Comparative Eminence: Foundation versus Domain-Specific Model for Cardiac Ultrasound Segmentation

Abstract

Importance A recently developed vision foundation model, "Segment Anything (SAM)," promises to segment any objects in images. However, the performance of SAM on clinical echocardiography images is yet to be investigated and compared against the domain-specific models. Objective To evaluate the performance of SAM on transthoracic echocardiography (TTE) and point-of-care ultrasound (POCUS) images. Design SAM was fine-tuned on the training set of EchoNet-Dynamic (TTE) and then evaluated on datasets containing TTE and POCUS images. Setting Multi-center, retrospective cohort study. Participants This study used two publicly available datasets (EchoNet-dynamic, Stanford University and CAMUS, University Hospital of St Etienne). The Mayo Clinic dataset contains a sample of 99 non-duplicated patients (58 TTE and 41 POCUS). Intervention/Exposure: not applicable. Main Outcomes and Measures Model segmentation performance: Dice similarity coefficient (DSC). Results Fine-tuned SAM had promising frame-level performance (SAM vs. EchoNet: DSC 0.911 +/- 0.045 vs. 0.915 +/- 0.047, p<0.0001), and consistent performance on the external datasets including TTE (Mayo Clinic: DSC 0.902 +/- 0.032 vs. 0.893 +/- 0.090, p<0.0001, CAMUS-A4C: DSC 0.897 +/- 0.036 vs. 0.850 +/- 0.097, p<0.0001, 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). Conclusions and Relevance Promising segmentation performance was observed after fine-tuning the SAM model on TTE. The strong generalization capability of SAM can facilitate the development of AI applications in cardiac ultrasound with less manual data curation.