Thyroid nodule ultrasound: prevalence-adjusted malignancy risk

Research prototype — not a medical device and not for clinical decisions. The analysis runs entirely in your browser; the image is not uploaded anywhere. Do not use images containing patient identifiers.

Accompanies the manuscript Artificial intelligence for thyroid nodule ultrasound overestimates malignancy risk when cancer prevalence is lower than in development cohorts (in preparation).

Load a B-mode image or choose an example.
Tap two opposite corners of the nodule.
Set how often nodules in your practice are malignant.
Analyse.

Ultrasound image Or an example from the TN5000 test set

Loading…

In your practice, what share of the nodules you evaluate turn out to be cancer? 10%

This is not the risk of this nodule — it is your clinic's malignancy rate (from your own FNA/surgery records). If unsure, keep 10%. The model was developed where 70.5% of nodules were malignant; this setting corrects for that difference.

Don't know your rate? Typical values from the literature: ≈10% — unselected nodules / nodules sent for FNA · ≈21% — patients who underwent thyroidectomy · 70% — the development cohort (no adjustment).

Performance in the study: internal test AUC 0.940; external test (TN3K) AUC 0.762. At 10% prevalence and a 10% threshold: sensitivity 88.6%, specificity 83.4%. Performance on other devices and populations may be lower — validate locally.

Method: lightweight CNN (0.29 M parameters) with 5-variant test-time augmentation after caliper-marker inpainting; prior (odds) correction logit(p′) = logit(p) + logit(π) − logit(0.705). Examples: TN5000 (Zhang et al., Sci Data 2025;12:1437; CC BY).

Typical malignancy rates: ≈10% of thyroid nodules harbour clinically significant cancer (Durante et al., JAMA 2018) and FNA series from Duke (10.3%) and Spain (11.1%) are in this range (Zhan et al., Front Oncol 2025); thyroidectomy cohorts in Spain (21.5%; Fernández Alba et al., Diagnostics 2025) and Portugal (20.9%; da Silva et al., Endocrine 2026).