dose reduction in tomography‐guided PT. Methods Mathematical considerations indicate that the required spatial resolution for dose computation is on the scale of the set‐up margins fed into the robust optimization. This hypothesis was tested by processing retrospectively 12 clinical PT cases, which reflect a variety of tumor localizations. Image sets were low‐pass filtered and were made noisy in a generic manner. Dose distributions on the modified CT scans were computed with a Monte‐Carlo dose engine. The similarity of these dose distributions with clinical ones was quantified with the gamma‐index (1 mm/1%). The potential reduction of the x‐ray exposure compared to the planning CT scan was estimated. Results Dose distributions within the irradiated volume were robust against low‐pass filtering of the CTs with kernels up to a full‐width‐at‐half‐maximum of 4 mm, that is, the gamma pass rate (1 mm/1%) was ≥$\ge$98%. The limit of the filter width was 6 mm for brain tumors and 8 mm for targets in the abdomen. These pass rates remained approximately unchanged if a limited amount of noise was added to the CT image sets. The estimated potential reductions of the x‐ray exposure were at least a factor of 20. Conclusions The requirements on IQ in terms of spatial resolution in combination with noise for computing the dose in PT are clearly lower than the IQ of current clinical planning. The results apply, for example, to ultra‐low dose x‐ray CTs, proton CTs with coarse spatial detection, and attenuation images from the joint reconstruction of time‐of‐flight PET scans.

Background The clinical translation of proton minibeam radiation therapy (pMBRT) presents significant challenges, particularly in developing an optimal treatment planning technique. A uniform target dose is crucial for maximizing anti‐tumor efficacy and facilitating the clinical acceptance of pMBRT. However, achieving a high peak‐to‐valley dose ratio (PVDR) in organs‐at‐risk (OAR) is essential for sparing normal tissue. This balance becomes particularly difficult when OARs are located distal to the beam entrance or require patient‐specific collimators. Purpose This work proposes a novel pMBRT treatment planning method that can achieve high PVDR at OAR and uniform dose at target simultaneously, via multi‐collimator pMBRT (MC‐pMBRT) treatment planning method with joint dose and PVDR optimization (JDPO). Methods MC‐pMBRT utilizes a set of generic and premade multi‐slit collimators with different center‐to‐center distances and does not need patient‐specific collimators. The collimator selection per field is OAR‐specific and tailored to maximize PVDR in OARs while preserving target dose uniformity. Then, the inverse optimization method JDPO is utilized to jointly optimize target dose uniformity, PVDR, and other dose‐volume‐histogram based dose objectives, which is solved by iterative convex relaxation optimization algorithm and alternating direction method of multipliers. Results The need and efficacy of MC‐pMBRT is demonstrated by comparing the single‐collimator (SC) approach with the multi‐collimator (MC) approach. While SC degraded either PVDR for OAR or dose uniformity for the target, MC provided a good balance of PVDR and target dose uniformity. The proposed JDPO method is validated in comparison with the dose‐only optimization (DO) method for MC‐pMBRT, in reference to the conventional (CONV) proton RT (no pMBRT). Compared to CONV, MC‐pMBRT (DO and JDPO) preserved target dose uniformity and plan quality, while providing unique PVDR in OAR. Compared to DO, JDPO further improved PVDR via PVDR optimization during treatment planning. Conclusion A novel pMBRT treatment planning method called MC‐pMBRT is proposed that utilizes a set of generic and premade collimators with joint dose and PVDR optimization algorithm to optimize OAR‐specific PVDR and target dose uniformity simultaneously.

Background The electronic portal imaging device (EPID) can be used in vivo, to detect on‐treatment errors by evaluating radiation exiting a patient. To detect deviations from the planning intent, image predictions need to be modeled based on the patient's anatomy and plan information. To date in vivo transit images have been predicted using Monte Carlo (MC) algorithms. A deep learning approach can make predictions faster than MC and only requires patient information for training. Purpose To test the feasibility and reliability of creating a deep‐learning model with patient data for predicting in vivo EPID images for IMRT treatments. Methods In our approach, the in vivo EPID image was separated into contributions from primary and scattered photons. A primary photon attenuation function was determined by measuring attenuation factors for various thicknesses of solid water. The scatter component of in vivo EPID images was estimated using a convolutional neural network (CNN). The CNN input was a 3‐channel image comprised of the non‐transit EPID image and ray tracing projections through a pretreatment CBCT. The predicted scatter component was added to the primary attenuation component to give the full predicted in vivo EPID image. We acquired 193 IMRT fields/images from 93 patients treated on the Varian Halcyon. Model training:validation:test dataset ratios were 133:20:40 images. Additional patient plans were delivered to anthropomorphic phantoms, yielding 75 images for further validation. We assessed model accuracy by comparing model‐calculated and measured in vivo images with a gamma comparison. Results Comparing the model‐calculated and measured in vivo images gives a mean gamma pass rate for the training:validation:test datasets of 95.4%:94.1%:92.9% for 3%/3 mm and 98.4%:98.4%:96.8% for 5%/3 mm. For images delivered to phantom data sets the average gamma pass rate was 96.4% (3%/3 mm criteria). In all data sets, the lower passing rates of some images were due to CBCT artifacts and patient motion that occurred between the time of CBCT and treatment. Conclusions The developed deep‐learning‐based model can generate in vivo EPID images with a mean gamma pass rate greater than 92% (3%/3 mm criteria). This approach provides an alternative to MC prediction algorithms. Image predictions can be made in 30 ms on a standard GPU. In future work, image predictions from this model can be used to detect in vivo treatment errors and on‐treatment changes in patient anatomy, providing an additional layer of patient‐specific quality assurance.

Background Nontuberculous mycobacterial lung disease (NTM‐LD) and Mycobacterium tuberculosis lung disease (MTB‐LD) are difficult to distinguish based on conventional imaging examinations. In recent years, radiomics has been used to discriminate them. However, existing radiomic methods mainly focus on specific lesion types, and have limitations in handling the presence of multiple lesion types that vary among different patients. Purpose We aimed to establish a radiomic model based on multiple lesion types in the patient's CT scans, and analyzed the importance of different lesion types in distinguishing the two diseases. Methods 120 NTM‐LD and 120 MTB‐LD patients were retrospectively enrolled in this study and randomly split into the training (168) and testing (72) sets. A total of 1037 radiomic features were extracted separately for each lesion type. The univariate analysis, least absolute shrinkage, and selection operator were used to select the significant radiomic features. The radiomic signature score (Radscore) from each lesion type was estimated and aggregated to construct the multi‐lesion feature vector for each patient. A multi‐lesion radiomic (MLR) model was then established using the random forest classifier, which can estimate importance coefficients for different lesion types. The performances of the MLR model and single radomic models were investigated by the receiver operating characteristic curve (ROC). The impact of the predicted lesion importance was also evaluated in subjective imaging diagnosis. Results The MLR model achieved an area under the curve (AUC) of 90.2% (95% CI: 86.2% 94.1%) in differentiating NTM‐LD and MTB‐LD, outperforming the models using specific lesion types following existing radiomic models by 1% to 13%. Among different lesion types, tree‐in‐bud pattern demonstrated the highest distinguishing value, followed by consolidation, nodules, and lymph node enlargement. Given the estimated lesion importance, two senior radiologists exhibited improved accuracy in diagnosis, with an increased accuracy of 8.33% and 8.34%, respectively. Conclusions This is the first radiomic study to use multiple lesion types to distinguish NTM‐LD and MTB‐LD. The developed MLR model performed well in differentiating the two diseases, and the lesion types with high importance exhibited the potential to assist experienced radiologists in clinical decision‐making.

Background Recent advances in deep learning have sparked new research interests in dynamic magnetic resonance imaging (MRI) reconstruction. However, existing deep learning‐based approaches suffer from insufficient reconstruction efficiency and accuracy due to the lack of time correlation modeling during the reconstruction procedure. Purpose Inappropriate tensor processing steps and deep learning models may lead to not only a lack of modeling in the time dimension but also an increase in the overall size of the network. Therefore, this study aims to find suitable tensor processing methods and deep learning models to achieve better reconstruction results and a smaller network size. Methods We propose a novel unrolling network method that enhances the reconstruction quality and reduces the parameter redundancy by introducing time correlation modeling into MRI reconstruction with low‐rank core matrix and convolutional long short‐term memory (ConvLSTM) unit. Results We conduct extensive experiments on AMRG Cardiac MRI dataset to evaluate our proposed approach. The results demonstrate that compared to other state‐of‐the‐art approaches, our approach achieves higher peak signal‐to‐noise ratios and structural similarity indices at different accelerator factors with significantly fewer parameters. Conclusions The improved reconstruction performance demonstrates that our proposed time correlation modeling is simple and effective for accelerating MRI reconstruction. We hope our approach can serve as a reference for future research in dynamic MRI reconstruction.

Background Pre‐clinical studies demonstrate that delivering a high dose at a high dose rate result in less toxicity while maintaining tumor control, known as the FLASH effect. In proton therapy, clinical trials have started using 250 MeV transmission beams and more trials are foreseen. A novel aspect of FLASH treatments, compared to conventional radiotherapy, is the importance of dose rate next to dose and geometry. Therefore, to ensure the safety and quality of FLASH treatments, patient‐specific dose‐rate verification before treatment is an important additional prerequisite. Various definitions of dose rate have been reported, however, the scanning proton beam (PBS) dose rate definition of Folkerts 2020 is currently the most used. It is the ratio between delta dose (ΔD) and delta time (Δt), subject to a predefined threshold, for a given position. Gafchromic film is a widely available detector used to perform relative and absolute integrated dose measurements. Since the response time of film is in the order of micro seconds it could also be suitable for pre‐treatment verification of FLASH proton therapy. Purpose Development of a patient‐specific pre‐treatment verification procedure for FLASH proton therapy based on time resolved film dosimetry. The detector design is presented and validated using three tests. Methods A dedicated setup was built that holds a Gafchromic film and a high‐speed camera to record the film during the irradiations. The red color channel of the camera's readings was converted into optical density (OD) and an OD‐to‐dose calibration curve was applied to determine the relative dose accumulation over time. To undo the film measurement (film response) of the post‐irradiation coloration process, it is assumed that each dose deposit (pulse) results in a similar film response function. The convolution of the film response function over the pulse provides the film response. First the film response function was obtained by fitting this parameter onto a known film response and corresponding pulse. Post‐irradiation coloration correction was performed by deconvoluting all film measurement by the obtained film response function. From the integral of each measured pulse, the Δt was obtained. Several validation tests were conducted: compare the Δt film measurement to a reference detector, exclude that revisiting spots result in an unwanted artefact on the dose accumulation measurement and thereby Δt, and compare Δt distributions of film measurement and simulation (local gamma evaluation, criteria 10%/2 mm) for nine QA fields (dose values; 12, 15, and 20 Gy, and, nozzle currents; 25, 120, and 215 nA). A similar analysis was performed for three dose optimized treatment beams, with and without scan patterns optimized on local dose rate. Results Good agreement was found for Δt comparing film to the reference detector, but for Δt values smaller than ∼20 ms the error becomes larger (≥15%). Dose accumulation measured with film over time from a single spot is independent of whether the dose is delivered at once, twice or thrice. All gamma evaluations resulted in a gamma pass rate of ≥90%. Conclusions Time resolved film dosimetry to perform patient‐specific pre‐treatment verification in FLASH proton therapy is feasible.

Background 124‐iodine (¹²⁴I) is used for positron emission tomography (PET) diagnostics and therapy planning in patients with differentiated thyroid cancer (DTC). Small lesion sizes (<10 mm) and low ¹²⁴I uptake are challenging conditions for the detection of DTC lymph node lesions. Purpose The aim of this study was to systematically investigate the lesion detectability and quantification performance under clinically challenging imaging conditions using non‐time‐of‐flight (TOF) PET/magnetic resonance imaging (MRI) in the clinical context of radionuclide therapy planning of DTC patients. Methods PET/MR measurements were performed on the Siemens Biograph mMR using a small lesion NEMA‐like phantom (six glass spheres, diameters 3.7–9.7 mm). 60 min list‐mode data were acquired for nine activity concentrations (AC) ranging from 25 kBq/mL to 0.25 kBq/mL using a sphere‐to‐background ratio of 20:1. PET list‐mode data were divided into five timeframes (60, 30, 16, 8, and 4 min) and reconstructed using either ordered‐subsets expectation maximization (OSEM) or OSEM+ point spread function (PSF) algorithm. For all reconstructions, the smallest detectable sphere size was investigated in a human observer study. Partial volume effect (PVE) corrected PET images (contour and oversize‐based approach) were analyzed considering a ± 30% deviation range between imaged and true AC as acceptable. Clinical data of eight DTC patients with small lymph node lesions were evaluated to assess agreement between the PVE correction approaches. Results Longer PET acquisition times, higher ACs, and PSF reconstructions resulted in improved PET image quality and overall improved lesion detectability. The smallest 3.7 mm sphere was only visible under the best imaging conditions. Using a typical clinical ¹²⁴I whole‐body PET/MRI protocol with an acquisition time of 8 min using OSEM reconstructions, all lesions of ≥ 6.5 mm in diameter could be detected and the quantification provided reliable results approximately above 5.0 kBq/mL. An accurate quantification of ACs in the 4.8 mm sphere was not feasible in this study. In the clinical evaluation of 10 lesions, a good agreement between oversize‐ and contour‐based PVE corrections was observed (<15% deviation). Conclusions The results showed that a reliable quantification of ¹²⁴I uptake with PET/MRI is feasible and, therefore, could be used to perform radioiodine pre‐therapy lesion dosimetry and individualized therapy planning in DTC patients.

Background Automatic segmentation of meibomian glands in near‐infrared meibography images is basis of morphological parameter analysis, which plays a crucial role in facilitating the diagnosis of meibomian gland dysfunction (MGD). The special strip shape and the adhesion between glands make the automatic segmentation of meibomian glands very challenging. Purpose A strip and boundary detection multi‐task learning network (SBD‐MTLNet) based on encoder‐decoder structure is proposed to realize the automatic segmentation of meibomian glands. Methods A strip mixed attention module (SMAM) is proposed to enhance the network's ability to recognize the strip shape of glands. To alleviate the problem of adhesion between glands, a boundary detection auxiliary network (BDA‐Net) is proposed, which introduces boundary features to assist gland segmentation. A self‐adaptive interactive information fusion module (SIIFM) based on reverse attention mechanism is proposed to realize information complementation between meibomian gland segmentation and boundary detection tasks. The proposed SBD‐MTLNet has been evaluated on an in‐house dataset (453 images) and a public dataset MGD‐1K (1000 images). Due to the limited number of images, a five‐fold cross validation strategy is adopted. Results Average dice coefficient of the proposed SBD‐MTLNet reaches 81.08% and 84.32% on the in‐house dataset and the public one, respectively. Comprehensive experimental results demonstrate the effectiveness the proposed SBD‐MTLNet, outperforming other state‐of‐the‐art methods. Conclusions The proposed SBD‐MTLNet can focus more on the shape characteristics of the meibomian glands and the boundary contour information between the adjacent glands via multi‐task learning strategy. The segmentation results of the proposed method can be used for the quantitative morphological characteristics analysis of meibomian glands, which has potential for the auxiliary diagnosis of MGD in clinic.

Background Bi‐parametric magnetic resonance imaging (bpMRI) has demonstrated promising results in prostate cancer (PCa) detection. Vision transformers have achieved competitive performance compared to convolutional neural network (CNN) in deep learning, but they need abundant annotated data for training. Self‐supervised learning can effectively leverage unlabeled data to extract useful semantic representations without annotation and its associated costs. Purpose This study proposes a novel self‐supervised learning framework and a transformer model to enhance PCa detection using prostate bpMRI. Methods and materials We introduce a novel end‐to‐end Cross‐Shaped windows (CSwin) transformer UNet model, CSwin UNet, to detect clinically significant prostate cancer (csPCa) in prostate bpMRI. We also propose a multitask self‐supervised learning framework to leverage unlabeled data and improve network generalizability. Using a large prostate bpMRI dataset (PI‐CAI) with 1476 patients, we first pretrain CSwin transformer using multitask self‐supervised learning to improve data‐efficiency and network generalizability. We then finetune using lesion annotations to perform csPCa detection. We also test the network generalization using a separate bpMRI dataset with 158 patients (Prostate158). Results Five‐fold cross validation shows that self‐supervised CSwin UNet achieves 0.888 ± 0.010 aread under receiver operating characterstics curve (AUC) and 0.545 ± 0.060 Average Precision (AP) on PI‐CAI dataset, significantly outperforming four comparable models (nnFormer, Swin UNETR, DynUNet, Attention UNet, UNet). On model generalizability, self‐supervised CSwin UNet achieves 0.79 AUC and 0.45 AP, still outperforming all other comparable methods and demonstrating good generalization to external data. Conclusions This study proposes CSwin UNet, a new transformer‐based model for end‐to‐end detection of csPCa, enhanced by self‐supervised pretraining to enhance network generalizability. We employ an automatic weighted loss (AWL) to unify pretext tasks, improving representation learning. Evaluated on two multi‐institutional public datasets, our method surpasses existing methods in detection metrics and demonstrates good generalization to external data.

Background Voxel‐based analysis (VBA) for population level radiotherapy (RT) outcomes modeling requires topology preserving inter‐patient deformable image registration (DIR) that preserves tumors on moving images while avoiding unrealistic deformations due to tumors occurring on fixed images. Purpose We developed a tumor‐aware recurrent registration (TRACER) deep learning (DL) method and evaluated its suitability for VBA. Methods TRACER consists of encoder layers implemented with stacked 3D convolutional long short term memory network (3D‐CLSTM) followed by decoder and spatial transform layers to compute dense deformation vector field (DVF). Multiple CLSTM steps are used to compute a progressive sequence of deformations. Input conditioning was applied by including tumor segmentations with 3D image pairs as input channels. Bidirectional tumor rigidity, image similarity, and deformation smoothness losses were used to optimize the network in an unsupervised manner. TRACER and multiple DL methods were trained with 204 3D computed tomography (CT) image pairs from patients with lung cancers (LC) and evaluated using (a) Dataset I (N = 308 pairs) with DL segmented LCs, (b) Dataset II (N = 765 pairs) with manually delineated LCs, and (c) Dataset III with 42 LC patients treated with RT. Results TRACER accurately aligned normal tissues. It best preserved tumors, indicated by the smallest tumor volume difference of 0.24%, 0.40%, and 0.13 % and mean square error in CT intensities of 0.005, 0.005, 0.004, computed between original and resampled moving image tumors, for Datasets I, II, and III, respectively. It resulted in the smallest planned RT tumor dose difference computed between original and resampled moving images of 0.01 and 0.013 Gy when using a female and a male reference. Conclusions TRACER is a suitable method for inter‐patient registration involving LC occurring in both fixed and moving images and applicable to voxel‐based analysis methods.

Purpose The US National Institutes of Health state that Sharing of clinical trial data has great potential to accelerate scientific progress and ultimately improve public health by generating better evidence on the safety and effectiveness of therapies for patients (https://www.ncbi.nlm.nih.gov/books/NBK285999/ accessed 2024‐01‐24.). Aligned with this initiative, the Trial Management Committee of the Trans‐Tasman Radiation Oncology Group (TROG) 15.01 Stereotactic Prostate Adaptive Radiotherapy utilizing Kilovoltage intrafraction monitoring (KIM) (SPARK) clinical trial supported the public sharing of the clinical trial data. Acquisition and Validation Methods The data originate from the TROG 15.01 SPARK clinical trial. The SPARK trial was a phase II prospective multi‐institutional clinical trial (NCT02397317). The aim of the SPARK clinical trial was to measure the geometric and dosimetric cancer targeting accuracy achieved with a real‐time image‐guided radiotherapy technology named KIM for 48 prostate cancer patients treated in 5 treatment sessions. During treatment, real‐time tumor translational and rotational motion were determined from x‐ray images using the KIM technology. A dose reconstruction method was used to evaluate the dose delivered to the target and organs‐at‐risk. Patient‐reported outcomes and toxicity data were monitored up to 2 years after the completion of the treatment. Data Format and Usage Notes The dataset contains planning CT images, treatment plans, structure sets, planned and motion‐included dose‐volume histograms, intrafraction kilovoltage, and megavoltage projection images, tumor translational and rotational motion determined by KIM, tumor motion ground truth data, the linear accelerator trajectory traces, and patient treatment outcomes. The dataset is publicly hosted by the University of Sydney eScholarship Repository at https://doi.org/10.25910/qg5d‐6058. Potential Applications The 3.6 Tb dataset, with approximately 1 million patient images, could be used for a variety of applications, including the development of real‐time image‐guided methods, adaptation strategies, tumor, and normal tissue control modeling, and prostate‐specific antigen kinetics.