A segment anything model for transesophageal echocardiography based on bidirectional spatiotemporal context fusion

Abstract

Accurate segmentation of the left atrial appendage (LAA) in transesophageal echocardiography is the foundation for clinical evaluation. However, the ambiguous boundaries of the LAA, together with ultrasound noise and complex cardiac motion, make it challenging to obtain temporally consistent and spatially reliable segmentation results. Furthermore, existing works often process spatial and temporal features in isolation, without effectively leveraging spatiotemporal context fusion to enhance segmentation performance. To address these challenges, we propose a Segment Anything Model Based on Bidirectional Spatiotemporal Context Fusion (BiSTC-SAM). First, we design a spatiotemporal context network that encodes effective pixels associated with target changes, thereby mining temporal cues from spatial features. Building on this, we further develop a multi-scale context memory network that performs dynamic feature alignment, thereby integrating temporal representations to refine spatial features. We evaluate the segmentation and generalization performance of our method on a self-constructed transesophageal echocardiography dataset, and further assess its adaptability to different modalities on two publicly available transthoracic echocardiography datasets. Experimental results demonstrate that our method outperforms competing methods in terms of boundary segmentation accuracy and temporal consistency.