Using Artificial Intelligence to Predict Risk of Thyroid Cancer on Ultrasound
When diagnosing whether something is harmful or not – like a lump – doctors have to look for patterns of features, like size, shape, expression of certain molecules, among other things. This process can be time-consuming and involve expensive procedures. Physicians are turning to artificial intelligence (AI) as an additional tool in their diagnostic artillery to help them make decisions faster about a lump or nodule whose risk of cancer is unknown. A new study from The Sidney Kimmel Cancer Center – Jefferson Health used a Google-platform machine-learning algorithm on ultrasound images of thyroid nodules, and found that the algorithm could predict high risk nodules with 97% specificity.
(Watch the animated video above explaining how machine learning works)
Thyroid nodules are small lumps that form within the thyroid gland and are quite common in the general population, with a prevalence as high as 67%. The great majority of thyroid nodules are not cancerous and cause no symptoms. However, reaching a diagnosis is tricky because standardized guidelines are limited. To improve the predictive power of the first-line diagnostic, the ultrasound, Jefferson researchers looked into machine learning models developed by Google. “The goal of our study was to see how well automated machine learning could predict the genetic risk of thyroid nodules,” says Kelly Daniels, first author of the study. First, they “trained” the machine-learning model or algorithm on ultrasound images of patients with genetically-confirmed high-risk thyroid nodules. The machine found patterns in the images that were indicative of high risk. Then the investigators tested the trained the algorithm on a different set of unlabeled images to see how closely it could classify high and low genetic risk nodules, compared to molecular — or genetic — test results.
The researchers found that their algorithm performed with 97% specificity, meaning that 97% of patients who truly have malignant nodules will have their ultrasound read as “malignant” by the algorithm. “This is the first use of machine learning combined with image-processing technology in the field of genetic risk stratification of thyroid nodules,” says John Eisenbrey, PhD, lead author of the study. “There’s already interest from other institutions to collaborate on data collection. The more data we feed the algorithm, the stronger and more predictive we’d expect it to become,” adds Elizabeth Cottrill, MD, the clinical leader of the study.
Though preliminary, the study suggests that automated machine learning shows promise as a rapid and inexpensive first screen for cancerous thyroid nodules.
Read more about the study here.