High angular diffusion tensor imaging estimation from minimal evenly distributed diffusion gradient directions

Abstract

Diffusion-weighted Imaging (DWI) is a non-invasive imaging technique based on Magnetic Resonance Imaging (MRI) principles to measure water diffusivity and reveal details of the underlying brain micro-structure. By fitting a tensor model to quantify the directionality of water diffusion a Diffusion Tensor Image (DTI) can be derived and scalar measures, such as fractional anisotropy (FA), can then be estimated from the DTI to summarise quantitative microstructural information for clinical studies. In particular, FA has been shown to be a useful research metric to identify tissue abnormalities in neurological disease (e.g. decreased anisotropy as a proxy for tissue damage). However, time constraints in clinical practice lead to low angular resolution diffusion imaging (LARDI) acquisitions that can cause inaccurate FA value estimates when compared to those generated from high angular resolution diffusion imaging (HARDI) acquisitions. In this work, we propose High Angular DTI Estimation Network (HADTI-Net) to estimate an enhanced DTI model from LARDI with a set of minimal and evenly distributed diffusion gradient directions. Extensive experiments have been conducted to show the reliability and generalisation of HADTI-Net to generate high angular DTI estimation from any minimal evenly distributed diffusion gradient directions and to explore the feasibility of applying a data-driven method for this task. The code repository of this work and other related works can be found at https://mri-synthesis.github.io/.

Keywords: DTI; DWI; MRI; deep learning; fractional anisotropy; high angular resolution.

Figure 1

Mean FA of WM in DTIs generated using different numbers of evenly distributed diffusion gradient directions with 95% CI. Due to its nature, the Kennard-Stone algorithm guarantees that the lower number of direction samples are always included in higher ones (e.g., the 6 evenly distributed directions are part of the 12 evenly distributed ones).

Figure 2

Detailed framework of the proposed HADTI-Net. HADTI-Net takes a concatenated patch of 3D T1-weighted, b0, and diffusion-weighted images with 6 minimal evenly distributed directions to predict an enhanced LAR-DTI.

Figure 3

Visualisations of (A) LAR-DTI, (B) HAR-DTI, enhanced LAR-DTI by (C) DeepDTI and (D) HADTI-Net using (A) as the input for a testing subject. From the top to bottom are the axial, coronal, and sagittal views, each including a zoomed region of interest. The color coding of the diffusion tensor indicates directionality, whereby red, green, and blue represent right-left, anterior-posterior, and inferior-superior directions, respectively.

Figure 4

Mean absolute FA differences for LAR-DTI and enhanced LAR-DTI by HADTI-Net in 72 WM tracts and entire WM region. Analysed WM tracts in order of apperance: arcuate fascicle (AF), anterior thalamic radiation (ATR), commissure anterior (CA), corpus callosum (CC) and its subregions (rostrum, genu, rostral body, anterior midbody, posterior midbody, isthmus, and splenium), cingulum (CG), corticospinal tract (CST), Middle longitudinal fascicle (MLF), fronto-pontine tract (FPT), fornix (FX), inferior cerebellar peduncle (ICP), inferior occipito-frontal fascicle (IFO), inferior longitudinal fascicle (ILF), middle cerebellar peduncle (MCP), optic radiation (OR), parieto-occipital pontine (POPT), superior cerebellar peduncle (SCP), superior longitudinal fascicle (SLF), superior thalamic radiation (STR), uncinate fascicle (UF), thalamo-prefrontal (T_PREF), thalamo-premotor (T_PREM), thalamo-precentral (T_PREC), thalamo-postcentral (T_POSTC), thalamo-parietal (T_PAR), thalamo-occipital (T_OCC), striato-fronto-orbital (ST_FO), striato-prefrontal (ST_PREF), striato-premotor (ST_PREM), striato-precentral (ST_PREC), striato-postcentral (ST_POSTC), striato-parietal (ST_PAR), and striato-occipital (ST_OCC).

Figure 5

The distribution of the absolute FA differences when comparing LAR-DTI and enhanced LAR-DTI by HADTI-Net to HAR-DTI for all the testing subjects in each individual WM tract. The reported FA differences for different brain disorders are marked with corresponding annotations. Significant differences between LARDI and enhanced LAR-DTI values inside the selected WM tracts are marked on the top of each violin plot with a $\ast$ ($p<0.01$).

Figure 6

The flow chart for the genralisation experimental design and data preparation. Each set of inputs is then fed to the trained HADTI-Net to generate the enhanced LAR-DTI estimates and the MAE of FA when compared to HAR-DTI is calculated.