Ultrasonography is the primary tool for diagnosis and initial risk stratification of thyroid nodules. Some open issues limit its usefulness: 1. limited diagnostic accuracy of single sonographic features; 2. inter-observer variability; 3. inconsistent definition of critical features. Furthermore, when cytology is performed, in some cases, it renders indeterminate results. In these cases, sonographic risk stratification may also provide guidance.
Several approaches have been proposed to improve the diagnostic accuracy and the reproducibility of thyroid sonographic evaluation. It was reported that the use of computer-aided diagnosis systems to differentiate malignant from benign nodules showed accuracy similar to that obtained by radiologists and may reduce intra- and inter-observer variability.
This study aims to develop a new thyroid nodule ultrasound classification system based on a deep learning approach. The process will include generating a sonographic image database large enough (10'000 images) to contain examples of all predictive features and adequately train the system. This approach could also lead to the discovery of new US features that could be identified during the training or a better definition of the known ones. The system will be subsequently validated in an independent validation cohort and compared to current TIRADS (Thyroid Imaging Reporting and Data systems).
