MoST-Net: Motion Memory Spatiotemporal Transformer Network for Myocardial Echocardiography Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Myocardial echocardiography segmentation is crucial for cardiac function assessment but is challenging due to motion, deformation, and speckle noise.
• Existing memory-based methods struggle to capture motion dynamics, while spatiotemporal Transformers often require full sequences, limiting online applications.

Purpose of the Study:

• To propose a novel Motion memory Spatiotemporal Transformer Network (MoST-Net) for accurate myocardial echocardiography segmentation.
• To address limitations of existing methods by integrating motion modulation and enabling online processing.

Main Methods:

• MoST-Net utilizes motion memory learning to modulate image features and capture short-term motion, extended via long-term memory modeling.
• A memory prompt encoder and spatiotemporal decoder with a boundary uncertainty enhancement module generate segmentations.
• An adaptive ranking strategy optimizes memory length based on temporal consistency and segmentation quality.

Main Results:

• MoST-Net achieved superior performance compared to state-of-the-art methods on multiple datasets, even with significant noise.
• The method demonstrated effectiveness in capturing motion dynamics and handling ambiguous boundaries.
• Analysis indicated that motion-aware memory learning reduces the need for excessively long memory banks.

Conclusions:

• MoST-Net offers a significant advancement in myocardial echocardiography segmentation, particularly for online clinical scenarios.
• The proposed motion memory approach enhances temporal dependency modeling and segmentation accuracy.
• MoST-Net shows potential for real-time cardiac function assessment in clinical practice.