AI makes biopsy recommendations comparable to expert radiologists

Deep learning algorithms can manage thyroid nodules on ultrasound (US) images at a level comparable to expert radiologists, according to new research published in Radiology.

“Imaging with US remains an accurate method to guide recommendation for management of thyroid nodules, although interpretation variability and overdiagnosis represent continual challenges,” wrote Mateusz Buda, department of radiology at Duke University School of Medicine, and colleagues.

Buda et al. developed a deep convolutional neural network that could provide biopsy recommendations for thyroid nodules detected on US images. The team’s training set included data from more than 1,300 nodules in more than 1,200 patients who underwent thyroid US examinations and a fine-needle aspiration (FNA) biopsy at a single institution from August 2006 to May 2010. The test set included 99 nodules and 91 patients. All images from the training set were interpreted by three radiologists who would go on to help develop the American College of Radiology (ACR) Thyroid Imaging Reporting and Data System (TI-RADS).

Overall, the algorithm had a sensitivity of 87% and specificity of 52% on the test set of 99 nodules. The consensus sensitivity and specificity of three ACR TI-RADS experts were 87% and 51%, respectively, almost exactly the same as the algorithm’s performance. The team’s algorithm also had a better sensitivity than five of the nine other radiologists who participated in the study, and its specificity was higher than seven of nine radiologists.

“In our study, we developed a deep learning algorithm to provide management recommendations for thyroid nodules observed on US images and compared its performance with radiologists who adhered to ACR TI-RADS guidelines,” the authors wrote. They also noted that their findings “add to the growing body of evidence demonstrating the potential power of deep learning when applied to thyroid US.”

The authors also noted that their work did have certain limitations. The final test set only included 15 malignant nodules, for instance, and a test set from an outside institution may have provided additional insight. Overall, however, Badu et al. believe their research shows that deep learning can play a key role in the management of thyroid nodules.