Reference-based super-resolution of Magnetic Resonance Imaging.

  • 13



Name of the Speaker: Ms. Madhu Mithra (EE19S019)
Guide: Dr. Mohanasankar Sivaprakasam
Venue/Online meeting link:
Date/Time: 13th January 2023 (Friday), 9:00 AM

Magnetic Resonance Imaging (MRI) is a valuable clinical diagnostic modality for spine pathologies with excellent characterization for infection, tumor, degenerations, fractures and herniations. However in surgery, image-guided procedures continue to rely on CT and fluoroscopy, as MRI slice resolutions are typically insufficient. Building upon state-of-the-art single image super-resolution, we propose a reference-based, unpaired texture-transfer strategy for deep learning based MRI super-resolution (SR). In this work, we implement reference based super-resolution both as an attachable module without a trainable feature extractor and as a complete end-to-end trainable SR network.

First, we use the scattering transform based module to relate the texture features of image patches to unpaired reference image patches, and we include a loss term for multi-contrast texture consistency. Secondly, we propose a fully trainable transformer network for the reference based super resolution (RefSR) with lifting scheme based texture extractor. The commonly used VGG feature extractors in RefSR contain redundant and irrelevant information and neglect the high frequency details of reference image. The usage of data-adaptive, lifting scheme based wavelet transform represents the texture information in multiple scales and frequency sub-bands. The texture matching is carried out in the low frequency sub-band and the multi-scale swapped features are incorporated into the model at different stages. We compare our models with state-of- the-art networks using three publicly available datasets and observe improvement in PSNR and SSIM metrics for 4x super-resolution.