Artificial intelligence for thyroid nodule ultrasound image analysis

Abstract

Deep learning (DL) as part of artificial intelligence (AI) is based on artificial neural networks, which use a multi-step process to automatically analyze features of an image, then classify them. Medical image analysis using DL has been rapidly adopted across multiple medical fields. Neck ultrasound (US) is a gold standard diagnostic modality in thyroid nodules, and suitable for DL diagnosis because the characteristics of the thyroid nodules can be captured in one representative image. Therefore, a number of studies applied DL in analyzing neck US and tried to predict malignancy risk of the thyroid nodules, and showed relatively high diagnostic performances. For successful DL analysis for the thyroid US images, several issues should be solved which includes image selection by experienced clinicians, proper quality of US, enough number of US images to train the DL, accurate labeling, and adequate hyperparameters. However, DL analysis is still in its early stages, and the questions are unanswered yet. We do not know how many images we need to obtain for proper training, and there is no standard of US quality. Moreover, DL analysis may not be able to classify indeterminate nodules into benign or malignant even in the near future, not to mention now. In this review, we will discuss the current status, limitations, and the future directions of DL for thyroid US image analysis from a clinician’s point of view.