ECHO-FUSE: a hybrid dual-encoder attention network for patient-level validated left-heart echocardiography segmentation

Manual tracing of left-heart borders in echocardiography is still affected by reader variation, and the images themselves often contain speckle, acoustic dropout, shadowing, and uneven contrast. In this work, we evaluate ECHO-FUSE, a dual-encoder attention U-Net for four-class segmentation of apical two-chamber end-diastolic frames from CAMUS. Before inference, each frame is resized in a deterministic way, filtered with an edge-preserving bilateral operation, and contrast-enhanced using CLAHE. The grayscale input is copied into three channels so that ResNet50 and Xception encoders can process the same echocardiographic frame in parallel. Their multi-scale skip features are aligned to matching spatial sizes, fused, and passed through attention-guided decoder blocks to predict background, LV endocardium, LV myocardium, and left atrium. The experimental protocol uses a locked patient-level split of 400 training, 50 validation, and 50 test subjects, with no patient overlap across subsets. The proposed model and four baseline architectures were trained under the same preprocessing, augmentation, checkpointing, and evaluation protocol. On the locked test set, the proposed dual-encoder attention U-Net achieved a patient-level foreground macro-Dice of 0.8607 [95% CI: 0.8355–0.8819], foreground macro IoU of 0.7700 [95% CI: 0.7407–0.7961], and foreground macro HD95 of 10.47 mm [95% CI: 8.02–13.79]. ECHO-FUSE’s performance was statistically comparable to the strongest Standard U-Net baseline for Dice (paired Wilcoxon p = 0.8109) and HD95 (p = 0.2571), while significantly outperforming DeepLabV3+ and EfficientNetB0 U-Net. Representative locked-test cases showed accurate ventricular delineation in good and average examples, while the difficult case exposed left atrial segmentation as the main failure mode. These findings support ECHO-FUSE as a reproducible segmentation framework and define clear limits for future multi-view, temporal, and externally validated echocardiographic analysis. Not applicable.