Thyroid Nodules Risk Stratification Through Deep Learning Based on Ultrasound Images
Ziyu Bai,Luchen Chang,Ruiguo Yu,Xuewei Li,Xi Wei,Mei Yu,Zhiqiang Liu,Jie Gao,Jialin Zhu,Yulin Zhang,Shuaijie Wang,Zhuo Zhang
DOI: https://doi.org/10.1002/mp.14543
2020-01-01
Abstract:PURPOSE Clinically, the risk stratification of thyroid nodules is usually used to formulate the next treatment plan. The American College of Radiology (ACR) Thyroid Imaging Reporting and Data System (TI-RADS) is a kind of medical standard widely used in classification diagnosis. It divides the nodule's ACR TI-RADS level into five levels by quantitative scoring, from benign to high suspicion of malignancy. However, such assessment often relies on the radiologists' experience and is time consuming. So Computer Aided Diagnosis is necessary. But many Deep Learning (DL) models are difficult for doctors to understand, limiting their applicability in clinical practice. In this work, we mainly focus on how to achieve automatic thyroid nodules risk stratification based on deep integration of deep learning and clinical experience. METHODS An automatic hierarchical method of thyroid nodules risk based on deep learning is proposed, called risk stratification network (RS-Net). It incorporates medical experience based on ACR TI-RADS. The convolutional neural network (CNN) is used to classify the five categories in ACR TI-RADS and assign their points respectively. According to the point totals, the level of risk can be obtained. In addition, a dataset involving 13984 thyroid ultrasound images is established to develop and evaluate the proposed method. RESULTS We have extensively compared the results of this paper with the evaluation results of sonographers. The accuracy of the risk stratification (TR1 to TR5) of the proposed method is 65%, and the mean absolute error (MAE) is 0.54. The MAE of the point totals (0 to 13 points) is 1.67. The Pearson's correlation between our method evaluation and doctor evaluation reached 0.84. For the benign and malignant classification, the performance indices accuracy, sensitivity, specificity, PPV, and NPV were 88.0%, 98.1%, 79.1%, 80.5%, and 97.9%, respectively. Our method's level of thyroid nodules risk stratification is comparable to that of a senior doctor. CONCLUSIONS This work provides a way to automate the risk stratification of thyroid nodules. Our method can effectively avoid missed diagnosis and misdiagnosis caused by the difference of observers so as to assist doctors to improve efficiency and diagnosis rate. And compared with the previous benign and malignant classification, the proposed method incorporates clinical experience. So it can greatly increase the clinicians' trust in the DL model, thereby improving the applicability of the model in clinical practice.