A new network for carotid artery plaque segmentation in ultrasound images

The plaque segmentation of ultrasound images is very important in the disease diagnosis of carotid artery, and it is hard to be realized due to the image noise, low contrast, and plaque shapes, etc.. This paper proposed an improved network model called RCS_UNet based on the classical UNet architecture. The residual convolution, CBAM and the multi-scale supervision modules improve the gradient vanishing, the model attention on key features and the performance of decoder part, respectively. RCS_UNet was trained and tested using a dataset containing 443 longitudinal carotid artery ultrasound images. RCS_UNet can accurately segment plaques in carotid artery ultrasound images, with DSC of 0.819 +/- 0.091, IoU of 0.703 +/- 0.123, recall of 0.845 +/- 0.100, and precision of 0.822 +/- 0.150, which was better than that of other advanced models. The trained model was used to segment plaques in carotid artery ultrasound images obtained from another ultrasound device, DSC was 0.832 +/- 0.075, IoU was 0.719 +/- 0.105, recall was 0.792 +/- 0.136, precision was 0.900 +/- 0.065, which proved that the model had a certain degree of generalization.