Multi-model Comparison of Cardiac Segmentation Model for Angiocardiography by Deep Learning

Angiocardiography is a key to the evaluation of cardiac function. The index of left ventricular ejection fraction (LVEF) also helps interpret cardiac hypertrophy, valve regurgitation, and regional myocardial systolic dysfunction for medical decision-making. However, it is difficult and time-consuming to calculate the LVEF. Therefore imaging techniques become much more important for assistance. In this research, three neural networks including DenseNet, EfficientNet, and ResNet are introduced for cardiac area calculation. EfficientNet turns out to be the best model with a mean dice accuracy of 0.91. The other two models also show significant performance with 0.958 for DenseNet and 0.955 for ResNet. With these network models, physicians can select and calculate the area of the heart. With the time-series images of cardiac catheterization, the size of the entire cardiac systole can be determined automatically. Therefore, it greatly helps medical professionals diagnose abnormal cardiac systole problems immediately.