Using neural networks to extend cropped medical images for deformable registration among images with differing scan extents.

PURPOSE: Missing or discrepant imaging volume is a common challenge in deformable image registration (DIR). To minimize the adverse impact, we train a neural network to synthesize cropped portions of head and neck CT's and then test its use in DIR. METHODS: Using a training dataset of 409 head and neck CT's, we trained a generative adversarial network to take in a cropped 3D image and output an image with synthesized anatomy in the cropped region. The network used a 3D U-Net generator along with Visual Geometry Group (VGG) deep feature losses. To test our technique, for each of the 53 test volumes, we used Elastix to deformably register combinations of a randomly cropped, full, and synthetically full volume to a single cropped, full, and synthetically full target volume. We additionally tested our method's robustness to crop extent by progressively increasing the amount of cropping, synthesizing the missing anatomy using our network, and then performing the same registration combinations. Registration performance was measured using 95% Hausdorff distance across 16 contours. RESULTS: We successfully trained a network to synthesize missing anatomy in superiorly and inferiorly cropped images. The network can estimate large regions in an incomplete image, far from the cropping boundary. Registration using our estimated full images was not significantly different from registration using the original full images. The average contour matching error for full image registration was 9.9 mm, whereas our method was 11.6, 12.1, and 13.6 mm for synthesized-to-full, full-to-synthesized, and synthesized-to-synthesized registrations, respectively. In comparison, registration using the cropped images had errors of 31.7 mm and higher. Plotting the registered image contour error as a function of initial preregistered error shows that our method is robust to registration difficulty. Synthesized-to-full registration was statistically independent of cropping extent up to 18.7 cm superiorly cropped. Synthesized-to-synthesized registration was nearly independent, with a -0.04 mm of change in average contour error for every additional millimeter of cropping. CONCLUSIONS: Different or inadequate in scan extent is a major cause of DIR inaccuracies. We address this challenge by training a neural network to complete cropped 3D images. We show that with image completion, the source of DIR inaccuracy is eliminated, and the method is robust to varying crop extent.

Multimodality image registration in the head-and-neck using a deep learning-derived synthetic CT as a bridge.

PURPOSE: To develop and demonstrate the efficacy of a novel head-and-neck multimodality image registration technique using deep-learning-based cross-modality synthesis. METHODS AND MATERIALS: Twenty-five head-and-neck patients received magnetic resonance (MR) and computed tomography (CT) (CTaligned ) scans on the same day with the same immobilization. Fivefold cross validation was used with all of the MR-CT pairs to train a neural network to generate synthetic CTs from MR images. Twenty-four of 25 patients also had a separate CT without immobilization (CTnon-aligned ) and were used for testing. CTnon-aligned 's were deformed to the synthetic CT, and compared to CTnon-aligned registered to MR. The same registrations were performed from MR to CTnon-aligned and from synthetic CT to CTnon-aligned . All registrations used B-splines for modeling the deformation, and mutual information for the objective. Results were evaluated using the 95% Hausdorff distance among spinal cord contours, landmark error, inverse consistency, and Jacobian determinant of the estimated deformation fields. RESULTS: When large initial rigid misalignment is present, registering CT to MRI-derived synthetic CT aligns the cord better than a direct registration. The average landmark error decreased from 9.8 ± 3.1 mm in MR→CTnon-aligned to 6.0 ± 2.1 mm in CTsynth →CTnon-aligned deformable registrations. In the CT to MR direction, the landmark error decreased from 10.0 ± 4.3 mm in CTnon-aligned →MR deformable registrations to 6.6 ± 2.0 mm in CTnon-aligned →CTsynth deformable registrations. The Jacobian determinant had an average value of 0.98. The proposed method also demonstrated improved inverse consistency over the direct method. CONCLUSIONS: We showed that using a deep learning-derived synthetic CT in lieu of an MR for MR→CT and CT→MR deformable registration offers superior results to direct multimodal registration.

Clinical experience using a video-guided spirometry system for deep inhalation breath-hold radiotherapy of left-sided breast cancer.

The purpose was to report clinical experience of a video-guided spirometry system in applying deep inhalation breath-hold (DIBH) radiotherapy for left-sided breast cancer, and to study the systematic and random uncertainties, intra- and interfraction motion and impact on cardiac dose associated with DIBH. The data from 28 left-sided breast cancer patients treated with spirometer-guided DIBH radiation were studied. Dosimetric comparisons between free-breathing (FB) and DIBH plans were performed. The distance between the heart and chest wall measured on the digitally reconstructed radiographs (DRR) and MV portal images, dDRR(DIBH) and dport(DIBH), respectively, was compared as a measure of DIBH setup uncertainty. The difference (Δd) between dDRR(DIBH) and dport(DIBH) was defined as the systematic uncertainty. The standard deviation of Δd for each patient was defined as the random uncertainty. MV cine images during radiation were acquired. Affine registrations of the cine images acquired during one fraction and multiple fractions were performed to study the intra- and interfraction motion of the chest wall. The median chest wall motion was used as the metric for intra- and interfraction analysis. Breast motions in superior-inferior (SI) direction and "AP" (defined on the DRR or MV portal image as the direction perpendicular to the SI direction) are reported. Systematic and random uncertainties of 3.8 mm and 2mm, respectively, were found for this spirometer-guided DIBH treatment. MV cine analysis showed that intrafraction chest wall motions during DIBH were 0.3mm in "AP" and 0.6 mm in SI. The interfraction chest wall motions were 3.6 mm in "AP" and 3.4 mm in SI. Utilization of DIBH with this spirometry system led to a statistically significant reduction of cardiac dose relative to FB treatment. The DIBH using video-guided spirometry provided reproducible cardiac sparing with minimal intra- and interfraction chest wall motion, and thus is a valuable adjunct to modern breast treatment techniques.

Toward optimizing patient-specific IMRT QA techniques in the accurate detection of dosimetrically acceptable and unacceptable patient plans.

PURPOSE: The authors investigated the performance of several patient-specific intensity-modulated radiation therapy (IMRT) quality assurance (QA) dosimeters in terms of their ability to correctly identify dosimetrically acceptable and unacceptable IMRT patient plans, as determined by an in-house-designed multiple ion chamber phantom used as the gold standard. A further goal was to examine optimal threshold criteria that were consistent and based on the same criteria among the various dosimeters. METHODS: The authors used receiver operating characteristic (ROC) curves to determine the sensitivity and specificity of (1) a 2D diode array undergoing anterior irradiation with field-by-field evaluation, (2) a 2D diode array undergoing anterior irradiation with composite evaluation, (3) a 2D diode array using planned irradiation angles with composite evaluation, (4) a helical diode array, (5) radiographic film, and (6) an ion chamber. This was done with a variety of evaluation criteria for a set of 15 dosimetrically unacceptable and 9 acceptable clinical IMRT patient plans, where acceptability was defined on the basis of multiple ion chamber measurements using independent ion chambers and a phantom. The area under the curve (AUC) on the ROC curves was used to compare dosimeter performance across all thresholds. Optimal threshold values were obtained from the ROC curves while incorporating considerations for cost and prevalence of unacceptable plans. RESULTS: Using common clinical acceptance thresholds, most devices performed very poorly in terms of identifying unacceptable plans. Grouping the detector performance based on AUC showed two significantly different groups. The ion chamber, radiographic film, helical diode array, and anterior-delivered composite 2D diode array were in the better-performing group, whereas the anterior-delivered field-by-field and planned gantry angle delivery using the 2D diode array performed less well. Additionally, based on the AUCs, there was no significant difference in the performance of any device between gamma criteria of 2%/2 mm, 3%/3 mm, and 5%/3 mm. Finally, optimal cutoffs (e.g., percent of pixels passing gamma) were determined for each device and while clinical practice commonly uses a threshold of 90% of pixels passing for most cases, these results showed variability in the optimal cutoff among devices. CONCLUSIONS: IMRT QA devices have differences in their ability to accurately detect dosimetrically acceptable and unacceptable plans. Field-by-field analysis with a MapCheck device and use of the MapCheck with a MapPhan phantom while delivering at planned rotational gantry angles resulted in a significantly poorer ability to accurately sort acceptable and unacceptable plans compared with the other techniques examined. Patient-specific IMRT QA techniques in general should be thoroughly evaluated for their ability to correctly differentiate acceptable and unacceptable plans. Additionally, optimal agreement thresholds should be identified and used as common clinical thresholds typically worked very poorly to identify unacceptable plans.

Effects of spatial resolution and noise on gamma analysis for IMRT QA.

We investigated the sensitivity of the gamma index to two factors: the spatial resolution and the noise level in the measured dose distribution. We also examined how the choice of reference distribution and analysis software affect the sensitivity of gamma analysis to these two factors for quality assurance (QA) of intensity-modulated radiation therapy (IMRT) treatment plans. For ten clinical IMRT plans, the dose delivered to a transverse dose plane was measured with EDR2 radiographic film. To evaluate the effects of spatial resolution, each irradiated film was digitized using three different resolutions (71, 142, and 285 dpi). To evaluate the effects of image noise, 1% and 2% local Gaussian noise was added to the film images. Gamma analysis was performed using 2%/2 mm and 3%/3 mm acceptance criteria and two commercial software packages, OmniPro I'mRT and DoseLab Pro. Dose comparisons were performed with the treatment planning system (TPS)-calculated dose as the reference, and then repeated with the film as the reference to evaluate how the choice of reference distribution affects the results of dose comparisons. When the TPS-calculated dose was designated as the reference distribution, the percentage of pixels with passing gamma values increased with both increasing resolution and noise. For 3%/3 mm acceptance criteria, increasing the film image resolution by a factor of two and by a factor of four caused a median increase of 0.9% and 2.6%, respectively, in the percentage of pixels passing. Increasing the noise level in the film image resulted in a median increase in percentage of pixels passing of 5.5% for 1% added local Gaussian noise and 5.8% for 2% added noise. In contrast, when the film was designated as the reference distribution, the percentage of pixels passing decreased with increased film noise, while increased resolution had no significant effect on passing rates. Furthermore, the sensitivity of gamma analysis to noise and resolution differed between OmniPro I'mRT and DoseLab Pro, with DoseLab Pro being less sensitive to the effects of noise and resolution. Noise and high scanning resolution can artificially increase the percentage of pixels with passing gamma values in IMRT QA. Thus, these factors, if not properly taken into account, can potentially affect the results of IMRT QA by causing a plan that should be classified as failing to be falsely classified as passing. In designing IMRT QA protocols, it is important to be aware that gamma analysis is sensitive to these parameters.

Reproducibility in patient-specific IMRT QA.

The purpose of this study was to determine the reproducibility of patient-specific, intensity-modulated radiation therapy (IMRT) quality assurance (QA) results in a clinical setting. Six clinical patient plans were delivered to a variety of devices and analyses, including 1) radiographic film; 2) ion chamber; 3) 2D diode array delivered and analyzed in three different configurations (AP delivery with field-by-field analysis, AP delivery with composite analysis, and planned gantry angle delivery); 4) helical diode array; and 5) in-house-designed multiple ion chamber phantom. The six clinical plans were selected from a range of treatment sites and were of various levels of complexity. Of note, three of the plans had failed at least preliminary evaluation with our in-house IMRT QA; the other three plans had passed QA. These plans were delivered three times sequentially without changing the setup, and then delivered two more times after breaking down and rebuilding the setup between each. This allowed for an investigation of reproducibility (in terms of dose, dose difference or percent of pixels passing gamma) of both the delivery and the physical setup. This study showed that the variability introduced from the setup was generally higher than the variability from redelivering the plan. Radiographic film showed the poorest reproducibility of the dosimeters investigated. In conclusion, the various IMRT QA systems demonstrated varying abilities to reproduce QA results consistently. All dosimetric devices demonstrated a reproducibility (coefficient of variation) of less than 4% in their QA results for all plans, with an average reproducibility of less than 2%. This work provides some quantification for the variability that may be seen for IMRT QA dosimeters.