AI with TI-RADS Can Improve Detection for Thyroid Cancer Patients

Saturday, Dec. 05, 2020
Aylin Tahmasebi, MD


Use of an artificial intelligence (AI) model in conjunction with the regular Thyroid Imaging Reporting and Data System (TI-RADS) can improve diagnostic performance and characterization, according to new research.

Aylin Tahmasebi, MD, a research fellow at Thomas Jefferson University in Philadelphia, presented her team's findings, "TI-RADS Modification Using Artificial Intelligence Techniques May Improve Radiologists Decision Making," in a Saturday session.

"Thyroid cancer is a common cancer, and 1.3% of the general population will be diagnosed with it at some point in their life," Dr. Tahmasebi said. "However, it has an excellent five-year survival rate of 98.2%."

Identifying High-Risk Nodules

While thyroid nodules are a common finding on US, many require more expensive and time-consuming imaging to determine if they are clinically significant and develop a path of treatment, she said.

The team at Thomas Jefferson University conducted a retrospective study of 258 patients who underwent US imaging for thyroid nodules and either follow up US-guided fine needle aspiration biopsy with next-generation sequencing or post-resection pathology. Of those cases, 211 were used to train an AI model. Of those, 47 were used to test the model.

The Google AI learning tool was trained to identify markers of high-risk nodules on US images such as specific mutations or molecular markers associated with malignant cancer.

As part of the project, three blinded radiologists scored the images using TI-RADS and assigned each nodule as either high or low risk based on the presence of the highly suspicious imaging features the AI model would also seek to identify.

"Our aims for this study were to evaluate the AI diagnostic accuracy in genetic risk stratification of thyroid nodules, evaluate AI performance in locating the thyroid nodule, and evaluate the radiologist's performance when using TI-RADS with and without AI assistance in thyroid nodule genetic risk stratification," Dr. Tahmasebi said.

Dr. Tahmasebi's team determined that the AI algorithm correctly located all nodules, or 100%, in the test set, and predicted malignancy risk with higher sensitivity and specificity than the three radiologists.

When predicting high risk, the AI model achieved sensitivity of 68% compared to 52% by the radiologists. In terms of specificity, the AI model had a rate of 70% compared to 62% with the radiologists. And when the radiologists used AI with TI-RADS, they performed better than with TI-RADS alone.

"Incorporation of AI into TI-RADS improved radiologists' performance and showed better genetic malignancy risk prediction," Dr. Tahmasebi said. "This approach may help more accurately identify truly high-risk nodules needing further evaluation and treatment."

Example prediction from the Google AutoML algorithm (AutoML Vision; Google LLC) Object Detection Model that correctly detected a nodule and correctly assigned a high-risk label with 0.94 (94%) certainty. The position of the high-risk nodule was marked by the orange color bounding box drawn by the model.

For More Information:

View the RSNA 2020 session, "TI-RADS Modification Using Artificial Intelligence Techniques May Improve Radiologists Decision Making" — SSNR16 at