Room: Karl Dean Ballroom B1
Purpose: High-resolution 4D-MRI requires scanning times that make the technique impractical in a clinical setting. We propose a deep learning-based super-resolution 4D-MRI (SR-4DMRI) reconstruction method from low resolution 4D MRI for MR-guided radiation therapy (MR-gRT) applications.
Methods: In the context of single image super-resolution, a generative adversarial network (GAN) is a convolutional neural net framework that consists of two models: 1) a generative model that produces high-resolution images from corresponding low-resolution images, and 2) a discriminative model that distinguishes with a certain probability if a given high-resolution image is drawn from the true distribution of high-resolution images or generated by the other network. Here, low- and high-resolution image pairs from a small number of volunteers were used to train a downsampling neural network. High-resolution clinical data sets were then downsampled to form low- and high-resolution image pairs for training. After training, the super-resolution network was applied to unseen low-resolution 4D-MRI datasets for testing. The efficacy of the method was evaluated in terms of the peak signal-to-noise ratio (PSNR).
Results: The downsampling network was trained on 480 pairs of low & high resolution pairs from 4 volunteers and the superresolution netweork was trained on 2487 slices of sagittal and coronal images form the treated patients. The resolution of low and high resolution images in the experiment are 6.0 mm x 6.0mm and 1.5mm x 1.5mm respectively. The network was trained on NVIDIA GTX Titan Xp for 1000 epochs in 14 hours. High quality and accurate Super Resolution MRI reconstruction was achieved without compromising image quality compared to ground truth high resolution MR. The reconstruction time was 34 ms/slice on average.
Conclusion: The proposed SR-4DMRI method provides sufficient image quality for future evaluation in the delineation and tracking of targets in MR-gRT applications without sacrificing spatial resolution for clinically feasible acquisition times.
Funding Support, Disclosures, and Conflict of Interest: This research was supported by Varian. J.S.K. acknowledges that this work was supported by Ministry of Science, ICT and Future Planning, Korea through the R&D program of NRF-2015M3A9E2067001.