Imaging Intravoxel Vessel Size Distribution in the Brain Using Susceptibility Contrast Enhanced MRI

Vascular remodelling is inherent to the pathogenesis of many diseases including cancer, neurodegeneration, fibrosis, hypertension, and diabetes. In this paper, a new susceptibility-contrast based MRI approach is established to analyse intravoxel vessel size distribution (VSD) enabling more comprehensive and quantitative assessment of vascular remodelling than existing clinical imaging modalities. We use segmented vascular structures from light-sheet fluorescence microscopy images of whole rodent brain to simulate gradient echo sampling of free induction decay and spin echo sequence (GESFIDE) and train a deep learning model to predict cerebral blood volume (CBV) and VSD from the simulated GESFIDE signal. The results from ex vivo experiments showed strong correlation (r=0.96) between the true and predicted CBV. Also, high similarity between true and predicted VSDs was observed with mean Bhattacharya Coefficient being 0.92. With further in vivo validation, intravoxel VSD imaging could become a transformative clinical tool for interrogating disease and treatment induced vascular remodelling.

💡 Research Summary

The paper introduces a novel susceptibility‑contrast MRI method for non‑invasively imaging the intravoxel vessel size distribution (VSD) in the brain, a metric that captures heterogeneity of the microvascular network beyond traditional mean vessel size or cerebral blood volume (CBV) measures. To overcome the limitations of existing approaches—most notably the reliance on simplified cylindrical vessel models and dictionary‑matching techniques—the authors combine high‑resolution light‑sheet fluorescence microscopy (LSFM) of whole rodent brains with physics‑based simulation of the GESFIDE (Gradient Echo Sampling of Free Induction Decay and Spin Echo) MRI sequence, and then train deep learning (DL) models to predict CBV and the full VSD from simulated MRI signals.

Data acquisition and ground‑truth generation
One rat, one mouse, and a rat bearing a patient‑derived GBM10 xenograft tumor were perfused, cleared, and imaged with LSFM at isotropic resolutions of 1 µm (mouse) and 1 × 1 × 3 µm (rat). After resampling to 1.8 µm, cubic volumes of interest (VOIs) of size 123 × 123 × 123 voxels were extracted. An image‑processing pipeline (CLAHE, Li‑Lee thresholding, morphological closing) produced binary vessel masks. Vessel radii were computed voxel‑wise, and the “true” VSD was defined as a normalized vessel‑volume‑fraction (vvf) weighted histogram of radii with 1 µm bins. CBV was calculated as the fraction of non‑zero voxels in the mask.

Signal simulation
For each VOI, the authors simulated pre‑ and post‑contrast GESFIDE signals using a Bloch‑based model that incorporates susceptibility differences (Δχ) of an iron‑oxide nanoparticle contrast agent, diffusion effects, and the combined GE, ASE, and SE contrasts. The ratio of post‑ to pre‑contrast signals forms a compact feature vector that encodes information about both blood volume and vessel size distribution.

Deep learning framework
Two fully‑connected neural networks (FCNs) were trained separately: one to regress CBV, the other to predict the full VSD histogram. Training employed mean‑squared error for CBV and a combined loss of MSE plus Bhattacharyya distance for VSD, encouraging accurate reconstruction of the distribution shape. The models were trained on 32 000 synthetic VOIs and validated on three independent test sets.

Performance evaluation

• Healthy brain VOIs (n = 3 132): CBV predictions achieved Pearson r = 0.95, and VSD predictions yielded a mean Bhattacharyya Coefficient (BC) of 0.87, indicating strong agreement with ground truth.
• Public mouse vasculature dataset (n = 1 000): Similar performance confirmed model generalizability across independently acquired vascular networks.
• Tumor VOIs (n = 706): Accuracy decreased modestly (CBV r = 0.78, VSD BC = 0.82), reflecting the more complex and heterogeneous tumor vasculature but still demonstrating practical utility.
• Noise robustness: Simulated SNR levels of 15, 30, 45, and 60 dB were tested. The DL models consistently outperformed traditional dictionary matching, maintaining higher correlation and BC values, especially at low SNR where dictionary methods degraded sharply.

Comparison with dictionary‑matching
The authors benchmarked their DL approach against a conventional dictionary‑matching pipeline that searches a pre‑computed library of GESFIDE signals generated from 2‑D or 3‑D cylindrical vessel models. The DL models not only reduced computational load (no exhaustive search) but also achieved superior accuracy, particularly when the simulated signals were corrupted by noise or when the underlying vascular geometry deviated from idealized cylinders.

Limitations and future directions
All validation was performed on ex‑vivo data and simulated MRI signals; in‑vivo translation will require accounting for physiological motion, variable contrast‑agent kinetics, and scanner‑specific sequence parameters. Moreover, the current FCN architecture treats each VOI independently and does not exploit spatial context; future work could explore 3‑D convolutional or graph‑based networks that directly ingest the binary vessel mask. Clinical feasibility will also depend on optimizing the GESFIDE acquisition for reasonable scan times in humans.

Conclusion
By integrating whole‑brain LSFM vascular atlases, realistic GESFIDE simulations, and data‑driven deep learning, the study provides a proof‑of‑concept for voxel‑wise quantification of vessel size heterogeneity. If validated in vivo, intravoxel VSD imaging could become a transformative tool for diagnosing, monitoring, and understanding vascular remodeling in a broad spectrum of neurological and systemic diseases.