A team led by Duke scientist Mateusz Buda has put together a “deep convolutional neural network” able to make biopsy recommendations based on ultrasound (US) images of thyroids with nodules on them. In early July, they published a study in Radiology based on their findings. The study found that their artificial intelligence (AI) algorithm was able to make recommendations at a level akin to that of experienced radiologists.
“Imaging with US remains an accurate method to guide recommendation for management of thyroid nodules, although interpretation variability and overdiagnosis represent continual challenges,” the team wrote.
For the study, the team used data from over 1,200 thyroid US scans, which included data for more than 1,300 thyroid nodules, which are often biopsied to be determined whether they should be considered cause for concern. The data, which was from scans performed between August 2006 and May 2010, was taken from one institution. Furthermore, all the images used to train the AI were also interpreted by three radiologists; interestingly enough, these same radiologists later developed the American College of Radiology (ACR) Thyroid Imaging Reporting and Data System (TI-RADS).
After training, the algorithm was tested on a sample size of 99 thyroid nodules collected from 91 patients. It was found to have a specificity of 52% and a sensitivity of 87%, while the three ACR TI-RADS experts demonstrated the same consensus values at 51% and 87%.
About the study’s results, the team wrote that they: “add to the growing body of evidence demonstrating the potential power of deep learning when applying to thyroid US.” They did, however, also acknowledge that their study had certain limitations, such as the fact that the final test set used only had 15 malignant nodules; using a test set from an additional, outside institution may have strengthened their data. They do, ultimately feel their research provides an illuminating insight into how AI can help maintain thyroid nodules.