Single patient convolutional neural networks for real-time MR reconstruction: coherent low-resolution versus incoherent undersampling

Abstract

Accelerated MRI involves undersampling k-space, creating unwanted artifacts when reconstructing the data. While the strategy of incoherent k-space acquisition is proven for techniques such as compressed sensing, it may not be optimal for all techniques. This study compares the use of coherent low-resolution (coherent-LR) and incoherent undersampling phase-encoding for real-time 3D CNN image reconstruction. Data were acquired with our 3 T Philips Achieva system. A retrospective analysis was performed on six non-small cell lung cancer patients who received dynamic acquisitions consisting of 650 free breathing images using a bSSFP sequence. We retrospectively undersampled the data by 5x and 10x acceleration using the two phase-encoding schemes. A quantitative analysis was conducted evaluating the tumor segmentations from the CNN reconstructed data using the Dice coefficient (DC) and centroid displacement. The reconstruction noise was evaluated using the structural similarity index (SSIM). Furthermore, we qualitatively investigated the CNN reconstruction using prospectively undersampled data, where the fully sampled training data set is acquired separately from the accelerated undersampled data. The patient averaged DC, centroid displacement, and SSIM for the tumor segmentation at 5x and 10x was superior using coherent low-resolution undersampling. Furthermore, the patient-specific CNN can be trained in under 6 h and the reconstruction time was 54 ms per image. Both the incoherent and coherent-LR prospective CNN reconstructions yielded qualitatively acceptable images; however, the coherent-LR reconstruction appeared superior to the incoherent reconstruction. We have demonstrated that coherent-LR undersampling for real-time CNN image reconstruction performs quantitatively better for the retrospective case of lung tumor segmentation, and qualitatively better for the prospective case. The tumor segmentation mean DC increased for all six patients at 5x acceleration and the temporal (dynamic) variance of the segmentation was reduced. The reconstruction speed achieved for our current implementation was 54 ms, providing an acceptable frame rate for real-time on-the-fly MR imaging.

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

“…Currently, none of these methods can achieve the required acceleration factor combined with low-latency reconstruction to estimate motion within 500 ms. 10 Recently, deep learning (DL) has been proposed to speed up MRI reconstruction and motion estimation, achieving performances on par, if not higher, than its non-DL counterparts. [20][21][22][23][24][25] Specifically, DL models allow for fast inference, leaving the time-consuming step to the training phase, which can take hours or days.…”

“…Recently, deep learning (DL) has been proposed to speed up MRI reconstruction and motion estimation, achieving performances on par, if not higher, than its non‐DL counterparts. 20 , 21 , 22 , 23 , 24 , 25 Specifically, DL models allow for fast inference, leaving the time‐consuming step to the training phase, which can take hours or days.…”

On Real-time Image Reconstruction with Neural Networks for MRI-guided Radiotherapy

“…Recently, neural networks have enabled fast, accurate reconstruction of undersampled MRI data. [17][18][19][20] However, despite these prospects, the successful deployment of neural networks for real-time imaging applications on systems including MRI-Linacs (see Fig. 1a) still hinges on the availability of training data and utilization of a reconstruction framework suitable for more challenging, non-uniformly sampled image reconstruction.…”

Intrafraction motion tracking for MR-Linac prostate radiotherapy

“…While training such networks is time consuming (ranging from hours to days), trained networks can be quickly applied to new imaging data resulting in reconstructed images in a matter of milliseconds. Dietz et al [167] described a method in which a convolutional neural network was used to reconstructed under-sampled 2D MR data. In this method they achieved a reconstruction speed of 54 ms, allowing for (near) real-time 2D MR imaging.…”