Abstract
Current deep learning approaches for diffusion MRI modeling circumvent the need for densely-sampled diffusion-weighted images (DWIs) by directly predicting microstructural indices from sparsely-sampled DWIs. However, they implicitly make unrealistic assumptions of static $q$-space sampling during training and reconstruction. Further, such approaches can restrict downstream usage of variably sampled DWIs for usages including the estimation of microstructural indices or tractography. We propose a generative adversarial translation framework for high-quality DWI synthesis with arbitrary $q$-space sampling given commonly acquired structural images (e.g., B0, T1, T2). Our translation network linearly modulates its internal representations conditioned on continuous $q$-space information, thus removing the need for fixed sampling schemes. Moreover, this approach enables downstream estimation of high-quality microstructural maps from arbitrarily subsampled DWIs, which may be particularly important in cases with sparsely sampled DWIs. Across several recent methodologies, the proposed approach yields improved DWI synthesis accuracy and fidelity with enhanced downstream utility as quantified by the accuracy of scalar microstructure indices estimated from the synthesized images.
Introduction
Diffusion MRI (dMRI) is a critical modality for investigating brain microstructure by measuring the diffusion of water molecules. Advanced models like Neurite Orientation Dispersion and Density Imaging (NODDI) and Diffusion Kurtosis Imaging (DKI) offer deep insights into tissue configuration but require densely sampled $q$-space data, leading to long scan times. These long scans are susceptible to patient motion and artifacts, often resulting in corrupted gradient directions that must be discarded, leaving "holes" in the dataset.
While deep learning has been employed to predict microstructural maps from undersampled data, most existing models assume a static sampling pattern. This lack of flexibility makes them unsuitable for clinical scans with varying sampling schemes. This paper introduces a generative adversarial network (GAN) that synthesizes high-quality DWIs from standard structural MRI scans (T1, T2, and B0) by conditioning the network on continuous $q$-space information, allowing for the generation of DWIs for any arbitrary gradient direction or b-value.
Methodology
The proposed framework utilizes a $q$-space Conditioned Translation Network within a GAN architecture. The system consists of two primary components:
- $Q$-space Conditioned Generator: The generator accepts T1, T2, and B0 images as input along with specific $q$-space coordinates (b-vector and b-value). It employs Feature-wise Linear Modulation (FiLM) to inject this coordinate information into the network's internal layers, enabling the synthesis of images across a continuous range of directions.
- Conditional U-Net Discriminator: A U-Net-based discriminator assesses both global realism (via the bottleneck) and pixel-level detail (via the output feature map) to ensure high-fidelity synthesis.
The model incorporates physics-informed constraints, such as recognizing antipodal symmetry in diffusion signals and ensuring that a b-value of zero correctly outputs the baseline (B0) image. The study utilized the HCP 500 dataset, training on 9 subjects and testing on 9 held-out subjects with DWIs at b-values of 1000, 2000, and 3000 $s/mm^2$.
Results
The proposed model (Model D: B0+T1+T2 + GAN) demonstrated superior performance in both image synthesis and downstream microstructural estimation. It achieved a Peak Signal-to-Noise Ratio (PSNR) of 29.77 and a Structural Similarity Index (SSIM) of 0.944, outperforming non-adversarial and single-modality baselines.
A key finding is the "downstream utility" of the synthesized images. When tested on heavily downsampled data (e.g., 13 or 26 DWIs), the synthesized images allowed for significantly more accurate estimation of NODDI and DKI parameters compared to standard model-fitting or existing deep learning approaches (q-DL). Qualitatively, the GAN-based approach recovered sharp high-frequency details, such as the external capsule, which were often blurred in simpler regression models.
Discussion and Conclusion
A major insight of this work is that synthesizing missing DWIs before calculating microstructural indices is more effective than direct regression of those indices from structural MRI. This approach preserves the flexibility to use any diffusion model post-synthesis. Furthermore, the use of multi-modal inputs (B0, T1, and T2) provides a clear "staircase" of improvement in structural fidelity.
While promising, future work must address the potential bias toward healthy cohorts, as the model's performance on pathological tissues (e.g., lesions) remains to be validated. Additionally, the framework's reliance on perfect inter-modality registration and resolution matching presents a challenge for real-world clinical application. Nonetheless, this $q$-space conditioned approach offers a robust solution for handling unpredictable data corruption and sparse sampling in diffusion MRI.
Summary updated