Machine Learning Project
Ejection Fraction Measurement from Heart Ultrasound Images
The result of left ventricle detection test during one cycle. (video)
This project aimed to detect the volume difference of the left ventricle during a cardiac cycle using YOLO-v7. Two datasets, HeartSeg and EchoNet-Dynamic, are transformed into the MS COCO dataset format, and the performance is tested. The results show that YOLO-v7 performs poorly on HeartSeg but well on EchoNet- Dynamic. However, detecting pixels of a target to compute the ejection fraction is found to be an unreliable method, as the standard deviation of the difference between the true value and prediction was significant, and the correlation value was minor. The reason for this is discussed, and a potential improvement method was provided. Additionally, an analysis of how YOLO-v7 works is also included.