Automatic Segmentation of Cardiac Structures from 2D Echocardiographic Images using Transformers

Segmentation of cardiac structures (CS) in echocardiographic images is an important step for coronary heart disease (CHD) diagnosis. Manual or semi-automatic delineation of CS is often time-consuming and prone to intra- and inter-observer variability. Thus, we propose to use a transformer based model for the automatic segmentation of CS in 2D echocardiographic frames of patients with a pathological risk of CHD. We analyse the performance of this model with different data augmentation settings, and suggest that there is an improved performance compared to the baseline UNet model. We conclude that the results could be accurate enough for bringing this automatic segmentation method of CS into clinical routine for CHD diagnosis.