Mean Distance To Agreement Research Articles

Methodological approach to create an atlas using a commercial auto-contouring software.

PurposeThe aim of this work was to establish a methodological approach for creation and optimization of an atlas for auto‐contouring, using the commercial software MIM MAESTRO (MIM Software Inc. Cleveland OH).MethodsA computed tomography (CT) male pelvis atlas was created and optimized to evaluate how different tools and options impact on the accuracy of automatic segmentation. Pelvic lymph nodes (PLN), rectum, bladder, and femurs of 55 subjects were reviewed for consistency by a senior consultant radiation oncologist with 15 yr of experience. Several atlas and workflow options were tuned to optimize the accuracy of auto‐contours. The deformable image registration (DIR), the finalization method, the k number of atlas best matching subjects, and several post‐processing options were studied. To test our atlas performances, automatic and reference manual contours of 20 test subjects were statistically compared based on dice similarity coefficient (DSC) and mean distance to agreement (MDA) indices. The effect of field of view (FOV) reduction on auto‐contouring time was also investigated.ResultsWith the optimized atlas and workflow, DSC and MDA median values of bladder, rectum, PLN, and femurs were 0.91 and 1.6 mm, 0.85 and 1.6 mm, 0.85 and 1.8 mm, and 0.96 and 0.5 mm, respectively. Auto‐contouring time was more than halved by strictly cropping the FOV of the subject to be contoured to the pelvic region.ConclusionA statistically significant improvement of auto‐contours accuracy was obtained using our atlas and optimized workflow instead of the MIM Software pelvic atlas.

A Patient-Specific Autosegmentation Strategy Using Multi-Input Deformable Image Registration for Magnetic Resonance Imaging–Guided Online Adaptive Radiation Therapy: A Feasibility Study

PurposeMagnetic resonance-guided online adaptive radiation therapy (MRgOART) requires accurate and efficient segmentation. However, the performance of current autosegmentation tools is generally poor for magnetic resonance imaging (MRI) owing to day-to-day variations in image intensity and patient anatomy. In this study, we propose a patient-specific autosegmentation strategy using multiple-input deformable image registration (DIR; PASSMID) to improve segmentation accuracy and efficiency for MRgOART. Methods and materialsLongitudinal MRI scans acquired on a 1.5T MRI-Linac for 10 patients with abdominal cancer were used. The proposed PASSMID includes 2 steps: applying a patient-specific image processing pipeline to longitudinal MRI scans, and populating all contours from previous sessions/fractions to a new fractional MRI using multiple DIRs and combining the resulted contours using simultaneous truth and performance level estimation (STAPLE) to obtain the final consensus segmentation. Five contour propagation strategies were compared: planning computed tomography to fractional MRI scans through rigid body registration (RDR), pretreatment MRI to fractional MRI scans through RDR and DIR, and the proposed multi-input DIR/STAPLE without preprocessing, and the PASSMID. Dice similarity coefficient (DSC) and mean distance to agreement (MDA) with ground truth contours were calculated slice by slice to quantify the contour accuracy. A quantitative index, defined as the ratio of acceptable slices, was introduced using a criterion of DSC > 0.8 and MDA < 2 mm. ResultsThe proposed PASSMID performed well with an average 2-dimensional DSC/MDA of 0.94/1.78 mm, 0.93/1.04 mm, 0.93/1.06 mm, 0.93/1.14 mm, 0.92/0.83 mm, 0.84/1.53 mm, 0.86/2.39 mm, 0.81/2.49 mm, 0.72/5.48 mm, and 0.70/5.03 mm for the liver, left kidney, right kidney, spleen, aorta, pancreas, stomach, duodenum, small bowel, and colon, respectively. Starting from the third fractions, the contour accuracy was significantly improved with PASSMID compared with the single-DIR strategy (P < .05). The mean ratio of acceptable slices were 13.9%, 17.5%, 60.8%, 70.6%, and 71.8% for the 5 strategies, respectively. ConclusionsThe proposed PASSMID solution, by combining image processing, multi-input DIRs, and STAPLE, can significantly improve the accuracy of autosegmentation for intrapatient MRI scans, reducing the time required for further contour editing, thereby facilitating the routine practice of MRgOART.

Assessment of Intrafraction Cervico-Uterine Motion in the Definitive Treatment of Locally Advanced Cervical Cancer Using MR-Guided Radiotherapy

To account for intrafraction motion, patients receiving intensity modulated radiation therapy (IMRT) for the definitive treatment of cervical cancer have an internal target volume (ITV) defined from bladder-full and bladder-empty CT simulation scans. We hypothesized that through daily volumetric magnetic resonance (MR) imaging for setup and treatment, we can assess cervico-uterine motion and create an intrafraction-ITV independent of interfraction bladder changes, and therefore dispense with the full-bladder to empty-bladder ITV paradigm. In this IRB-approved, study patients with locally advanced cervical cancer received IMRT with brachytherapy. IMRT was delivered on a clinical MR-guided radiotherapy (MRgRT) Linac. All patients received a pre-treatment (pre-tx) and post-treatment (post-tx) 0.35T balanced steady-state free precession MR scans in treatment position. The intrafraction cervico-uterine motion was assessed through segmentation of the clinical target volume (CTV) by a gynecological radiation oncologist on each scan. The pre-tx and post-tx MR scans were in the same frame of reference, and any residual patient motion was removed through a rigid registration of the post-tx MR scan to the pre-tx MR scan based on bony anatomy. All post-tx contours were rigidly copied to the pre-tx MR scans. The maximum and mean difference in spatial position between the pre-tx and post-tx MR scans were quantified using maximum Hausdorff distance (HD) and mean distance to agreement (MDA), respectively. Mean, standard deviation, and 95% confidence interval were calculated to determine the intrafraction MR-based internal target volume (ITVMRgRT). In total, 19 same day sessions (N = 38 scans) of intrafraction motion were assessed. Minimal patient voluntary motion was observed across rigid registrations with respect to bony anatomy. The average time between the pre-tx and post-tx MR scans was 14.3 ± 7.1 minutes (mean ± standard deviation), resembling a hypothetical time of on-table adaptation and treatment. The average maximum HD was 0.88 ± 0.46 cm (95% CI 0.668 - 1.08) and average MDA was 0.08 ± 0.07 cm (95% CI 0.05- 0.118). Due to HD’s sensitivity to outliers, a more robust metric (i.e., MDA) was utilized for margin calculation. The ITVMRgRT margin necessary to encompass the MDA for 95% of the intrafraction motion was 0.12 cm; therefore, we propose an ITVMRgRT of 0.15 cm given the non-partial voxelization of segmentation. This work supports margin reduction of the ITV to ITVMRgRT of 0.15 cm when using daily MR-guided online adaptive radiotherapy. Given the maximum setup uncertainty is limited in MRgRT to spatial resolution of the real-time cine imaging sequence, we propose an ITVMRgRT-PTV margin of 0.3 cm. Therefore, our total CTV to PTV margin is 0.45 cm.

Development of an automated radiotherapy dose accumulation workflow for locally advanced high-risk prostate cancer - A technical report.

An automated dose accumulation and contour propagation workflow using daily cone beam computed tomography (CBCTs) images for prostate cases that require pelvic lymph nodes irradiation (PLNs) was developed. This workflow was constructed using MIM® software with the intention to provide accurate dose transformations for plans with two different isocentres, whereby two sequential treatment phases were prescribed. The pre‐processing steps for data extractions from treatment plans, CBCTs, determination of couch shift information and management of missing CBCTs are described. To ensure that the imported translational couch shifts were in the correct orientation and readable in MIM, phantom commissioning was performed. For dose transformation, rigid registration with corrected setup shifts and scaled fractional dose was performed for pCT to daily CBCTs, which were then deformed onto CBCT1. Fractional dose summation resulted in the final accumulated dose for the patient allowing differences in dosimetry between the planned and accumulated dose to be analysed. Contour propagations of the prostate, bladder and rectum were performed within the same workflow. Transformed contours were then deformed onto daily CBCTs to generate trending reports for analysis, including Dice Similarity Coefficient (DSC) and Mean Distance to Agreement (MDA). Results obtained from phantom commissioning (DSC = 0.96, MDA = 0.89 mm) and geometrical analysis of the propagated contours for twenty patients; prostate (DSC: 0.9 ± 0.0, MDA: 1.0 ± 0.3 mm), rectum (DSC: 0.8 ± 0.1, mm, MDA: 1.7 ± 0.6 mm) and bladder (DSC: 0.8 ± 0.1, MDA: 2.8 ± 1.0 mm) were within clinically accepted tolerances for both DSC (>0.8) and MDA (< 0.3 mm). The developed workflow is being performed on a larger patient cohort for predictive model building, with the goal of correlating observed toxicity with the actual accumulated dose received by the patient.

Evaluation of Automatic Segmentation Model With Dosimetric Metrics for Radiotherapy of Esophageal Cancer

Background and Purpose: Automatic segmentation model is proven to be efficient in delineation of organs at risk (OARs) in radiotherapy; its performance is usually evaluated with geometric differences between automatic and manual delineations. However, dosimetric differences attract more interests than geometric differences in the clinic. Therefore, this study aimed to evaluate the performance of automatic segmentation with dosimetric metrics for volumetric modulated arc therapy of esophageal cancer patients.Methods: Nineteen esophageal cancer cases were included in this study. Clinicians manually delineated the target volumes and the OARs for each case. Another set of OARs was automatically generated using convolutional neural network models. The radiotherapy plans were optimized with the manually delineated targets and the automatically delineated OARs separately. Segmentation accuracy was evaluated by Dice similarity coefficient (DSC) and mean distance to agreement (MDA). Dosimetric metrics of manually and automatically delineated OARs were obtained and compared. The clinically acceptable dose difference and volume difference of OARs between manual and automatic delineations are supposed to be within 1 Gy and 1%, respectively.Results: Average DSC values were greater than 0.92 except for the spinal cord (0.82), and average MDA values were <0.90 mm except for the heart (1.74 mm). Eleven of the 20 dosimetric metrics of the OARs were not significant (P > 0.05). Although there were significant differences (P < 0.05) for the spinal cord (D2%), left lung (V10, V20, V30, and mean dose), and bilateral lung (V10, V20, V30, and mean dose), their absolute differences were small and acceptable for the clinic. The maximum dosimetric metrics differences of OARs between manual and automatic delineations were ΔD2% = 0.35 Gy for the spinal cord and ΔV30 = 0.4% for the bilateral lung, which were within the clinical criteria in this study.Conclusion: Dosimetric metrics were proposed to evaluate the automatic delineation in radiotherapy planning of esophageal cancer. Consequently, the automatic delineation could substitute the manual delineation for esophageal cancer radiotherapy planning based on the dosimetric evaluation in this study.

Manual and semi-automated delineation of locally advanced rectal cancer subvolumes with diffusion-weighted MRI.

To evaluate interobserver agreement for T2 weighted (T2W) and diffusion-weighted MRI (DW-MRI) contours of locally advanced rectal cancer (LARC); and to evaluate manual and semi-automated delineations of restricted diffusion tumour subvolumes. 20 cases of LARC were reviewed by 2 radiation oncologists and 2 radiologists. Contours of gross tumour volume (GTV) on T2W, DW-MRI and co-registered T2W/DW-MRI were independently delineated and compared using Dice Similarity Coefficient (DSC), mean distance to agreement (MDA) and other metrics of interobserver agreement. Restricted diffusion subvolumes within GTVs were manually delineated and compared to semi-automatically generated contours corresponding to intratumoral apparent diffusion coefficient (ADC) centile values. Observers were able to delineate subvolumes of restricted diffusion with moderate agreement (DSC 0.666, MDA 1.92 mm). Semi-automated segmentation based on the 40th centile intratumoral ADC value demonstrated moderate average agreement with consensus delineations (DSC 0.581, MDA 2.44 mm), with errors noted in image registration and luminal variation between acquisitions. A small validation set of four cases with optimised planning MRI demonstrated improvement (DSC 0.669, MDA 1.91 mm). Contours based on co-registered T2W and DW-MRI could be used for delineation of biologically relevant tumour subvolumes. Semi-automated delineation based on patient-specific intratumoral ADC thresholds may standardise subvolume delineation if registration between acquisitions is sufficiently accurate. This is the first study to evaluate the feasibility of semi-automated diffusion-based subvolume delineation in LARC. This approach could be applied to dose escalation or 'dose painting' protocols to improve delineation reproducibility.

Influence of different urinary bladder filling levels and controlling regions of interest selection on deformable image registration algorithms

PurposeEvaluation of Raystation ANAtomically CONstrained Deformation Algorithm (ANACONDA) performance to different urinary bladder filling levels in male pelvis anatomic site varying the controlling Regions Of Interest (ROIs). MethodsDifferent image datasets were obtained with ImSimQA (Oncology System Limited, Shrewsbury, UK) to evaluate ANACONDA performances (RaySearch Laboratories, Stockholm, Sweden). Deformation vector fields were applied to a synthetic man pelvis and a real patient computed tomography (CT) dataset (reference CTs) resulting in deformed CTs (target CTs) with various bladder filling levels. Different deformable image registrations (DIRs) were generated between each target CTs and reference CTs varying the controlling ROIs subset. Deformed ROIs were mapped from target CT to reference CT and then compared to reference ROIs. Evaluation was performed by Dice Similarity Coefficient (DSC), Correlation Coefficient (CC), Mean Distance to Agreement (MDA), maximum Distance to Agreement (maxDA) and with the introduction of global DSC (global_DSC) and global CC (global_CC) parameters. ResultsIn both synthetic and real patient CT cases, DSC scored less than 0.75 and MDA greater than 3 mm when no ROIs or only bladder were exploited as controlling ROI. DSC and CC increased by increasing the number of controlling ROIs selected whereas, an opposite behavior was observed for MDA and maxDA. ConclusionsANACONDA performances can be influenced by bladder filling fluctuation if no controlling ROIs are selected. Global_DSC and global_CC are useful parameters to quantitatively compare DIR algorithms. DIR performances improve by increasing the number of controlling ROIs selected, reaching a saturation level after a defined ROIs subset selection.

A Patient-Specific Autosegmentation Strategy Using Multi-Input Deformable Image Registration for Magnetic Resonance Imaging–Guided Online Adaptive Radiation Therapy: A Feasibility Study

PurposeMagnetic resonance-guided online adaptive radiation therapy (MRgOART) requires accurate and efficient segmentation. However, the performance of current autosegmentation tools is generally poor for magnetic resonance imaging (MRI) owing to day-to-day variations in image intensity and patient anatomy. In this study, we propose a patient-specific autosegmentation strategy using multiple-input deformable image registration (DIR; PASSMID) to improve segmentation accuracy and efficiency for MRgOART. Methods and materialsLongitudinal MRI scans acquired on a 1.5T MRI-Linac for 10 patients with abdominal cancer were used. The proposed PASSMID includes 2 steps: applying a patient-specific image processing pipeline to longitudinal MRI scans, and populating all contours from previous sessions/fractions to a new fractional MRI using multiple DIRs and combining the resulted contours using simultaneous truth and performance level estimation (STAPLE) to obtain the final consensus segmentation. Five contour propagation strategies were compared: planning computed tomography to fractional MRI scans through rigid body registration (RDR), pretreatment MRI to fractional MRI scans through RDR and DIR, and the proposed multi-input DIR/STAPLE without preprocessing, and the PASSMID. Dice similarity coefficient (DSC) and mean distance to agreement (MDA) with ground truth contours were calculated slice by slice to quantify the contour accuracy. A quantitative index, defined as the ratio of acceptable slices, was introduced using a criterion of DSC > 0.8 and MDA < 2 mm. ResultsThe proposed PASSMID performed well with an average 2-dimensional DSC/MDA of 0.94/1.78 mm, 0.93/1.04 mm, 0.93/1.06 mm, 0.93/1.14 mm, 0.92/0.83 mm, 0.84/1.53 mm, 0.86/2.39 mm, 0.81/2.49 mm, 0.72/5.48 mm, and 0.70/5.03 mm for the liver, left kidney, right kidney, spleen, aorta, pancreas, stomach, duodenum, small bowel, and colon, respectively. Starting from the third fractions, the contour accuracy was significantly improved with PASSMID compared with the single-DIR strategy (P < .05). The mean ratio of acceptable slices were 13.9%, 17.5%, 60.8%, 70.6%, and 71.8% for the 5 strategies, respectively. ConclusionsThe proposed PASSMID solution, by combining image processing, multi-input DIRs, and STAPLE, can significantly improve the accuracy of autosegmentation for intrapatient MRI scans, reducing the time required for further contour editing, thereby facilitating the routine practice of MRgOART.

Validation of an MR-guided online adaptive radiotherapy (MRgoART) program: Deformation accuracy in a heterogeneous, deformable, anthropomorphic phantom

Background and purposeTo investigate deformable image registration (DIR) and multi-fractional dose accumulation accuracy of a clinical MR-guided online adaptive radiotherapy (MRgoART) program, utilizing clinically-based magnitudes of abdominal deformation vector fields (DVFs). Materials and methodsA heterogeneous anthropomorphic multi-modality abdominal deformable phantom was comprised of MR and CT anatomically-relevant materials. Thermoluminescent dosimeters (TLDs) were affixed within regions of interest (ROIs). CT and MR simulation scans were acquired. CT was deformed to MR for dose calculations. MRgoART was executed on a MR-linac (MRIdian) for 5 Gy/5 fractions. Before each fraction, a deformation was applied. Ground truth was known for ROI volume, TLD position, and TLD dose measured by an accredited dosimetry calibration laboratory. To validate the range of applied deformations, phantom DVFs were compared to DVFs of clinical abdominal MRgoART fractions. MR-MR deformation accuracy was quantified through dice similarity coefficient (DSC), Hausdorff distance (HD), mean distance-to-agreement (MDA), and as mean-absolute-error (MAE) for CT-MR-MR deformation. Arithmetic-summation of calculated dose at respective TLD positions and deform-accumulated dose (MIM) was compared to TLD measured dose, respectively. MR-MR deformation statistics were quantified for MRIdian and MIM. ResultsMean phantom DVFs were 5.0 ± 2.9 mm compared to mean DVF of clinical abdominal patients at 5.2 ± 3.0 mm. Respective mean DSC, HD, MDA was 0.93 ± 0.03, 0.74 ± 0.80 cm, 0.08 ± 0.03 cm for MRIdian and 0.93 ± 0.03, 0.54 ± 0.27 cm, 0.08 ± 0.03 cm for MIM (N = 80 ROIs). Mean MAE was 20.5 HU. Respective mean and median dose differences were 0.3%, −0.3% for arithmetic-summation and 4.1%, 0.6% for deformed-accumulation. Maximum differences were 0.21 Gy (arithmetic-summation), 0.31 Gy (deformed-accumulation). ConclusionsMRgoART deformation and dosimetric accuracy has been benchmarked for mean fractional DVFs of 5 mm in a multiple-rigid-body deformable phantom. Deformation accuracy was within TG132 criteria and clinically acceptable end-to-end MRgoART dosimetric agreement was observed for this phantom. Further efforts are needed in validation of deform-accumulated dose.

Atlas based segmentation in prone breast cancer radiation therapy

The evidence supporting atlas based contour generation is growing and includes breast, prostate, central nervous system, gastrointestinal, gynecologic, and head and neck cancer patient populations. We sought to investigate atlas based segmentation (ABS) in patients with early stage breast cancer status post breast conserving surgery treated with adjuvant radiation therapy in the prone position. An initial atlas library was generated by uploading 20 previously treated patients. Subsequently, a group of 20 consecutive patients underwent treatment planning. Heart, right lung, left lung, total lungs, and breast clinical target volume (CTV) targets were manually contoured per our standard workflow. ABS was then incorporated into our radiation planning workflow and differences in contouring time were recorded, including time needed for ABS volume refinement. ABS generated volumes were compared subjectively in the unedited and edited stages by an independent radiation oncologist to reduce bias by incorporating an interobserver quality analysis. Various objective measurements were used to compare target volume quality including mean distance to agreement (MDA), Dice Coefficient (DC), logit transformation of DC (logit(DC)). The contouring physician edited 88.75% of organ at risk (OAR) volumes on average per patient, whereas the independent reviewing physician recommended revision of 27.5% OAR on average volumes per patient. CTV editing was performed in 20/20 (100%) of cases by the contouring physician, whereas CTV revision was recommended by the independent reviewing physician in 4/20 (20%) of cases. Our atlas performed well with DC values of >0.909 and logit(DC) of >2.344 across heart, lung, and breast volumes when compared to manually generated volumes. All objective measurements demonstrated improvement with physician refinement of ABS generated volumes. The largest absolute improvement was seen in the heart and breast CTV targets. There was 100% acceptance of the edited ABS generated volumes by the independent reviewing physician. The average time saved using ABS was 6.27 minutes (56.92%) per patient. This study confirms ABS offers improvements in efficiency without sacrificing contour quality in the early stage breast cancer patient population and demonstrates the functionality of ABS with prone patient positioning.

Cardiac substructure segmentation with deep learning for improved cardiac sparing.

Radiation dose to cardiac substructures is related to radiation-induced heart disease. However, substructures are not considered in radiation therapy planning (RTP) due to poor visualization on CT. Therefore, we developed a novel deep learning (DL) pipeline leveraging MRI's soft tissue contrast coupled with CT for state-of-the-art cardiac substructure segmentation requiring a single, non-contrast CT input. Thirty-two left-sided whole-breast cancer patients underwent cardiac T2 MRI and CT-simulation. A rigid cardiac-confined MR/CT registration enabled ground truth delineations of 12 substructures (chambers, great vessels (GVs), coronary arteries (CAs), etc.). Paired MRI/CT data (25 patients) were placed into separate image channels to train a three-dimensional (3D) neural network using the entire 3D image. Deep supervision and a Dice-weighted multi-class loss function were applied. Results were assessed pre/post augmentation and post-processing (3D conditional random field (CRF)). Results for 11 test CTs (seven unique patients) were compared to ground truth and a multi-atlas method (MA) via Dice similarity coefficient (DSC), mean distance to agreement (MDA), and Wilcoxon signed-ranks tests. Three physicians evaluated clinical acceptance via consensus scoring (5-point scale). The model stabilized in ~19h (200 epochs, training error <0.001). Augmentation and CRF increased DSC 5.0±7.9% and 1.2±2.5%, across substructures, respectively. DL provided accurate segmentations for chambers (DSC=0.88±0.03), GVs (DSC=0.85±0.03), and pulmonary veins (DSC=0.77±0.04). Combined DSC for CAs was 0.50±0.14. MDA across substructures was <2.0mm (GV MDA=1.24±0.31mm). No substructures had statistical volume differences (P>0.05) to ground truth. In four cases, DL yielded left main CA contours, whereas MA segmentation failed, and provided improved consensus scores in 44/60 comparisons to MA. DL provided clinically acceptable segmentations for all graded patients for 3/4 chambers. DL contour generation took ~14s per patient. These promising results suggest DL poses major efficiency and accuracy gains for cardiac substructure segmentation offering high potential for rapid implementation into RTP for improved cardiac sparing.

Structure guided deformable image registration for treatment planning CT and post stereotactic body radiation therapy (SBRT) Primovist® (Gd-EOB-DTPA) enhanced MRI.

The purpose of this study was to assess the performance of structure‐guided deformable image registration (SG‐DIR) relative to rigid registration and DIR using TG‐132 recommendations. This assessment was performed for image registration of treatment planning computed tomography (CT) and magnetic resonance imaging (MRI) scans with Primovist® contrast agent acquired post stereotactic body radiation therapy (SBRT). SBRT treatment planning CT scans and posttreatment Primovist® MRI scans were obtained for 14 patients. The liver was delineated on both sets of images and matching anatomical landmarks were chosen by a radiation oncologist. Rigid registration, DIR, and two types of SG‐DIR (using liver contours only; and using liver structures along with anatomical landmarks) were performed for each set of scans. TG‐132 recommended metrics were estimated which included Dice Similarity Coefficient (DSC), Mean Distance to Agreement (MDA), Target Registration Error (TRE), and Jacobian determinant. Statistical analysis was performed using Wilcoxon Signed Rank test. The median (range) DSC for rigid registration was 0.88 (0.77–0.89), 0.89 (0.81–0.93) for DIR, and 0.90 (0.86–0.94) for both types of SG‐DIR tested in this study. The median MDA was 4.8 mm (3.7–6.8 mm) for rigid registration, 3.4 mm (2.4–8.7 mm) for DIR, 3.2 mm (2.0–5.2 mm) for SG‐DIR where liver structures were used to guide the registration, and 2.8 mm (2.1–4.2 mm) for the SG‐DIR where liver structures and anatomical landmarks were used to guide the registration. The median TRE for rigid registration was 7.2 mm (0.5–23 mm), 6.8 mm (0.7–30.7 mm) for DIR, 6.1 mm (1.1–20.5 mm) for the SG‐DIR guided by only the liver structures, and 4.1 mm (0.8–19.7 mm) for SG‐DIR guided by liver contours and anatomical landmarks. The SG‐DIR shows higher liver conformality as per TG‐132 metrics and lowest TRE compared to rigid registration and DIR in Velocity AI software for the purpose of registering treatment planning CT and post‐SBRT MRI for the liver region. It was found that TRE decreases when liver contours and corresponding anatomical landmarks guide SG‐DIR.

Deep Learning Improved Clinical Target Volume Contouring Quality and Efficiency for Postoperative Radiation Therapy in Non-small Cell Lung Cancer.

Purpose: To investigate whether a deep learning-assisted contour (DLAC) could provide greater accuracy, inter-observer consistency, and efficiency compared with a manual contour (MC) of the clinical target volume (CTV) for non-small cell lung cancer (NSCLC) receiving postoperative radiotherapy (PORT).Materials and Methods: A deep dilated residual network was used to achieve the effective automatic contour of the CTV. Eleven junior physicians contoured CTVs on 19 patients by using both MC and DLAC methods independently. Compared with the ground truth, the accuracy of the contour was evaluated by using the Dice coefficient and mean distance to agreement (MDTA). The coefficient of variation (CV) and standard distance deviation (SDD) were rendered to measure the inter-observer variability or consistency. The time consumed for each of the two contouring methods was also compared.Results: A total of 418 CTV sets were generated. DLAC improved contour accuracy when compared with MC and was associated with a larger Dice coefficient (mean ± SD: 0.75 ± 0.06 vs. 0.72 ± 0.07, p < 0.001) and smaller MDTA (mean ± SD: 2.97 ± 0.91 mm vs. 3.07 ± 0.98 mm, p < 0.001). The DLAC was also associated with decreased inter-observer variability, with a smaller CV (mean ± SD: 0.129 ± 0.040 vs. 0.183 ± 0.043, p < 0.001) and SDD (mean ± SD: 0.47 ± 0.22 mm vs. 0.72 ± 0.41 mm, p < 0.001). In addition, a value of 35% of time saving was provided by the DLAC (median: 14.81 min vs. 9.59 min, p < 0.001).Conclusions: Compared with MC, the DLAC is a promising strategy to obtain superior accuracy, consistency, and efficiency for the PORT-CTV in NSCLC.

Quantifying the accuracy of deformable image registration for cone-beam computed tomography with a physical phantom.

PurposeKilo‐voltage cone‐beam computed tomography (CBCT) is widely used for patient alignment, contour propagation, and adaptive treatment planning in radiation therapy. In this study, we evaluated the accuracy of deformable image registration (DIR) for CBCT under various imaging protocols with different noise and patient dose levels.MethodsA physical phantom previously developed to facilitate end‐to‐end testing of the DIR accuracy was used with Varian Velocity v4.0 software to evaluate the performance of image registration from CT to CT, CBCT to CT, and CBCT to CBCT. The phantom is acrylic and includes several inserts that simulate different tissue shapes and properties. Deformations and anatomic changes were simulated by changing the rotations of both the phantom and the inserts. CT images (from a head and neck protocol) and CBCT images (from pelvis, head and “Image Gently” protocols) were obtained with different image noise and dose levels. Large inserts were filled with Mobil DTE oil to simulate soft tissue, and small inserts were filled with bone materials. All inserts were contoured before the DIR process to provide a ground truth contour size and shape for comparison. After the DIR process, all deformed contours were compared with the originals using Dice similarity coefficient (DSC) and mean distance to agreement (MDA). Both large and small volume of interests (VOIs) for DIR volume selection were tested by simulating a DIR process that included whole patient image volume and clinical target volumes (CTV) only (for CTVs propagation).ResultsFor cross‐modality DIR registration (CT to CBCT), the DSC were >0.8 and the MDA were <3 mm for CBCT pelvis, and CBCT head protocols. For CBCT to CBCT and CT to CT, the DIR accuracy was improved relative to the cross‐modality tests. For smaller VOIs, the DSC were >0.8 and MDA <2 mm for all modalities.ConclusionsThe accuracy of DIR depends on the quality of the CBCT image at different dose and noise levels.

Evaluation of Deformable Dose Accumulation Accuracy for Head and Neck Cancer

Accurate dose summation is crucial to adaptive radiotherapy, but moderate and large anatomic changes (e.g., tumor shrinkage, weight loss) present challenges to daily dose accumulation based on deformable image registration (DIR). This study sought to evaluate the accuracy of deformable dose accumulation commonly encountered during radiation therapy (RT) for head and neck cancer (HNC). Data for eight HNC patients were selected in this study. For each case, a planning kVCT and 15 daily MVCTs acquired during MVCT-guided RT on a helical tomotherapy system were analyzed. A commercial DIR algorithm was used to register the planning kVCT to each MVCT. The registration accuracy was evaluated using mean distance to agreement (MDA) and Dice coefficient (DC). The resultant displacement vector fields (DVFs) were used to map the doses calculated on daily MVCT to the planning image. The dose in the daily MVCTs and the deformed dose in the planning CT were converted to radiation energy by multiplying these doses by mass density in each voxel, where the mass density at each image voxel is converted from the Hounsfield units (HU) using HU-to-mass-density tables calibrated for the kVCT and MVCT scanners, respectively. Changes in dose, volume and energy before and after mapping were compared to each other for both planning target volume (PTV) and organ-at-risk (OAR), and the correlations of these changes were quantified to assess uncertainties in the dose mapping operations. For the eight cases, mean MDA and standard deviation were 0.45±0.09, 0.46±0.06, 0.44±0.01 mm and the DC were 0.958±0.02, 0.946±0.02, 0.949±0.01 for the PTV, left and right parotid respectively. Each of those result conforms to the AAPM task group 132 (TG-132) recommendations (DC > 0.8, MDA < CT voxel width: 2 mm). The average volume changes from fraction 1 to fraction 15 for PTV, left and right parotids are 6.3±3.0%, 5.3±1.5%, and 5.7±3.6%, respectively. The changes of V30 for dose calculated on daily MVCT and warped dose on planning CT were 10.0% and 7.8% for left parotid, and 28.4% and 27.1% for right parotid. The radiation dose (mean) mapped to the planning PTV is within 99% of the dose calculated in daily MVCT images, and radiation energy calculated in PTV and deformed PTVs differ 4.6±3.2% on average. In the left and right parotids, average energy losses were 7.6±2.5%, and 7.9±3.3%, and their mean dose differences are 1.2±0.3% (0.29 cGy) and 1.3±0.5% (0.58 cGy). For the 8 patients, the dose changes are positively correlated to the energy differences with their correlation coefficients equal to 0.56, 0.52 and 0.54 for the PTV, left and right parotids, respectively. The DIR algorithm performed led to acceptable results based on AAPM TG-132. The mean dose, DVH and total deposited energy showed sensitivity to volume changes in an increasing order. It is valuable to include these metrics in a comprehensive QA protocol for adaptive radiotherapy.

Evaluation of a Deep Learning-Based Auto-Segmentation Tool for Online Adaptive Radiation Therapy

Fast and consistent segmentation of tumors and organs at risk on daily images is vital for online adaptive radiation therapy (OART). For this purpose, most sought-after method to-date is auto-segmentation, built upon machine learning algorithms. This study evaluates the performance of a Deep Learning based automatic segmentation model for OART. A research version of Deep Learning (DL) auto-segmentation model, using Deep Convolutional Neural Network (DCNN) for auto-segmentation on CT and MRI is evaluated based on CT and MRI of randomly selected 50 prostate cancer patients. The model’s architecture belongs to Fully Convolutional Neural Network, with end-to-end mapping and sequential application of 2.5D and 3D models for 3D image segmentation. Enhanced convergence is achieved via short- and long-range residue and U-Net skip connections. Data augmentation through random transformations of input images and respective segmentation maps was used in model training. For testing CT sets, contours of prostate, rectum and bladder were drawn manually based on the CT and MRI registration by a physician or were verified by physician or physician dictated criteria. The clinical viability of the DL contours was checked by qualitative and quantitative comparison with manuals, ground truth contours, and AAPM TG-132 accuracy tolerance recommendations for Dice Similarity Coefficient (DSC) > 0.8 to 0.9 and Mean Distance to Agreement (MDA) < 2 to 3 mm were enforced. The auto-segmentation tool was run on an Intel Xeon CPU (2.4 GHz, 64 GB RAM). The computing times for all the auto-segmentation tests were recorded. For the cases tested, 98% of the auto-segmented contours remained within TG-132 dictated tolerances indicating clinical acceptability of the contours. The mean and standard deviations of DSC, MDA and Hausdorff Distance (HD) values were tabulated and were found to be (0.87 ± 0.037, 1.38 ± 0.393, 7.5 ± 3.07), (0.89 ± 0.053, 1.13 ± 0.752, 11.2 ± 6.37) and (0.95 ± 0.019, 0.91 ± 0.413, 6.86 ± 5.04) for prostate, rectum and bladder, respectively. A detailed slice-by-slice visual inspection indicated that approximately 20% of the DL contours were fully acceptable, requiring no edits. For the rest, accommodation of anatomical challenges, like abnormal bladder shapes and indistinct organ boundaries (35%), or edits subjective to physician’s discretion, required minor manual editing, with average editing time < 2 minutes. On average the execution of the auto-segmentations took about 0.5 minutes. The qualitative and quantitative evaluation of the newly developed DL-based auto-segmentation model prove it to be accurate and efficient, with clinically acceptable structures for male pelvis. The auto-segmentation tool can be used to generate contours for male pelvic organs with or without minor editing within 3 minutes, particularly applicable for online adaptive replanning.

Can Convolutional Neural Network Delineate OAR in Lung Cancer More Accurately and Efficiently Compared with Atlas Based Method？

Delineation of organ at risk (OAR) is important but time-consuming for radiotherapy planning. Automatic segmentation of OAR based on a convolutional neural networks (CNN) with 101 layers has been established for lung cancer patients in our institution. The aim of this study was to compare the efficacy and accuracy of two automatic OAR contouring methods, CNN based segmentation and atlas-based contouring. The OARs, including the lungs, esophagus, heart and spinal cord, of ten NSCLC patients were delineated using three methods: (1) automatic segmentation based on CNN(AS-CNN); (2) automatic segmentation based on atlas by treatment planning software; (3) manually delineated(MD) by a senior radiation oncologist. The MD was used the ground truth for reference. The results of the automated contouring methods were compared by delineation time, Hausdorff Distance(HD), Mean Distance to Agreement(MDA), and Dice similarity coefficient (DSC).Paired t test was used to compare the accuracy between different methods. The manual delineation time was 21.7±2.9 min. Automatic segmentation time of AS-CNN and AS-Atlas was 1.6±0.2min(on a Titan XP graphics card) and 2.2±0.1min, which were only 7% and 10% of manual delineation time, respectively. AS-CNN was more efficiency than AS-Atlas(P<0.001). In terms of accuracy, the MDA and DSC were listed in table 1. The esophagus AS-CNN performed well, with a DSC of 0.77, while the AS-Atlas of esophagus was unavailable. AS-CNN showed marginal better than AS-Atlas in the segmentation of heart. The AS-CNN of spinal cord performed slightly worse the AS-Atlas. The results were similar in contouring of both lungs. AS-CNN significantly reduces contouring time compared with AS-Atlas. AS-CNN is viable and performed well, especially for heart and esophagus, while AS-Atlas performed better for the spinal cord. However, careful review and manually modifying is still required for both methods to assure the safety of radiotherapy. AS-CNN was in continuous evolution as more clinical data was collected in the database for AS-CNN, and it could provide more benefit.Abstract 3246; Table 1MDA and DSC of AS-CNN and AS-Atlas in automatic segmentationOrgansParameterAS-CNNAS-AtlasP valueLeft lungMDA(mm)1.14±0.152.08±2.170.207DSC0.95±0.010.92±0.050.076Right lungMDA(mm)2.87±3.092.78±3.270.404DSC0.94±0.020.94±0.020.892HeartMDA(mm)1.67±0.463.63±3.370.107DSC0.87±0.030.77±0.140.067Spinal cordMDA(mm)1.02±0.180.55±0.100.000DSC0.80±0.050.88±0.020.000 Open table in a new tab

From Multimodality to Monomodality: Head and Neck MR-CT Registration with Synthetic Image Bridge

To develop and demonstrate the efficacy of a novel head-and-neck multimodality image registration technique using generative adversarial networks, and show its improvement over traditional direct multimodal registration. 25 head-and-neck patients received MR and CT (CTaligned) scans on the same day with the same immobilization. 19 of the MR-CT pairs were used to train a neural network to generate synthetic CTs from MR images while the registration testing was performed on the remaining 6 patients, who also had a separate CT without immobilization (CTnon-aligned). The CTnon-aligned images exhibited large neck motion. CTnon-aligned’s were deformed to the synthetic CT using b-splines and mutual information, thus effectively becoming a monomodal registration. These results were compared to CTnon-aligned directly registered to MR (multimodal). The same registrations were performed in the opposite direction. Results were evaluated using the mean distance to agreement (MDA) among spinal cord contours, landmark error, and Jacobian determinant of the estimated deformation fields. The cord contours are better matched when a synthetic CT replaces the MR, compared to a direct deformable registration with the MR. At larger initial misalignment (2-15mm MDA), the synthetic CT decreases the MDA by 1 - 10mm and 0.4 - 3mm for MR→CT and CT→MR directions, respectively. The average landmark error decreased from 10.95mm to 6.22mm when going from an MR→CTnon-aligned to a CTsynth→CTnon-aligned deformable registration. In the CT to MR direction, the average landmark error decreased from 8.19mm to 6.71mm when going from a CTnon-aligned→MR to a CTnon-aligned→CTsynth deformable registration. The proposed method demonstrated inverse consistency for both the contour and landmark analysis, while the direct method does not. Additionally, the Jacobian determinant showed feasible expansion and shrinking, with no infeasible space folding (negative values). We showed that using a synthetic CT in lieu of an MR for MR→CT and CT→MR deformable registration offers superior results to direct registration.

Volumetric and Voxel-Wise Analysis of Dominant Intraprostatic Lesions on Multiparametric MRI.

Introduction: Multiparametric MR imaging (mpMRI) has shown promising results in the diagnosis and localization of prostate cancer. Furthermore, mpMRI may play an important role in identifying the dominant intraprostatic lesion (DIL) for radiotherapy boost. We sought to investigate the level of correlation between dominant tumor foci contoured on various mpMRI sequences.Methods: mpMRI data from 90 patients with MR-guided biopsy-proven prostate cancer were obtained from the SPIE-AAPM-NCI Prostate MR Classification Challenge. Each case consisted of T2-weighted (T2W), apparent diffusion coefficient (ADC), and Ktrans images computed from dynamic contrast-enhanced sequences. All image sets were rigidly co-registered, and the dominant tumor foci were identified and contoured for each MRI sequence. Hausdorff distance (HD), mean distance to agreement (MDA), and Dice and Jaccard coefficients were calculated between the contours for each pair of MRI sequences (i.e., T2 vs. ADC, T2 vs. Ktrans, and ADC vs. Ktrans). The voxel wise spearman correlation was also obtained between these image pairs.Results: The DILs were located in the anterior fibromuscular stroma, central zone, peripheral zone, and transition zone in 35.2, 5.6, 32.4, and 25.4% of patients, respectively. Gleason grade groups 1–5 represented 29.6, 40.8, 15.5, and 14.1% of the study population, respectively (with group grades 4 and 5 analyzed together). The mean contour volumes for the T2W images, and the ADC and Ktrans maps were 2.14 ± 2.1, 2.22 ± 2.2, and 1.84 ± 1.5 mL, respectively. Ktrans values were indistinguishable between cancerous regions and the rest of prostatic regions for 19 patients. The Dice coefficient and Jaccard index were 0.74 ± 0.13, 0.60 ± 0.15 for T2W-ADC and 0.61 ± 0.16, 0.46 ± 0.16 for T2W-Ktrans. The voxel-based Spearman correlations were 0.20 ± 0.20 for T2W-ADC and 0.13 ± 0.25 for T2W-Ktrans.Conclusions: The DIL contoured on T2W images had a high level of agreement with those contoured on ADC maps, but there was little to no quantitative correlation of these results with tumor location and Gleason grade group. Technical hurdles are yet to be solved for precision radiotherapy to target the DILs based on physiological imaging. A Boolean sum volume (BSV) incorporating all available MR sequences may be reasonable in delineating the DIL boost volume.

Inter-observer variability of clinical target volume delineation in definitive radiotherapy of neck lymph node metastases from unknown primary. A cooperative study of the Italian Association of Radiotherapy and Clinical Oncology (AIRO) Head and Neck Group.

This study, promoted by Italian Association of Radiotherapy and Clinical Oncology (AIRO) Head and Neck Group, aimed to assess the current national practice of target volume delineation on a case of neck lymph node metastases from unknown primary evaluating inter-observer variability, in a setting of primary radiotherapy. A case of metastatic neck lymph node from occult primary was proposed to 17 radiation oncologists. A national reference RT center was identified and considered as benchmark. Participants were requested to delineate target volumes. A structured questionnaire was administered. A comparison between following parameters of the CTVs was performed: centroids distances, Dice similarity index (DSI), Jaccard index and mean distance to agreement (MDA). Volume expressed in cubic centimeters and CTVs cranio-caudal extension were evaluated. Sixteen of 17 radiation oncologists recommended three CTVs dose levels. (CTV HD, CTV ID and CTV LD); CTV ID was not delineated by one of the participants and by the reference center. The distance between the reference centroid and the mean centroid of CTVs HD was 1.09cm (0.36-3.99cm); for CTV LD, a mean centroids distance of 2.45 (0.27-4.83cm) was found, and for CTV HD, mean DSI is 0.48 and mean Jaccard index is 0.32 and MDA was 8.89mm. CTV LD showed a mean DSI of 0.46, mean Jaccard index of 0.31 and MDA of 14.87 when compared to the reference. Many aspects concerning treatment optimization of cervical nodes metastases from occult primary remain unclear, and we found a notable heterogeneity of global radiotherapy management reporting discordances both in target volume delineation and volume prescription.