Key result
Fine-tuning EchoNet-Dynamic cuts LVEF mean absolute error by ~2.6 percentage points in dilated cardiomyopathy.
Why the study?
Deep learning models for echocardiography may lose performance when acquisition characteristics differ between training and deployment settings, posing a challenge in low- and middle-income countries. The study evaluated how models developed on established datasets generalise to a local clinical dataset in Tanzania and whether transfer learning could mitigate domain shift.
Does transfer learning with a U-Net deep learning framework improve automated left ventricular segmentation and ejection fraction estimation in patients with dilated cardiomyopathy in a resource-limited setting?
Observational (n=186)
No
Does transfer learning with a U-Net deep learning framework improve automated left ventricular segmentation and ejection fraction estimation in patients with dilated cardiomyopathy in a resource-limited setting?
Fine-tuning an established video-based deep learning model using transfer learning significantly improves the accuracy of automated left ventricular ejection fraction estimation and segmentation in patients with dilated cardiomyopathy in a resource-limited setting.
Deep learning models may lose performance when echocardiographic acquisition characteristics differ between training and deployment settings, a challenge for AI echocardiography in low- and middle-income countries (LMICs). We evaluated a U-Net deep learning framework for automated left ventricular (LV) segmentation and ejection fraction (EF) estimation using real-world echocardiographic data from 186 patients with dilated cardiomyopathy (DCM) at the Jakaya Kikwete Cardiac Institute (JKCI), Tanzania. We assessed how models developed on established echocardiographic datasets, the CAMUS still-image dataset and the EchoNet-Dynamic video dataset, would generalise to this local clinical dataset, and to what extent transfer learning could mitigate domain shift. The CAMUS-trained still-image model produced fragmented predictions on JKCI data that were unsuitable for LV volume estimation. Fine-tuning the EchoNet-Dynamic model reduced EF mean absolute error (MAE) from 7.21% to 4.63% and root mean square error (RMSE) from 9.84% to 6.66%. LV segmentation achieved Dice similarity coefficients of 0.90 at end-systole and 0.84 at end-diastole. Because the models differed in architecture, input modality, and adaptation, these results do not isolate the independent effect of video input or temporal encoding. The findings support further evaluation of locally adapted echocardiography models in resource-limited settings.
No takes yet. Share an insight, caveat, or question.
Mzurikwao et al. (2026) conducted an observational in Dilated cardiomyopathy (n=186). Fine-tuned EchoNet-Dynamic U-Net deep learning model vs. Unadapted EchoNet-Dynamic video model and CAMUS-trained still-image model was evaluated on Left ventricular ejection fraction (EF) mean absolute error (MAE) and segmentation Dice similarity coefficient. Fine-tuning the EchoNet-Dynamic deep learning model reduced ejection fraction mean absolute error from 7.21% to 4.63% and root mean square error from 9.84% to 6.66% in dilated cardiomyopathy.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: