Prediction of Central Lymph Node Metastasis in Papillary Thyroid Microcarcinoma Using a Deep Learning Radiomics Model Based on SAM3 Automatic Segmentation of Ultrasound Images: A Multicenter Cohort Study.

RATIONALE AND OBJECTIVES: We aimed to establish a Segment Anything Model 3 (SAM3) based on ultrasound images for automatic papillary thyroid microcarcinoma (PTMC) segmentation and develop and validate a deep learning radiomics (DLR) model based on ultrasound images for noninvasive prediction of central lymph node metastasis (CLNM) in PTMC. MATERIALS AND METHODS: We retrospectively collected data from 1859 patients with PTMC (1674 and 185 patients in training and validation groups, respectively) who underwent thyroidectomy and lymph node dissection from four medical centers between June 2017 and December 2025. To test the generalizability of the model, we collected data from 140 patients from another facility between June 2017 and December 2025. We automatically segmented tumors in PTMC ultrasound images using SAM3. We then extracted deep learning (DL) and radiomics features from 2D ultrasound images and established a DLR model following dimensionality reduction. We evaluated model utility using receiver operating characteristics, calibration, and decision curve analyses. RESULTS: We evaluated 1999 patients. The Dice similarity coefficient was 0.882 ± 0.145 and 0.859 ± 0.200 in the validation and external testing groups, respectively. The areas under the curve of the radiomics, DL, and DLR models were 0.825, 0.844, and 0.901 in the training group; 0.794, 0.825, and 0.875 in the validation group; and 0.779, 0.812, and 0.853 in the external testing group, respectively. Decision curve analysis validated the utility of the DLR model. CONCLUSION: DL-based segmentation built upon SAM3 achieves satisfactory outcomes in automatic PTMC delineation from ultrasound images. The DLR model noninvasively predicts CLNM, supporting clinical decision-making for PTMC.