Patient-Paired Anatomical Consistency Learning for Cross-View Echocardiographic Segmentation

Echocardiographic segmentation remains difficult in low-resource clinical settings because pixel-level annotations are expensive and imaging views vary substantially. Standard cardiac exams routinely acquire multiple complementary views of the same patient - most commonly apical two-chamber (2CH) and four-chamber (4CH) - yet most segmentation methods process each view independently and discard this naturally occurring patient-level correspondence. We introduce Patient-Paired Anatomical Consistency Learning, a framework that treats paired multi-view acquisitions as a free source of supervision for cross-view segmentation, rather than relying on generic domain-adaptation assumptions. Given a labeled source view (4CH) and an unlabeled, patient-matched target view (2CH), a shared U-Net encoder-decoder produces a segmentation mask and a latent anatomical representation for each image. We add two consistency objectives on top of standard Dice/cross-entropy supervision: an anatomical representation consistency loss, a patient-level contrastive objective that pulls representations of paired images from the same patient together while pushing apart representations from other patients; and a functional consistency loss that penalizes disagreement between left-ventricular ejection fraction estimated from the two views' predicted masks. Both consistency terms require no target-view annotation. On the CAMUS dataset, under a 4CH→2CH cross-view generalization setting where target-view labels are withheld entirely during training, our method improves LV, MYO, and LA Dice by 0.9, 1.9, and 1.9 points respectively over adversarial feature alignment, and reduces ejection-fraction mean absolute error by 1.1 points, closing most of the gap to a fully supervised pooled-view upper bound. An ablation replacing true patient pairing with random cross-patient pairing recovers only part of the gain, indicating that genuine patient-level anatomical correspondence, not consistency regularization alone,drives the improvement. This work points to an annotation-efficient direction for medical image analysis: many clinical workflows already produce multiple correlated observations of the same patient, and this structure can be directly exploited as supervision rather than treated as incidental.