Semi-supervised Anatomy Tracking with Contrastive Representation Learning in Ultrasound Sequences

Anatomy tracking plays an essential role in real-time US image interpretation. However, the low US image quality poses a challenge to this task. Intensive manual labeling is often required to obtain acceptable performance. In this work, we propose a Semi-Supervised US Anatomy Tracking method (SSAT-US) with minimum supervision for each US sequence. Self-supervised contrastive learning (SSCL) along with US-specific data augmentation is first applied to warm up a discriminative patch feature representation. To take full advantage of the unlabeled data, a confidence-aware pseudo-labeling module is presented to select reliable training pairs for semi-supervised training. Evaluations on clinical and public datasets of different anatomies were conducted. Results showed that SSAT-US could achieve accurate tracking performance with 0.832 (EAO), 0.000 (Rob), 0.812 (Acc), and generalize to the challenging unseen dataset with 0.472 (EAO), 0.329 (Rob), 0.789 (Acc), demonstrating that our method could perform accurate and robust anatomy-agnostic real-time tracking in US sequences.