Artificial Intelligence Gains Ground in Thyroid Nodule Diagnosis

Nancy A. Melville

October 10, 2018

WASHINGTON, DC — Artificial intelligence (AI) is making strides in improving the tricky task of evaluation of thyroid nodules, with new research out of China showing an AI-assisted ultrasound diagnostic system to have similar — and in some cases even better — accuracy in diagnostic evaluation of nodules compared with physicians.

"Our cases show that the AI-assisted ultrasound diagnostic system can recognize image features of...thyroid nodules and assess their malignant risk," said first author Yu-kun Luo, MD, of the Ultrasound Department, General Hospital of the People's Liberation Army, Beijing, China, presenting here at the 2018 Annual Meeting of the American Thyroid Association (ATA).

"Furthermore, the system can help physicians to improve diagnostic efficiency and accuracy."

Luo and colleagues programmed the AI-assisted system with "deep learning" and a convolutional neural network to autonomously process benign and malignant thyroid nodule images using a database of more than 6000 images for which they already had pathological results.

"The system used static image segmentation to identify clear margins, irregular margins, and position adjacent to a carotid artery and major nodules," Luo said. Dynamic image segmentation allowed for further evaluation.

Asked for her opinion of the work, session moderator Mabel Ryder, MD, Mayo Clinic, Rochester, Minnesota, told Medscape Medical News that the research suggests intriguing possibilities for thyroid evaluation in the future.

"I think it's certainly interesting...and with thyroid nodules being so common, it's exciting to think that tools outside the human brain can help us risk stratify thyroid nodules and hopefully prevent a number of unnecessary surgeries of benign nodules," said Ryder, who is co-chair of the ATA program committee.

"Nodules affect 70% of the population. Many who read ultrasound reports may not be experienced thyroid radiologists or endocrinologists, who are good at picking out suspicious from nonsuspicious nodules, so to have an AI approach reliably do that would be very interesting and exciting," she added.

AI Readings Compared With, and in Addition to, Physician Judgment

To validate the system's accuracy, Luo's team selected 789 static images of thyroid nodules, including 444 that were benign and 345 that were malignant, which had clear biopsies or surgical pathology results. The images used were from 10 different types of ultrasound equipment, with line array probes ranging from 4.5 to 13 MHz.

The images were stored as consecutively longitudinal and transverse sections.

Comparison of the results of the AI readings with those of junior and senior physicians showed that the AI system had a sensitivity of 81.2% compared with 67.3% for junior physicians, 87.4% for senior physicians, and 83.2% when AI was combined with the judgment of junior physicians.

The corresponding rates in terms of specificity were 83.0% for AI alone, 94.3% for junior physicians, 83.2% for senior physicians, and 86.7% for AI combined with junior physicians.

The AI-assisted system had a diagnostic accuracy of 82.3% compared with 77.6% for junior physicians alone, 85.6% for senior physicians, and 85.2% when AI was combined with junior physicians.

Further tests assessing the diagnostic efficiency of AI on thyroid nodules of different sizes showed an accuracy of 83.7% for those less than 1 cm (sensitivity, 87.6%; specificity, 79.5%) and 84.9% accuracy for nodules ranging from 1 to 2 cm (sensitivity, 88.3%; specificity, 78.9%).

For AI combined with junior physicians, the accuracy was 81.7% for nodules less than 1 cm and 88.5% for nodules 1 to 2 cm.

However, accuracy for AI alone for nodules above 2 cm was lower, at 70.8% (P < .001), with 57.7% sensitivity and 91.8% specificity.

When AI was combined with junior physicians, the accuracy for nodules larger than 2 cm was significantly improved, at 86.6%.  

AI System Picks Up Subtle Image Details Elusive to the Human Eye

Luo said several cases of nodules were misdiagnosed by physicians.

"The physician's diagnosis was benign, but the AI system was able to diagnose it as malignant, and the pathological diagnosis did show it was a papillary thyroid cancer," she said.

"So it was determined that the AI system likely could recognize the subtle image information that could not be distinguished by the human eye."

Luo noted that AI is also making headway in other specialties in diagnostic applications for a range of conditions, including vascular, musculoskeletal, cardiac applications, and in maternal medicine. The technology is also being used to develop a multi-image fusion system that will combine ultrasound, CT, MRI, and AI.

"Although AI has a long way to go towards its application in clinical practice, with interdisciplinary cooperation, the development of imaging medicine (with AI) could help us solve more medical problems and benefit more patients," she concluded.

Ryder said that "the big question" will be the system's ability to identify more challenging nodules.

"I think the most important question that we all struggle with, whether it's a computer or a human, is: What are the nodules diagnosed as indeterminate nodules on fine-needle aspiration — follicular neoplasms, Hurthle cell neoplasms, for example? Those are either adenomas or cancers, and they don't necessarily have classic features on ultrasound as papillary thyroid cancer," Ryder explained.

"So, while the human eye has done many studies into specific features, the question will be: Can AI independently figure out which [nodules] are suspicious and which are not?" she concluded.

The authors and Ryder have reported no relevant financial relationships.

2018 Annual Meeting of the American Thyroid Association. October 4, 2018; Washington, DC. Abstract Oral 1.

For more diabetes and endocrinology news, follow us on Twitter and on Facebook.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.