FOD-Swin-Net: angular super resolution of fiber orientation distribution using a transformer-based deep model

TL;DR

This work addresses the challenge of obtaining high-angular-resolution diffusion information for fiber orientation estimation from limited-angle DW-MRI data. It introduces FOD-Swin-Net, a transformer-based, patch-wise Swin UNETR model that refines FODs reconstructed from single-shell data with $32$ directions toward the quality of multi-shell reconstructions with $288$ directions. Using SS3T for initial FOD, the network learns to translate 3D patches of SH coefficients to higher-angular-detail outputs, trained on HCP data and evaluated with ACC across white-matter tissues. Results show superior angular accuracy and significantly faster inference compared with prior methods, enabling practical single-shell acquisitions to achieve near multi-shell performance. This approach has implications for speeding diffusion imaging and improving tractography in clinical and research settings.

Abstract

Identifying and characterizing brain fiber bundles can help to understand many diseases and conditions. An important step in this process is the estimation of fiber orientations using Diffusion-Weighted Magnetic Resonance Imaging (DW-MRI). However, obtaining robust orientation estimates demands high-resolution data, leading to lengthy acquisitions that are not always clinically available. In this work, we explore the use of automated angular super resolution from faster acquisitions to overcome this challenge. Using the publicly available Human Connectome Project (HCP) DW-MRI data, we trained a transformer-based deep learning architecture to achieve angular super resolution in fiber orientation distribution (FOD). Our patch-based methodology, FOD-Swin-Net, is able to bring a single-shell reconstruction driven from 32 directions to be comparable to a multi-shell 288 direction FOD reconstruction, greatly reducing the number of required directions on initial acquisition. Evaluations of the reconstructed FOD with Angular Correlation Coefficient and qualitative visualizations reveal superior performance than the state-of-the-art in HCP testing data. Open source code for reproducibility is available at https://github.com/MICLab-Unicamp/FOD-Swin-Net.

Figures (4)

• Figure 1: Outline of the method. The downsampled DWI is turned into single-shell FOD using SS3T. 3D patches with significant brain white/gray matter presence are selected to train a Swin UNETR architecture. Targets for FOD reconstruction learned with MSE loss are derived from multi-shell 288 directions data. A whole volume is constructed through sliding windows.
• Figure 2: Half-violin distributions and boxplots for each method's ACC in the test set, in different regions: WM, WM with CGM, and WM with SGM.
• Figure 3: Qualitative visualizations in (b-d) showing FOD reconstruction in the region depicted in (a) for SS3T, FOD-Net, and FOD-Swin-Net (this work), with (e) being the multishell-based ground truth. The zoomed-in area corresponds to an intersection between the Corpus Callosum, the Superior Longitudinal Fascicle, and the Arcuate Fascicle - as indicated by the segmentation of the FODs into different fiber bundles using the TractSeg software tractseg
• Figure 4: ACC heatmap in the same middle WM axial slice from a subject for SS3T (a), FOD-Net (b) and FOD-Swin-Net (c).