A deep learning computer model based on ultrasound (US) images could be used as an adjuvant diagnostic technique that could aid in the differentiation between medullary thyroid carcinoma (MTC) and follicular thyroid adenoma (FTA), according to a recently published study in the European Journal of Neurology.  

Currently, the rate of cervical lymph node metastasis at the moment of diagnosis in MTC is over 70%; this occurs, in part, due to the lack of specific US manifestations of MTC. According to several studies, approximately 33% of MTCs display benign US characteristics similar to those observed in FTAs. Misdiagnosis of MTC as benign nodules often leads to treatment delay and a higher risk of metastasis.

Deep learning is a machine learning technique that uses artificial neural networks to extract better features from raw data. Previous studies have confirmed the good performance of deep learning models in classifying thyroid nodules. However, no studies assess these models’ ability to distinguish between MTC and FTA.

Continue Reading

Read more about MTC diagnosis

The authors aimed to create a deep learning model for differentiation between MTC and FTA that could outperform a junior sonographer. The study included files from 385 patients with either MTC or FTA confirmed by immunohistochemistry and pathology. 

All images were classified by 2 junior sonographers as highly suspicion, intermediate suspicion, low suspicion, very low suspicion, and benign. The senior sonographer had 2 years of experience, and the junior sonographer had 5 years of experience. They posteriorly assessed images.

The diagnostic performance of the sonographers was compared with that of 2 separate deep learning models (ResNet-18 and Res-Net34) using the area under the receiver operating characteristic curve. Both deep learning models outperformed the junior sonographers; there was no significant statistical difference between the deep learning models and the senior sonographer.

Interestingly, the Res-Net34 had the best sensitivity, while the senior sonographer had the better specificity. Although the senior sonographer diagnosed all typical MTCs correctly, he misdiagnosed 11 MTCs out of 16, while Res-Net34 misdiagnosed only 4 atypical MTCs.

“The results of our study show that deep learning can be used as an auxiliary diagnostic tool to differentiate MTC from FTA on the gray scale US,” the authors concluded. 


Zhang R, Yi G, Pu S, et al. Deep learning based on ultrasound to differentiate pathologically proven atypical and typical medullary thyroid carcinoma from follicular thyroid adenoma. Eur J  Radiol. Published online October 3,  2022. doi:10.1016/j.ejrad.2022.110547