Michael Monostra , 2025-05-02 12:46:00
Key takeaways:
- An AI model was trained to determine prognosis of indeterminate thyroid nodules using ultrasound images and molecular testing.
- Researchers said the model could be used to provide treatment guidance in practice.
An AI model trained on analyzing ultrasound images and molecular testing results may assist health care professionals treating indeterminate thyroid nodules, researchers stated in a letter to the editor published in Thyroid.
Researchers created an attention-multiple-instance deep learning AI model to predict whether indeterminate thyroid nodules were benign or malignant. After creating a baseline model that performed about as well as models reported in previous studies, researchers made adjustments and eventually created an ensemble model that significantly improved upon the base model.
“We were able to improve suspicion scores from molecular testing reports by integrating them with features from diagnostic ultrasound studies,” William Speier, PhD, assistant professor of radiology, bioengineering and medical informatics and associate director of the Biomedical AI Research lab at the University of California, Los Angeles, told Healio. “Because overtreatment is a significant concern in thyroid cancer management, our goal was to avoid false positives without missing malignancies. We tuned our model to match clinical sensitivity and found that we were able to reduce false positives by 11.5%.”
Speier and colleagues collected data from patients with indeterminate thyroid nodules who presented at the UCLA Medical Center from May 2016 to February 2022. The initial model incorporated ultrasound imaging, molecular testing results and Bethesda category. The model’s goal was to reduce false positives among indeterminate thyroid nodules.
Of the 333 patients, 259 had a benign thyroid nodule (median age, 56 years; 19.1% men) and 74 had a malignant nodule (median age, 47 years; 21.1% men). Of those with malignancy, 47% had papillary thyroid cancer, 35% had noninvasive follicular thyroid neoplasm with papillary-like nuclear features and 15% had follicular thyroid cancer.
The baseline AI model had an area under the receiver operating characteristic curve (AUROC) of 0.728, a sensitivity of 0.946 and a specificity of 0.664. The positive predictive value for the baseline model was 0.448. The researchers wrote that the AUROC with the base model was consistent with what was observed with other models in prior studies.
“We were surprised that including the Bethesda score from biopsies did not improve diagnostic outcome,” Speier said. “We expected that this score would be helpful information for our AI model because of the published difference in malignancy rate between Bethesda III and Bethesda IV biopsies. However, including this information did not give significant results when used alone or in combination with imaging or molecular testing.”
Researchers tested further models that excluded Bethesda category. In a final ensemble model, AUROC rose to 0.831 and specificity increased to 0.703 (P = .0008 for both) while sensitivity remained the same as the base model. Positive predictive value increased slightly to 0.477 in the final model.
“This algorithm could be used to provide guidance for physicians choosing treatment options after molecular testing of indeterminate nodules,” Speier said. “This model can identify patients who have had a suspicious molecular test, but who do not have a malignancy. This knowledge can potentially assist physicians to choose a less drastic treatment without increasing the risk of missing cancer.”
The findings are promising, but a larger study using data from multiple centers is needed to validate the model, and researchers are continuing to alter the model and are looking to incorporate more data, Speier said.
“We are working to expand our model to include additional information and modalities, including radiology and pathology reports, digitized cytopathology slides and specific mutation information from the molecular test reports,” Speier said.
For more information:
William Speier, PhD, can be reached at speier@ucla.edu.
