FCG-Net: An innovative full-scale connected network for thyroid nodule segmentation in ultrasound images

Abstract

Ultrasonic imaging is more susceptible to poor contrast and speckle noise than other imaging modalities, which poses certain challenges to computer-aided automatic segmentation of thyroid nodules. Nine popular segmentation networks for thyroid nodule in ultrasound images were well trained with the same training strategy and the performances on the public dataset and the clinical dataset were well compared using different indicators. Among them, UNet3+ achieved the best performance. However, it brings more parameters compared to other lightweight networks, which limits its application in portable diagnosis. This paper proposed an innovative thyroid nodule segmentation network FCG-Net that using full-scale skip connection to extracted multi resolution features, and Ghost bottleneck was introduced in the encoder and decoder process to reduce computational complexity of feature maps. Compared to UNet3+, the parameters of the model were reduced from 27 million to 11 million, and the inference speed was improved by 1.07 FPS. Meanwhile, the segmentation performance of FCG-Net exceeds that of UNet3+. 729 ultrasound images from the public dataset and clinical ultrasound images were used to validate the performance of the network by 5-fold cross-validation. The accuracy, precision, sensitivity, specificity and Dice of the FCG-Net were 0.9488, 0.8579, 0.8984, 0.9627 and 0.8670 respectively. The network not only improves the accuracy of ultrasonic diagnosis but also reduces the requirements of hardware and calculation, which is significant to promote the development of portable ultrasonic diagnosis of thyroid nodule.