Improving multiple sclerosis lesion segmentation across clinical sites: A federated learning approach with noise-resilient training.

Accurately measuring the evolution of Multiple Sclerosis (MS) with magnetic resonance imaging (MRI) critically informs understanding of disease progression and helps to direct therapeutic strategy. Deep learning models have shown promise for automatically segmenting MS lesions, but the scarcity of accurately annotated data hinders progress in this area. Obtaining sufficient data from a single clinical site is challenging and does not address the heterogeneous need for model robustness. Conversely, the collection of data from multiple sites introduces data privacy concerns and potential label noise due to varying annotation standards. To address this dilemma, we explore the use of the federated learning framework while considering label noise. Our approach enables collaboration among multiple clinical sites without compromising data privacy under a federated learning paradigm that incorporates a noise-robust training strategy based on label correction. Specifically, we introduce a Decoupled Hard Label Correction (DHLC) strategy that considers the imbalanced distribution and fuzzy boundaries of MS lesions, enabling the correction of false annotations based on prediction confidence. We also introduce a Centrally Enhanced Label Correction (CELC) strategy, which leverages the aggregated central model as a correction teacher for all sites, enhancing the reliability of the correction process. Extensive experiments conducted on two multi-site datasets demonstrate the effectiveness and robustness of our proposed methods, indicating their potential for clinical applications in multi-site collaborations to train better deep learning models with lower cost in data collection and annotation.

Real world validation of an AI-based CT hemorrhage detection tool.

Introduction: Intracranial hemorrhage (ICH) is a potentially life-threatening medical event that requires expedited diagnosis with computed tomography (CT). Automated medical imaging triaging tools can rapidly bring scans containing critical abnormalities, such as ICH, to the attention of radiologists and clinicians. Here, we retrospectively investigated the real-world performance of VeriScout™, an artificial intelligence-based CT hemorrhage detection and triage tool. Methods: Ground truth for the presence or absence of ICH was iteratively determined by expert consensus in an unselected dataset of 527 consecutively acquired non-contrast head CT scans, which were sub-grouped according to the presence of artefact, post-operative features and referral source. The performance of VeriScout™ was compared with the ground truths for all groups. Results: VeriScout™ detected hemorrhage with a sensitivity of 0.92 (CI 0.84-0.96) and a specificity of 0.96 (CI 0.94-0.98) in the global dataset, exceeding the sensitivity of general radiologists (0.88) with only a minor relative decrement in specificity (0.98). Crucially, the AI tool detected 13/14 cases of subarachnoid hemorrhage, a potentially fatal condition that is often missed in emergency department settings. There was no decrement in the performance of VeriScout™ in scans containing artefact or postoperative change. Using an integrated informatics platform, VeriScout™ was deployed into the existing radiology workflow. Detected hemorrhage cases were flagged in the hospital radiology information system (RIS) and relevant, annotated, preview images made available in the picture archiving and communications system (PACS) within 10 min. Conclusion: AI-based radiology worklist prioritization for critical abnormalities, such as ICH, may enhance patient care without adding to radiologist or clinician burden.

Multiple sclerosis lesion segmentation: revisiting weighting mechanisms for federated learning.

Background and introduction: Federated learning (FL) has been widely employed for medical image analysis to facilitate multi-client collaborative learning without sharing raw data. Despite great success, FL's applications remain suboptimal in neuroimage analysis tasks such as lesion segmentation in multiple sclerosis (MS), due to variance in lesion characteristics imparted by different scanners and acquisition parameters. Methods: In this work, we propose the first FL MS lesion segmentation framework via two effective re-weighting mechanisms. Specifically, a learnable weight is assigned to each local node during the aggregation process, based on its segmentation performance. In addition, the segmentation loss function in each client is also re-weighted according to the lesion volume for the data during training. Results: The proposed method has been validated on two FL MS segmentation scenarios using public and clinical datasets. Specifically, the case-wise and voxel-wise Dice score of the proposed method under the first public dataset is 65.20 and 74.30, respectively. On the second in-house dataset, the case-wise and voxel-wise Dice score is 53.66, and 62.31, respectively. Discussions and conclusions: The Comparison experiments on two FL MS segmentation scenarios using public and clinical datasets have demonstrated the effectiveness of the proposed method by significantly outperforming other FL methods. Furthermore, the segmentation performance of FL incorporating our proposed aggregation mechanism can achieve comparable performance to that from centralized training with all the raw data.

A patient-specific deep learning framework for 3D motion estimation and volumetric imaging during lung cancer radiotherapy.

Objective. Respiration introduces a constant source of irregular motion that poses a significant challenge for the precise irradiation of thoracic and abdominal cancers. Current real-time motion management strategies require dedicated systems that are not available in most radiotherapy centers. We sought to develop a system that estimates and visualises the impact of respiratory motion in 3D given the 2D images acquired on a standard linear accelerator.Approach. In this paper we introduceVoxelmap, a patient-specific deep learning framework that achieves 3D motion estimation and volumetric imaging using the data and resources available in standard clinical settings. Here we perform a simulation study of this framework using imaging data from two lung cancer patients.Main results. Using 2D images as input and 3D-3DElastixregistrations as ground-truth,Voxelmapwas able to continuously predict 3D tumor motion with mean errors of 0.1 ± 0.5, -0.6 ± 0.8, and 0.0 ± 0.2 mm along the left-right, superior-inferior, and anterior-posterior axes respectively.Voxelmapalso predicted 3D thoracoabdominal motion with mean errors of -0.1 ± 0.3, -0.1 ± 0.6, and -0.2 ± 0.2 mm respectively. Moreover, volumetric imaging was achieved with mean average error 0.0003, root-mean-squared error 0.0007, structural similarity 1.0 and peak-signal-to-noise ratio 65.8.Significance. The results of this study demonstrate the possibility of achieving 3D motion estimation and volumetric imaging during lung cancer treatments on a standard linear accelerator.

MArkerless image Guidance using Intrafraction Kilovoltage x-ray imaging (MAGIK): study protocol for a phase I interventional study for lung cancer radiotherapy.

INTRODUCTION: In radiotherapy, tumour tracking leads the radiation beam to accurately target the tumour while it moves in a complex and unpredictable way due to respiration. Several tumour tracking techniques require the implantation of fiducial markers around the tumour, a procedure that involves unnecessary risks and costs. Markerless tumour tracking (MTT) negates the need for implanted markers, potentially enabling accurate and optimal radiotherapy in a non-invasive way. METHODS AND ANALYSIS: We will perform a phase I interventional trial called MArkerless image Guidance using Intrafraction Kilovoltage x-ray imaging (MAGIK) to investigate the technical feasibility of the MTT technology developed at the University of Sydney (sponsor). 30 participants will undergo the current standard of care lung stereotactic ablative radiation therapy, with the exception that kilovoltage X-ray images will be acquired continuously during treatment delivery to enable MTT. If MTT indicates that the mean lung tumour position has shifted >3 mm, a warning message will be displayed to indicate the need for a treatment intervention. The radiation therapist will then pause the treatment, shift the treatment couch to account for the shift in tumour position and resume the treatment. Participants will be implanted with fiducial markers, which act as the ground truth for evaluating the accuracy of MTT. MTT is considered feasible if the tracking accuracy is <3 mm in each dimension for >80% of the treatment time. ETHICS AND DISSEMINATION: The MAGIK trial has received ethical approval from The Alfred Human Research Ethics Committee and has been registered with ClinicalTrials.gov with the Identifier: NCT04086082. Estimated time of first recruitment is early 2022. The study recruitment and data analysis phases will be performed concurrently. Treatment for all 30 participants is expected to be completed within 2 years and participant follow-up within a total duration of 7 years. Findings will be disseminated through peer-reviewed publications and conference presentations. TRIAL REGISTRATION NUMBER: NCT04086082; Pre-result.

Proof-of-concept for x-ray based real-time image guidance during cardiac radioablation.

Cardiac radioablation offers non-invasive treatments for refractory arrhythmias. However, treatment delivery for this technique remains challenging. In this paper, we introduce the first method for real-time image guidance during cardiac radioablation for refractory atrial fibrillation on a standard linear accelerator. Our proposed method utilizes direct diaphragm tracking on intrafraction images to estimate the respiratory component of cardiac substructure motion. We compare this method to treatment scenarios without real-time image guidance using the 4D-XCAT digital phantom. Pre-treatment and intrafraction imaging was simulated for 8 phantoms with unique anatomies programmed using cardiorespiratory motion from healthy volunteers. As every voxel in the 4D-XCAT phantom is labelled precisely according to the corresponding anatomical structure, this provided ground-truth for quantitative evaluation. Tracking performance was compared to the ground-truth for simulations with and without real-time image guidance using the left atrium as an exemplar target. Differences in target volume size, mean volumetric coverage, minimum volumetric coverage and geometric error were recorded for each simulation. We observed that differences in target volume size were statistically significant (p < 0.001) across treatment scenarios and that real-time image guidance enabled reductions in target volume size ranging from 11% to 24%. Differences in mean and minimum volumetric coverage were statistically insignificant (bothp = 0.35) while differences in geometric error were statistically significant (p = 0.039). The results of this study provide proof-of-concept for x-ray based real-time image guidance during cardiac radioablation.

A real-time IGRT method using a Kalman filter framework to extract 3D positions from 2D projections.

Purpose.To estimate 3D prostate motion in real-time during irradiation from 2D prostate positions acquired from a kV imager on a standard linear accelerator utilising a Kalman filter (KF) framework. The advantage of this novel method is threefold: (1) eliminating the need of an initial learning period, therefore reducing patient imaging dose, (2) more robust against measurement noise and (3) more computationally efficient. In this paper, the novel KF method was evaluatedin silicousing patients' 3D prostate motion and simulated 2D projections.Methods.A KF framework was implemented to estimate 3D motion from 2D projection measurements in real-time during prostate cancer treatments. The noise covariance matrix was adaptively estimated from the previous 10 measurements. This method did not require an initial learning period as the KF process distribution was initialised using a population covariance matrix. This method was evaluated using a ground-truth motion dataset of 17 prostate cancer patients (536 trajectories) measured with electromagnetic transponders. 3D motion was projected onto a rotating imager (SID = 180 cm) (pixel size = 0.388 mm) and rotation speed of 6°/s and 2°/s to simulate VMAT treatments. Gantry-varying additive random noise (≤5 mm) was added to ground-truth measurements to simulate segmentation error and image quality degradation due to the patient's pelvic bones. For comparison, motion was also estimated using the clinically implemented Gaussian probability density function (PDF) method initialised with 600 projections.Results.Without noise, the 3D root mean square-errors (3D RMSEs) of motion estimated by the KF method were 0.4 ± 0.1 mm and 0.3 ± 0.2 mm for 2°/s and 6°/s gantry rotation, respectively. With noise, 3D RMSEs of KF estimated motion were 1.1 ± 0.1 mm for both slow and fast gantry rotation scenarios. In comparison, using a Gaussian PDF method, with noise, 3D RMSE was 2 ± 0.1 mm for both gantry rotation scenarios.Conclusion.This work presents a fast and accurate method for real-time 2D to 3D motion estimation using a KF approach to handle the random-walk component of prostate cancer motion. This method has sub-mm accuracy and is highly robust against measurement noise.

The adaptation and investigation of cone-beam CT reconstruction algorithms for horizontal rotation fixed-gantry scans of rabbits.

Fixed-gantry radiation therapy has been proposed as a low-cost alternative to the conventional rotating-gantry radiation therapy, that may help meet the rising global treatment demand. Fixed-gantry systems require gravitational motion compensated reconstruction algorithms to produce cone-beam CT (CBCT) images of sufficient quality for image guidance. The aim of this work was to adapt and investigate five CBCT reconstruction algorithms for fixed-gantry CBCT images. The five algorithms investigated were Feldkamp-Davis-Kress (FDK), prior image constrained compressed sensing (PICCS), gravitational motion compensated FDK (GMCFDK), motion compensated PICCS (MCPICCS) (a novel CBCT reconstruction algorithm) and simultaneous motion estimation and iterative reconstruction (SMEIR). Fixed-gantry and rotating-gantry CBCT scans were acquired of 3 rabbits, with the rotating-gantry scans used as a reference. Projections were sorted into rotation bins, based on the angle of rotation of the rabbit during image acquisition. The algorithms were compared using the structural similarity index measure root mean square error, and reconstruction time. Evaluation of the reconstructed volumes showed that, when compared with the reference rotating-gantry volume, the conventional FDK algorithm did not accurately reconstruct fixed-gantry CBCT scans. Whilst the PICCS reconstruction algorithm reduced some motion artefacts, the motion estimation reconstruction methods (GMCFDK, MCPICCS and SMEIR) were able to greatly reduce the effect of motion artefacts on the reconstructed volumes. This finding was verified quantitatively, with GMCFDK, MCPICCS and SMEIR reconstructions having RMSE 17%-19% lower and SSIM 1% higher than a conventional FDK. However, all motion compensated fixed-gantry CBCT reconstructions had a 56%-61% higher RMSE and 1.5% lower SSIM than FDK reconstructions of conventional rotating-gantry CBCT scans. The results show that motion compensation is required to reduce motion artefacts for fixed-gantry CBCT reconstructions. This paper further demonstrates the feasibility of fixed-gantry CBCT scans, and the ability of CBCT reconstruction algorithms to compensate for motion due to horizontal rotation.

Evaluating reconstruction algorithms for respiratory motion guided acquisition.

Conventional thoracic 4DCBCT scans take 1320 projections over 4 min. This paper investigates which reconstruction algorithms best leverage Respiratory-Motion-Guided (RMG) acquisition in order to reduce scan time and dose while maintaining image quality. We investigated a 200 projection, on average 1 min RMG acquisition. RMG acquisition ensures even angular separation between projections at each respiratory phase by adjusting the imaging gantry rotation to the patient respiratory signal in real time. Conventional 1320 projection data and RMG 200 projection data were simulated from 4DCT volumes of 14 patients. Each patient had an initial 4DCT reconstruction, treated as a planning 4DCT, and a 4DCT reconstruction acquired later, used for 4DCBCT data simulation and evaluation. Reconstructions were computed using the Feldkamp-David-Kress (FDK), McKinnon-Bates (MKB), RecOnstructiOn using Spatial and TEmporal Regularization (ROOSTER), and Motion Compensated FDK (MCFDK) algorithms. We also introduced and evaluated a novel MCMKB algorithm. Image quality was evaluated with Root-Mean-Square Error (RMSE), Structural SIMilarity index (SSIM) and Tissue Interface Sharpness (TIS). Rigid registration of the tumor volume regions between the reconstruction and the ground truth was used to evaluate geometric accuracy. Relative to conventional 4DCBCT acquisition, the RMG acquisition delivered 80% less dose and was on average 70% faster. The conventional-acquisition 4DFDK-reconstruction volumes had mean RMSE, SSIM, TIS and geometric error of 94, 0.9987, 2.69 and 1.19 mm respectively. The RMG-acquisition MCFDK-reconstruction volumes had mean RMSE, SSIM, TIS and geometric error of 113, 0.9986, 1.76 and 1.77 mm respectively with minimal increase in computational cost. These results suggest scan time and dose can be significantly reduced with minimal impact on reconstruction quality by implementing RMG acquisition and motion compensated reconstruction.

The first prospective implementation of markerless lung target tracking in an experimental quality assurance procedure on a standard linear accelerator.

The ability to track tumour motion without implanted markers on a standard linear accelerator (linac) could enable wide access to real-time adaptive radiotherapy for cancer patients. We previously have retrospectively validated a method for 3D markerless target tracking using intra-fractional kilovoltage (kV) projections acquired on a standard linac. This paper presents the first prospective implementation of markerless lung target tracking on a standard linac and its quality assurance (QA) procedure. The workflow and the algorithm developed to track the 3D target position during volumetric modulated arc therapy treatment delivery were optimised. The linac was operated in clinical QA mode, while kV projections were streamed to a dedicated computer using a frame-grabber software. The markerless target tracking accuracy and precision were measured in a lung phantom experiment under the following conditions: static localisation of seven distinct positions, dynamic localisation of five patient-measured motion traces, and dynamic localisation with treatment interruption. The QA guidelines were developed following the AAPM Task Group 147 report with the requirement that the tracking margin components, the margins required to account for tracking errors, did not exceed 5 mm in any direction. The mean tracking error ranged from 0.0 to 0.9 mm (left-right), -0.6 to -0.1 mm (superior-inferior) and -0.7 to 0.1 mm (anterior-posterior) over the three tests. Larger errors were found in cases with large left-right or anterior-posterior and small superior-inferior motion. The tracking margin components did not exceed 5 mm in any direction and ranged from 0.4 to 3.2 mm (left-right), 0.7 to 1.6 mm (superior-inferior) and 0.8 to 1.5 mm (anterior-posterior). This study presents the first prospective implementation of markerless lung target tracking on a standard linac and provides a QA procedure for its safe clinical implementation, potentially enabling real-time adaptive radiotherapy for a large population of lung cancer patients.

Real-time direct diaphragm tracking using kV imaging on a standard linear accelerator.

PURPOSE: As the predominant driver of respiratory motion, the diaphragm represents a key surrogate for motion management during the irradiation of thoracic cancers. Existing approaches to diaphragm tracking often produce phase-based estimates, suffer from lateral side failures or are not executable in real-time. In this paper, we present an algorithm that continuously produces real-time estimates of three-dimensional (3D) diaphragm position using kV images acquired on a standard linear accelerator. METHODS: Patient-specific 3D diaphragm models were generated via automatic segmentation on end-exhale four-dimensional-computed tomography (4D-CT) images. The estimated trajectory of diaphragmatic motion, referred to as the principal motion vector, was obtained by registering end-exhale to end-inhale 4D-CT images. Two-dimensional (2D) diaphragm masks were generated by forward-projecting 3D models over the complement of angles spanned during kV image acquisition. For each kV image, diaphragm position was determined by shifting angle-matched 2D masks along the principal motion vector and selecting the position of highest contrast on a vertical difference image. Retrospective analysis was performed using 22 cone beam CT (CBCT) image sequences for six lung cancer patients across two datasets. Given the current lack of objective ground truth for diaphragm position, our algorithm was evaluated by examining its ability to track implanted markers. Simple linear regression was used to construct 3D marker motion models and estimation errors were computed as the difference between estimated and ground truth marker positions. Additionally, Pearson correlation coefficients were used to characterize diaphragm-marker correlation. RESULTS: The mean ± standard deviation of the estimation errors across all image sequences was -0.1 ± 0.7 mm, -0.1 ± 1.8 mm and 0.2 ± 1.4 mm in the LR, SI, and AP directions respectively. The 95th percentile of the absolute errors ranged over 0.5-3.1 mm, 1.6-6.7 mm, and 1.2-4.0 mm in the LR, SI, and AP directions, respectively. The mean ± standard deviation of diaphragm-marker correlations over all image sequences was -0.07 ± 0.57, 0.67 ± 0.49, and 0.29 ± 0.52 in the LR, SI, and AP directions, respectively. Diaphragm-marker correlation was observed to be highly dependent on marker position. Mean correlation along the SI axis ranged over 0.91-0.93 for markers situated in the lower lobes of the lung, while correlations ranging over -0.51-0.79 were observed for markers situated in the upper and middle lobes. CONCLUSION: This work advances a new approach to real-time direct diaphragm tracking in realistic treatment scenarios. By achieving continuous estimates of diaphragmatic motion, the proposed method has applications for both markerless tumor tracking and respiratory binning in 4D-CBCT.

SPARE: Sparse-view reconstruction challenge for 4D cone-beam CT from a 1-min scan.

PURPOSE: Currently, four-dimensional (4D) cone-beam computed tomography (CBCT) requires a 3-4 min full-fan scan to ensure usable image quality. Recent advancements in sparse-view 4D-CBCT reconstruction have opened the possibility to reduce scan time and dose. The aim of this study is to provide a common framework for systematically evaluating algorithms for 4D-CBCT reconstruction from a 1-min scan. Using this framework, the AAPM-sponsored SPARE Challenge was conducted in 2018 to identify and compare state-of-the-art algorithms. METHODS: A clinically realistic CBCT dataset was simulated using patient CT volumes from the 4D-Lung database. The selected patients had multiple 4D-CT sessions, where the first 4D-CT was used as the prior CT, and the rest were used as the ground truth volumes for simulating CBCT projections. A GPU-based Monte Carlo tool was used to simulate the primary, scatter, and quantum noise signals. A total of 32 CBCT scans of nine patients were generated. Additional qualitative analysis was performed on a clinical Varian and clinical Elekta dataset to validate the simulation study. Participants were blinded from the ground truth, and were given 3 months to apply their reconstruction algorithms to the projection data. The submitted reconstructions were analyzed in terms of root-mean-squared-error (RMSE) and structural similarity index (SSIM) with the ground truth within four different region-of-interests (ROI) - patient body, lungs, planning target volume (PTV), and bony anatomy. Geometric accuracy was quantified as the alignment error of the PTV. RESULTS: Twenty teams participated in the challenge, with five teams completing the challenge. Techniques involved in the five methods included iterative optimization, motion-compensation, and deformation of the prior 4D-CT. All five methods rendered significant reduction in noise and streaking artifacts when compared to the conventional Feldkamp-Davis-Kress (FDK) algorithm. The RMS of the three-dimensional (3D) target registration error of the five methods ranged from 1.79 to 3.00 mm. Qualitative observations from the Varian and Elekta datasets mostly concur with those from the simulation dataset. Each of the methods was found to have its own strengths and weaknesses. Overall, the MA-ROOSTER method, which utilizes a 4D-CT motion model for temporal regularization, had the best and most consistent image quality and accuracy. CONCLUSION: The SPARE Challenge represents the first framework for systematically evaluating state-of-the-art algorithms for 4D-CBCT reconstruction from a 1-min scan. Results suggest the potential for reducing scan time and dose for 4D-CBCT. The challenge dataset and analysis framework are publicly available for benchmarking future reconstruction algorithms.

Dual cardiac and respiratory gated thoracic imaging via adaptive gantry velocity and projection rate modulation on a linear accelerator: A Proof-of-Concept Simulation Study.

PURPOSE: Cardiac motion is typically not accounted for during pretreatment imaging for central lung and mediastinal tumors. However, cardiac induced tumor motion averages 5.8 mm for esophageal tumors and 3-5 mm for some lung tumors, which can result in positioning errors. Our aim is to reduce both cardiac- and respiratory-induced motion artifacts in thoracic cone beam computed tomography (CBCT) images through gantry velocity and projection rate modulation on a standard linear accelerator (linac). METHODS: The acquisition of dual cardiac and respiratory gated CBCT thoracic images was simulated using the XCAT phantom with patient-measured respiratory and ECG traces. The gantry velocity and projection rate were modulated based on the cardiac and respiratory signals to maximize the angular consistency between adjacent projections in the gated cardiac-respiratory bin. The mechanical limitations of a gantry-mounted CBCT system were investigated. For our protocol, images were acquired during the 60%-80% window of cardiac phase and 20% displacement either side of peak exhale of the respiratory cycle. The comparator method was the respiratory-only gated CBCT acquisition with constant gantry speed and projection rate in routine use for respiratory correlated four-dimensional (4D) CBCT. All images were reconstructed using the Feldkamp-Davis-Kress (FDK) algorithm. The methods were compared in terms of image sharpness as measured using the edge response width (ERW) and contrast as measured using the contrast to noise ratio (CNR). The effects of the total number of projections acquired and magnitude of cardiac motion on scan time and image quality were also investigated. RESULTS: Median total scan times with our protocol ranged from 117 s (40 projections) through to 296 s (100 projections), compared with 240 s for the conventional protocol (1320 projections). The scan times were dictated by the number of projections acquired, heart rate, length of the respiratory cycle and mechanical constraints of the CBCT system. Our protocol was able to provide between 8% and 43% decrease in the median value of the ERW in the anterior/posterior (AP) direction across the 17 traces when there was 0.5 cm of cardiac motion and between 35% and 64% decrease when there was 1.0 cm of cardiac motion over conventional acquisition. In the superior-inferior (SI) direction, our protocol was able to provide between 22% and 26% decrease in the median value of the ERW across the 17 traces when there was 0.5 cm of cardiac motion and between 30% and 35% decrease when there was 1.0 cm of cardiac motion over conventional acquisition. The magnitude of the cardiac motion did not have an observable effect on the median value of the CNR. Across all 17 traces, our adaptive protocol produced noticeably more consistent, albeit lower CNR values compared with conventional acquisition. CONCLUSION: For the first time, the potential of adapting CBCT image acquisition to changes in the patient's cardiac and respiratory rates simultaneously for applications in radiotherapy was investigated. This work represents a step towards thoracic imaging that reduces the effects of both cardiac and respiratory motion on image quality.

Both four-dimensional computed tomography and four-dimensional cone beam computed tomography under-predict lung target motion during radiotherapy.

BACKGROUND AND PURPOSE: To test the hypothesis that 4DCT and 4DCBCT-measured target motion ranges predict target motion ranges during lung cancer SABR. MATERIALS AND METHODS: Ten lung SABR patients were implanted with Calypso beacons. 4DCBCT was reconstructed for 29 fractions (1-4fx/patient) from a 1 min CBCT scan. The beacon centroid motion segmented for all 4DCT and 4DCBCT bins was compared with the real-time imaging and treatment beacon centroid ("target") motion range (4SDs) for each fraction. We tested the hypotheses that (1) 4DCT and 4CBCT predict treatment motion range and (2) there is no difference between 4DCT and 4DCBCT for predicting treatment motion range. Phase-wise root-mean-square errors (RMSEs) between imaging and treatment motion and reconstructed motion (4DCT, 4DCBCT) were calculated. Relationships between motion ranges in 4DCT and 4DCBCT and imaging and treatment motion ranges were investigated for the superior-inferior (SI), left-right (LR) and anterior-posterior (AP) directions. Baseline drifts and amplitude variability were investigated as potential factors leading to motion misrepresentation. RESULTS: SI 4DCT, 4DCBCT, imaging and treatment motion ranges were 6.3 ±â€¯3.6 mm, 7.1 ±â€¯4.5 mm, 11.1 ±â€¯7.5 mm and 10.9 ±â€¯6.9 mm, respectively. Similar 4DCT and 4DCBCT under-predictions were observed in the LR and AP directions. Hypothesis (1) was rejected (p < 0.0001). Treatment target motion range was under-predicted in 4DCT by factors of 1.7, 1.9 and 1.7 and in 4DCBCT by factors of 1.5, 1.6 and 1.6 in the SI, LR, and AP directions, respectively. RMSEs were generally lower for end-exhale than inhale. 4DCBCT showed higher correlations with the imaging and treatment target motion than 4DCT and testing hypothesis (2) a statistically significant difference between 4DCT and 4DCBCT was shown in the SI direction (p = 0.03). CONCLUSION: For lung SABR patients both 4DCT and 4DCBCT significantly under-predict treatment target motion ranges.

A deep learning framework for automatic detection of arbitrarily shaped fiducial markers in intrafraction fluoroscopic images.

PURPOSE: Real-time image-guided adaptive radiation therapy (IGART) requires accurate marker segmentation to resolve three-dimensional (3D) motion based on two-dimensional (2D) fluoroscopic images. Most common marker segmentation methods require prior knowledge of marker properties to construct a template. If marker properties are not known, an additional learning period is required to build the template which exposes the patient to an additional imaging dose. This work investigates a deep learning-based fiducial marker classifier for use in real-time IGART that requires no prior patient-specific data or additional learning periods. The proposed tracking system uses convolutional neural network (CNN) models to segment cylindrical and arbitrarily shaped fiducial markers. METHODS: The tracking system uses a tracking window approach to perform sliding window classification of each implanted marker. Three cylindrical marker training datasets were generated from phantom kilovoltage (kV) and patient intrafraction images with increasing levels of megavoltage (MV) scatter. The cylindrical shaped marker CNNs were validated on unseen kV fluoroscopic images from 12 fractions of 10 prostate cancer patients with implanted gold fiducials. For the training and validation of the arbitrarily shaped marker CNNs, cone beam computed tomography (CBCT) projection images from ten fractions of seven lung cancer patients with implanted coiled markers were used. The arbitrarily shaped marker CNNs were trained using three patients and the other four unseen patients were used for validation. The effects of full training using a compact CNN (four layers with learnable weights) and transfer learning using a pretrained CNN (AlexNet, eight layers with learnable weights) were analyzed. Each CNN was evaluated using a Precision-Recall curve (PRC), the area under the PRC plot (AUC), and by the calculation of sensitivity and specificity. The tracking system was assessed using the validation data and the accuracy was quantified by calculating the mean error, root-mean-square error (RMSE) and the 1st and 99th percentiles of the error. RESULTS: The fully trained CNN on the dataset with moderate noise levels had a sensitivity of 99.00% and specificity of 98.92%. Transfer learning of AlexNet resulted in a sensitivity and specificity of 99.42% and 98.13%, respectively, for the same datasets. For the arbitrarily shaped marker CNNs, the sensitivity was 98.58% and specificity was 98.97% for the fully trained CNN. The transfer learning CNN had a sensitivity and specificity of 98.49% and 99.56%, respectively. The CNNs were successfully incorporated into a multiple object tracking system for both cylindrical and arbitrarily shaped markers. The cylindrical shaped marker tracking had a mean RMSE of 1.6 ± 0.2 pixels and 1.3 ± 0.4 pixels in the x- and y-directions, respectively. The arbitrarily shaped marker tracking had a mean RMSE of 3.0 ± 0.5 pixels and 2.2 ± 0.4 pixels in the x- and y-directions, respectively. CONCLUSION: With deep learning CNNs, high classification performances on unseen patient images were achieved for both cylindrical and arbitrarily shaped markers. Furthermore, the application of CNN models to intrafraction monitoring was demonstrated using a simple tracking system. The results demonstrate that CNN models can be used to track markers without prior knowledge of the marker properties or an additional learning period.

Towards patient connected imaging with ACROBEAT: Adaptive CaRdiac cOne BEAm computed Tomography.

Robotic C-arm cone beam computed tomography (CBCT) systems are playing an increasingly pivotal role in interventional cardiac procedures and high precision radiotherapy treatments. One of the main challenges in any form of cardiac imaging is mitigating the intrinsic motion of the heart, which causes blurring and artefacts in the 3D reconstructed image. Most conventional 3D cardiac CBCT acquisition techniques attempt to combat heart motion through retrospective gating techniques, whereby acquired projections are sorted into the desired cardiac phase after the completion of the scan. However, this results in streaking artefacts and unnecessary radiation exposure to the patient. Here, we present our Adaptive CaRdiac cOne BEAm computed Tomography (ACROBEAT) acquisition protocol that uses the patient's electrocardiogram (ECG) signal to adaptively regulate the gantry velocity and projection time interval in real-time. It enables prospectively gated patient connected imaging in a single sweep of the gantry. The XCAT digital software phantom was used to complete a simulation study to compare ACROBEAT to a conventional multi-sweep retrospective ECG gated acquisition, under a variety of different acquisition conditions. The effect of location and length of the acquisition window and total number of projections acquired on image quality and total scan time were examined. Overall, ACROBEAT enables up to a 5 times average improvement in the contrast-to-noise ratio, a 40% reduction in edge response width and an 80% reduction in total projections acquired compared to conventional multi-sweep retrospective ECG gated acquisition.

Real-time respiratory triggered four dimensional cone-beam CT halves imaging dose compared to conventional 4D CBCT.

Four dimensional cone-beam computed tomography (4D CBCT) improves patient positioning and the accuracy of radiation therapy for patients with mobile tumours. Generally, 4D CBCT requires many hundreds of x-ray projections to measure target trajectories and the imaging frequency is not adapted to the patient's respiratory signal resulting in over-sampling. In contrast, respiratory triggered 4D CBCT (RT 4D CBCT) is an acquisition technique that has been experimentally implemented and has shown to reduce the number of x-ray projections and thus 4D CBCT dose with minimal impact on image quality. The aim of this work is to experimentally investigate RT 4D CBCT in situ and measure target trajectory mean position, image quality and imaging dose from this approach. A commercially available phantom with programmable target motion was programmed with nine target trajectories derived from patient-measured respiratory traces known to span the range of image quality when used for 4D CBCT reconstruction. 4D CBCT datasets were acquired for each target trajectory using the RT 4D CBCT acquisition technique and the conventional 4D CBCT acquisition technique. From the reconstructed 4D CBCT datasets, target trajectory mean positions, imaging dose and image quality metrics were calculated and compared between the two techniques. Target trajectory and mean position were measured by tracking the target's displacement in the phantom; imaging dose was measured by counting the total number of x-ray projections acquired; and image quality was assessed by calculating the contrast-to-noise ratio (CNR), signal-to-noise ration (SNR) and edge response width (ERW). For each of the nine cases, the target trajectory mean position as determined by RT 4D CBCT and conventional 4D CBCT varied from the reference source trajectory mean position by 0.7 mm or less except for one case where a conventional 4D CBCT mean position varied by 1.3 mm. On the average of these nine studies, RT 4D CBCT required half as many projections as conventional 4D CBCT giving a 50% reduction in imaging dose. Overall, the image quality metrics (CNR and SNR) were marginally worse for RT 4D CBCT; ERW metric showed no statistically significant difference between the RT 4D CBCT and conventional 4D CBCT reconstructed datasets. Respiratory triggered 4D CBCT couples the real-time respiratory signal to the 4D CBCT image acquisition system and requires less imaging dose than conventional 4D CBCT to determine target trajectory mean positions.

Cone-beam CT reconstruction with gravity-induced motion.

Fixed-gantry cone-beam computed tomography (CBCT), where the imaging hardware is fixed while the subject is continuously rotated 360° in the horizontal position, has implications for building compact and affordable fixed-gantry linear accelerators (linacs). Fixed-gantry imaging with a rotating subject presents a challenging image reconstruction problem where the gravity-induced motion is coupled to the subject's rotation angle. This study is the first to investigate the feasibility of fixed-gantry CBCT using imaging data of three live rabbits in an ethics-approved study. A novel data-driven motion correction method that combines partial-view reconstruction and motion compensation was developed to overcome this challenge. Fixed-gantry CBCT scans of three live rabbits were acquired on a standard radiotherapy system with the imaging beam fixed and the rabbits continuously rotated using an in-house programmable rotation cradle. The reconstructed images of the thoracic region were validated against conventional CBCT scans acquired at different cradle rotation angles. Results showed that gravity-induced motion caused severe motion blur in all of the cases if unaccounted for. The proposed motion correction method yielded clinically usable image quality with <1 mm gravity-induced motion blur for rabbits that were securely immobilized on the rotation cradle. Shapes of the anatomic structures were correctly reconstructed with <0.5 mm accuracy. Translational motion accounted for the majority of gravity-induced motion. The motion-corrected reconstruction represented the time-averaged location of the thoracic region over a 360° rotation. The feasibility of fixed-gantry CBCT has been demonstrated. Future work involves the validation of imaging accuracy for human subjects, which will be useful for emerging compact fixed-gantry radiotherapy systems.

4-Dimensional Cone Beam Computed Tomography-Measured Target Motion Underrepresents Actual Motion.

PURPOSE: Four-dimensional cone beam computed tomography (4DCBCT) facilitates verification of lung tumor motion before each treatment fraction and enables accurate patient setup in lung stereotactic ablative body radiation therapy. This work aims to quantify the real-time motion represented in 4DCBCT, depending on the reconstruction algorithm and the respiratory signal utilized for reconstruction. METHODS AND MATERIALS: Eight lung cancer patients were implanted with electromagnetic Calypso beacons in airways close to the tumor, enabling real-time motion measurements. 4DCBCT scans were reconstructed from projections for treatment setup CBCT for 1 to 2 fractions of 8 patients with the Feldkamp-Davis-Kress (FDK) algorithm or the prior image constrained compressed sensing (PICCS) method and internal real-time Calypso beacon trajectories or an external respiratory signal (bellows belt). The real-time beacon centroid ("target") motion was compared with beacon centroid positions segmented in the 4DCBCT reconstructions. We tested the hypotheses that (1) the actual target motion was accurately represented in the reconstructions and (2) the reconstruction/respiratory signal combinations performed similarly in the representation of the real-time motion. RESULTS: On average the target motion was significantly underrepresented and exceeded the 4DCBCT motion for 48%, 25%, and 40% of the time in the left-right (LR), superior-inferior (SI), and anterior-posterior (AP) directions, respectively. The average underrepresentation for the LR, SI, and AP direction was 1.7 mm, 4.2 mm, and 2.5 mm, respectively. No difference could be shown between the reconstruction algorithms or respiratory signals in LR direction (FDK vs PICCS: P = .47, Calypso vs bellows: P = .19), SI direction (FDK vs PICCS: P = .49, Calypso vs bellows: P = .22), and AP direction (FDK vs PICCS: P = .62, Calypso vs bellows: P = .34). CONCLUSIONS: The 4DCBCT scans all underrepresented the real-time target motion. The selection of the reconstruction algorithm and respiratory signal for the 4DCBCT reconstruction does not have an impact on the reconstructed motion range.

A CBCT study of the gravity-induced movement in rotating rabbits.

Fixed-beam radiotherapy systems with subjects rotating about a longitudinal (horizontal) axis are subject to gravity-induced motion. Limited reports on the degree of this motion, and any deformation, has been reported previously. The purpose of this study is to quantify the degree of anatomical motion caused by rotating a subject around a longitudinal axis, using cone-beam CT (CBCT). In the current study, a purpose-made longitudinal rotating was aligned to a Varian TrueBeam kV imaging system. CBCT images of three live rabbits were acquired at fixed rotational offsets of the cradle. Rigid and deformable image registrations back to the original position were used to quantify the motion experienced by the subjects under rotation. In the rotation offset CBCTs, the mean magnitude of rigid translations was 5.7 ± 2.7 mm across all rabbits and all rotations. The translation motion was reproducible between multiple rotations within 2.1 mm, 1.1 mm, and 2.8 mm difference for rabbit 1, 2, and 3, respectively. The magnitude of the mean and absolute maximum deformation vectors were 0.2 ± 0.1 mm and 5.4 ± 2.0 mm respectively, indicating small residual deformations after rigid registration. In the non-rotated rabbit 4DCBCT, respiratory diaphragm motion up to 5 mm was observed, and the variation in respiratory motion as measured from a series of 4DCBCT scans acquired at each rotation position was small. The principle motion of the rotated subjects was rigid translational motion. The deformation of the anatomy under rotation was found to be similar in scale to normal respiratory motion. This indicates imaging and treatment of rotated subjects with fixed-beam systems can use rigid registration as the primary mode of motion estimation. While the scaling of deformation from rabbits to humans is uncertain, these proof-of-principle results indicate promise for fixed-beam treatment systems.