Ejection fraction refers to the measurement, expressed in percentage, of the volume of blood being pumped out of the left ventricle when our heart muscles contract. A healthy, fully-functioning heart will eject more than half of the blood, making an ejection fraction to be over 50%. Typically, an ejection fraction is obtained by outlining the left ventricle on a digital image to estimate the blood volume at the start and end of a heartbeat.
The need for a more accurate cardiac pump function measurement
An experienced physician may make an intelligent guess of an ejection fraction by watching several loops of heartbeat ultrasound videos. However, he or she may find it difficult to do so if only a frame showing the start or end of an ejection is presented. Therefore, it’s often recommended for physicians to estimate an ejection fraction based on several heartbeats even though in an actual clinical setting, physicians tend to churn out the percentage based on just one heartbeat.
A group of researchers from Stanford University and Cedars-Sinai Medical Center is now relying on artificial intelligence (AI) to examine cardiac ultrasound videos and generate non-stopping, beat-by-beat measurement of our cardiac pump function. They believe if the accuracy of ejection fraction estimation can be improved, it may shed light on many intricacies of heart function and disease. For example, arrhythmia or the condition when one’s heart beats in irregular rhythm.
When arrhythmia happens, the volume of blood filling up and pumping out of the left ventricle will vary, leading to a fluctuation of ejection fraction. Atrial fibrillation, a form of arrhythmia, is projected to affect more than six million people in the US by 2050 and 17.9 million people in Europe by 2060. Since more than one-third of those who have atrial fibrillation will develop heart failure, their ejection fraction need to be closely monitored frequently.
3D Convolutional Neural Network
The research team fed the convolutional neural network (CNN) with more than 10,000 cardiac ultrasound videos and images outlining the inner border of the left ventricle at the start and end of a heartbeat. They then trained it to recognize the left ventricle in the video frame as well as to outline the border of the ventricle in the course of each heartbeat cycle. Strictly speaking, this is not something novel, but what makes the project special is the research team made use of three-dimensional CNN, which encompasses 2D video images that change over time to retrieve information that’s required to assess a heart’s function.
3D CNN has been used to analyze medical imaging and physical activities but was never used for cardiac ultrasounds. The research team found that the 3D CNN has a relatively lower reported error as compared to other CNN that were used to estimate ejection fraction in other studies. They also realized the method has an overall mean error of 4.1% and 6% respectively for two separate sets of data that were used for validation purpose.
Furthermore, when the research team tested 55 more patients that had been assessed independently by two ultrasound specialists, they found that the 3D CNN method generated a more consistent ejection fraction for images taken by different ultrasound machines and under different circumstances. With that, the research team is confident that the network will be of value to keep track of a patient’s ejection fraction when he or she undergoes complicated medical procedures.
Nevertheless, the research team did warn that its 3D CNN needs to be refined and trained substantially before it can be used clinically. This is especially so when in reality, there may be suboptimal images found within ultrasound videos. Besides, more studies are required to determine whether the 3D CNN is capable of tracking measures that are more sensitive than ejection fraction. Overall, the model showed potential in shouldering physicians’ workload while rendering high-quality care at the same time.