What is the clinical evidence from this study?

Study design: Cohort. Population: Left ventricular segmentation on cardiac ultrasound (n=10629). Intervention: Fine-tuned Segment Anything Model (SAM) vs. EchoNet model (domain-specific model). Primary outcome: Dice similarity coefficient (DSC) on EchoNet test set (p=<0.0001).

September 19, 2023Open Access

Foundation versus Domain-Specific Models for Left Ventricular Segmentation on Cardiac Ultrasound

Q: What are the key findings of this study?

Fine-tuned SAM achieved excellent left ventricular segmentation performance on the EchoNet dataset (DSC 0.911) compared to the domain-specific EchoNet model (DSC 0.915, p<0.0001).

Key Result

Fine-tuned SAM achieved excellent left ventricular segmentation performance on the EchoNet dataset (DSC 0.911) compared to the domain-specific EchoNet model (DSC 0.915, p<0.0001).

Study Design

Type

Cohort (n=10,629)

Multicenter

Yes

Structured PICO

Does a fine-tuned Segment Anything Model (SAM) improve left ventricular segmentation performance compared to domain-specific models in transthoracic echocardiography and point-of-care ultrasound images?

Population

Echocardiography datasets including EchoNet-Dynamic (10,030 videos from Stanford), Mayo Clinic dataset (n=99 patients: 58 TTE, 41 POCUS), and CAMUS dataset (500 cases from University Hospital of St Etienne).

Intervention

Fine-tuned Segment Anything Model (SAM) for left ventricular segmentation

Comparator

Zero-shot SAM and domain-specific EchoNet model

Outcome

Model segmentation performance measured by Dice similarity coefficient (DSC)surrogate

Fine-tuning the Segment Anything Model (SAM) on echocardiography data yields strong generalization for left ventricular segmentation across different institutions and ultrasound modalities, including POCUS, reducing the need for extensive manual data curation.

Main Result

Absolute Event Rate: 0.911% vs 0.915%

p-value: p=<0.0001

Limitations

Retrospective nature subject to selection bias
Performance evaluated on pre-selected end-diastolic and end-systolic frames rather than beat-to-beat assessment
Did not specifically explore adapters for SAM
Real-world effects and clinical workflow integration need to be validated in a prospective setting

Abstract

Abstract The Segment Anything Model (SAM) was fine-tuned on the EchoNet-Dynamic dataset and evaluated on external transthoracic echocardiography (TTE) and Point-of-Care Ultrasound (POCUS) datasets from CAMUS (University Hospital of St Etienne) and Mayo Clinic (99 patients: 58 TTE, 41 POCUS). Fine-tuned SAM was superior or comparable to MedSAM. The fine-tuned SAM also outperformed EchoNet and U-Net models, demonstrating strong generalization, especially on apical 2-chamber (A2C) images (fine-tuned SAM vs. EchoNet: CAMUS-A2C: DSC 0.891 ± 0.040 vs. 0.752 ± 0.196, p<0.0001) and POCUS (DSC 0.857 ± 0.047 vs. 0.667 ± 0.279, p<0.0001). Additionally, SAM-enhanced workflow reduced annotation time by 50% (11.6 ± 4.5 sec vs. 5.7 ± 1.7 sec, p<0.0001) while maintaining segmentation quality. We demonstrated an effective strategy for fine-tuning a vision foundation model for enhancing clinical workflow efficiency and supporting human-AI collaboration.

Mark Helpful

Bookmark

Relay

View Full Paper