Current Research Projects

3.HemoVision: Unmasking SCTA’s Dynamic Blood Flow with Deep Learning

Cardiovascular diseases (CVDs) encompass various pathological conditions in the human vascular system, including aneurysms, arterio-venous malformations, vasospasms, atherosclerosis, and peripheral artery disease. Nearly half of the U.S. adult population is affected by CVDs. While preventive surgical and endovascular interventions are considered to avert catastrophic events like subarachnoid hemorrhages and strokes, these interventions carry significant risks. Current clinical decisions rely on anatomical measures derived from large population studies, lacking patient-specific disease states. Consequently, decisions based on such measures may result in unnecessary or incorrect interventions, leading to increased costs and patient morbidities. This  research project is focused on Scientific Machine Learning (SciML). the project integrates domain-specific models with cutting-edge artificial intelligence techniques. the main goal of this project is estimate time-resolved blood flow velocities and relative pressures from computed tomographic angiography (sCTA) imaging (as shown in figure below).

Overview of the project

The objective is to reconstruct dynamic blood flow velocity and relative pressure maps from time-stamped sCTA sinograms. Despite only having dynamic contrast concentration data in the sinograms, the method aims to estimate blood flow velocities and pressure maps by incorporating the coupled physics of contrast transport and blood flow during the reconstruction process.
The project has two main research goals to achieve:
  1. To estimate blood flow velocities and relative pressures from sCTA.
  2. Show that  physics-informed machine learning is a reliable and accurate is  in such estimation.
A 2D preliminary numerical test on a cerebral aneurysm  was used as a prove of concept.

Recovering hemodynamics from spatio-temporally averaged contrast data. 
Recovering contrast dynamics.

As shown in the figure above, the left panel is the noisy spatio-temporally averaged time-resolved contrast concentration (This is the input to the algorithm). The middle panel is ground truth. The right most panel is the predicted concentration. it can be shown from the figure that the method (physics informed neural network have great potential, as the accuracy is apparent from the figure shown above.

Recovery of the underlying hidden velocity

In the figure shown above represents the predicted underlying hidden velocity. the left panel shows the velocity magnitude of the ground truth. The right panel shows the predicted velocity magnitude. These results show and prove the potential and ability of this method to predict hemodynamics such as velocity and pressure.

 

 

Pages: 1 2 3 4