Thyroid nodules: a model of radiomics that knows how to say "I don't know" and refers 37% of cases to the doctor

Researchers combine radiomics, interpretable additive modeling, and conformal prediction to reduce unnecessary biopsies while flagging ambiguous ultrasounds.

The assessment of the risk of malignancy of a thyroid nodule is based on ultrasound, but the current grading grids vary greatly from observer to observer, contributing to many unnecessary biopsies. A team of researchers proposes a framework that extracts quantitative characteristics of the images with PyRadiomics, then classes them with a generalized additive model (GAM), selected among six candidates for its good compromise between performance and readability.

Explicit uncertainty

Tested on TN5000, a public base of nodule ultrasounds, the model retains 36 characteristics and reaches a AUC of 0.89, with a sensitivity of 0.83 and a specificity of 0.82. Above all, the authors add a cross-consistent prediction, guaranteed separately for each class: instead of forcing a binary response, the system only slices when it is sufficiently secure.

With the chosen threshold, 63% of nodules receive a single prediction, with an accuracy of 0.94 for malignant cases and a negative predictive value of 0.73 for those considered benign. The remaining 37% are reported as ambiguous and returned to the clinician.