DeepVolume: Brain Structure and Spatial Connection-Aware Network for Brain MRI Super-Resolution | JoVE Visualize

Area of Science:

Medical Imaging
Artificial Intelligence
Neuroscience

Background:

Thin-section magnetic resonance imaging (MRI) offers superior anatomical detail but is often unavailable due to cost.
Thick-section MRI data, common in multicenter studies, presents challenges for detailed analysis due to lower resolution and varying section thickness.
Accurate reconstruction of thin-section MRI from thick-section data is crucial for advancing large-scale neuroimaging research.

Purpose of the Study:

To introduce DeepVolume, a two-step deep learning architecture for accurate thin-section MRI reconstruction from thick-section images.
To enhance the precision and anatomical accuracy of reconstructed MRI by integrating structural and spatial information.
To validate the clinical utility of DeepVolume for applications like brain volume estimation and morphometry.

Main Methods:

A two-stage deep learning approach: a brain structure-aware network (multitask 3-D U-net) fuses axial and sagittal thick-section MRI with segmentation priors.
A spatial connection-aware network (recurrent convolutional network with LSTM) refines reconstructions slice-by-slice using sagittal information.
Utilized 305 paired brain MRI samples with 1.0 mm and 6.5 mm section thicknesses for training and validation.

Main Results:

DeepVolume achieved state-of-the-art reconstruction results, outperforming traditional methods by embedding anatomical knowledge.
The method successfully enhanced the precision of thin-section MRI reconstruction.
Validated clinical utility through improved brain volume estimation and voxel-based morphometry on thick-section data.

Conclusions:

DeepVolume effectively reconstructs high-quality thin-section MRI from thick-section data, addressing limitations of current imaging practices.
The integrated deep learning architecture leverages anatomical priors and spatial context for superior reconstruction accuracy.
DeepVolume offers significant potential for more reliable neuroimaging analysis, particularly in large-scale retrospective studies using existing thick-section MRI data.