Semi-supervised deep neural networks for left atrial segmentation from cardiac MRI

Abstract Funding Acknowledgements Type of funding sources: None. Introduction Left atrium size and function is an important but less studied prognostic marker for many cardiovascular diseases [1]. Cardiac Magnetic Resonance Imaging (MRI) can capture atrial structure with a high temporal spatial resolution. However, diagnoses accurate measurement volumes from MRI often rely on time-consuming manual processing, or needs correction software segmentation error by clinicians. Robust, automated methods are highly desirable to assist morphological analysis dynamic changes in the heart. Ultimately this facilitates wider application cardiac early identification prevention potential disease. Purpose To introduce semi-supervised teacher-student deep neural networks (requiring labelled data) left images, compare their performance conventional supervised learning. Methods The basic mean teacher network shown Fig.1. Network training starts manner. At each batch input, both student model evaluate images applied noise. updated gradient descent loss obtained comparing its prediction label. exponential moving average (EMA) weight then migrated update network. Consistency between student’s teacher’s (filtered uncertainty map Uncertainty-Aware Mean Teacher, UA-MT [2]) jointly student. Unlabeled continually improve without using loss. We compared three UA-MT, Hierarchical Regularized Teacher (HCR-MT) Semi-supervised Contrastive (SCC), TransUNnet [3], benchmark dataset MICCAI 2018 Atrial Segmentation Challenge [4]. Out 20 scans (1760 images), 16 were used training, 4 testing. For learning, 75% set was as data 25% assumed be unlabeled. In all labels enabled. Results Example results models Fig 2. comparison, fully produced accuracy 76%. Benefiting effective achieved superior 89%, while also requiring data. Conclusion proposed outperforms TransUNet segmenting MRI. Our approach uses data, reducing tedious labelling tasks, ultimately unlocking ability effectively utilise full power learning medical applications.