MediViSTA: Medical Video Segmentation Via Temporal Fusion SAM Adaptation for Echocardiography.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• The Segment Anything Model (SAM) excels in natural image segmentation but lacks precision in medical imaging.
• SAM's 2D architecture limits its application to volumetric or video medical data.
• Medical video segmentation, especially echocardiography, requires specialized adaptation for accuracy.

Purpose of the Study:

• To introduce MediViSTA, a parameter-efficient fine-tuning method for adapting vision foundation models to medical video segmentation.
• To specifically address the challenges of echocardiography segmentation using advanced adaptation techniques.
• To improve the precision and stability of automated segmentation in cardiac imaging.

Main Methods:

• MediViSTA employs parameter-efficient fine-tuning to adapt foundation models for medical video.
• Spatial adaptation is achieved through frequency feature fusion incorporating CNN-derived spatial frequency information.
• Temporal adaptation is integrated using temporal adapters within transformer blocks for video sequence processing.

Main Results:

• MediViSTA outperforms state-of-the-art methods in echocardiography segmentation without prompts.
• The model demonstrates strong generalization on unseen datasets, improving Dice scores by 2.15% and temporal consistency by 0.09.
• Evaluated on three diverse datasets, MediViSTA shows consistent superiority.

Conclusions:

• MediViSTA offers an effective and efficient approach for medical video segmentation, particularly for echocardiography.
• The method significantly advances the accuracy and robustness of cardiac assessment applications.
• Parameter-efficient fine-tuning combined with spatial and temporal adaptation is key to successful medical video segmentation.