Ultrasound Plus Machine Learning Identifies Thyroid Cancer

November 4, 2019

Machine learning of ultrasound scans could provide a preliminary method of screening for thyroid cancer.

Machine learning of ultrasound scans could provide a preliminary method of screening for thyroid cancer, according to a new study.

The artificial intelligence (AI) was able to learn the hallmarks of lesions among more than 100 patients, and then make predictions based on the details of the images with an overall accuracy of 77.4% (95% CI, 63.8%-97.7%), according study results published in JAMA Otolaryngology-Head and Neck Surgery.

“The study found that the model developed through automated machine learning could produce high specificity for identifying nodules with high-risk mutations on molecular testing,” the study authors wrote. “This finding shows promise for the diagnostic applications of machine learning interpretation of sonographic imaging of indeterminate thyroid nodules.”

The machine was the commercial platform AutoML Vision made by Google.

The patient group was taken from a single site at the Thomas Jefferson University Hospital in Philadelphia.

The electronic health records of all patients who underwent ultrasound with the standard follow-up molecular testing of a thyroid nodule between January 2017 and August 2018 were included in the group.

The molecular testing of the entire group was an institution-specific panel contain 23 gene mutations and 5 gene rearrangements, the data showed.

The tally of patients and images was: 134 suspicious lesions in 121 patients, captured in 683 diagnostic ultrasound images, the authors wrote.

The computer training and testing process relied on roughly 80% of images for training, 10% used for validation-and the last 10% set aside for resting the accuracy of the model at the end, the investigators wrote.

Of the 134 lesions, 75 had no mutation, 43 had a high-risk mutation, and a further 16 were unknown or had low-risk mutations, according to the study.

The machine had an overall accuracy of 77.4%, according to the findings.

The breakdown included a sensitivity of 45% (95% CI, 23.1%-68.5%), a specificity of 97% (95% CI, 84.2%-99.9%), a positive predictive value of 90% (95% CI, 55.2%-98.5%), and a negative predictive value of 74.4%, the data showed.

Limitations exist, however. Among others: multiple images of a single nodule were used in the entire grouping of 683 images, meaning “learner bias” was possible if the same nodule ended up in the training and testing sets, according to the paper.

But the authors said they felt it was a positive first step toward having a new thyroid nodule screening tool.

“The more data we fed the algorithm the stronger and more predictive we’d expect it to become,” said Elizabeth Cottrill, MD, clinical leader of the study, who is a otolaryngologist at Thomas Jefferson.

REFERENCES:
Daniels K, Gummadi S, Zhu Z, et al. Machine Learning by Ultrasonography for Genetic Risk Stratification of Thyroid Nodules. JAMA Otolaryngol Head Neck Surg. 2019 Oct 24:1-6. doi:10.1001/jamaoto.2019.3073.