Deformation Vector Fields Research Articles

Inter- and intrafraction motion assessment and accumulated dose quantification of upper gastrointestinal organs during magnetic resonance-guided ablative radiation therapy of pancreas patients.

Background and purposeStereotactic body radiation therapy (SBRT) of locally advanced pancreatic cancer (LAPC) is challenging due to significant motion of gastrointestinal (GI) organs. The goal of our study was to quantify inter and intrafraction deformations and dose accumulation of upper GI organs in LAPC patients.Materials and methodsFive LAPC patients undergoing five-fraction magnetic resonance-guided radiation therapy (MRgRT) using abdominal compression and daily online plan adaptation to 50 Gy were analyzed. A pre-treatment, verification, and post-treatment MR imaging (MRI) for each of the five fractions (75 total) were used to calculate intra and interfraction motion. The MRIs were registered using Large Deformation Diffeomorphic Metric Mapping (LDDMM) deformable image registration (DIR) method and total dose delivered to stomach_duodenum, small bowel (SB) and large bowel (LB) were accumulated. Deformations were quantified using gradient magnitude and Jacobian integral of the Deformation Vector Fields (DVF). Registration DVFs were geometrically assessed using Dice and 95th percentile Hausdorff distance (HD95) between the deformed and physician’s contours. Accumulated doses were then calculated from the DVFs.ResultsMedian Dice and HD95 were: Stomach_duodenum (0.9, 1.0 mm), SB (0.9, 3.6 mm), and LB (0.9, 2.0 mm). Median (max) interfraction deformation for stomach_duodenum, SB and LB was 6.4 (25.8) mm, 7.9 (40.5) mm and 7.6 (35.9) mm. Median intrafraction deformation was 5.5 (22.6) mm, 8.2 (37.8) mm and 7.2 (26.5) mm. Accumulated doses for two patients exceeded institutional constraints for stomach_duodenum, one of whom experienced Grade1 acute and late abdominal toxicity.ConclusionLDDMM method indicates feasibility to measure large GI motion and accumulate dose. Further validation on larger cohort will allow quantitative dose accumulation to more reliably optimize online MRgRT.

VoxelMorph++: a convolutional neural network architecture for unsupervised CBCT to CT deformable image registration

We use an unsupervised method based on the VoxelMorph architecture for Cone-beam computed tomography (CBCT) to CT deformable image registration (DIR), and propose VoxelMorph++, a new architecture for predicting the deformation vector field (DVF). The proposed architecture (1) overcomes the limitation that the optimal depth of encoder-decoder is unknown, by forming a nested structure where each feature with varying depth in the encoder path has a corresponding depth decoder; (2) fuses features of varying semantic scales more flexibly by redesigning skip connections. In the testing phase, we used ITK-SNAP software to semi-automatically segment the patients’ lung regions as labels to solve the problem of expensive manual labelling. We evaluated these two architectures using lung region registration results from 10 patients’ CBCT and CT images. After registration, the mean Dice score improved from 0.8556 to 0.9412 and 0.9430 for VoxelMorph and the proposed architecture, respectively. The results show that both architectures perform well in our dataset and the proposed architecture outperforms VoxelMorph in terms of registration accuracy.

Gastric deformation models for adaptive radiotherapy: Personalized vs population-based strategy

Background and purposeTo create a library of plans (LoP) for gastric cancer adaptive radiotherapy, accurate predictions of shape changes due to filling variations are essential. The ability of two strategies (personalized and population-based) to predict stomach shape based on filling was evaluated for volunteer and patient data to explore the potential for use in a LoP. Materials and methodsFor 19 healthy volunteers, stomachs were delineated on MRIs with empty (ES), half-full (HFS) and full stomach (FS). For the personalized strategy, a deformation vector field from HFS to corresponding ES was acquired and extrapolated to predict FS. For the population-based strategy, the average deformation vectors from HFS to FS of 18 volunteers were applied to the HFS of the remaining volunteer to predict FS (leave-one-out principle); thus, predictions were made for each volunteer. Reversed processes were performed to predict ES. To validate, for seven gastric cancer patients, the volunteer population-based model was applied to their pre-treatment CT to predict stomach shape on 2–3 repeat CTs. For all predictions, volume was made equal to true stomach volume. ResultsFS predictions were satisfactory, with median Dice similarity coefficient (mDSC) of 0.91 (population-based) and 0.89 (personalized). ES predictions were poorer: mDSC = 0.82 for population-based; personalized strategy yielded unachievable volumes. Population-based shape predictions (both ES and FS) were comparable between patients (mDSC = 0.87) and volunteers (0.88). ConclusionThe population-based model outperformed the personalized model and demonstrated its ability in predicting filling-dependent stomach shape changes and, therefore, its potential for use in a gastric cancer LoP.

Identification of modes of tumor regression in non-small cell lung cancer patients during radiotherapy.

Observed gross tumor volume (GTV) shrinkage during radiotherapy (RT) raises the question of whether to adapt treatment to changes observed on the acquired images. In the literature, two modes of tumor regression have been described: elastic and non-elastic. These modes of tumor regression will affect the safety of treatment adaptation. This study applies a novel approach, using routine cone-beam computed tomography (CBCT) and deformable image registration to automatically distinguish between elastic and non-elastic tumor regression. In this retrospective study, 150 locally advanced non-small cell lung cancer patients treated with 55 Gray of radiotherapy were included. First, the two modes of tumor regression were simulated. For each mode of tumor regression, one timepoint was simulated. Based on the results of simulated data, the approach used for analysis in real patients was developed. CBCTs were non-rigidly registered to the baseline CBCT using a cubic B-spline algorithm, NiftyReg. Next, the Jacobian determinants were computed from the deformation vector fields. To capture local volume changes, 10 Jacobian values were sampled perpendicular to the surface of the GTV, across the lung-tumor boundary. From the simulated data, we can distinguish elastic from non-elastic tumor regression by comparing the Jacobian values samples between 5 and 12.5mm inside and 5 and 12.5mm outside the planning GTV. Finally, morphometric results were compared between tumors of different histologies. Most patients (92.3%) in our cohort showed stable disease in the first week of treatment and non-elastic shrinkage in the later weeks of treatment. At week 2, 125 patients (88%) showed stable disease, three patients (2.1%) disease progression, and 11 patients (8%) regression. By treatment completion, 91 patients (64%) had stable disease, one patient (0.7%) progression and 46 patients (32%) regression. A slight difference in the mode of tumor change was observed between tumors of different histologies. Our novel approach shows that it may be possible to automatically quantify and identify global changes in lung cancer patients during RT, using routine CBCT images. Our results show that different regions of the tumor change in different ways. Therefore, careful consideration should be taken when adapting RT.

Automatic liver tumor localization using deep learning-based liver boundary motion estimation and biomechanical modeling (DL-Bio).

Recently, two-dimensional-to-three-dimensional (2D-3D) deformable registration has been applied to deform liver tumor contours from prior reference images onto estimated cone-beam computed tomography (CBCT) target images to automate on-board tumor localizations. Biomechanical modeling has also been introduced to fine-tune the intra-liver deformation-vector-fields (DVFs) solved by 2D-3D deformable registration, especially at low-contrast regions, using tissue elasticity information and liver boundary DVFs. However, the caudal liver boundary shows low contrast from surrounding tissues in the cone-beam projections, which degrades the accuracy of the intensity-based 2D-3D deformable registration there and results in less accurate boundary conditions for biomechanical modeling. We developed a deep-learning (DL)-based method to optimize the liver boundary DVFs after 2D-3D deformable registration to further improve the accuracy of subsequent biomechanical modeling and liver tumor localization. The DL-based network was built based on the U-Net architecture. The network was trained in a supervised fashion to learn motion correlation between cranial and caudal liver boundaries to optimize the liver boundary DVFs. Inputs of the network had three channels, and each channel featured the 3D DVFs estimated by the 2D-3D deformable registration along one Cartesian direction (x, y, z). To incorporate patient-specific liver boundary information into the DVFs, the DVFs were masked by a liver boundary ring structure generated from the liver contour of the prior reference image. The network outputs were the optimized DVFs along the liver boundary with higher accuracy. From these optimized DVFs, boundary conditions were extracted for biomechanical modeling to further optimize the solution of intra-liver tumor motion. We evaluated the method using 34 liver cancer patient cases, with 24 for training and 10 for testing. We evaluated and compared the performance of three methods: 2D-3D deformable registration, 2D-3D-Bio (2D-3D deformable registration with biomechanical modeling), and DL-Bio (DL model prediction with biomechanical modeling). The tumor localization errors were quantified through calculating the center-of-mass-errors (COMEs), DICE coefficients, and Hausdorff distance between deformed liver tumor contours and manually segmented "gold-standard" contours. The predicted DVFs by the DL model showed improved accuracy at the liver boundary, which translated into more accurate liver tumor localizations through biomechanical modeling. On a total of 90 evaluated images and tumor contours, the average (± sd) liver tumor COMEs of the 2D-3D, 2D-3D-Bio, and DL-Bio techniques were 4.7 ± 1.9mm, 2.9 ± 1.0mm, and 1.7 ± 0.4mm. The corresponding average (± sd) DICE coefficients were 0.60 ± 0.12, 0.71 ± 0.07, and 0.78 ± 0.03; and the average (± sd) Hausdorff distances were 7.0 ± 2.6mm, 5.4 ± 1.5mm, and 4.5 ± 1.3mm, respectively. DL-Bio solves a general correlation model to improve the accuracy of the DVFs at the liver boundary. With improved boundary conditions, the accuracy of biomechanical modeling can be further increased for accurate intra-liver low-contrast tumor localization.

Multi-contrast four-dimensional magnetic resonance imaging (MC-4D-MRI): Development and initial evaluation in liver tumor patients.

To develop a novel multi-contrast four-dimensional magnetic resonance imaging (MC-4D-MRI) technique that expands single image contrast 4D-MRI to a spectrum of native and synthetic image contrasts and to evaluate its feasibility in liver tumor patients. The MC-4D-MRI technique integrates multi-parametric MRI fusion, 4D-MRI, and deformable image registration (DIR) techniques. The fusion technique consists of native MRI as input, image pre-processing, fusion algorithm, adaptation, and fused multi-contrast MRI as output. Four-dimensional deformation vector fields (4D-DVF) were generated from an original T2/T1-w 4D-MRI by deforming end-of-inhalation (EOI) to nine other phase volumes via DIR. The 4D-DVF were applied to multi-contrast MRI to generate a spectrum of 4D-MRI in different image contrasts. The MC-4D-MRI technique was evaluated in five liver tumor patients on tumor contrast-to-noise ratio (CNR), internal target volume (ITV) contouring consistency, diaphragm motion range, and tumor motion trajectory; and in digital anthropomorphic phantoms on 4D-DIR introduced errors in tumor motion range, centroid location, extent, and volume. MC-4D-MRI consisting of 4D-MRIs in native image contrasts (T1-w, T2-w, and T2/T1-w) and synthetic image contrasts, such as tumor-enhanced contrast (TEC) were generated in five liver tumor patients. Patient tumor CNR increased from 2.6±1.8 in the T2/T1-w MRI, to -4.4±2.4, 6.6±3.0, and 9.6±3.9 in the T1-w, T2-w, and TEC MRI, respectively. Patient ITV inter-observer mean Dice similarity coefficient (mDSC) increased from 0.65±0.10 in the original T2/T1-w 4D-MRI, to 0.76±0.14, 0.77±0.12, and 0.86±0.05 in the T1-w, T2-w, and TEC 4D-MRI, respectively. Patient diaphragm motion range absolute differences between the three new 4D-MRIs and original T2/T1-w 4D-MRI were 1.2±1.3, 0.3±0.7, and 0.5±0.5mm, respectively. Patient tumor displacement phase-averaged absolute differences between the three 4D-MRIs and the original 4D-MRI were 0.72±0.33, 0.62±0.54, and 0.74±0.43mm in the superior-inferior (SI) direction, and 0.59±0.36, 0.51±0.30, and 0.50±0.24mm in the anterior-posterior (AP) direction, respectively. In the digital phantoms, phase-averaged absolute tumor centroid shift caused by the 4D-DIR were at or below 0.5mm in SI, AP, and left-right (LR) directions. We developed an MC-4D-MRI technique capable of expanding single image contrast 4D-MRI along a new dimension of image contrast. Initial evaluations in liver tumor patients showed enhancements in image contrast variety, tumor contrast, and ITV contouring consistencies using MC-4D-MRI. The technique might offer new perspectives on the image contrast of MRI and 4D-MRI in MR-guided radiotherapy.

Deep learning-based motion tracking using ultrasound images.

Ultrasound (US) imaging is an established imaging modality capable of offering video-rate volumetric images without ionizing radiation. It has the potential for intra-fraction motion tracking in radiation therapy. In this study, a deep learning-based method has been developed to tackle the challenges in motion tracking using US imaging. We present a Markov-like network, which is implemented via generative adversarial networks, to extract features from sequential US frames (one tracked frame followed by untracked frames) and thereby estimate a set of deformation vector fields (DVFs) through the registration of the tracked frame and the untracked frames. The positions of the landmarks in the untracked frames are finally determined by shifting landmarks in the tracked frame according to the estimated DVFs. The performance of the proposed method was evaluated on the testing dataset by calculating the tracking error (TE) between the predicted and ground truth landmarks on each frame. The proposed method was evaluated using the MICCAI CLUST 2015 dataset which was collected using seven US scanners with eight types of transducers and the Cardiac Acquisitions for Multi-structure Ultrasound Segmentation (CAMUS) dataset which was acquired using GE Vivid E95 ultrasound scanners. The CLUST dataset contains 63 2D and 22 3D US image sequences respectively from 42 and 18 subjects, and the CAMUS dataset includes 2D US images from 450 patients. On CLUST dataset, our proposed method achieved a mean tracking error of 0.70 ± 0.38mm for the 2D sequences and 1.71 ± 0.84mm for the 3D sequences for those public available annotations. And on CAMUS dataset, a mean tracking error of 0.54 ± 1.24mm for the landmarks in the left atrium was achieved. A novel motion tracking algorithm using US images based on modern deep learning techniques has been demonstrated in this study. The proposed method can offer millimeter-level tumor motion prediction in real time, which has the potential to be adopted into routine tumor motion management in radiation therapy.

Real-time 3D motion estimation from undersampled MRI using multi-resolution neural networks.

Purpose: To enable real‐time adaptive magnetic resonance imaging–guided radiotherapy (MRIgRT) by obtaining time‐resolved three‐dimensional (3D) deformation vector fields (DVFs) with high spatiotemporal resolution and low latency (<500 ms). Theory and Methods: Respiratory‐resolved T1‐weighted 4D‐MRI of 27 patients with lung cancer were acquired using a golden‐angle radial stack‐of‐stars readout. A multiresolution convolutional neural network (CNN) called TEMPEST was trained on up to 32× retrospectively undersampled MRI of 17 patients, reconstructed with a nonuniform fast Fourier transform, to learn optical flow DVFs. TEMPEST was validated using 4D respiratory‐resolved MRI, a digital phantom, and a physical motion phantom. The time‐resolved motion estimation was evaluated in‐vivo using two volunteer scans, acquired on a hybrid MR‐scanner with integrated linear accelerator. Finally, we evaluated the model robustness on a publicly‐available four‐dimensional computed tomography (4D‐CT) dataset. Results: TEMPEST produced accurate DVFs on respiratory‐resolved MRI at 20‐fold acceleration, with the average end‐point‐error <2 mm, both on respiratory‐sorted MRI and on a digital phantom. TEMPEST estimated accurate time‐resolved DVFs on MRI of a motion phantom, with an error <2 mm at 28× undersampling. On two volunteer scans, TEMPEST accurately estimated motion compared to the self‐navigation signal using 50 spokes per dynamic (366× undersampling). At this undersampling factor, DVFs were estimated within 200 ms, including MRI acquisition. On fully sampled CT data, we achieved a target registration error of 1.87±1.65 mm without retraining the model. Conclusion: A CNN trained on undersampled MRI produced accurate 3D DVFs with high spatiotemporal resolution for MRIgRT.

Cone beam CT-based dose accumulation and analysis of delivered dose to the dominant intraprostatic lesion in primary radiotherapy of prostate cancer

BackgroundEvaluation of delivered dose to the dominant intraprostatic lesion (DIL) for moderately hypofractionated radiotherapy of prostate cancer by cone beam computed tomography (CBCT)-based dose accumulation and target coverage analysis.MethodsTwenty-three patients with localized prostate cancer treated with moderately hypofractionated prostate radiotherapy with simultaneous integrated boost (SIB) between December 2016 and February 2020 were retrospectively analyzed. Included patients were required to have an identifiable DIL on bi-parametric planning magnetic resonance imaging (MRI). After import into the RayStation treatment planning system and application of a step-wise density override, the fractional doses were computed on each CBCT and were consecutively mapped onto the planning CT via a deformation vector field derived from deformable image registration. Fractional doses were accumulated for all CBCTs and interpolated for missing CBCTs, resulting in the delivered dose for PTVDIL, PTVBoost, PTV, and the organs at risk. The location of the index lesions was recorded according to the sector map of the Prostate Imaging Reporting and Data System (PIRADS) Version 2.1. Target coverage of the index lesions was evaluated and stratified for location.ResultsIn total, 338 CBCTs were available for analysis. Dose accumulation target coverage of PTVDIL, PTVBoost, and PTV was excellent and no cases of underdosage in DMean, D95%, D02%, and D98% could be detected. Delivered rectum DMean did not significantly differ from the planned dose. Bladder mean DMean was higher than planned with 19.4 ± 7.4 Gy versus 18.8 ± 7.5 Gy, p < 0.001. The penile bulb showed a decreased delivered mean DMean with 29.1 ± 14.0 Gy versus 29.8 ± 14.4 Gy, p < 0.001. Dorsal DILs, defined as DILs in the posterior medial peripheral zone of the prostate, showed a significantly lower delivered dose with a mean DMean difference of 2.2 Gy (95% CI 1.3–3.1 Gy, p < 0.001) compared to ventral lesions.ConclusionsCBCT-based dose accumulation showed an adequate delivered dose to the dominant intraprostatic lesion and organs at risk within planning limits. Cautious evaluation of the target coverage for index lesions adjacent to the rectum is warranted to avoid underdosage.

Unsupervised Learning-Based CBCT-CT Deformable Image Registration for CBCT-Guided Abdominal Radiotherapy

Daily Cone beam CT (CBCT) imaging provides necessary anatomical information for accurate patient setup. Image quality of CBCT is usually far inferior to simulation CT scans. A workaround is to register the CT to the CBCT such that the contours and Hounsfield Unit (HU) values of the CT can be propagated to the CBCT. However, the inconsistent HU values across CT and CBCT make it less effective to use conventional image similarity measures. We aim to develop an unsupervised registration network to overcome this challenge in multimodal CT-CBCT image registration.We propose to integrate directional local structural similarity into an unsupervised learning framework to perform abdominal CT-CBCT image registration. Directional local structural similarity measures the image's self-similarity which reflects the underlying structural similarity regardless of the modality in use. The CBCT and CT images were separately processed to extract directional local structural similarity feature maps in different directions. We concatenated the directional local structural similarity feature maps and the original images as network input. Taking both the original images and their respective structural similarity feature maps as input allows the network to fully explore the potential correlations between CBCT and CT for accurate deformation vector field (DVF) prediction. Salient features learnt through previous iterations were highlighted by attention gates across layers to expedite the learning process. A 3D bicubic interpolation was used to up-sample and smooth the predicted DVF. We performed a leave-one-out cross validation with an image dataset of 45 patients to evaluate the proposed registration method. Normalized cross correlation (NCC) and target registration error (TRE) between CBCT and deformed CT were calculated to quantify the registration accuracy.Our results show that the alignment between the abdominal soft tissues has been greatly improved after registration for all patients. The mean and standard deviation of NCC and TRE were 0.97 (range 0.95-0.99) and 1.88 (range 1.03-2.67) mm. The proposed network allows for many datasets to be used as training datasets since ground truth DVF is not needed for the training process. The proposed network can predict a final DVF via a single forward prediction, which is faster than the conventional iterative registration algorithms.We have developed a novel unsupervised multimodal image registration method for CT-CBCT abdominal image registration, which does not need ground truth DVF for training, and demonstrated its feasibility. Taking both CBCT and CT images and their respective directional local structural similarity features as input, the proposed network performs direct DVF prediction to register the abdominal CT to CBCT images. This tool will be useful for future CBCT-guided radiotherapy of abdominal malignancies.

Cardiac Motion and its Impact on Target Coverage During SBRT for Refractory Ventricular Tachycardia

<h3>Purpose/Objective(s)</h3> Refractory ventricular tachycardia (VT) is a common cause of death and morbidity in the heart failure population. Stereotactic body radiation therapy (SBRT) is a new and promising approach to treating VT, but there are practical concerns about SBRT delivery to the heart, which undergoes motion from both cardiac and respiratory cycles. In this study, we investigate motion magnitudes from both cycles and their impact on treatment margins and target coverage for VT SBRT. <h3>Materials/Methods</h3> Twenty-one patients were treated with SBRT for VT at our institution between April 2018 and December 2020 under an IRB-approved registry protocol. All patients had respiratory-correlated 4DCT for treatment planning and 7 patients had diagnostic cardiac-gated CT scans taken less than two weeks prior to treatment. Target structures were delineated on free-breathing CT and transferred to respiratory average (RAVG) CT via rigid registration. Dose was initially calculated on the RAVG per current institutional standard of care. For this retrospective study, deformable image registration (DIR) was applied between RAVG and the cardiac average CT (CAVG), systolic, and diastolic CTs from the cardiac dataset. Target contours were transferred to the registration target image. Motion magnitudes and distributions from the deformation vector fields were analyzed. Internal target volumes (ITVs) were created on the respiratory average images for quantitative analysis of target motion and coverage changes. From these 7 patients, 11 distinct target volumes were treated, with 8 being epicardial and 3 endocardial. Motion and dosimetry for all 11 targets were analyzed separately. <h3>Results</h3> The average (SD) target shift magnitudes (in mm) from the RAVG to CAVG CTs were 2.0 ± 1.0, 2.0 ± 1.3, and 4.0 ± 2.0 in the left-right (LR), anterior/posterior (AP), and superior/inferior (SI) directions. Endocardial targets had greater motion magnitude than epicardial targets, with mean shifts of 2.2 ± 1.0, 2.3 ± 1.9, and 4.2 ± 2.4 in the LR, AP, and SI directions versus 1.4 ± 0.9, 1.3 ± 0.8, and 3.6 ± 0.5. Cardiac cycle motion showed more variability for epicardial targets with shifts of 1.4 ± 2.5, 2.4 ± 3.0, and 2.5 ± 3.7 in the LR, AP and SI directions, versus 2.0 ± 0.6, 2.1 ± 0.8, and 1.6 ± 0.7 for endocardial targets. Creation of motion encompassing ITVs led to an average target size increase of 100% (25 cm³). V95 for initial target volumes was 99.9 ± 0.4% while the average V95 for ITVs was 84.1 ± 15.6%. Similarly, D95 for the target dropped from 108.2 ± 5.7% to 75.3 ± 28.4%. <h3>Conclusion</h3> In this cohort of patients undergoing SBRT for VT, cardiac shifts were location-dependent and were largest in the SI direction. These shifts typically exceeded the 1-5 mm margins used for SBRT treatment planning and ITV generation to encompass all target motion led to a doubling in target size. Our data suggest that cardiac-gated delivery may be helpful to minimize treatment volumes and spare cardiac tissue.

Impact of Inter-and Intra-Fraction Motion on Delivered Dose for Prostate SBRT Patients Treated on an MRL-Linac Using a Novel Semi-Automated Dose Accumulation Framework

Online MR-guided adaptive RT enables generation of an adapted treatment plan at each fraction to account for daily anatomical changes, with both inter- and intrafraction internal motion impacting the accumulated dose. Our objective is to quantify the impact of motion on the accumulated delivered dose for prostate SBRT patients treated on an MRL using a novel semi-automated process for dose accumulation.We included 18 prostate SBRT patients treated on a prospective clinical trial with 30 Gy in 5 fraction SBRT on an integrated 1.5 T MRL with consecutive 15 Gy in 1 brachytherapy boost. We performed dose accumulation across the MRL treatment using a novel semi-automated workflow to import, deformably register, and deform each fraction dose to a reference MR. All deformable image registration (DIR) steps were performed automatically without manual input. We accumulated both the adapted plan dose, as assessed on localization MR image, and the final delivered dose, assessed on beam-on images acquired during treatment delivery. We qualitatively evaluated the DIR for all registrations using a 3-point scale and extracted organ motion from the deformation vector field. To assess the impact of intrafraction motion during online adaptation, we compared target and organ-at-risk dose-volume (DVH) metrics on the total accumulated dose for adapted plans versus delivered dose via paired t-tests with Bonferroni Correction.We qualitatively scored the DIR agreement for CTV, proximal bladder, rectum, and femurs to be within 2 mm agreement for all fractions, however, the distal bladder was unacceptable (> 5 mm differences). Mean absolute intrafraction motion was evaluated over all patients and fractions to be 0.7 ± 0.6 mm LR, 2.0 ± 1.4 mm AP and 1.8 ± 1.3 mm SI for the CTV, and 1.7 ± 0.7 mm LR, 2.5 ± 1.1 mm AP and 2.6 ± 1.0 mm SI for the rectum. The CTV D98 for accumulated delivered dose met the clinical goal of > 2850 cGy for all patients, with no significant difference compared to the adapted plans. The accumulated rectum D1cc DVH metric exhibited a significant decrease for delivered dose (2406 cGy) compared to the adapted plan (2548 cGy, P < 0.005), whereas bladder D5cc exhibited a significant increase for delivered dose (2905 cGy) compared to the adapted plan (2757 cGy, P < 0.001). No differences existed for rectum D20, D50, or femur D5.Here we demonstrated the feasibility of MRL dose accumulation for both online adapted plan distributions and estimated delivered dose for prostate SBRT treatments. In the presence of intrafraction motion between adapted plan generation and beam delivery, our results suggest only a minor impact on the accumulated dose. In the future we plan to use this dose accumulation framework to investigate different treatment approaches, such as adaptive intraprostatic dose escalation.

Few-shot learning for deformable image registration in 4DCT images.

To develop a rapid and accurate 4D deformable image registration (DIR) approach for online adaptive radiotherapy. We propose a deep learning (DL)-based few-shot registration network (FR-Net) to generate deformation vector fields from each respiratory phase to an implicit reference image, thereby mitigating the bias introduced by the selection of reference images. The proposed FR-Net is pretrained with limited unlabeled 4D data and further optimized by maximizing the intensity similarity of one specific four-dimensional computed tomography (4DCT) scan. Because of the learning ability of DL models, the few-shot learning strategy facilitates the generalization of the model to other 4D data sets and the acceleration of the optimization process. The proposed FR-Net is evaluated for 4D groupwise and 3D pairwise registration on thoracic 4DCT data sets DIR-Lab and POPI. FR-Net displays an averaged target registration error of 1.48 mm and 1.16 mm between the maximum inhalation and exhalation phases in the 4DCT of DIR-Lab and POPI, respectively, with approximately 2 min required to optimize one 4DCT. Overall, FR-Net outperforms state-of-the-art methods in terms of registration accuracy and exhibits a low computational time. We develop a few-shot groupwise DIR algorithm for 4DCT images. The promising registration performance and computational efficiency demonstrate the prospective applications of this approach in registration tasks for online adaptive radiotherapy. This work exploits DL models to solve the optimization problem in registering 4DCT scans while combining groupwise registration and few-shot learning strategy to solve the problem of consuming computational time and inferior registration accuracy.

First clinical experience of correcting phantom-based image distortion related to gantry position on a 0.35T MR-Linac.

MR‐guided radiotherapy requires strong imaging spatial integrity to deliver high quality plans and provide accurate dose calculation. The MRI system, however, can be compromised by the integrated linear accelerator (Linac), resulting in inaccurate imaging isocenter position and geometric distortion. Dependence on gantry position further complicates the correction of distortions. This work presents a new clinical application of a commercial phantom and software system that quantifies isocenter alignment and geometric distortion, as well as providing a deformation vector field (DVF). A large distortion phantom and a smaller grid phantom were imaged at multiple gantry angles from 0 to 330° on a 0.35 T integrated MR‐Linac. The software package was used to assess geometric distortion and generate DVFs to correct distortions within the phantom volume. The DVFs were applied to the grid phantom with resampling software then evaluated using structural similarity index measure (SSIM). Scans were also performed with a ferromagnetic clip near the phantom to investigate the correction of more severe artifacts. The mean magnitude isocenter shift was 0.67 mm, ranging from 0.25 to 1.04 mm across all angles. The DVF had a mean component value of 0.27 ± 0.02, 0.24 ± 0.01, and 0.19 ± 0.01 mm in the right‐left (RL), anterior‐posterior (AP), and superior‐inferior (SI) directions. The ferromagnetic clip increased isocenter position error from 1.98 mm to 2.20 mm and increased mean DVF component values in the RL and AP directions. The resampled grid phantom had an increased SSIM for all gantry angles compared to original images, increasing from 0.26 ± 0.001 to 0.70 ± 0.004. Through this clinical assessment, we were able to correct geometric distortion and isocenter shift related to gantry position on a 0.35 T MR‐Linac using the distortion phantom and software package. This provides encouragement that it could be used for quality assurance and clinically to correct systematic distortion caused by imaging at different gantry angles.

Adaptation to an illusory aspect ratio distorted by motion patches in a deformation vector field

Adaptation to an illusory aspect ratio distorted by motion patches in a deformation vector field

Lung tumor segmentation in 4D CT images using motion convolutional neural networks.

Manual delineation on all breathing phases of lung cancer 4D CT image datasets can be challenging, exhaustive, and prone to subjective errors because of both the large number of images in the datasets and variations in the spatial location of tumors secondary to respiratory motion. The purpose of this work is to present a new deep learning-based framework for fast and accurate segmentation of lung tumors on 4D CT image sets. The proposed DL framework leverages motion region convolutional neural network (R-CNN). Through integration of global and local motion estimation network architectures, the network can learn both major and minor changes caused by tumor motion. Our network design first extracts tumor motion information by feeding 4D CT images with consecutive phases into an integrated backbone network architecture, locating volume-of-interest (VOIs) via a regional proposal network and removing irrelevant information via a regional convolutional neural network. Extracted motion information is then advanced into the subsequent global and local motion head network architecture to predict corresponding deformation vector fields (DVFs) and further adjust tumor VOIs. Binary masks of tumors are then segmented within adjusted VOIs via a mask head. A self-attention strategy is incorporated in the mask head network to remove any noisy features that might impact segmentation performance. We performed two sets of experiments. In the first experiment, a five-fold cross-validation on 20 4D CT datasets, each consisting of 10 breathing phases (i.e., 200 3D image volumes in total). The network performance was also evaluated on an additional unseen 200 3D images volumes from 20 hold-out 4D CT datasets. In the second experiment, we trained another model with 40 patients' 4D CT datasets from experiment 1 and evaluated on additional unseen nine patients' 4D CT datasets. The Dice similarity coefficient (DSC), center of mass distance (CMD), 95th percentile Hausdorff distance (HD95 ), mean surface distance (MSD), and volume difference (VD) between the manual and segmented tumor contour were computed to evaluate tumor detection and segmentation accuracy. The performance of our method was quantitatively evaluated against four different methods (VoxelMorph, U-Net, network without global and local networks, and network without attention gate strategy) across all evaluation metrics through a paired t-test. The proposed fully automated DL method yielded good overall agreement with the ground truth for contoured tumor volume and segmentation accuracy. Our model yielded significantly better values of evaluation metrics (p<0.05) than all four competing methods in both experiments. On hold-out datasets of experiment 1 and 2, our method yielded DSC of 0.86 and 0.90 compared to 0.82 and 0.87, 0.75 and 0.83, 081 and 0.89, and 0.81 and 0.89 yielded by VoxelMorph, U-Net, network without global and local networks, and networks without attention gate strategy. Tumor VD between ground truth and our method was the smallest with the value of 0.50 compared to 0.99, 1.01, 0.92, and 0.93 for between ground truth and VoxelMorph, U-Net, network without global and local networks, and networks without attention gate strategy, respectively. Our proposed DL framework of tumor segmentation on lung cancer 4D CT datasets demonstrates a significant promise for fully automated delineation. The promising results of this work provide impetus for its integration into the 4D CT treatment planning workflow to improve the accuracy and efficiency of lung radiotherapy.

Ventilation measurements using fast-helical free-breathing computed tomography.

To examine the use of multiple fast-helical free breathing computed tomography (FHFBCT) scans for ventilation measurement. Ten patients were scanned 25 times in alternating directions using a FHFBCT protocol. Simultaneously, an abdominal pneumatic bellows was used as a real-time breathing surrogate. Regions-of-interest (ROIs) were selected from the upper right lungs of each patient for analysis. The ROIs were first registered using a published registration technique (pTV). A subsequent follow-up registration employed an objective function with two terms, a ventilation-adjusted Hounsfield Unit difference and a conservation-of-mass term labeled ΔΓ that denoted the difference between the deformation Jacobian and the tissue density ratio. The ventilations were calculated voxel-by-voxel as the slope of a first-order fit of the Jacobian as a function of the breathing amplitude. The ventilations of the 10 patients showed different patterns and magnitudes. The average ventilation calculated from the deformation vector fields (DVFs) of the pTV and secondary registration was nearly identical, but the standard deviation of the voxel-to-voxel differences was approximately 0.1. The mean of the 90th percentile values of ΔΓ was reduced from 0.153 to 0.079 between the pTV and secondary registration, implying first that the secondary registration improved the conservation-of-mass criterion by almost 50% and that on average the correspondence between the Jacobian and density ratios as demonstrated by ΔΓ was less than 0.1. This improvement occurred in spite of the average of the 90th percentile changes in the DVF magnitudes being only 0.58mm. This work introduces the use of multiple free-breathing CT scans for free-breathing ventilation measurements. The approach has some benefits over the traditional use of 4-dimensional CT (4DCT) or breath-hold scans. The benefit over 4DCT is that FHFBCT does not have sorting artifacts. The benefits over breath-hold scans include the relatively small motion induced by quiet respiration versus deep-inspiration breath hold and the potential for characterizing dynamic breathing processes that disappear during breath hold.

Validation of proton dose calculation on scatter corrected 4D cone beam computed tomography using a porcine lung phantom

Proton therapy treatment for lungs remains challenging as images enabling the detection of inter- and intra-fractional motion, which could be used for proton dose adaptation, are not readily available. 4D computed tomography (4DCT) provides high image quality but is rarely available in-room, while in-room 4D cone beam computed tomography (4DCBCT) suffers from image quality limitations stemming mostly from scatter detection. This study investigated the feasibility of using virtual 4D computed tomography (4DvCT) as a prior for a phase-per-phase scatter correction algorithm yielding a 4D scatter corrected cone beam computed tomography image (4DCBCTcor), which can be used for proton dose calculation. 4DCT and 4DCBCT scans of a porcine lung phantom, which generated reproducible ventilation, were acquired with matching breathing patterns. Diffeomorphic Morphons, a deformable image registration algorithm, was used to register the mid-position 4DCT to the mid-position 4DCBCT and yield a 4DvCT. The 4DCBCT was reconstructed using motion-aware reconstruction based on spatial and temporal regularization (MA-ROOSTER). Successively for each phase, digitally reconstructed radiographs of the 4DvCT, simulated without scatter, were exploited to correct scatter in the corresponding CBCT projections. The 4DCBCTcor was then reconstructed with MA-ROOSTER using the corrected CBCT projections and the same settings and deformation vector fields as those already used for reconstructing the 4DCBCT. The 4DCBCTcor and the 4DvCT were evaluated phase-by-phase, performing proton dose calculations and comparison to those of a ground truth 4DCT by means of dose-volume-histograms (DVH) and gamma pass-rates (PR). For accumulated doses, DVH parameters deviated by at most 1.7% in the 4DvCT and 2.0% in the 4DCBCTcor case. The gamma PR for a (2%, 2 mm) criterion with 10% threshold were at least 93.2% (4DvCT) and 94.2% (4DCBCTcor), respectively. The 4DCBCTcor technique enabled accurate proton dose calculation, which indicates the potential for applicability to clinical 4DCBCT scans.

Deformable scintillation dosimeter: II. Real-time simultaneous measurements of dose and tracking of deformation vector fields

Anatomical motion and deformation pose challenges to the understanding of the delivered dose distribution during radiotherapy treatments. Hence, deformable image registration (DIR) algorithms are increasingly used to map contours and dose distributions from one image set to another. However, the lack of validation tools slows their clinical adoption, despite their commercial availability. This work presents a novel water-equivalent deformable dosimeter that simultaneously measures the dose distribution and tracks deformation vector fields (DVF). The dosimeter in made of an array of 19 scintillating fiber detectors embedded in a cylindrical elastomer matrix. It is imaged by two pairs of stereoscopic cameras tracking the position and angulation of the scintillators, while measuring the dose. The resulting system provides a precision of 0.3 mm on DVF measurements. The dosimeter was irradiated with 5 × 3, 4 × 3 and 3 × 3 cm2 6 MV photon beams in both fixed and deformed conditions. The measured DVF was compared to the one computed with a DIR algorithm (Plastimatch). The deviations between the computed and measured DVFs was below 1.5 mm. As for dose measurements, the dosimeter acquired the dose distribution in fixed and deformed conditions within 1% of the treatment planning system calculation and complementary dose validation using the Hyperscint dosimetry system. Using the demonstrated qualities of scintillating detectors, we developed a real-time, water-equivalent deformable dosimeter. Given it’s sensor tracking position precision and dose measurements accuracy, the developed detector is a promising tools for the validation of DIR algorithms as well as dose distribution measurements under fixed and deformed conditions.

Deformation driven Seq2Seq longitudinal tumor and organs-at-risk prediction for radiotherapy.

Radiotherapy presents unique challenges and clinical requirements for longitudinal tumor and organ-at-risk (OAR) prediction during treatment. The challenges include tumor inflammation/edema and radiation-induced changes in organ geometry, whereas the clinical requirements demand flexibility in input/output sequence timepoints to update the predictions on rolling basis and the grounding of all predictions in relationship to the pre-treatment imaging information for response and toxicity assessment in adaptive radiotherapy. To deal with the aforementioned challenges and to comply with the clinical requirements, we present a novel 3D sequence-to-sequence model based on Convolution Long Short-Term Memory (ConvLSTM) that makes use of series of deformation vector fields (DVFs) between individual timepoints and reference pre-treatment/planning CTs to predict future anatomical deformations and changes in gross tumor volume as well as critical OARs. High-quality DVF training data are created by employing hyper-parameter optimization on the subset of the training data with DICE coefficient and mutual information metric. We validated our model on two radiotherapy datasets: a publicly available head-and-neck dataset (28 patients with manually contoured pre-, mid-, and post-treatment CTs), and an internal non-small cell lung cancer dataset (63 patients with manually contoured planning CT and 6weekly CBCTs). The use of DVF representation and skip connections overcomes the blurring issue of ConvLSTM prediction with the traditional image representation. The mean and standard deviation of DICE for predictions of lung GTV at weeks 4, 5, and 6 were 0.83±0.09, 0.82±0.08, and 0.81±0.10, respectively, and for post-treatment ipsilateral and contralateral parotids, were 0.81±0.06 and 0.85±0.02. We presented a novel DVF-based Seq2Seq model for medical images, leveraging the complete 3D imaging information of a relatively large longitudinal clinical dataset, to carry out longitudinal GTV/OAR predictions for anatomical changes in HN and lung radiotherapy patients, which has potential to improve RT outcomes.