Cerebrovascular super-resolution 4D Flow MRI – using deep learning to non-invasively quantify velocity, flow, and relative pressure

Edward Ferdian, David Marlevi, Jonas Schollenberger, Maria Aristova, Elazer R. Edelman, Susanne Schnell, C. Alberto Figueroa, David Nordsletten, Alistair A. YoungbioRxiv 2021 · DOI

TL;DR

This paper introduces a deep learning framework to create super-resolution 4D Flow MRI images of the brain's blood vessels. This enhanced resolution allows for more accurate, non-invasive measurement of blood flow and pressure, which is crucial for diagnosing and monitoring cerebrovascular diseases.

What Problem Does This Paper Address?

Cerebrovascular diseases, like stroke and aneurysms, are closely linked to how blood flows and the pressures within the brain's intricate network of vessels. Monitoring these hemodynamics non-invasively is key for diagnosis and treatment. However, current 4D Flow MRI techniques often lack the spatial resolution needed to accurately capture the fine details of these narrow, winding vessels. This low resolution can lead to significant biases in measurements of blood flow and, critically, relative pressure, hindering accurate clinical assessment.

Methods

The core of this work is a re-trained deep learning model, "4DFlowNet," originally developed for aortic flows. The authors adapted this network for cerebrovascular applications by:

  • Modifying the Architecture: Adjusting the patch input size to better suit smaller cerebrovascular vessels and switching activation functions to improve training stability.
  • Physics-Informed Pressure Estimation: Combining the super-resolved flow data with a physics-driven algorithm called "virtual work-energy relative pressure" (vWERP) to calculate relative pressure. This method is known for its ability to probe complex vascular structures.
  • Synthetic Data Generation: Creating realistic training data by simulating blood flow in patient-specific cerebrovascular models derived from MRI and CFD. This synthetic data was augmented with realistic noise characteristics from actual clinical scans.

Datasets

  • In-silico (Synthetic) Data: Generated from patient-specific computational fluid dynamics (CFD) models of the cerebrovasculature. Data was sampled at various resolutions (1.5 mm down to 0.375 mm) to create low-resolution/high-resolution pairs for training and validation. Four different subject models were used, representing various cerebrovascular conditions.
  • In-vivo (Clinical) Data: Acquired from 8 healthy volunteers using 3T MRI. Data was collected at two resolutions (1.1 mm and 0.8 mm isotropic) to serve as a basis for generating clinical-quality training data and for final validation.

Key Insights and Ideas

  • Retraining is Crucial: The original 4DFlowNet, trained on aortic flows, performed poorly on cerebrovascular data. Re-training the network with cerebrovascular-specific data significantly improved its performance, highlighting the importance of domain-specific training for deep learning models in medical imaging.
  • Super-Resolution Enhances Pressure Estimation: The most significant finding is that super-resolution imaging, when combined with the vWERP algorithm, substantially reduces the bias in relative pressure estimations that is present in low-resolution data. This is critical because pressure gradients are vital indicators of cerebrovascular disease.
  • Noise Reduction is a Bonus: Beyond just increasing resolution, the super-resolution process also effectively reduces noise in the flow images, leading to a better velocity-to-noise ratio (VNR).

Key Results

  • Velocity and Flow Accuracy: The super-resolved velocities showed excellent correlation with high-resolution reference data (k>0.91, R2>0.96 for core regions). Peak velocity magnitude estimation achieved a relative error of 12.0 ± 0.07%. Flow rate relative error was 6.6 ± 4.7%.
  • Pressure Recovery: Relative pressure estimation showed a significant improvement. Peak relative pressure error decreased from 23.3 ± 14.9% at low resolution (1 mm) to 11.0 ± 7.3% with super-resolution (0.5 mm), approaching the accuracy of high-resolution data (5.1 ± 2.3%). The RMSE for relative pressure also dropped from 1.1 ± 1.7 mmHg (low-res) to 0.3 ± 0.2 mmHg (super-res).
  • In-vivo Performance: The method was successfully applied to in-vivo data, generating images at <0.5 mm resolution and showing improved VNR (up to 4-times increase). Relative pressure estimates in the in-vivo cohort also showed reduced bias compared to base resolutions.

Limitations and Future Work

  • In-vivo Pressure Validation: Direct comparison with invasive catheter-based pressure measurements in vivo is still pending due to regulatory challenges.
  • Limited In-silico Data: The number of in-silico models used for training was modest; more diverse data could enhance network versatility.
  • CFD Model Accuracy: The accuracy of the super-resolution results relies on the accuracy of the underlying CFD models used for training data generation.
  • Data Storage: Super-resolution conversion significantly increases data storage requirements.
  • Future Work: The authors suggest expanding training data, modifying network architecture for better near-wall velocity prediction, and conducting more extensive clinical validation. They also propose exploring adaptable grid representations or graph-based networks to address data storage issues. Future clinical studies could evaluate the impact of improved relative pressure estimations on clinical risk stratification.

Summary updated

Abstract

ABSTRACT The development of cerebrovascular disease is tightly coupled to changes in cerebrovascular hemodynamics, with altered flow and relative pressure indicative of the onset, development, and acute manifestation of pathology. Image-based monitoring of cerebrovascular hemodynamics is, however, complicated by the narrow and tortuous vasculature, where accurate output directly depends on sufficient spatial resolution. To address this, we present a method combining dedicated deep learning and state-of-the-art 4D Flow MRI to generate super-resolution full-field images with coupled quantification of relative pressure using a physics-driven image processing approach. The method is trained and validated in a patient-specific in-silico cohort, showing good accuracy in estimating velocity (relative error: 12.0 ± 0.1%, mean absolute error (MAE): 0.07 ± 0.06 m/s at peak velocity), flow (relative error: 6.6 ± 4.7%, root mean square error (RMSE): 0.5 ± 0.1 mL/s at peak flow), and with maintained recovery of relative pressure through the circle of Willis (relative error: 11.0 ± 7.3%, RMSE: 0.3 ± 0.2 mmHg). Furthermore, the method is applied to an in-vivo volunteer cohort, effectively generating data at &lt;0.5mm resolution and showing potential in reducing low-resolution bias in relative pressure estimation. Our approach presents a promising method to non-invasively quantify cerebrovascular hemodynamics, applicable to dedicated clinical cohorts in the future.
BibTeX
@misc{ferdian2021cerebrovascular,
  author = {Edward Ferdian and David Marlevi and Jonas Schollenberger and Maria Aristova and Elazer R. Edelman and Susanne Schnell and C. Alberto Figueroa and David Nordsletten and Alistair A. Young},
  title = {{Cerebrovascular super-resolution 4D Flow MRI – using deep learning to non-invasively quantify velocity, flow, and relative pressure}},
  year = {2021},
  doi = {10.1101/2021.08.25.457611},
  howpublished = {bioRxiv},
}

Loading related content...

Figure