Cost-efficient method to predict 3D hemodynamics

A review on Prediction of 3D Cardiovascular hemodynamics before and after coronary artery bypass surgery via deep learning


Photo by Francesco Ungaro from Pexels


One of the biggest challenges in the medical industry is to provide medical assistance at a low cost. Some of the techniques tend to be very expensive and cannot be afforded by a lot of patients. For instance, in the case of coronary heart disease, most the death are caused by coronary stenosis—the narrowing or blockage of the coronary arteries. In order to perform a proper treatment, a proper form of diagnosis is required. Currently, pressure fluid-based fractional flow reserve (FFR) holds the high standards for diagnosing patients with myocardial ischemia — damage or disease in the heart's major blood vessels—caused by coronary stenosis, and the treatment followed by it is efficient and safe.


Image of describing the narrowing of the coronary veins

Source

However, patients with severe myocardial ischemia have to undergo revascularization. It is a process of restoring the blood flow in an operated organ or body part. There two main procedures for revascularization: coronary artery bypass grafting (CABG) and angioplasty. CABG is the most widely used revascularization procedure for myocardial ischemia. Usually, a large vein is taken from any other body part and then grafted to the concerned coronary vein in the heart. After these veins are grafted in the coronary veins, the point of measurement for the stability of CABG postoperation remains the blood flow at the distill end of the stenotic coronary artery and the blood flow in the grafted vein. This usually comes under hemodynamics the study of blood flow.


Image of describing the bypass grating

Source

As it turns out, the application rate of hemodynamic parameters in clinical practice is low, mainly due to its high measurement cost and potential risks during catheter insertion. But the need of accessing the stability of CABG is still important.


Image of describing the catheter insertion

Source

A large number of studies have been conducted that use computational fluid dynamics (CFD) to improve cardiovascular hemodynamics by reducing the risks as well as the cost. Medical imaging (MRI and CT scans) can provide cardiovascular geometry to obtain solutions for the velocity and pressure of the blood flow at relatively low costs but again the cost of modeling cardiovascular hemodynamics is very high. Therefore, it has become necessary to develop a method to calculate cardiovascular hemodynamic that can reduce calculation costs while ensuring model accuracy and patient stability after postoperation.


In the era of artificial intelligence, and deep learning in specific, amazing breakthroughs are being made especially in the medical and biological sciences. A lot of problems can be solved using the deep learning algorithm. Deep learning can extract any information as complex as it can be and then using that information we can develop systems or hypothetical models to address an issue or problem. Deep learning uses mathematical linear equations followed by a non-linear activation function to extract patterns that can be used for classification, regression, learning distributions of the given, or use extracted information to generate new data with new conditions or distributions. Thus, deep learning can solve any problem if given enough data.


The data itself should contain high spatial resolution so that complex features of both velocity and pressure could be extracted. This data can be collected through MRI or CT scan which provides a 3D perception of cardiovascular geometry of patients with coronary stenosis (pre and postoperation). Once enough data is collected deep learning model could consume it to predict pressure and velocity based upon the geometry.

Following the previous step, to make them clear and accurate segmentation of different parts of the heart, the output results could be passed through another network called the PointNet that could be used to introduce spatial relationships. By extracting and integrating global and local features of the point cloud, the network could analyze and reproduce the relationship between vessel geometry in the point cloud datasets and the corresponding hemodynamics.


Results (as per the original paper)


Prediction results of the velocity field


Two images of preoperation and postoperation are shown while the former is CFD generated and later is DL predicted

Source

The 3d model obtained from the deep learning algorithm confirms that the image is clearer and almost identical to the image obtained from computational fluid dynamics. Hence, it is proved that the velocity detected by CFD is similar to the velocity predicted by the DL algorithm.


Prediction results of the pressure field


Two images of preoperation and postoperation are shown while the former is CFD generated and later is DL predicted

Source

The same goes while predicting pressure in cardiovascular data obtained from the CT scans. The 3d model obtained from the deep learning algorithm confirms that the image is clearer and almost identical to the image obtained from computational fluid dynamics.

The results show that the deep learning methods are accurate enough to predict the right information.


But what about the cost?

As it turns out the DL method is inexpensive compare to modeling a hemodynamic system. The energy consumption is also very less or a one-time requirement i.e. only during training and the same model can use for multiple patients, on the other hand, CFD needs to iterated for every new patient. This not only saves manpower but also the cost of operating such instruments.


If a sufficient amount of data is present then the DL model can perform tasks with higher accuracy.


Deep learning can break many obstacles in the medical field and provide faster ways to understand and treat patients which will be safer, faster, efficient, and low at cost.


Reference:

  1. Prediction of 3D Cardiovascular hemodynamics before and after coronary artery bypass surgery via deep learning

  2. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation