MTN-Net: A Multi-Task Network for Detection and Segmentation of Thyroid Nodules in Ultrasound Images

Abstract

AbstractAutomatic detection and segmentation of thyroid nodules is crucial for the identification of benign and malignant nodules in computer-aided diagnosis (CAD) systems. However, the diverse sizes of thyroid nodules in ultrasound images, nodules with complex internal textures, and multiple nodules pose many challenges for automatic detection and segmentation of thyroid nodules in ultrasound images. In this paper we propose a multi-task network based on Trident network, called MTN-Net, to accurately detect and segment the thyroid nodules in ultrasound images. The backbone of MTN-Net can generate scale-specific feature maps through trident blocks with different receptive fields to detect thyroid nodules with different sizes. In addition, a novel semantic segmentation branch is embedded into the detection network for the task of segmenting thyroid nodules, which is also valid for the complete segmentation of nodules with complex textures. Furthermore, we propose an improved NMS method, named TN-NMS, for combining thyroid detection results from multiple branches, which can effectively suppress falsely detected internal nodules. The experimental results show that MTN-Net outperforms the State-of-the-Arts methods in terms of detection and segmentation accuracy on both the public TN3K dataset and the public DDTI dataset, which indicates that our method can be applied to CAD systems with practical clinical significance.KeywordsUltrasound imageThyroid noduleDetectionSegmentationMulti-task network