In the Journals

Machine learning algorithm may diagnose thyroid nodules better than radiologists

A machine learning algorithm provided thyroid nodule diagnoses with higher sensitivity, specificity and accuracy than an experienced radiologist, suggesting that ultrasound diagnosis could be enhanced by adding artificial intelligence to human examination, according to findings presented in Thyroid.

“Improved diagnostic methods are needed for identifying malignant thyroid nodules that would be less subjective and more robust than conventional [ultrasound] and that would reduce the need for biopsy or diagnostic surgery,” Shuixing Zhang, MD, PhD, of the department of radiology at The First Affiliated Hospital of Jinan University in Guangzhou, Guangdong, China, and colleagues wrote. “It may be possible to develop a superior noninvasive diagnostic method by taking advantage of the ability of artificial intelligence to ‘learn’ from clinical data sets and predict diagnoses, in some cases more accurately than human experts.”

Zhang and colleagues conducted a retrospective study weighing the effectiveness of using preselected algorithms to identify and diagnose 2,064 thyroid nodules from 2,032 patients (mean age, 45.3 years; 65.8% women) who were consecutively enrolled from April 2011 to June 2016. All patients had a thyroid nodule measuring 2.5 cm or less without a history of surgery and who, at least 1 month before thyroidectomy, had both ultrasound and real-time elastography performed.

Using the data from the ultrasound and real-time elastography procedures, the researchers established two data sets that nine algorithms interpreted 1,000 times. Results from each algorithm were compared with one another, and the superior of the nine was then compared with the diagnostic area under the curve, sensitivity, specificity and accuracy of a thyroid radiologist with at least 17 years in the field.

The algorithm had a better AUC (0.924; 95% CI, 0.895-0.953) than the radiologist (0.834; 95% CI, 0.815-0.853) when using ultrasound data alone. According to the researchers, the algorithm was more sensitive (88.1%; 95% CI, 81.3-95 vs. 83.5%; 95% CI, 80-86.6), more specific (84.2%; 95% CI, 80.9-87.5 vs. 83.2%; 95% CI, 80.9-85.5) and more accurate (84.6%; 95% CI, 81.6-87.5 vs. 83.3%; 95% CI, 81.5-85.2) compared with the radiologist in this instance.

The researchers noted that including real-time elastography data improved the outcomes from both the algorithm and the radiologist. However, the algorithm still produced superior readings in AUC (0.938; 95% CI, 0.914-0.961 vs. 0.843; 95% CI, 0.829-0.857), sensitivity (89.1%; 95% CI, 82.5-95.8 vs. 85%; 95% CI, 83-87), specificity (85.3%; 95% CI, 83-87.6 vs. 83.6%; 95% CI, 81.8-85.4) and accuracy (85.7%; 95% CI, 83.6-87.8 vs. 84.1%; 95% CI, 82.7-85.5) compared with the radiologist.

“These results demonstrate the feasibility of strengthening diagnosis of malignant thyroid nodules using machine learning,” the researchers wrote. “They have also led to an online diagnostic tool that can guide diagnosis and clinical decisionmaking.” – by Phil Neuffer

Disclosures: The authors report no relevant financial disclosures.

A machine learning algorithm provided thyroid nodule diagnoses with higher sensitivity, specificity and accuracy than an experienced radiologist, suggesting that ultrasound diagnosis could be enhanced by adding artificial intelligence to human examination, according to findings presented in Thyroid.

“Improved diagnostic methods are needed for identifying malignant thyroid nodules that would be less subjective and more robust than conventional [ultrasound] and that would reduce the need for biopsy or diagnostic surgery,” Shuixing Zhang, MD, PhD, of the department of radiology at The First Affiliated Hospital of Jinan University in Guangzhou, Guangdong, China, and colleagues wrote. “It may be possible to develop a superior noninvasive diagnostic method by taking advantage of the ability of artificial intelligence to ‘learn’ from clinical data sets and predict diagnoses, in some cases more accurately than human experts.”

Zhang and colleagues conducted a retrospective study weighing the effectiveness of using preselected algorithms to identify and diagnose 2,064 thyroid nodules from 2,032 patients (mean age, 45.3 years; 65.8% women) who were consecutively enrolled from April 2011 to June 2016. All patients had a thyroid nodule measuring 2.5 cm or less without a history of surgery and who, at least 1 month before thyroidectomy, had both ultrasound and real-time elastography performed.

Using the data from the ultrasound and real-time elastography procedures, the researchers established two data sets that nine algorithms interpreted 1,000 times. Results from each algorithm were compared with one another, and the superior of the nine was then compared with the diagnostic area under the curve, sensitivity, specificity and accuracy of a thyroid radiologist with at least 17 years in the field.

The algorithm had a better AUC (0.924; 95% CI, 0.895-0.953) than the radiologist (0.834; 95% CI, 0.815-0.853) when using ultrasound data alone. According to the researchers, the algorithm was more sensitive (88.1%; 95% CI, 81.3-95 vs. 83.5%; 95% CI, 80-86.6), more specific (84.2%; 95% CI, 80.9-87.5 vs. 83.2%; 95% CI, 80.9-85.5) and more accurate (84.6%; 95% CI, 81.6-87.5 vs. 83.3%; 95% CI, 81.5-85.2) compared with the radiologist in this instance.

The researchers noted that including real-time elastography data improved the outcomes from both the algorithm and the radiologist. However, the algorithm still produced superior readings in AUC (0.938; 95% CI, 0.914-0.961 vs. 0.843; 95% CI, 0.829-0.857), sensitivity (89.1%; 95% CI, 82.5-95.8 vs. 85%; 95% CI, 83-87), specificity (85.3%; 95% CI, 83-87.6 vs. 83.6%; 95% CI, 81.8-85.4) and accuracy (85.7%; 95% CI, 83.6-87.8 vs. 84.1%; 95% CI, 82.7-85.5) compared with the radiologist.

“These results demonstrate the feasibility of strengthening diagnosis of malignant thyroid nodules using machine learning,” the researchers wrote. “They have also led to an online diagnostic tool that can guide diagnosis and clinical decisionmaking.” – by Phil Neuffer

Disclosures: The authors report no relevant financial disclosures.