Endocrine Abstracts

Background

Artificial intelligence algorithms can be used in classification of thyroid nodules to reduce subjectivity. External validation on images collected from different ultrasound machines and other institutions are vital before wider use of these algorithms. We retrospectively analyzed the performance of AIBx on thyroid nodules with surgically proven pathology.

Materials and methods

Patients harboring thyroid nodules 1–4 cm in size, who underwent thyroid surgery from 2014 to 2016 in a single institution, were included in this study. Medullary thyroid cancer, metastasis from other cancers, thyroid lymphomas, and purely cystic nodules were excluded. Results were compared with TI-RADS calculated by experienced physicians. A subgroup analysis was done on cytologically indeterminate nodules.

Results

Out of 329 patients, 257 nodules from 209 individuals met the eligibility criteria. 51 nodules (19.8%) were malignant. AIBx had a negative predictive value (NPV) of 89.2%. Sensitivity, specificity, and positive predictive values (PPV) were 78.4%, 44.2% and 25.8% respectively. Considering both TI-RADS 4 and TI-RADS 5 nodules as malignant lesions resulted in an NPV of 93%, while PPV and specificity were only 22.4% and 19.4%, respectively. NPV was 89.6% if only TI-RADS 5 nodules were considered malignant. TI-RADS predicted all Bethesda category III nodules as malignant, despite the fact that none of them were malignant on histology. In contrast, only 25% of nodules in the Bethesda category III were predicted to be malignant by AIBx. On subgroup analysis, AIBx had a NPV of 95.8% in classical papillary thyroid cancer, while TI-RADS had an NPV of 93.0%.

Conclusion

When applied to an external dataset consisting of ultrasound images obtained in a different setting than used for training, AIBx had comparable negative predictive values to TI-RADS. AIBx performed even better than TI-RADS in Bethesda category III and classical papillary thyroid cancer. These data prove the concept of AIBx for thyroid nodules, and this tool may help less experienced operators by reducing the subjectivity inherent to thyroid ultrasound interpretation.