Abstract: The ultrasound image segmentation of thyroid is very important in clinical ultrasonography. In view of the problems of the low signal-to-noise ratio, high speckle noise and uncertain thyroid morphology in thyroid ultrasound images, an improved MultiResUNet segmentation network, named as Oct-MRU-Net, is proposed by combining the basic structure of MultiResUNet network and Octave convolution. And, the improved Inception module is used to learn the features of different spatial scales. The feature map in the training process is divided into high and low frequency features according to the channel direction, in which the high frequency features describe the image details and edge information and the low frequency features describe the overall image contour information. Then, the method of ultrasound image segmentation can focus on high-frequency information to reduce spatial redundancy, so more precise edge segmentation can be achieved. The experimental results show that the performance of Oct-MRU-Net network is better than that of U-Net network and MultiResUNet network, and so the Oct-MRU-Net network has better segmentation effect on thyroid ultrasound images.