Title: YOLO-TN: ultrasound diagnosis of thyroid nodules based on transfer learning and convolutional neural networks
Authors: Xun Wang; Ning Zhang; Mao Ding; Nuo Xu
Addresses: College of Computer Science and Technology, China University of Petroleum (East China), Qingdao, 266580, Shandong, China ' College of Computer Science and Technology, China University of Petroleum (East China), Qingdao, 266580, Shandong, China ' Department of Neurology Medicine, Cheeloo College of Medicine, Shandong University, Jinan, 250033, Shandong, China ' Department of Ultrasound, First Medical University Jinan, 250033, Shandong, China
Abstract: Thyroid nodules are a common clinical condition. Currently, radiologists commonly use ultrasound as a method of diagnosing them. However, the varying experience of radiologists leads them to describe the features of nodules in the same ultrasound image differently. Incorrect descriptions will lead to misdiagnosis for the patient. To address the problem of difficult diagnosis of benign and malignant thyroid nodules in ultrasound images, we proposed a model for the diagnosis of thyroid nodules, which is named YOLO-TN (thyroid nodules). In this work, we used transfer learning and data augmentation to alleviate the data dependency of the model and designed CSPSE-DarkNet to simulate clinical diagnosis. In addition, we enhanced the detection of small objects and proposed soft ROI selection to enrich the contextual information of thyroid nodules. The model achieved a mAP value of 91.3% on our dataset, providing better performance than some of the popular networks currently available.
Keywords: ultrasound images; thyroid nodules; benign and malignant diagnosis; transfer learning; convolutional neural network.
International Journal of Adaptive and Innovative Systems, 2022 Vol.3 No.2, pp.87 - 99
Received: 12 Apr 2021
Accepted: 26 Apr 2021
Published online: 25 Jul 2022 *