An online real-time reconstruction technique that combines compressed sensing with principal component analysis (CS-PCA) was developed for the purpose of adaptive radiotherapy using our Linac-MR system. Our technique uses a database of images, acquired prior to an incoherently accelerated acquisition, to fill in the missing lines of k-space using PCA. Our technique can reconstruct images ranging from 5-20 frames per second with minimal artefacts.
Six retrospective fully sampled dynamic data sets of patients with non-small cell lung cancer were used to investigate the CS-PCA algorithm. Fully sampled data was required as it allowed for a ground truth to compare tumour segmentation at different accelerations. The data sets were acquired at 3T and consisted of 650 free-breathing dynamic frames, acquired using a bSSFP pulse sequence (TE/TR = 1.1ms/2.2ms, FOV: 40x40cm2, 128x128). The CS-PCA algorithm is implemented in k-space and is shown schematically in Figure 1. Our CS-PCA implementation used a database of the initial 30 images, which were used to calculate the principal components (PCs) for reconstruction of the following 620 under-sampled images. The 620 images were incoherently under-sampled along the PE direction using a Monte Carlo based sampling scheme. The missing k-space data was calculated by projecting the current k-space data onto the PC's to generate the corresponding PC weights. The weighted PC's were summed together along with the mean database, and the missing k-space was iteratively updated. Figure 2 displays example mean database images, along with two (unweighted) PC's for two patients.
Acceleration factors ranging from 2-10x were investigated using both CS-PCA, and Split Bregman CS3. CS-PCA parameters were investigated and include the number of PCA iterations (updated PC weights per reconstruction), the PC threshold (dictates number of discarded PC's), and the database size. Metrics to determine the reconstruction quality include the artefact power (AP) and dice coefficients (DC) of the tumour segmentations. The DC compares the accelerated reconstruction contour to the fully sampled data contour for each image; a DC of 1 indicates a perfect segmentation. The contours are generated using a neural-network based segmentation algorithm4. As our Linac-MR system uses a 0.5T bi-planar magnet, simulated 0.5T images were generated by adding Gaussian noise to the 3T data sets5,6.
1. PJ Keall, et al. The management of respiratory motion in radiation oncology report of AAPM Task Group 76a. Med. Phys. 2006;33(10):3874-3900.
2. F Zong, M Nogueira, and P Galvosas. Fast reconstruction of highly undersampled MR images using one and two dimensional principal component analysis. MRI. 2016;34:227-237.
3. Goldstein T and Osher S.The Split Bregman Method for L1-Regularized Problems. SIAM. 2009;2(2):323-343.
4. J Yun, E Yip, Z Gabos, K Wachowicz, S Rathee, and BG Fallone. Neural-network based autocontouring algorithm for intrafractional lung-tumour tracking using Linac-MR. Med. Phys. 2015;42(5):2296-2310.
5. E Yip, et al. Prior data assisted compressed sensing: A novel MR imaging strategy for real time tracking of lung tumors. Med. Phys. 2014;41(8):082301 (12pp.).
6. J Yun, et al. Evaluation of a lung tumor autocontouring algorithm for intrafractional tumor tracking using low-field MRI: A phantom study. Med. Phys. 2012;39(3)1481-1494.