Improving deep learning organ segmentation accuracy in cone-beam CT-guided radiotherapy using a robust scatter suppression method Authors: Rupesh Dotel, Farhang Bayat, Junxiao Hu, Uttam Pyakurel, Ryan Sabounchi, Roy Bliley, Ryan Lanning, Brian Kavanagh, Tyler Robin, Cem Altunbas Published: 2026 ABSTRACT Background and purposeDeep learning–based segmentation of organs and tumors in cone-beam computed tomography (CBCT) is critical for improving workflow efficiency and accuracy in dose delivery monitoring and adaptive radiotherapy. However, degraded CBCT image quality, primarily due to scatter, can adversely affect segmentation performance. This study evaluated whether improving CBCT image quality through enhanced scatter mitigation improves deep learning (DL) segmentation accuracy.Materials and methodsA quantitative CBCT method incorporating a novel antiscatter grid and dedicated reconstruction pipeline was evaluated in a prospective study of 26 patients with cancers in the prostate, head and neck (H&N), and pelvis/abdomen regions. Each patient underwent both standard-of-care CBCT and quantitative CBCT scans. A foundation model, agnostic to CBCT images, was used to segment 13 organs and targets across CBCT and planning computed tomography (CT) images. DL segmentation accuracy was evaluated using Dice similarity coefficient and maximum Hausdorff distance with respect to clinician-drawn contours.ResultsMean Dice coefficients were higher for quantitative CBCT (0.66) compared to standard-of-care CBCT (0.61–0.63, p < 0.05). For the prostate, bladder, and rectum, Dice coefficients improved from 0.68 to 0.81 to 0.80–0.86, with significant gains for the bladder and rectum (p < 0.05). Hausdorff distances were lower for quantitative CBCT in the prostate and pelvis/abdomen regions, indicating improved boundary agreement. In the H&N region, quantitative CBCT did not improve segmentation accuracy.ConclusionsImproved quantitative accuracy in CBCT through robust scatter suppression enhances deep learning-based segmentation, particularly for large organs, supporting its role in improving adaptive radiotherapy workflows. Introduction The role of cone-beam computed tomography (CBCT) in radiotherapy has expanded beyond patient setup correction to include dose monitoring, online treatment plan adaptation, and personalized response assessment through evaluation of changes in normal tissues and targets [1], [2], [3], [4], [5]. CBCT image quality plays an important role in enabling these applications [6], [7], [8]. In addition to increasing confidence in clinical decision making, reduced artifacts and accurate CT numbers are required for direct dose calculation using CBCT [9], [10], [11]. Automated tissue segmentation is another key step for accurate dose estimation across targets and surrounding organs during adaptive radiotherapy [4], [12]. Beyond plan adaptation, changes in the volume and shape of automatically segmented structures can also be used to assess response to radiotherapy [13], [14], [15]. Contouring of anatomical structures is one of the most time-consuming steps during an adaptive treatment session [16], [17]. Manual contouring of all organs and targets by clinicians is often not feasible in this setting. As a result, automated segmentation methods have been widely used and studied, including deformable image registration approaches that propagate contours from planning CT to CBCT acquired on the day of treatment, as well as deep learning (DL)-based methods. DL-based approaches have shown strong performance in tissue segmentation, with most organ delineation performed automatically by trained models, and several methods are now integrated into adaptive radiotherapy platforms [18], [19], [20], [21]. Despite the success of DL-based tissue segmentation, manual editing of DL contours is often required, increasing both treatment time and the complexity of the adaptive radiotherapy workflow. For example, Moazzezi et al. reported that 96% of fractions required edits to auto-segmented contours [2]. Others have reached similar conclusions [5], [22], [23], [24], [25]. One of the major factors contributing to suboptimal DL-generated segmentation is degraded CBCT image quality. Compared to multidetector CT images used for treatment planning, CBCT images typically exhibit lower soft tissue contrast, increased image artifacts, and reduced CT number accuracy. X-ray scatter is a primary source of this degradation [26]. To mitigate scatter, current approaches commonly use a combination of radiographic one-dimensional antiscatter grids and model-based correction methods [27]. However, despite these advances, scatter-related image degradation persists in CBCT images [28]. To address the limitations of CBCT image quality, DL-based synthetic CT generation methods have also been extensively investigated [29]. An alternative approach to further mitigate scatter is to enhance hardware-based scatter rejection, such as the two-dimensional antiscatter grid (2D ASG) technique [30]. The 2D ASG has been shown to effectively reduce scatter fluence incident on the kV flat panel detector and improve CT number accuracy, thereby decreasing reliance on correction algorithms that may lead to under or overcorrection in CBCT images [31]. However, the impact of improved image quality achieved with the 2D ASG technique on DL-based tissue segmentation accuracy has not been previously investigated. We hypothesize that improved CT number accuracy and reduced image artifacts with the 2D ASG can enhance dose calculation accuracy [10] and DL-based tissue segmentation performance. To test this hypothesis, a 2D ASG-based scatter mitigation approach was implemented on a CBCT linac system [28], [32]. Due to its high CT number accuracy [28], [32], [33], [34], this method is referred to as quantitative CBCT (qCBCT). The qCBCT approach is still experimental and is not yet suitable for clinical workflows. Therefore, a prospective, early-phase clinical trial was conducted to evaluate the impact of qCBCT on DL-based tissue segmentation accuracy, compared to standard-of-care CBCT (SOC CBCT) methods. Materials and Methods Acquisition of CBCT images In this prospective institutional review board-approved study (NCT04565457), 26 patients undergoing image-guided radiotherapy were scanned using both the qCBCT technique and the SOC CBCT protocol. Participants were stratified into prostate, H&N, and upper pelvis and abdomen groups, with 9, 9, and 8 participants in each subgroup, respectively. These regions were selected because they present challenges in soft tissue visualization due to scatter and motion artifacts. The prostate region is dominated by scatter with limited motion; the H&N region by breathing and swallowing motion plus a large scatter fraction at the level of the humeral heads; and the upper pelvis/abdomen by both substantial motion and high scatter. Each participant was scanned once with SOC CBCT and once with qCBCT within a two-hour window on the same day using a Varian TrueBeam CBCT system (Version 2.7, Varian Medical Systems, Palo Alto, USA) [34]. On the day of imaging, SOC CBCT was acquired first as part of the standard radiotherapy workflow. The conventional radiographic grid was then removed, and a 2D ASG was installed on the kV imager. A qCBCT scan was subsequently acquired using the same imaging parameters, dose, immobilization devices and patient positioning as for the SOC CBCT scan. The study participant was removed from the treatment table between the two scans. Each CBCT acquisition lasted 60 s, and participants were free breathing during all scans. The 2D ASG used in this work was an experimental prototype developed by our group [30]. It was 3D printed from tungsten, with a grid pitch of 2 mm and a grid ratio of 12 (Fig. 1). The active area was 40 cm by 10 cm and was designed for the TrueBeam large field of view scan geometry.Fig. 1A photo of the 2D antiscatter grid installed on the flat panel detector, which was used for qCBCT acquisitions.Fig. 1 A photo of the 2D antiscatter grid installed on the flat panel detector, which was used for qCBCT acquisitions. SOC CBCT scans were reconstructed twice: once using the standard protocol, which employs scatter kernel superposition-based scatter correction and filtered backprojection [35], and once using an advanced method known as iCBCT, which uses Acuros-based scatter correction and iterative reconstruction [36]. These two approaches are referred to as “standard” and “advanced” SOC CBCT throughout this work. Since the qCBCT approach is an experimental method not implemented in clinical workflows, qCBCT images were processed using an offline raw data correction pipeline followed by filtered backprojection reconstruction [10], [28]. For each participant, the three CBCT image sets and the planning CT (planCT) were rigidly coregistered. The axial field of view (FOV) was 46 cm in all CBCT images, while the qCBCT and SOC CBCT images had 6.5 and 17 cm scan length in the superior-inferior direction, respectively. The reconstruction FOV and voxel dimensions of all CBCT and planCT images were matched to those of the qCBCT images. The voxel size was 0.9 × 0.9 × 2 mm3 in all images. DL segmentation of CBCT and planCT images The primary goal of this study was to investigate the effect of CBCT image quality variability across different CBCT methods on DL segmentation performance. This required a segmentation model that was largely agnostic to CBCT image type and capable of generalizing across different CBCT systems without retraining. Large organs such as the bladder, rectum, and liver were not always fully captured within the CBCT FOV, in part due to the limited 6.5 cm superior-inferior scan length of our qCBCT method. The model therefore needed to support slice-wise segmentation of individual axial CBCT slices. To meet these requirements, we used the Segment Anything in Medical Images (MedSAM) to segment anatomical structures. MedSAM is a Vision Transformer-based promptable segmentation model derived from the Segment Anything Model and trained on approximately 1.5 million medical image mask pairs, including a large number of CT images [37], [38], [39]. In this study, the pretrained MedSAM model was applied in a semi-automatic fashion. For each organ on each axial slice, a bounding box prompt was manually defined by the operator through visual inspection of the corresponding organ of interest, and MedSAM generated the segmentation mask from that prompt. Image preprocessing consisted of applying a soft-tissue windowing function (window center of 40 HU, window width of 400 HU) followed by normalization to the 0–255 intensity range expected by the model. MedSAM was deployed as a plugin within 3D Slicer [40], an open-source medical image computing platform, providing an interactive interface for prompt-based segmentation. Axial slices with excessive metal artifacts were excluded, as these severely degrade segmentation accuracy. The most superior and inferior slices of each structure were also excluded, as preliminary results showed poor segmentation across all image types, partly attributed to partial volume effects. Analysis of DL segmentation accuracy Across the three CBCT types and planCT images, anatomical structures and targets were manually delineated by a radiation oncologist (clinician-drawn contours) and segmented using MedSAM with prompt-based guidance (Table 1). The clinician-drawn contours were delineated on the advanced SOC CBCT, qCBCT, and planCT image sets, and those on the advanced SOC CBCT were rigidly copied to the standard SOC CBCT, as both image sets were reconstructed using the same SOC CBCT scan. Based on preliminary observations of the structure-size dependence of DL segmentation performance, structures with average cross-sectional areas greater than 300 mm2 were classified as large (L) organs (Table I), and the remaining structures as small (S) organs. This threshold was chosen based on a gap in the distribution between the largest small organ (aorta, ∼232 mm2) and the smallest large organ (rectum, ∼315 mm2) (Fig. S1 in the supplementary material).Table 1Segmented organ characteristics categorized as large (L) or small (S) based on the mean cross-sectional area in axial planning CT slices. n represents the number of 2D slices analyzed for each structure, aggregated over all study participants. Left and right kidney contours and left and right submandibular gland contours are each counted as one structure. Structures with average cross-sectional areas greater than 300 mm2 were classified as large organs.OrganMean area ± Standard deviation (SD) (mm2)Size nAnatomical siteLiver2460 ± 494L 18Pelvis/abdomenBladder1875 ± 643L 112ProstateKidneys1305 ± 234L 28Pelvis/abdomenProstate802 ± 248L 52ProstateRectum315 ± 106L 98ProstateAorta232 ± 23S 30Pelvis/abdomenLymph Nodes109 ± 99S 88Pelvis/abdomenSubmandibular Glands94 ± 44S 20H&NRight Common Iliac Artery87 ± 28S 26Pelvis/abdomenGross target volume50 ± 28S 14H&NLeft External Iliac Vein48 ± 22S 72Pelvis/abdomenCervical Esophagus47 ± 22S 8H&NCarotid Artery31 ± 11S 108H&N Table 1Segmented organ characteristics categorized as large (L) or small (S) based on the mean cross-sectional area in axial planning CT slices. n represents the number of 2D slices analyzed for each structure, aggregated over all study participants. Left and right kidney contours and left and right submandibular gland contours are each counted as one structure. Structures with average cross-sectional areas greater than 300 mm2 were classified as large organs.OrganMean area ± Standard deviation (SD) (mm2)Size nAnatomical siteLiver2460 ± 494L 18Pelvis/abdomenBladder1875 ± 643L 112ProstateKidneys1305 ± 234L 28Pelvis/abdomenProstate802 ± 248L 52ProstateRectum315 ± 106L 98ProstateAorta232 ± 23S 30Pelvis/abdomenLymph Nodes109 ± 99S 88Pelvis/abdomenSubmandibular Glands94 ± 44S 20H&NRight Common Iliac Artery87 ± 28S 26Pelvis/abdomenGross target volume50 ± 28S 14H&NLeft External Iliac Vein48 ± 22S 72Pelvis/abdomenCervical Esophagus47 ± 22S 8H&NCarotid Artery31 ± 11S 108H&N Segmented organ characteristics categorized as large (L) or small (S) based on the mean cross-sectional area in axial planning CT slices. n represents the number of 2D slices analyzed for each structure, aggregated over all study participants. Left and right kidney contours and left and right submandibular gland contours are each counted as one structure. Structures with average cross-sectional areas greater than 300 mm2 were classified as large organs. Segmentation accuracy was evaluated by comparing DL contours with clinician-drawn contours using the Dice similarity coefficient and the maximum Hausdorff distance. The Dice coefficient quantifies area overlap between DL-based and clinician-drawn contours on individual axial slices, where a value of 1 indicates perfect agreement [41], [42]. The Hausdorff distance measures the maximum distance between DL-based and clinician- contours, capturing the largest boundary discrepancy between the two contours for a given organ on individual axial slices [43]. An ideal Hausdorff distance of 0 mm indicates that DL and clinician-drawn contours are identical. For subgroup analyses, contour similarity metrics were stratified by three anatomical sites: prostate, H&N, and pelvis/abdomen, to assess whether anatomical region influenced segmentation accuracy. In a second analysis, segmentation accuracy was compared between the small- and large-structure groups. In addition to size-specific differences, organ-specific differences in segmentation accuracy between SOC CBCT and qCBCT were evaluated. Segmentation accuracy in planCT images was also assessed as the reference standard. Statistical analysis Linear mixed-effects models [44] were used to compare Dice similarity coefficients and Hausdorff distances across CBCT types, accounting for the hierarchical structure of the data by including random effects for slices nested within patients. Dice coefficients were analyzed after a squared transformation to improve adherence to the normality assumption of model residuals. Least-squares means and corresponding 95% confidence intervals (CI) were estimated from the fitted models and back-transformed to the original scale for presentation. Prespecified subgroup analyses were performed to evaluate differences among CBCT types within each organ and within each treatment site separately. Multiplicity was addressed using Tukey's honestly significant difference adjustment to control the family-wise error rate at a two-sided α level of 0.05 [45]. In pairwise comparisons of CBCT types, a p-value of less than 0.05 was considered statistically significant. Results Examples of DL contours and clinician-drawn contours from a prostate cohort participant are shown in Fig. 2, qualitatively showing the segmentation performance for bladder, prostate, and rectum. When all anatomical sites were combined, standard SOC CBCT, advanced SOC CBCT, and qCBCT yielded mean Dice coefficients of 0.63 (95% CI: 0.57–0.68), 0.61 (95% CI: 0.54–0.66), and 0.66 (95% CI: 0.60–0.71), respectively (Fig. 3). Mean Hausdorff distances for the respective CBCT methods were 2.8 (95% CI: 2.4–3.0), 2.8 (95% CI: 2.5–3.1), and 2.6 (95% CI: 2.3–2.9) mm, indicating that DL contours in qCBCT images were more similar to clinician-drawn contours (p < 0.05). DL contouring accuracy was highest in planCT images based on both Dice coefficient (p < 0.001) and Hausdorff distance (p < 0.05).Fig. 2Example DL contours (red) and clinician-drawn contours (green) for a prostate patient. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)Fig. 2Fig. 3Mean Dice coefficients (a) and Hausdorff distances (b) for all anatomical sites aggregated. Whiskers represent 95% confidence intervals. ****: p < 0.0001, ***: p < 0.001, **: p < 0.01, *: p < 0.05, ns: not significant, p ≥ 0.05.Fig. 3 Example DL contours (red) and clinician-drawn contours (green) for a prostate patient. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Mean Dice coefficients (a) and Hausdorff distances (b) for all anatomical sites aggregated. Whiskers represent 95% confidence intervals. ****: p < 0.0001, ***: p < 0.001, **: p < 0.01, *: p < 0.05, ns: not significant, p ≥ 0.05. When contours were stratified by size, DL contours in qCBCT showed significantly higher Dice coefficients than those in SOC CBCT for large organs (Fig. 4a). Mean Dice coefficients for advanced SOC CBCT and qCBCT were 0.69 and 0.76, respectively (p < 0.001). Mean Hausdorff distances for advanced SOC CBCT and qCBCT were 3.4 mm and 3.1 mm, respectively (Fig. 4b), indicating improved agreement with clinician-drawn contours for qCBCT (p < 0.001). For small organs, differences in segmentation accuracy between qCBCT and SOC CBCT were not statistically significant (p ≥ 0.05).Fig. 4Dice coefficients and Hausdorff distances stratified by organ size are shown in (a) and (b). Dice coefficients and Hausdorff distances stratified by organ site are shown in (c) and (d). Whiskers represent 95% confidence intervals. ****: p < 0.0001, ***: p < 0.001, **: p < 0.01, *: p < 0.05, ns: not significant, p ≥ 0.05.Fig. 4 Dice coefficients and Hausdorff distances stratified by organ size are shown in (a) and (b). Dice coefficients and Hausdorff distances stratified by organ site are shown in (c) and (d). Whiskers represent 95% confidence intervals. ****: p < 0.0001, ***: p < 0.001, **: p < 0.01, *: p < 0.05, ns: not significant, p ≥ 0.05. Overall, lower Dice coefficients were observed in the H&N group (Fig. 4c). Dice coefficients for standard SOC CBCT and qCBCT were 0.54 and 0.45, respectively (p < 0.01). Mean Hausdorff distances were 2.0 and 2.3 mm for the standard SOC CBCT and qCBCT, respectively (p ≥ 0.05). The highest Dice coefficients were observed in the prostate group. Mean Dice coefficients were 0.73 and 0.79 for advanced SOC CBCT and qCBCT, respectively (p < 0.001). The largest Hausdorff distances were also observed in the prostate group (Fig. 4d). Mean Hausdorff distances were 3.5 and 3.2 mm for advanced SOC CBCT and qCBCT, respectively (p < 0.001). Lower Hausdorff distances in qCBCT images indicate improved DL contouring accuracy. The largest boundary deviations occurred at the prostate–bladder interface, where low intrinsic soft-tissue contrast in CBCT hampers reliable boundary delineation. In the pelvis and abdomen group, qCBCT yielded higher segmentation accuracy than SOC CBCT. Dice coefficients for standard SOC CBCT and qCBCT were 0.62 and 0.67, respectively (p < 0.05). Mean Hausdorff distances were 2.5 and 2.2 mm for standard SOC CBCT and qCBCT, respectively (p < 0.01). DL contours on planCT demonstrated consistently higher accuracy than those from all CBCT methods across all analyzed groups. However, the differences between qCBCT and planCT were generally small. Dice coefficients for qCBCT and planCT were 0.66 and 0.70, respectively (p < 0.001). The largest statistically significant differences were observed in the small organ group and in the H&N region. No significant differences between qCBCT and planCT were observed in the pelvis/abdomen or prostate groups (p ≥ 0.05). In the organ-specific analysis, statistically significant differences in DL contouring accuracy between qCBCT and SOC CBCT were observed primarily for organs in the large organ group (Fig. 5, Table S1). Dice coefficients were highest for the liver and bladder across all CBCT methods, reflecting strong contour area agreement between DL and clinician-drawn contours. However, Hausdorff distances were also largest for these organs, pointing to greater localized deviations between DL and clinician-drawn contours. qCBCT showed significantly higher Dice coefficients than advanced SOC CBCT, particularly for the rectum (p < 0.001). Hausdorff distances were generally lower in qCBCT. For most small organs, Dice coefficients were below 0.6 and the differences in Hausdorff distances between SOC CBCT and qCBCT were not statistically significant (Fig. 5, Table S2).Fig. 5Dice coefficients and Hausdorff distances for 13 structures are shown in heat maps. A black vertical line separates large organs (left) from small organs and structures (right). Statistical significance of the differences is shown in Tables S1 and S2.Fig. 5 Dice coefficients and Hausdorff distances for 13 structures are shown in heat maps. A black vertical line separates large organs (left) from small organs and structures (right). Statistical significance of the differences is shown in Tables S1 and S2. In comparisons between standard and advanced SOC CBCT, standard SOC CBCT yielded higher Dice coefficients only in the pelvis/abdomen group and for small organs (p < 0.01), and Dice coefficient differences were not statistically significant for other subgroups. Differences in Hausdorff distances were small and not statistically significant. Discussion This study evaluated the effect of an experimental quantitative CBCT technique, qCBCT, on DL-based segmentation accuracy in a prospective imaging trial. The main innovation of the qCBCT technique is the use of a 2D ASG to improve CT number accuracy and reduce image artifacts. Overall, qCBCT improved DL contouring accuracy relative to SOC CBCT, highlighting the impact of improved quantitative accuracy on DL-based contouring performance. Scatter, noise, and reduced soft tissue contrast in CBCT can degrade DL contouring accuracy [37], [46]. Consistent with prior reports [19], [47], DL contouring accuracy was higher in planning CT images than in CBCT images, while improved CT number accuracy and reduced artifacts in qCBCT [34] narrowed the DL contouring accuracy gap between CBCT and planning CT. No statistically significant differences in contouring accuracy were observed between standard and advanced SOC CBCT, except for the pelvis/abdomen region. To further improve contouring accuracy, DL-based synthetic CT generation methods can be combined with SOC CBCT as well as qCBCT [21], [48], [49], [50]. Dice coefficients in our study were slightly lower than those reported in prior works [19], [51]. For example, in SOC CBCT images, Dice coefficients ranged from 0.68 to 0.81 for the prostate, bladder, and rectum. In prior studies, Dice coefficients ranged from 0.83 to 0.95 for these organs in similar CBCT scans [18], [52]. These differences may be attributable to differences in model architecture and training strategies, as the cited studies used 3D segmentation models trained on CBCT images. In contrast, our study used a foundation DL-based model that was not trained on CBCT images to reduce potential training-related bias and enable fair comparison between qCBCT and SOC CBCT. This study has several limitations. The qCBCT approach used here is experimental and not yet suitable for clinical workflows. To reduce prototyping costs, the superior–inferior scan length of the 2D grid prototype was limited to 6.5 cm, which truncated several organs. Increasing the scan length is one of the goals of future 2D grid fabrication efforts. The 2D, bounding-box-based segmentation approach was chosen to address these data limitations and avoid training bias, not intended for clinical implementation; bounding-box position and size can also bias segmentation accuracy. In the future, a larger-field-of-view qCBCT approach and additional patient data may allow 3D segmentation methods, eliminate the need for bounding boxes, and better differentiate performance among CBCT types. Reference contours were drawn by one clinician, which may introduce bias. We used the Hausdorff distance to measure the maximum distance between two contours, which is sensitive to outliers; metrics such as the 95th-percentile Hausdorff distance and mean surface distance could reduce this effect in future work. In addition, qCBCT and SOC CBCT used different processing and reconstruction pipelines, which may have contributed to the observed differences in segmentation accuracy. Finally, since our DL contouring was performed slice-wise on 2D images rather than by the volumetric 3D segmentation more commonly used in CBCT studies, and since slices with severe metal artifacts and the most superior and inferior slices of each structure were excluded, direct quantitative comparison of our Dice coefficients and Hausdorff distances with those from conventional 3D segmentation studies is not feasible. In summary, qCBCT improved DL segmentation accuracy compared to SOC CBCT, particularly for large organs and in the pelvis/abdomen, and prostate regions. Overall, these findings emphasize the importance of CBCT image quality for DL segmentation performance and support the use of robust scatter reduction approaches to improve DL contouring in CBCT-guided adaptive radiotherapy. CRediT authorship contribution statement Rupesh Dotel: Writing – review & editing, Visualization, Validation, Software, Methodology, Formal analysis, Data curation. Farhang Bayat: Writing – review & editing, Software, Investigation, Data curation. Junxiao Hu: Writing – review & editing, Visualization, Validation, Investigation, Formal analysis. Uttam Pyakurel: Writing – review & editing, Software, Investigation, Data curation. Ryan Sabounchi: Writing – review & editing, Investigation, Data curation. Roy Bliley: Writing – review & editing, Investigation, Data curation. Ryan Lanning: Writing – review & editing, Resources, Investigation. Brian Kavanagh: Writing – review & editing, Resources, Investigation. Tyler Robin: Writing – review & editing, Resources, Investigation. Cem Altunbas: Writing – review & editing, Writing – original draft, Supervision, Software, Resources, Project administration, Methodology, Investigation, Funding acquisition, Conceptualization. Ethics declaration Informed consent and patient details Written informed consent to take part in the study and to publish the article has been obtained from all participants or their legal representatives. The privacy rights of participants have been observed. Studies in human This study was performed in compliance with relevant laws, regulatory frameworks and guidelines where the research took place. This study was approved by the Colorado Multiple Institutional Review Board (COMIRB) (Approval No. 20–1684) Clinical trials The results of this clinical trial and any associated work have been posted in a registry. This clinical trial was registered with number NCT04565457 (clinicaltrials.gov). Funding This work was funded in part by grants from NIH/NCI R01CA245270 and P30CA046934. Declaration of competing interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Cem Altunbas is the founder of M2 Technologies and holds patents on 2D antiscatter grids. M2 Technologies was not involved with any of the methods developed or investigations presented in this manuscript. All remaining authors have declared no conflicts of interest.