Work

B-Mode Echocardiogram Segmentation Using Fusion Framework of Artificial Intelligence and Machine-Vision (FAIM-Echo)

Year: 2026

Type: article

Abstract: FAIM-Echo is a fully automated hybrid framework that fuses a task-optimized UNet (EchoUNet) with a physiologic, peak-prominence-weighted graph search (ProID) to delineate the left-ventricular (LV) endocardium in B-mode echocardiography. EchoUNet is first trained on the public CAMUS dataset of adult subjects to produce probability maps for the LV and left atrium; a feature-detection block then infers the apex and mitral-annulus landmarks, which initialize a per-frame Dijkstra search that refines the contour along radial spokes without user clicks or temporal carry-over. This modular design decouples appearance learning from physiologic path enforcement, improving interpretability and reducing dependence on training-set idiosyncrasies. On in-distribution CAMUS test data, EchoUNet achieves the highest Dice similarity scores and lowest Hausdorff and mean absolute distances across image-quality strata, while FAIM-Echo remains clinically competitive with > 90% median Dice in good and fair quality subsets. On an unseen clinical cohort from the University of Nebraska of pediatric subjects (total 66, 373 cycles spanning dilated and hypertrophic cardiomyopathy and age-matched controls), FAIM-Echo consistently outperforms EchoUNet and UNet-2 across Dice (4-7% points higher in diseased ESV), Hausdorff distance (> 2-3 mm lower in diseased ESV), and mean absolute distance (~0.4-0.7 mm lower in diseased ESV) for both end-diastolic and end-systolic phases and across healthy and diseased subgroups, with the largest gains in diseased systolic cases. These results demonstrate that coupling learned appearance priors with an explicit physiologic graph search improves cross-domain generalization under the evaluated domain shift: adult benchmark to pediatric clinical cohort, GE to Philips, multi-view to A4C-only, and mitigates cross-domain degradation, supporting real-world deployment of automated echocardiographic quantification.

Source: IEEE Access

DOI: 10.1109/access.2026.3689751

Authors: Ruhi Sharmin, Brett A. Meyers, Ioannis Dimitroglou, Pavlos P. Vlachos

Institution: Purdue University West Lafayette

Language: English

Topic: Medical Image Segmentation Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: gold

Estimated APC paid: $2,075