Measurement of Echocardiographic Parameters in Various Cardiac Phases: A Comparative Study of Deep Learning Models and Keypoint Representations

The assessment of the parasternal long axis (PLAX) view is a critical task during echocardiographic examination. Accurate measurement and quantification of anatomical structures in PLAX view is essential for the diagnosis, treatment, and prognosis of cardiac diseases. Yet, manual measurement is limited by its time-consuming nature and susceptibility to inter-individual variability and image quality fluctuations. Automated keypoint detection of anatomical structures using deep learning offers a robust solution, potentially assisting physicians in the precise measurement of parameters within specified views. A systematic comparison of various deep learning models and keypoint representations is conducted in this paper for keypoint detection within the PLAX view, with the aim of exploring the best models and representations. Three datasets, comprising 442 patients and including data from end-diastole, mid-systole, and end-systole, are constructed to fully evaluate the accuracy and generalizability of different deep learning methods. In the 12-keypoint detection task conducted on the end-diastole dataset, the HRNet-Heatmap method achieved the highest success detection rate (SDR) of 68.77% and the lowest mean absolute error (MAE) of 3.03 mm among the evaluated approaches. For the 2-keypoint detection task on the mid-systole dataset, the HRFormer-Heatmap method performed the best, with an SDR of 89.09% and an MAE of 1.90 mm. In the 2-keypoint detection task on the end-systole dataset, the CSPNeXt-SimCC method demonstrated the best performance, achieving an SDR of 94.46% and an MAE of 2.95 mm.