Applying simultaneous super-resolution and contrast synthesis to routine clinical magnetic resonance images for automated segmentation of knee joint cartilage

Ales Neubert, Pierrick Bourgeat, Jason Wood, Craig Engstrom, Shekhar S. Chandra, Stuart Crozier, Jurgen Fripp

Keywords: super-resolution, contrast synthesis, knee cartilage segmentation

TL;DR: Use a 3D Unet to generate 3D synthetic DESS from low resolution TSE of the knee, and evaluation of their utility for cartilage segmentation.

Abstract: High resolution 3D MR images are well suited for automated cartilage segmentation in the human knee joint. However, volumetric scans such as 3D Double-Echo Steady-State (DESS) images, are not routinely acquired. Instead, typical clinical knee MR imaging exams involve acquisition of a series of 2D turbo spin echo (TSE) sequences. TSE images typically have high in-plane resolution (\textit{e.g.}\ 0.4 mm), but large slice thickness (\textit{e.g.}\ 3 mm). The cartilage visualization in the individual 2D TSE images is prone to partial volume artifacts due to the thick slices and high cartilage curvature, often resulting in cartilage appearing thinner than it actually is (\figureref{fig:segmentation}). Consequently, 2D TSE images of the human knee joint are not well suited for automatic cartilage segmentation. In this work, a patch-based UNet convolutional neural network is employed for synthesizing artificial 3D DESS scans (Syn-DESS) from 2D TSE. An automatic segmentation method is then employed to assess the suitability of the Syn-DESS images for knee cartilage segmentation.

Code Of Conduct: I have read and accept the code of conduct.

Remove If Rejected: Remove submission from public view if paper is rejected.