Deep Learning Reconstruction Helps to Achieve Diagnostic-like Image Quality in MR Radiation Therapy of Brain.

PURPOSE/OBJECTIVE(S) In order to obtain reliable clinical images for treatment planning, MR Simulation in Radiation Therapy (RT) has different requirements than diagnostic MR imaging. The images acquired for treatment planning purposes need to have a higher spatial resolution (for better delineation of tumors and other structures) and higher geometric fidelity (for more accurate treatment planning and delivery). Higher geometric accuracy can be achieved by increasing the RF receiver bandwidth (BW). Both increasing spatial resolution and BW result in a reduction of Signal to Noise Ratio (SNR) compared to diagnostic MR imaging. In addition, patient positioning and coil set up are altered in RT to accommodate patient positioning devices. For instance, in brain imaging, the thermoplastic mask does not fit in the diagnostic head coils and other coil solutions such as flex coils need to be employed. This compromised coil solution typically reduces SNR as well. Deep Learning Reconstruction (DLR) techniques have shown great potential to reduce the noise in MR images and improve SNR. The study aimed to evaluate DLR in MR RT imaging protocol in the brain to improve the SNR and achieve comparable images to diagnostic imaging. MATERIALS/METHODS A routine brain protocol recommended by AAPM task group 284 was scanned on a 3T MR scanner using both diagnostic 32 channel head coil and RT setup with two 16 channel flex coils. For diagnostic imaging thermoplastic mask was not used but images in RT setup were acquired with flat tabletop and thermoplastic mask. The protocol included 3D T1 FFE, 3D T2 FSE, and 3D T2 FLAIR sequences. The common parameters between diagnostic and RT setup for all sequences were FOV = 24×24 cm2, matrix size = 256×256, spatial resolution = 0.9×0.9 mm2, and slice thickness = 1 mm. The BW of 3D T1 FFE sequence using diagnostic head coil BW was 326 Hz/pixel, but it was increased to 651 Hz/pixel for RT protocol to improve the geometric accuracy. RESULTS An image quality comparison of 3D T1 FFE images acquired using the diagnostic setup with BW 326, RT setup with BW 651 with a standard reconstruction technique, and RT setup with BW 651 and DLR technique was shown a significant improvement using DLR in higher BW sequences. The RT setup images with standard reconstruction show increased noise in the images compared to diagnostic setup due to higher BW and compromised coil setup. The DLR significantly improved the image quality of RT setup by removing the noise while maintaining the contrast and anatomical structures. In addition, three sequences acquired using RT setup and reconstructed using standard reconstruction were compared to DLR. The images reconstructed using DLR attained higher SNR and improved image quality. This might help clinicians to better detect small lesions in the brain. CONCLUSION Deep learning reconstruction can be effectively employed to reduce the noise and improve SNR in MR RT protocols where the SNR is typically reduced and image quality is degraded due to specific requirements in RT setup and imaging.