Single-Shell to MSMT fODF Reconstruction via a Convolutional ODE-Based Model

N.krishna, Vyshnavi Munagala, Harshini Pasupula, Deepa Pandey, Ranjeet Ranjan Jha

Keywords: Diffusion MRI fiber orientation distribution function, multi-shell reconstruction

Abstract:

Diffusion MRI (dMRI) is the most popular and powerful non-invasive medical imaging technology for neurological diagnosis. However, to accurately reconstruct the fiber orientation distribution functions and related microstructural biomarkers, multi-shell HARDI acquisitions with long scan times and high sensitivity to noise are still required. Current deep learning and spherical-harmonic reconstruction methods reduce acquisition requirements but remain constrained by the requirement of multi-shell data, large, labeled datasets, or limited interpretability. This work addresses these limitations by incorporating bio numerical modeling with modern learning-based reconstruction Framework. We improved the model’s accuracy by modifying the previous architecture, in which the Residual Block contain only three RK3 and Adam-Bashforth ODE solver layers. Added three modules in this improved model, Squeeze-Excitation, Directional Embedding, and a Masked Channel Transformer with Multi-Head Attention inside the Residual Block. This optimization cuts down on I/O overhead and accelerates the entire training pipeline. These improvements allowed our model to enhance feature representation and brought higher accuracy compared to the previous research model.Together,these methods demonstrate highly accurate reconstructions of high b-values diffusion signals, fODFs, tractography, and NODDI parameters from single-shell or low-resolution data. Empirical evaluations using HCP datasets demonstrate strong agreements with multi-shell references, improved noise robustness, and substantial scan duration reductions. Taken together, these developments indicate that bio numerically constrained and Learning-driven reconstruction pipelines can overcome major practical barriers in clinical dMRI, allowing fast, reliable, and interpretable microstructural estimation
from limited acquisitions.

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Application: Neuroimaging

Registration Requirement: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 264