Impact of CT reconstruction algorithm on auto‐segmentation performance

Abstract Model‐based iterative reconstruction (MBIR) reduces CT imaging dose while maintaining image quality. However, MBIR reduces noise while preserving edges which may impact intensity‐based tasks such as auto‐segmentation. This work evaluates the sensitivity of an auto‐contouring prostate atlas across multiple MBIR reconstruction protocols and benchmarks the results against filtered back projection (FBP). Images were created from raw projection data for 11 prostate cancer cases using FBP and nine different MBIR reconstructions (3 protocols/3 noise reduction levels) yielding 10 reconstructions/patient. Five bony structures, bladder, rectum, prostate, and seminal vesicles (SVs) were segmented using an auto‐segmentation pipeline that renders 3D binary masks for analysis. Performance was evaluated for volume percent difference (VPD) and Dice similarity coefficient (DSC), using FBP as the gold standard. Nonparametric Friedman tests plus post hoc all pairwise comparisons were employed to test for significant differences (P < 0.05) for soft tissue organs and protocol/level combinations. A physician performed qualitative grading of 396 MBIR contours across the prostate, bladder, SVs, and rectum in comparison to FBP using a six‐point scale. MBIR contours agreed with FBP for bony anatomy (DSC ≥ 0.98), bladder (DSC ≥ 0.94, VPD < 8.5%), and prostate (DSC = 0.94 ± 0.03, VPD = 4.50 ± 4.77% (range: 0.07–26.39%). Increased variability was observed for rectum (VPD = 7.50 ± 7.56% and DSC = 0.90 ± 0.08) and SVs (VPD and DSC of 8.23 ± 9.86% range (0.00–35.80%) and 0.87 ± 0.11, respectively). Over the all protocol/level comparisons, a significant difference was observed for the prostate VPD between BSPL1 and BSTL2 (adjusted P‐value = 0.039). Nevertheless, 300 of 396 (75.8%) of the four soft tissue structures using MBIR were graded as equivalent or better than FBP, suggesting that MBIR offered potential improvements in auto‐segmentation performance when compared to FBP. Future work may involve tuning organ‐specific MBIR parameters to further improve auto‐segmentation performance. Running title: Impact of CT Reconstruction Algorithm on Auto‐segmentation Performance.


| INTRODUCTION
One of the largest sources of uncertainty in radiation therapy planning (RTP) is the delineation of the target and organs at risk (OARs) using computed tomography (CT) datasets. 1 Aside from the uncertainty introduced in the delineation process, manual delineation of OARs is time-consuming. 2 Thus, efforts to implement auto-segmentation routines are advantageous and have shown promise for several disease sites, most commonly for head and neck, and prostate cancers. [3][4][5] Current approaches to automated segmentation are commonly atlas-based or a combination of atlas-and modelbased techniques. 3 In the pelvis, auto-segmentation has yielded good overall performance (~3 mm distance to mean surface in prostate segmentation from mean expert delineations). 6,7 Atlasbased segmentation algorithms have appeared promising for the segmentation of bladder, rectum, and prostate (Dice similarity coefficient (DSC) > 0.70 with respect to radiation oncologist ground truth delineations) for prostate cancer treatment planning. 8 Another auto-segmentation toolkit, Smart Probabilistic Image Contouring Engine (SPICE), has been applied to CT scans in various disease sites and has demonstrated promise for clinical utility. 9,10 Current efforts are being made to move toward lower dose CT scanning, although image noise can be a rate-limiting step due to the use of filtered back projection (FBP) for image reconstruction in CT simulation (CT-SIM) datasets. 11 One potential way to overcome image noise is to employ advanced reconstruction algorithms to maintain the same image quality while increasing the contrast to noise ratio. 12,13 Advanced CT reconstruction methods such as hybrid iterative reconstructions (HIR), model-based iterative reconstruction (MBIR), and adaptive statistical iterative reconstruction (ASIR) have been integrated into clinical diagnostic CT scanners, 14,15 whereas their application in radiation oncology has been limited to date. In a study conducted by Price et al., HIR was found to maintain image quality with dose reduction protocols of up to~70% when compared to FBP for CT-SIM datasets in the female pelvis. 16 While dose reduction is possible and MBIR has shown improvements in diagnosis and image quality, 17 advanced iterative reconstruction algorithms have been shown to change the overall texture of datasets compared to standard FBP, 13,18,19 which may lead to differences in intensity-dependent automatic segmentation routines.
This study aims to evaluate the sensitivity of an auto-segmentation algorithm to MBIR CT reconstructions for prostate cancer treatment planning and compare the results to the standard of care in CT-SIM (FBP). With a better understanding of the impact of reconstruction methods on auto-segmentation performance, the utility and potential application of advanced reconstruction algorithms may be integrated into the RTP workflow.

2.A | CT-Simulation and patient cohort
Eleven prostate cancer patients underwent CT simulation to generate a patient model for external beam treatment planning using a Brilliance Big Bore (Philips Health Care, Cleveland, OH) scanner positioned supine with the following parameters: 120-140 kVp, 500 mAs, 512 × 512 in-plane image dimensions, 1.28 × 1.28 mm 2 in-plane spatial resolution, and 3 mm slice thickness. Patients were immobilized using bands placed around the feet, a ring to hold on their chest, and a shaped foam pad for leg immobilization. Raw sinogram data were exported from the clinical scanner and de-identified for further processing.

2.B | Model-based image reconstructions
Raw sinogram data were retrospectively reconstructed using FBP and MBIR algorithms with varied parameters using research reconstruction software (IMR, Philips Medical Systems, Cleveland, OH).
The Philips Big Bore scanner specifies three user MBIR reconstruction protocols: Body Soft Tissue (BST), Body Routine (BR), and Body Sharp Plus (BSP). Each protocol is distinguished primarily by the reconstruction filter. Choice of filter reflects a tradeoff between noise and contrast resolution, BSP having the highest noise and resolution. 20,21 The MBIR reconstruction optimization equation is defined as: where the function F(x) is composed of a data fit term D(x) and a noise reducing (but edge preserving) regularization term R(x) with strength controlled by the factor β. For FBP, image noise is known to scale with well-defined ratios based on slice thickness, patient size, and mAs. While noise reductions arising from the MBIR reconstruction levels (L1-L3) depend on acquisition parameters including slice thickness, mAs, and patient size, they do not scale in proportion to the FBP noise. 22 Nevertheless, as the underlying raw data acquisition parameters were fixed in this study, higher level reconstructions (i.e., L3) were expected to have the largest reduction in noise, particularly for low dose protocols (e.g.,~60% noise reduction between L1 and L3 for an abdomen CT acquired at 300 mAs, 1 mm, Body Routine filter acquired with an Ingenuity CT scanner 22 ).

2.C | Organ auto-segmentation
For each reconstructed CT image, auto-segmentation was performed using a research prototype version of SPICE software (Philips, Cleveland, OH). The prostate cancer cases produced auto-segmentations for nine OARs (four soft tissue (prostate, bladder, rectum, and seminal vesicles) and five bony structures (left and right sides of the pelvis and femur, as well as the sacrum). The segmentation software pipeline includes three main steps: (a) global positioning, (b) organ-specific positioning, and (c) structure refinement, as described by Bzdusek et al. 23 Briefly, the first step rotates, translates, and performs an isotropic scaling registration of a tissue probability atlas to a tissue classified target image. In the second step, the organs are positioned while using both the tissue classified image and the organ-specific probability maps.
Lastly, in the third step, model-based segmentation is used to bring the surface mesh triangles to trained image features. 23 Comparisons were performed to investigate the differences in each organ segmentation (e.g., prostate, rectum) based upon the MBIR protocol and level. The auto-segmentations were compared for the different MBIR reconstructions of the identical raw dataset, as well as between patients to further understand the impact of the CT reconstruction on auto-segmentation of organs, and patientspecific differences.

2.D | Physician qualitative contour grading
Eleven prostate cancer patient datasets (396 generated contours (99 prostate, rectum, seminal vesicles, and bladder)) were qualitatively evaluated by a physician instructed to grade the generated MBIR contours at each protocol/level in comparison to those generated on the reference dataset (i.e., FBP) using a six-point scale (1: Better than reference, 2: Slightly better than reference, 3: Equivalent to reference, 4: Slightly worse than reference, 5: Worse than reference, 6: Clinically unacceptable). 16,18 A score of six was also assigned to contours that incorrectly segmented the organ regardless of agreement to FBP. During the contour review, the physician was blinded to reconstruction protocol/level. After review of the grading results, any contour that was deemed clinically unacceptable (score of 6) was removed from the patient cohort for any further statistical evaluation to ensure that only clinically acceptable data fit for review were maintained within the cohort.

2.E | Analyses and quantitative comparisons
The auto-segmentations from each of the nine different MBIR reconstructions and FBP were quantitatively evaluated using volume percent difference (VPD) and DSC. For each prostate cancer case, the auto-segmentations were analyzed using an in-house MATLAB program (Mathworks, Natick, MA) to measure the volume of the OARs. VPD was calculated with the FBP being the reference as defined: where MBIR volume and FBP volume are the measured auto-segmented volumes of an organ using the MBIR and FBP reconstructions, respectively.
The resulting segmentations were also quantitatively compared using the DSC to assess the regional overlap between the organs according to: In this equation, A and B are the volumes of the reference (i.e., FBP) and MBIR auto-segmented organ respectively, and A∩B is defined as the volume of the intersection of the auto-segmented organ from the two reconstruction methods. DSC equal to zero describes no overlap and DSC equal to 1 demonstrates complete agreement of the auto-segmentations. Finally, center of mass (COM) comparisons were made between the auto-segmented organs from FBP and MBIR reconstructions to elucidate potential location differences across each major axes (X-(right-left), Y-(anterior-posterior), and Z-(superior-inferior)).     These sensitivity results demonstrate differences between the FBP and MBIR CT reconstruction algorithms for auto-segmentation, further described by Table 2 for the COM differences in the X-  Auto-segmentation performance has also been shown to be impacted by organ size, as found by Kumarasiri et al. where soft tissue structures were classified by size for CT auto-segmentation. 26 In studies focusing on the pelvic region, similar results were found in that larger structures, such as the bladder, had smaller volume variations when compared to smaller structures similar to the prostate. 27 This was also observed in our study where the bladder was more accurately contoured in comparison to the smaller structures (rectum, prostate, and SVs), as shown by the DSC results in Table 1    An additional limitation of this work is that all MBIR segmentation results were reported in reference to FBP. Although FBP is considered the gold standard for radiation oncology delineations, 32 FBP may be limited by sensitivity to noise, motion, metal, and streak artifacts. 33 Nevertheless, this work incorporated qualitative scoring by a physician to assess differences between FBP and MBIR results.

2.F | Analyses for statistical comparisons
Work to integrate advanced reconstruction algorithms into CT-SIM platforms is ongoing. 16 Future work can build on our qualitative grading by incorporating physician-based ground truth delineations to find the most promising MBIR algorithm combinations for different contouring endpoints. Although our study did not use physician delineations, the quantitative differences found between FBP and MBIR demonstrated for both VPD and DSC show that the advanced reconstruction algorithms are providing images with different characteristics than the FBP reconstruction and thus have an impact on auto-segmentation for lower contrast organs such as the prostate and SVs. Additionally, more research on MBIR reconstruction is needed to further investigate how it can be used or improved to be implemented in the treatment or location detection of a specific organ. As MBIR makes its way into Radiation Oncology CT-SIM platforms to enable reductions in imaging dose, the impact on auto-segmentation task performance will be of increasing importance for clinical efficiency. This work revealed that auto-segmentation performance on MBIR images was comparable or better than FBP for 75% of the generated soft tissue contours, although more complex structures, such as the SVs may still require manual edits.

| CONCLUSION
Automatic segmentation for MBIR on high contrast structures was successful and offered improved segmentation quality for 30-40% of the bladder, prostate, and SV contours as compared to FBP.
Although manual modifications may still be necessary, when coupling MBIR with auto-segmentation, both imaging dose and treatment planning time are reduced. Future work may involve selecting organspecific MBIR parameters to improve auto-segmentation performance.

ACKNOWLEDG MENTS
The submitting institution holds research agreements with Philips Healthcare, ViewRay, Inc., and Modus Medical. Research was partially supported by the National Cancer Institute of the National Institutes of Health under Award Number R01CA204189.

CONFLI CT OF INTEREST
The submitting institution holds research agreements with Philips Healthcare. Noel Black, Paul Klahr, Heinrich Schulz, and Liran Goshen are clinical scientist employees of Philips Healthcare.