Improving Diagnostic Precision in Thyroid Pathology by Synergistic Use of AI and Molecular Markers

For indeterminate thyroid nodules, molecular tests offer high negative predictive value (NPV), reducing missed malignancies, but have limited positive predictive value (PPV), potentially leading to unnecessary surgeries. We evaluated how integrating artificial intelligence-based imaging (AIBx V2) with ThyroSeq v3 could enhance diagnostic accuracy.

We retrospectively analyzed 108 indeterminate thyroid nodules. Surgical pathology was available for 42 nodules for primary analysis; the remaining 66, without surgical pathology, were deemed benign for analytic purposes and included in the total cohort for secondary analysis reflecting real-world practice. We calculated test performance for AIBx V2 alone, ThyroSeq v3 alone, and a combined approach (AIBx V2+Mol). In the combined approach, when ThyroSeq v3 reported "Malignant", but the estimated malignancy probability was intermediate or lower, the final classification deferred to AIBx V2.

In the surgical pathology subset (n = 42), ThyroSeq v3 demonstrated high sensitivity (0.95) but moderate specificity (0.45), leading to a PPV of 0.65. AIBx V2 improved specificity (0.60) but had lower sensitivity (0.77). The AIBx V2+Mol approach retained high sensitivity (0.95) while raising specificity to 0.60, improving PPV to 0.72 and AUC from 0.70 to 0.77. In the entire cohort (n = 108), AIBx V2+Mol maintained excellent sensitivity (0.95) and further enhanced specificity, 0.90 vs0.87, PPV 0.72 vs0.65, and AUC 0.93 vs0.91.

Integrating AIBx V2 imaging model with ThyroSeq v3 preserved the high sensitivity of molecular testing while improving specificity and PPV. These exploratory results need validation in larger studies before the combined model is incorporated into clinical practice.