Super-Resolution of 3D Brain MRI With Filter Learning Using Tensor Feature Clustering

Abstract

Surface-based analysis of magnetic resonance imaging (MRI) data of the brain plays an important role in clinical and research applications. To achieve accurate three-dimensional (3D) surface reconstruction, high-resolution (HR) MR image acquisition is needed. However, HR image acquisition is hindered by hardware limitations that result in long acquisition time and low spatial coverage. Single image super-resolution (SISR) can alleviate these problems by converting a low-resolution (LR) image to an HR image. However, unlike 2D SISR methods, conventional 3D methods incur a large computational cost and require abundant data. Further, 3D boundaries for surface reconstruction based on MR images have not been sufficiently investigated. We herein propose a cost-efficient novel regression-based framework for full 3D super-resolution imaging that directly analyzes 3D features by introducing a tensor using gradient information. We initially cluster features using tensors to create labels for both the training and testing stages. In the training stage, for each label, we collect LR patches and corresponding HR intensities to compute filters. In the testing stage, for each voxel, we construct a tensor to obtain a feature and predict the HR intensity using trained filters. We also propose a patch span reduction method by limiting patch orientation to reduce the orientation span and increase shape variety. Using only 30 masked T1-weighted brain MR volumes from the Human Connectome Project (HCP) 900 dataset, the proposed algorithm exhibited superior performance in terms of HR boundary recovery in the cerebral cortex as well as improved overall quality compared to conventional methods.

Improving Structural MRI Preprocessing with Hybrid Transformer GANs

“…Then, additionally, authors have mentioned that they computed the gradient modulus (GM)-or, in other words, the image gradient-with a Scharr gradient operator. Partial derivatives for the image are calculated as in Equations ( 17) and (18).…”

“…The main difference was that they utilized adjacent MRI slices in the network layers. There have also been attempts to divide 3D volume into patches and then learn filters that are capable of upscaling patches, which can be combined back with the whole volume at the end, as described in [ 18 ]. We can also find other methods, such as deep 3D CNNs with skip-connections, like in [ 19 ] or [ 20 ].…”

Control Method of Complex Environmental Engineering Machinery Cooling System Integrating Deep Learning and Feature Clustering

“…The network uses massive training samples to learn more complex feature expressions and complete accurate recognition. The trained model has strong generalization ability [7][8] .…”