Vendor-Agnostic 3D Mitral Valve Segmentation Using Semi-Supervised Learning

Mitral valve assessment for disease diagnosis and treatment is commonly guided by ultrasound imaging, with 3D transesophageal echocardiography being the de facto standard modality. The complex 3D structure of the mitral valve poses a challenge for its time-efficient and accurate quantification. Deep neural networks have been proposed for automatic mitral valve segmentation, but variations in imaging devices and protocols across ultrasound vendors cause significant domain differences in echocardiography datasets, leading to suboptimal dataset-dependent segmentation techniques. In this work, we apply a semi-supervised learning method to develop a vendor-agnostic, automatic segmentation tool for 3D transesophageal echocardiography. We learn from one manually annotated ultrasound domain, to generate reliable pseudo-labels using an ensemble of three models, for another unseen ultrasound domain. Our teacher-student framework is validated on images from two of the biggest cardiac ultrasound manufacturers. In our experiments, the student model outperforms each teacher, with a Dice score of 82.20 ± 4.62% on known data and 71.32 ± 6.95% on unseen data, and an Average Surface Distance of 0.37 ± 0.07 mm and 0.82 ± 0.15 mm for the mitral valve on known and unknown domain respectively. These results enable efficient cross-domain analysis by reducing the need for manual annotation and ensuring consistent mitral valve analysis across different vendors.