Fine-tuned Medical Vision-Language Model (MedGemma 1.5 4B) with and without Structured Radiologist Descriptors versus EU-TIRADS for Malignancy Risk Stratification of Thyroid Nodules at Ultrasound: A Tunisian Multicenter Diagnostic Accuracy Study
Journal:
medRxiv
Published Date:
Oct 24, 2026
Abstract
Purpose. To evaluate a medical vision-language foundation model fine-tuned on thyroid ultrasound images, with or without the radiologist's structured descriptors, against expert-rated EU-TIRADS for the classification of thyroid nodules as benign or malignant. Materials and Methods. This retrospective study included nodules that underwent ultrasound-guided fine-needle aspiration (FNA) between January 2019 and August 2025 in two Tunisian centers, with consecutive recruitment, together with a selection of images from an international public dataset. Of 1,580 nodules, 1,456 (272 malignant; 1,280 patients) with an interpretable cytological or histological reference standard were allocated at patient level to training (n = 942), validation (n = 258) and a test set sealed before training (n = 256; 51 malignant). MedGemma 1.5 4B was fine-tuned with low-rank adaptation under four input configurations: the image with nodule biometry, with or without age and sex (A, B); A with chain-of-thought supervision (C); and A plus the eleven qualitative descriptors of the radiologist's structured report (D). The EU-TIRADS category was never a model input. Areas under the receiver operating characteristic curve (AUROCs) were compared with DeLong's test and a prespecified non-inferiority margin of 0.05; EU-TIRADS >= 4 was the prespecified comparator, with >= 5 as a sensitivity analysis. Results. Config D reached an AUROC of 0.953 (95% CI 0.909-0.976), versus 0.948 (0.894-0.972) for EU-TIRADS (difference +0.004, 95% CI -0.016 to +0.025), which met the non-inferiority criterion; Configs A-C, without the qualitative descriptors, reached 0.883-0.892 (Holm-adjusted p = 0.045 versus EU-TIRADS). At >= 4, EU-TIRADS had a sensitivity of 98.0% and produced 94 false positives among 205 benign nodules; at a threshold fixed on validation data, Config D reclassified 72 of them as negative while missing two additional cancers. Against EU-TIRADS >= 5 (sensitivity 92.2%, specificity 95.1%), Config D at its native threshold (88.2%, 94.1%) did not classify more nodules correctly (McNemar p = 0.22), and at its two other thresholds it classified fewer nodules correctly (p = 0.004 and 0.007). Raw probabilities were miscalibrated (expected calibration error 0.117 for Config D), and Platt scaling within the test set reduced this error to 0.019-0.040. Conclusion. Given the ultrasound image together with the radiologist's structured descriptors, the fine-tuned model matched expert EU-TIRADS discrimination and, against the >= 4 threshold, reclassified most false positives as negative, a gain concentrated in category 4 and based on only three cancers in category 4. It is best positioned as a report-standardizing second reader in the EU-TIRADS 4 gray zone, pending prospective external validation.
