Fine-tuned MedGemma model with structured descriptors matches EU-TIRADS for thyroid nodule risk stratification

[medRxiv] A retrospective diagnostic accuracy study at two Tunisian centers found that a fine-tuned MedGemma 1.5 4B model given ultrasound images plus radiologists' structured descriptors (Config D) reached an AUROC of 0.953 (95% CI 0.909-0.976) on a test set of 256 thyroid nodules sealed before training, versus 0.948 (0.894-0.972) for expert EU-TIRADS, a difference of +0.004 (95% CI -0.016 to +0.025) that met the prespecified 0.05 non-inferiority margin, according to the preprint. Configurations without the qualitative descriptors reached 0.883-0.892 (Holm-adjusted p=0.045 versus EU-TIRADS). Against EU-TIRADS 4 or higher, which had 98.0% sensitivity and produced 94 false positives among 205 benign nodules, Config D reclassified 72 of those false positives as negative at a threshold fixed on validation data, at the cost of two additional missed cancers. The authors said the model is best positioned as a report-standardizing second reader in the EU-TIRADS 4 gray zone, pending prospective external validation.