The following is a summary of “Combining Image Similarity and Predictive AI Models to Decrease Subjectivity in Thyroid Nodule Diagnosis and Improve Malignancy Prediction,” published in the August 2024 issue of Endocrinology by Nair et al.
This study aimed to assess the efficacy of integrating predictive artificial intelligence (AI) with image similarity models for risk-stratifying thyroid nodules through a retrospective external validation approach.
Two distinct datasets were utilized to evaluate the AI application’s performance: the Stanford dataset, comprising ultrasound images of 192 thyroid nodules collected between April 2017 and May 2018, and a private practice dataset, including 118 nodule images from January 2018 to December 2023. These nodules were definitively diagnosed via cytology or surgical pathology. The AI application was employed to predict the diagnosis and the American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS) score.
In the Stanford dataset, the AI application demonstrated a sensitivity of 1.0 and a specificity of 0.55 for predicting malignancies, with a positive predictive value (PPV) of 0.18 and a negative predictive value (NPV) of 1.0. The Area Under the Curve – Receiver Operating Characteristic (AUC-ROC) was 0.78. The ACR TI-RADS-based clinical recommendations exhibited a polychoric correlation of 0.67. Conversely, in the private practice dataset, the AI application achieved a sensitivity of 0.91 and a specificity of 0.95, with a PPV of 0.80 and an NPV of 0.98. The AUC-ROC was 0.93, and the accuracy was 0.94, while the ACR TI-RADS score showed a polychoric correlation of 0.94.
The AI application demonstrated strong performance in sensitivity and NPV across both datasets and showed potential for a 61.5% reduction in the necessity for fine needle aspiration (FNA) procedures. It also exhibited a robust correlation with ACR TI-RADS scoring. However, variability in PPV between the datasets, likely due to differences in image selection and malignancy prevalence, was noted. If this AI application is implemented consistently across diverse clinical settings, it could significantly reduce patient burden associated with invasive procedures and potentially decrease healthcare costs.
Source: sciencedirect.com/science/article/abs/pii/S1530891X24006475
The post Improving Thyroid Nodule Diagnosis with AI and Image Similarity Models first appeared on Physician's Weekly.