1883
A hybrid PCA acceleration method for rapid real-time 2D target tracking1Oncology, University of Alberta, Edmonton, AB, Canada, 2Medical Physics, Cross Cancer Institute, Edmonton, AB, Canada
Synopsis
Keywords: Image Reconstruction, Radiotherapy, real-time, MR-Linac
Motivation: Most standard MR sequences are too slow for real-time applications. Accelerated acquisitions are one method to achieve the required frame rates. Neural-network reconstruction and parallel imaging are two methods that can achieve the necessary frame rates but are computationally expensive or require extensive coil arrays.
Goal(s): To develop an accelerated reconstruction method suitable for real-time applications which is computationally inexpensive and simple to implement.
Approach: Our method was tested retrospectively on lung, liver, and prostate patients for image quality and auto-contourability using a range of metrics.
Results: Image quality and contourability improved over similar methods while maintaining good reconstruction times for real-time applications.
Impact: An accelerated PCA-based reconstruction method was developed suitable for real-time applications, and in particular, target tracking. It has improved image quality and auto-contourability compared to similar methods while still maintaining simplicity in its implementation (low-cost computing, single coil arrays).
Introduction
The development of hybrid MRI/radiotherapy systems has resulted in the need for real-time MR imaging methods to be developed for target tracking purposes during radiotherapy treatments. To accurately capture motion detail during respiratory cycles, a minimum of four frames per second is targeted.1 Further to this it is recommended to have a maximum total system latency of 500ms for real-time applications.2 This includes everything from MR acquisition/reconstruction to auto-contouring and beam adjustments. Accelerated acquisitions can expand sequence options for the desired temporal resolution while maintaining good spatial resolution. Previous work investigated utilizing Principal Component Analysis (PCA) in the reconstruction of undersampled data.1,3 While these methods worked well, they suffered from temporal instability or was limited by instability when using higher order PCs for reconstruction. Compressed sensing, parallel imaging and neural-network reconstruction methods have been developed and can achieve the necessary frame rates for real-time applications. These methods however require large amounts of data for training, take long periods of time to train, are computationally expensive or require complex coil arrays. The goal of this work is to produce a reconstruction method that is low-latency for real-time accelerated applications while also making it suitable for systems without extensive coil arrays or a lot of computing power (i.e. GPUs).Methods
A new hybrid spatial/temporal PCA-based technique was developed for robust accelerated real-time image reconstruction. Data is undersampled such that a core set of phase encodes located in the central low-frequency region of k-space is acquired in every frame. Outside of this core, the outer high-frequency data is undersampled in such a way that all of k-space is acquired in a pre-determined number of frames. A sliding window of previous frames is used for the reconstruction of the current frame. Two separate techniques for reconstruction, one with PCA applied along the temporal direction of the core data, and another with PCA applied in the intra-frame dimension, were developed and optimized independently (Figures 1-2). Both involve fitting PCA characterizations of data within the sliding window to fit to available data within the final frame, allowing unsampled data to be estimated. In the combined hybridized approach, the data is first estimated by the intra-frame technique, followed by a secondary fitting and estimate using the time-domain method for improved robustness. Image quality and auto-contouring metrics were applied to 15 lung, 10 liver and 10 prostate patients to test the abilities of these methods at a variety of different acceleration rates. Contouring was performed using an in-house developed neural-network based software.4Results
The proposed hybrid method outperformed the independent techniques particularly at higher accelerations (Figure 3). SSIM values for lung and liver sites ranged from 0.97 at acceleration 3x to 0.91 at acceleration 8x. Similarly, NMSE ranged from .009 to .032 over the different acceleration rates. As seen in Figure 4, dice coefficient values for the hybrid method ranged from 0.94 for liver, 0.95 for lung and 0.97 for prostate at an acceleration of 3x to 0.88, 0.905 and 0.97 at the same sites for an acceleration of 8x. Hausdorff distance and centroid displacement showed similar trends.Discussion
Benefits were seen in image quality: NMSE values for the hybrid PCA method were found to be 0.012 compared to 0.025 for a previous PCA method3 at an acceleration of 4x. Auto-contourability also improved. For example, at an acceleration of 8x, dice coefficient improved from 0.88 for the previous PCA method to 0.92 for the hybrid method when looking at lung tumour contours. These metrics are qualitatively comparable to neural-network AUTOMAP based reconstruction software.5 The referenced AUTOMAP based work reported an average SSIM of 0.92 at a 4x acceleration, compared to 0.95 for our hybrid method, although direct comparison cannot be made due to different data sets. Three different sites were tested for the contourability of the reconstructed images from the hybrid method. Liver and lung were chosen due to the anatomical motion found in those regions. Similarly liver and prostate sites were selected due to the low contrast found in those regions. It was found that the contours performed well, within the range of expert intra-observer variability, when compared to a gold standard contour generated on a fully-sampled image (Figure 5).6Conclusion
A real-time hybrid PCA acceleration method was generated. It was found that image quality and auto-contouring accuracy improved compared to previous PCA methods as shown with metrics such as NMSE, SSIM and Dice coefficient. Reconstruction times were found to be ~50ms/frame (Intel Core i5-1135G7 CPU @ 2.40 GHz, 16GB RAM), well within the recommended parameters for real-time applications.Acknowledgements
The authors would like to acknowledge the Yau Family Foundation for helping fund this project.References
1. Dietz B, Yip E, Yun J, Fallone BG, Wachowicz K. Real-time dynamic MR image reconstruction using compressed sensing and principal component analysis (CS - PCA): demonstration in lung tumor tracking. Med Phys. 2017;44(8):3978-3989.
2. Keall PJ, Mageras GS, Balter JM, et al. The management of respiratory motion in radiation oncology report of ANMSEM Task Group 76. Med Phys. 2006;33:3874–3900.
3. Wright M, Dietz B, Yip E, Yun J, Gabos Z, Fallone, BG, Wachowicz K. Time domain principal component analysis for rapid, real-time 2D MRI reconstruction from undersampled data. Med Phys. 2021;48:6724-6739.
4. Han G, Wachowicz K, Usmani N, Yee D, Wong J, Elangovan A. Fallone BG, Yun J. Deep learning-based autocontouring algorithm for non-invasive intrafractional tumour-tracked radiotherapy (nifteRT) on Linac-MR. Medical Physics. 2022;49(8):5634.
5. Waddington DEJ, Hindley N, Koonjoo N, et al. Real-time radial reconstruction with domain transform manifold learning for MRI-guided radiotherapy. Med Phys. 2023;50:1962–1974. https://doi.org/10.1002/mp.16224
6. Yip E, Yun J, Gabos Z, et al. Evaluating performance of a user-trained MR lung tumor autocontouring algorithm in the context of intra- and inter-observer variations. Medical Physics.2018;45(1):307-313.
Figures
Figure 1
a) An example sliding window and how the k-space data is acquired. Highlighted is the core set of phase encodes from which PCA is applied and a PC matrix is generated
b) The inverse truncated PC matrix is multiplied with the acquired/estimated k-space data from the previous NWIN-1 frames (DOUTER) to calculate weights (A). This method assumes that fluctuations in outer k-space can be represented from fluctuations in central k-space due to common physiological motion
c) These weights (A) are multiplied with the PCs from the frame of interest (PCEND) to fill in the missing k-space data
Figure 2
a) PCA is applied to all acquired data within the sliding window
b) The portion of the PCs corresponding to the core data (PCCORE) is inverted and multiplied with the core data from the frame of interest to calculate weights (A)
c) These weights are then multiplied with the PCs representative of the acquired outer k-space data of a particular undersampling pattern to estimate the missing data in the frame of interest for the phase encodes of a particular undersampling pattern. The entire process is repeated NCOMP-1 times so that all of the missing data is filled in
Figure 3
Normalized Mean Square Error (NMSE), Peak SNR (PSNR) and Strucutral Similarity Index (SSIM) for the dynamic intra-frame and hybrid reconstruction methods over a range of accelerations. Results were generated on 15 lung and 5 liver patient data sets
Figure 4
Average contour metrics over ten lung, ten liver, and ten prostate tumour patients over a range of acceleration rates. Images were reconstructed using the hybrid reconstruction method.
Figure 5
A visual representation of the contour developed on the accelerated reconstructions and the original fully-sampled image for one lung (top), liver (middle) and prostate (bottom) patient.