An Efficient Tracker for Thyroid Nodule Detection and Tracking During Ultrasound Screening.

Thyroid tumor is a common disease in clinic. Junior doctors could easily miss or get false detection due to the unclear boundary and similarity between nodules and tissues during thyroid screening. In this paper, we propose an efficient tracker for simultaneously detecting and tracking nodules to assist doctors in examination and improve their work efficiency. An attention based fusion block which adaptively combines the features of previous and current frames is introduced to acquire better detection and tracking result. To increase the detection accuracy, we propose an advanced post-processing strategy instead of using general post-processing methods to train the network to obtain the best prediction. Moreover, a minibatch self-supervised learning module is embedded to reduce the false positive rate (FPR) by strengthening the ability of distinguishing nodules from similar tissues. The proposed framework is validated on a dataset of 1555 thyroid ultrasound movies with 13314 frames. The result of 91% recall with 3.8% FPR running at 30 fps demonstrates the effectiveness of our method.