J Yun

Evaluation of a lung tumor autocontouring algorithm for intrafractional tumor tracking using low‐field MRI: A phantom study

PURPOSE The first aim of this study is to investigate the feasibility of online autocontouring of tumor in low field MR images (0.2 and 0.5 T) by means of a phantom and simulation study for tumor-tracking in linac-MR systems. The second aim of this study is to develop an MR compatible, lung tumor motion phantom. METHODS An autocontouring algorithm was developed to determine both the position and shape of a lung tumor from each intra fractional MR image. To initiate the algorithm, an expert user contours the tumor and its maximum anticipated range of motion (herein termed the Background) using pretreatment scan data. During treatment, the algorithm processes each intrafractional MR image and automatically contours the tumor. To evaluate this algorithm, the authors built a phantom that replicates the low field contrast parameters (proton density, T(1), T(2)) of lung tumors and healthy lung parenchyma. This phantom allows simulation of MR images with the expected lung tumor CNR at 0.2 and 0.5 T by using a single 3 T scanner. Dynamic bSSFP images (approximately 4 images per second) are acquired while the phantom undergoes a series of preprogrammed motions based on patient lung tumor motion data. These images are autocontoured off-line using our algorithm. The fidelity of autocontouring is assessed by comparing autocontoured tumor shape and its centroid position to the actual tumor shape and its position. RESULTS The algorithm successfully contoured the shape of a moving tumor model from dynamic MR images acquired every 275 ms. Dices coefficients of > 0.96 and > 0.93 are achieved in 0.5 and 0.2 T equivalent images, respectively. Also, the algorithm tracked tumor position during dynamic studies, with root mean squared error (RMSE) values of < 0.55 and < 0.92 mm for 0.5 and 0.2 T equivalent images, respectively. Autocontouring speed is approximately 5 ms for each image. CONCLUSIONS Dices coefficients of > 0.96 and > 0.93 are achieved between autocontoured and real tumor shapes, and the position of a tumor can be tracked with RMSE values of < 0.55 and < 0.92 mm in 0.5 and 0.2 T equivalent images, respectively. These results demonstrate the feasibility of lung tumor autocontouring in low field MR images, and, by extension, intrafractional lung tumor tracking with our laboratorys linac-MR system.

Neural-network based autocontouring algorithm for intrafractional lung-tumor tracking using Linac-MR.

PURPOSE To develop a neural-network based autocontouring algorithm for intrafractional lung-tumor tracking using Linac-MR and evaluate its performance with phantom and in-vivo MR images. METHODS An autocontouring algorithm was developed to determine both the shape and position of a lung tumor from each intrafractional MR image. A pulse-coupled neural network was implemented in the algorithm for contrast improvement of the tumor region. Prior to treatment, to initiate the algorithm, an expert user needs to contour the tumor and its maximum anticipated range of motion in pretreatment MR images. During treatment, however, the algorithm processes each intrafractional MR image and automatically generates a tumor contour without further user input. The algorithm is designed to produce a tumor contour that is the most similar to the experts manual one. To evaluate the autocontouring algorithm in the authors Linac-MR environment which utilizes a 0.5 T MRI, a motion phantom and four lung cancer patients were imaged with 3 T MRI during normal breathing, and the image noise was degraded to reflect the image noise at 0.5 T. Each of the pseudo-0.5 T images was autocontoured using the authors algorithm. In each test image, the Dice similarity index (DSI) and Hausdorff distance (HD) between the experts manual contour and the algorithm generated contour were calculated, and their centroid positions were compared (Δd centroid). RESULTS The algorithm successfully contoured the shape of a moving tumor from dynamic MR images acquired every 275 ms. From the phantom study, mean DSI of 0.95-0.96, mean HD of 2.61-2.82 mm, and mean Δd centroid of 0.68-0.93 mm were achieved. From the in-vivo study, the authors algorithm achieved mean DSI of 0.87-0.92, mean HD of 3.12-4.35 mm, as well as Δd centroid of 1.03-1.35 mm. Autocontouring speed was less than 20 ms for each image. CONCLUSIONS The authors have developed and evaluated a lung tumor autocontouring algorithm for intrafractional tumor tracking using Linac-MR. The autocontouring performance in the Linac-MR environment was evaluated using phantom and in-vivo MR images. From the in-vivo study, the authors algorithm achieved 87%-92% of contouring agreement and centroid tracking accuracy of 1.03-1.35 mm. These results demonstrate the feasibility of lung tumor autocontouring in the authors laboratorys Linac-MR environment.

An artificial neural network (ANN)-based lung-tumor motion predictor for intrafractional MR tumor tracking.

PURPOSE To address practical issues of implementing artificial neural networks (ANN) for lung-tumor motion prediction in MRI-based intrafractional lung-tumor tracking. METHODS A feedforward four-layered ANN structure is used to predict future tumor positions. A back-propagation algorithm is used for ANN learning. Adaptive learning is incorporated by continuously updating weights and learning rate during prediction. An ANN training scheme specific for MRI-based tracking is developed. A multiple-ANN structure is developed to reduce tracking failures caused by the lower imaging rates of MRI. We used particle swarm optimization to optimize the ANN structure and initial weights (IW) for each patient and treatment fraction. Prediction accuracy is evaluated using the 1D superior-inferior lung-tumor motions of 29 lung cancer patients for system delays of 120-520 ms, in increments of 80 ms. The result is compared with four different scenarios: (1), (2) ANN structure optimization + with∕without IW optimization, and (3), (4) no ANN structure optimization + with∕without IW optimization, respectively. An additional simulation is performed to assess the value of optimizing the ANN structure for each treatment fraction. RESULTS For 120-520 ms system delays, mean RMSE values (ranges 0.0-2.8 mm from 29 patients) of 0.5-0.9 mm are observed, respectively. Using patient specific ANN structures, a 30%-60% decrease in mean RMSE values is observed as a result of IW optimization, alone. No significant advantages in prediction performance are observed, however, by optimizing for each fraction. CONCLUSIONS A new ANN-based lung-tumor motion predictor is developed for MRI-based intrafractional tumor tracking. The prediction accuracy of our predictor is evaluated using a realistic simulated MR imaging rate and system delays. For 120-520 ms system delays, mean RMSE values of 0.5-0.9 mm (ranges 0.0-2.8 mm from 29 patients) are achieved. Further, the advantage of patient specific ANN structure and IW in lung-tumor motion prediction is demonstrated by a 30%-60% decrease in mean RMSE values.PURPOSE To address practical issues of implementing artificial neural networks (ANN) for lung-tumor motion prediction in MRI-based intrafractional lung-tumor tracking. METHODS A feedforward four-layered ANN structure is used to predict future tumor positions. A back-propagation algorithm is used for ANN learning. Adaptive learning is incorporated by continuously updating weights and learning rate during prediction. An ANN training scheme specific for MRI-based tracking is developed. A multiple-ANN structure is developed to reduce tracking failures caused by the lower imaging rates of MRI. We used particle swarm optimization to optimize the ANN structure and initial weights (IW) for each patient and treatment fraction. Prediction accuracy is evaluated using the 1D superior-inferior lung-tumor motions of 29 lung cancer patients for system delays of 120-520 ms, in increments of 80 ms. The result is compared with four different scenarios: (1), (2) ANN structure optimization + with/without IW optimization, and (3), (4) no ANN structure optimization + with/without IW optimization, respectively. An additional simulation is performed to assess the value of optimizing the ANN structure for each treatment fraction. RESULTS For 120-520 ms system delays, mean RMSE values (ranges 0.0-2.8 mm from 29 patients) of 0.5-0.9 mm are observed, respectively. Using patient specific ANN structures, a 30%-60% decrease in mean RMSE values is observed as a result of IW optimization, alone. No significant advantages in prediction performance are observed, however, by optimizing for each fraction. CONCLUSIONS A new ANN-based lung-tumor motion predictor is developed for MRI-based intrafractional tumor tracking. The prediction accuracy of our predictor is evaluated using a realistic simulated MR imaging rate and system delays. For 120-520 ms system delays, mean RMSE values of 0.5-0.9 mm (ranges 0.0-2.8 mm from 29 patients) are achieved. Further, the advantage of patient specific ANN structure and IW in lung-tumor motion prediction is demonstrated by a 30%-60% decrease in mean RMSE values.

Radio frequency noise from an MLC: a feasibility study of the use of an MLC for linac-MR systems.

Currently several groups are actively researching the integration of a megavoltage teletherapy unit with magnetic resonance (MR) imaging for real-time image-guided radiotherapy. The use of a multileaf collimator (MLC) for intensity-modulated radiotherapy for linac-MR units must be investigated. The MLC itself will likely reside in the fringe field of the MR and the motors will produce radio frequency (RF) noise. The RF noise power spectral density from a Varian 52-leaf MLC motor, a Varian Millennium MLC motor and a brushless fan motor has been measured as a function of the applied magnetic field using a near field probe set. For the Varian 52-leaf MLC system, the RF noise produced by 13 of 52 motors is studied as a function of distance from the MLC. Data are reported in the frequency range suitable for 0.2-1.5 T linac-MR systems. Below 40 MHz the Millennium MLC motor tested showed more noise than the Varian 52-leaf motor or the brushless fan motor. The brushless motor showed a small dependence on the applied magnetic field. Images of a phantom were taken by the prototype linac-MR system with the MLC placed in close proximity to the magnet. Several orientations of the MLC in both shielded and non-shielded configurations were studied. For the case of a non-shielded MLC and associated cables, the signal-to-noise ratio (SNR) was reduced when 13 of 52 MLC leaves were moved during imaging. When the MLC and associated cables were shielded, the measured SNR of the images with 13 MLC leaves moving was experimentally the same as the SNR of the stationary MLC image. When the MLC and cables are shielded, subtraction images acquired with and without MLC motion contains no systematic signal. This study illustrates that the small RF noise produced by functioning MLC motors can be effectively shielded to avoid SNR degradation. A functioning MLC can be incorporated into a linac-MR unit.

Prior data assisted compressed sensing: A novel MR imaging strategy for real time tracking of lung tumors

PURPOSE Hybrid radiotherapy-MRI devices promise real time tracking of moving tumors to focus the radiation portals to the tumor during irradiation. This approach will benefit from the increased temporal resolution of MRIs data acquisition and reconstruction. In this work, the authors propose a novel spatial-temporal compressed sensing (CS) imaging strategy for the real time MRI--prior data assisted compressed sensing (PDACS), which aims to improve the image quality of the conventional CS without significantly increasing reconstruction times. METHODS Conventional 2D CS requires a random sampling of partial k-space data, as well as an iterative reconstruction that simultaneously enforces the images sparsity in a transform domain as well as maintains the fidelity to the acquired k-space. PDACS method requires the additional acquisition of the prior data, and for reconstruction, it additionally enforces fidelity to the prior k-space domain similar to viewsharing. In this work, the authors evaluated the proposed PDACS method by comparing its results to those obtained from the 2D CS and viewsharing methods when performed individually. All three methods are used to reconstruct images from lung cancer patients whose tumors move and who are likely to benefit from lung tumor tracking. The patients are scanned, using a 3T MRI, under free breathing using the fully sampled k-space with 2D dynamic bSSFP sequence in a sagittal plane containing lung tumor. These images form a reference set for the evaluation of the partial k-space methods. To create partial k-space, the fully sampled k-space is retrospectively undersampled to obtain a range of acquisition acceleration factors, and reconstructed with 2D-CS, PDACS, and viewshare methods. For evaluation, metrics assessing global image artifacts as well as tumor contour shape fidelity are determined from the reconstructed images. These analyses are performed both for the original 3T images and those at a simulated 0.5T equivalent noise level. RESULTS In the 3.0T images, the PDACS strategy is shown to give superior results compared to viewshare and conventional 2D CS using all metrics. The 2D-CS tends to perform better than viewshare at the low acceleration factors, while the opposite is true at the high acceleration factors. At simulated 0.5T images, PDACS method performs only marginally better than the viewsharing method, both of which are superior compared to 2D CS. The PDACS image reconstruction time (0.3 s/image) is similar to that of the conventional 2D CS. CONCLUSIONS The PDACS method can potentially improve the real time tracking of moving tumors by significantly increasing MRIs data acquisition speeds. In 3T images, the PDACS method does provide a benefit over the other two methods in terms of both the overall image quality and the ability to accurately and automatically contour the tumor shape. MRIs data acquisition may be accelerated using the simpler viewsharing strategy at the lower, 0.5T magnetic field, as the marginal benefit of the PDACS method may not justify its additional reconstruction times.

Brushed permanent magnet DC MLC motor operation in an external magnetic field

PURPOSE Linac-MR systems for real-time image-guided radiotherapy will utilize the multileaf collimators (MLCs) to perform conformal radiotherapy and tumor tracking. The MLCs would be exposed to the external fringe magnetic fields of the linac-MR hybrid systems. Therefore, an experimental investigation of the effect of an external magnetic field on the brushed permanent magnet DC motors used in some MLC systems was performed. METHODS The changes in motor speed and current were measured for varying external magnetic field strengths up to 2000 G generated by an EEV electromagnet. These changes in motor characteristics were measured for three orientations of the motor in the external magnetic field, mimicking changes in motor orientations due to installation and/or collimator rotations. In addition, the functionality of the associated magnetic motor encoder was tested. The tested motors are used with the Varian 120 leaf Millennium MLC (Maxon Motor half leaf and full leaf motors) and the Varian 52 leaf MKII MLC (MicroMo Electronics leaf motor) including a carriage motor (MicroMo Electronics). RESULTS In most cases, the magnetic encoder of the motors failed prior to any damage to the gearbox or the permanent magnet motor itself. This sets an upper limit of the external magnetic field strength on the motor function. The measured limits of the external magnetic fields were found to vary by the motor type. The leaf motor used with a Varian 52 leaf MKII MLC system tolerated up to 450 +/- 10 G. The carriage motor tolerated up to 2000 +/- 10 G field. The motors used with the Varian 120 leaf Millennium MLC system were found to tolerate a maximum of 600 +/- 10 G. CONCLUSIONS The current Varian MLC system motors can be used for real-time image-guided radiotherapy coupled to a linac-MR system, provided the fringe magnetic fields at their locations are below the determined tolerance levels. With the fringe magnetic fields of linac-MR systems expected to be larger than the tolerance levels determined, some form of magnetic shielding would be required.

Real-time dynamic MR image reconstruction using compressed sensing and principal component analysis (CS-PCA): Demonstration in lung tumor tracking

Purpose: This work presents a real‐time dynamic image reconstruction technique, which combines compressed sensing and principal component analysis (CS‐PCA), to achieve real‐time adaptive radiotherapy with the use of a linac‐magnetic resonance imaging system. Methods: Six retrospective fully sampled dynamic data sets of patients diagnosed with non–small‐cell lung cancer were used to investigate the CS‐PCA algorithm. Using a database of fully sampled k‐space, principal components (PCs) were calculated to aid in the reconstruction of undersampled images. Missing k‐space data were calculated by projecting the current undersampled k‐space data onto the PCs to generate the corresponding PC weights. The weighted PCs were summed together, and the missing k‐space was iteratively updated. To gain insight into how the reconstruction might proceed at lower fields, 6× noise was added to the 3T data to investigate how the algorithm handles noisy data. Acceleration factors ranging from 2 to 10× were investigated using CS‐PCA and Split Bregman CS for comparison. Metrics to determine the reconstruction quality included the normalized mean square error (NMSE), as well as the dice coefficients (DC) and centroid displacement of the tumor segmentations. Results: Our results demonstrate that CS‐PCA performed superior than CS alone. The CS‐PCA patient averaged DC for 3T and 6× noise added data remained above 0.9 for acceleration factors up to 10× The patient averaged NMSE gradually increased with increasing acceleration; however, it remained below 0.06 up to an acceleration factor of 10× for both 3T and 6× noise added data. The CS‐PCA reconstruction speed ranged from 5 to 20 ms (Intel i7‐4710HQ CPU {L‐End}@ 2.5 GHz), depending on the chosen parameters. Conclusions: A real‐time reconstruction technique was developed for adaptive radiotherapy using a Linac‐MRI system. Our CS‐PCA algorithm can achieve tumor contours with DC greater than 0.9 and NMSE less than 0.06 at acceleration factors of up to, and including, 10× The reconstruction speed for the Split Bregman CS ranged from 200 to 260 ms, whereas the CS‐PCA reconstruction speed ranged from 5 to 20 ms implemented using nonoptimized MATLAB code.

Improved lung tumor autocontouring algorithm for intrafractional tumor tracking using 0.5 T linac-MR

To add an intelligent parameter optimization capability to our autocontouring algorithm, and evaluate its performance using in-vivo data. Methods An autocontouring algorithm for intrafractional lung-tumor tracking using linac-MR was previously developed based on pulse-coupled neural networks. The algorithms contouring performance is dependent on eight parameters (including four integer parameters). Previously, the parameters were optimized using a time-consuming, exhaustive method. To avoid this inefficiency, adaptive particle swarm optimization (APSO) was adopted in this study, which is a stochastic, non-gradient based optimization algorithm that can handle integer variables. For this study, six non-small cell lung cancer patients were imaged with 3T MRI at ~4 frames per second (2D sagittal plane, free breathing). For each patient, an expert delineated a gold standard contour (ROIstd) of the lung tumor in 130 consecutive images. The first 30 ROIstd were used for parameter optimization, and the rest 100 ROIstd were used to validate autocontours (ROIauto). In each image, Dice similarity index, Hausdorff distance, and centroid position difference (Δdcentroid) were calculated between ROIstd and ROIauto to measure their similarity. Results & Conclusion An efficient, fully automatic parameter optimization was added to our autocontouring algorithm. Using the six patients data, approximately 1/24 time reduction was achieved in parameter optimization (63–125 hrs to 2–4 hrs per patient), while maintaining the same or slightly improved performance.

Sliding window prior data assisted compressed sensing for MRI tracking of lung tumors

Purpose: Hybrid magnetic resonance imaging and radiation therapy devices are capable of imaging in real‐time to track intrafractional lung tumor motion during radiotherapy. Highly accelerated magnetic resonance (MR) imaging methods can potentially reduce system delay time and/or improves imaging spatial resolution, and provide flexibility in imaging parameters. Prior Data Assisted Compressed Sensing (PDACS) has previously been proposed as an acceleration method that combines the advantages of 2D compressed sensing and the KEYHOLE view‐sharing technique. However, as PDACS relies on prior data acquired at the beginning of a dynamic imaging sequence, decline in image quality occurs for longer duration scans due to drifts in MR signal. Novel sliding window‐based techniques for refreshing prior data are proposed as a solution to this problem. Methods: MR acceleration is performed by retrospective removal of data from the fully sampled sets. Six patients with lung tumors are scanned with a clinical 3 T MRI using a balanced steady‐state free precession (bSSFP) sequence for 3 min at approximately 4 frames per second, for a total of 650 dynamics. A series of distinct pseudo‐random patterns of partial k‐space acquisition is generated such that, when combined with other dynamics within a sliding window of 100 dynamics, covers the entire k‐space. The prior data in the sliding window are continuously refreshed to reduce the impact of MR signal drifts. We intended to demonstrate two different ways to utilize the sliding window data: a simple averaging method and a navigator‐based method. These two sliding window methods are quantitatively compared against the original PDACS method using three metrics: artifact power, centroid displacement error, and Dices coefficient. The study is repeated with pseudo 0.5 T images by adding complex, normally distributed noise with a standard deviation that reduces image SNR, relative to original 3 T images, by a factor of 6. Results: Without sliding window implemented, PDACS‐reconstructed dynamic datasets showed progressive increases in image artifact power as the 3 min scan progresses. With sliding windows implemented, this increase in artifact power is eliminated. Near the end of a 3 min scan at 3 T SNR and 5× acceleration, implementation of an averaging (navigator) sliding window method improves our metrics by the following ways: artifact power decreases from 0.065 without sliding window to 0.030 (0.031), centroid error decreases from 2.64 to 1.41 mm (1.28 mm), and Dice coefficient agreement increases from 0.860 to 0.912 (0.915). At pseudo 0.5 T SNR, the improvements in metrics are as follows: artifact power decreases from 0.110 without sliding window to 0.0897 (0.0985), centroid error decreases from 2.92 mm to 1.36 mm (1.32 mm), and Dice coefficient agreements increases from 0.851 to 0.894 (0.896). Conclusions: In this work we demonstrated the negative impact of slow changes in MR signal for longer duration PDACS dynamic scans, namely increases in image artifact power and reductions of tumor tracking accuracy. We have also demonstrated sliding window implementations (i.e., refreshing of prior data) of PDACS are effective solutions to this problem at both 3 T and simulated 0.5 T bSSFP images.

Tracking tumor boundary using point correspondence for adaptive radio therapy

BACKGROUND AND OBJECTIVE Tracking mobile tumor regions during the treatment is a crucial part of image-guided radiation therapy because of two main reasons which negatively affect the treatment process: (1) a tiny error will lead to some healthy tissues being irradiated; and (2) some cancerous cells may survive if the beam is not accurately positioned as it may not cover the entire cancerous region. However, tracking or delineation of such a tumor region from magnetic resonance imaging (MRI) is challenging due to photometric similarities of the region of interest and surrounding area as well as the influence of motion in the organs. The purpose of this work is to develop an approach to track the center and boundary of tumor region by auto-contouring the region of interest in moving organs for radiotherapy. METHODS We utilize a nonrigid registration method as well as a publicly available RealTITracker algorithm for MRI to delineate and track tumor regions from a sequence of MRI images. The location and shape of the tumor region in the MRI image sequence varies over time due to breathing. We investigate two approaches: the first one uses manual segmentation of the first frame during the pretreatment stage; and the second one utilizes manual segmentation of all the frames during the pretreatment stage. RESULTS We evaluated the proposed approaches over a sequence of 600 images acquired from 6 patients. The method that utilizes all the frames in the pretreatment stage with moving mesh based registration yielded the best performance with an average Dice Score of 0.89 ± 0.04 and Hausdorff Distance of 3.38 ± 0.10 mm. CONCLUSIONS This study demonstrates a promising boundary tracking tool for delineating the tumor region that can deal with respiratory movement and the constraints of adaptive radiation therapy.