Research on Ultrasonic Image Segmentation of Thyroid Nodules Based on Improved U-net++

Abstract

The ultrasound image of thyroid nodules has low contrast and severe speckle noise, and the morphology of thyroid nodules in different patients is quite different, which makes it extremely difficult for doctors to accurately and quickly diagnose the nodules. In order to accurately segment the thyroid nodules from the ultrasound image, the paper improves the U-Net++ network. Based on the U-Net++ model, EfficientDet is used as the encoder, and the CSSE block is merged in the encoder and decoder to improve performance. In addition, the paper improves the network structure and reduces the number of model parameters. After testing 720 ultrasound images of thyroid nodules, the improved U-Net++ image segmentation has an average Dice coefficient of 0.9213, an average accuracy of 0.9262, an average recall rate of 0.9011, and an average F1 score of 0.9202. The Dice coefficient of the improved algorithm segmentation is 9.01% higher than that of U-Net++. The improved algorithm is of great significance for the application of automatic segmentation of ultrasound images of thyroid nodules in actual clinical medicine.