Semi-supervised learning with double head approach for carotid artery detection

Abstract

The detection of carotid artery in ultrasound medical images helps locating other organs to build an efficient computer-aided diagnosis system. Due to the differences in the speed, direction and angle of continuous scanning of the carotid artery, its imaging shape in the ultrasound image is complex and easy to be distorted. Meanwhile, the labeled medical image datasets are limited. So it's hard to make carotid artery detection accurately. Although existing methods such as Mask R-CNN attempt to achieve carotid artery segmentation by ultrasound data, the results are not ideal. We propose a novel architecture called SSL-DH-Faster RCNN, which is based on a semi-supervised learning approach using unlabeled medical images to improve our model performance. In our framework, we adopt double head detection architecture to solve the problem that single head structure performs poorly on handling both classification and localization task at the same time. Concretely, a fully connected head(fc-head) for classification task and a convolution head(conv-head) for regression is adopted based on the reason that fc-head got better performance on classification task and conv-head is more suitable for localization task. Simultaneously, we combine PAFPN module into our framework to make low-layer information easier to propagate with above two methods and improve model performance further. Experiments show that SSL-DH-Faster RCNN method proposed in this paper achieves superior performance, and outperforms several popular methods. Experiments show that compared with existing popular methods, our method achieves the best performance on AP50, AP75 and AP@[0.50:0.95] metrics.