Multi-scale Supervised U-Net Ultrasound Image Segmentation of Thyroid Nodule

Abstract

To tackle the misdiagnosis and missed diagnosis problem in the process of manual diagnosis and screening due to the variable size of thyroid nodule and fuzzy edge of nodule in ultrasonic image, in this paper, a mutil-scale attention UNet (MSA-UNet) method for ultrasonic segmentation of thyroid nodule was proposed. The algorithm first extracts the characteristic information of thyroid lesions by using cavity convolution with different void rates, and then fuses the characteristic information of different scales to solve the influence of different thyroid nodule sizes on ultrasonic image segmentation. Considering the influence of location relation information learning and deep semantic feature screening on the segmentation model, channel attention mechanism is used to make the network model more focused on more useful feature information, to improve the segmentation accuracy of thyroid nodes and achieve fine segmentation of lesion region. Experimental results show that the recall rate in thyroid ultrasound image data set is 87%, the segmentation accuracy is 86.1%, and the Dice value is 84.6%, higher than those of existing deep learning method. This work can provide a new research idea for the detection and diagnosis of thyroid nodules.

Keywords

• image segmentation
• deep learning
• attention mechanism
• neural network
• u network