Accurate preoperative localization and characterization of parathyroid lesions is essential for successful surgical outcomes in patients with hyperparathyroidism. Ultrasonography is the most commonly used first-line modality owing to its noninvasiveness and accessibility. However, its diagnostic accuracy is limited by high operator dependency and difficulty in distinguishing parathyroid lesions from thyroid nodules, lymph nodes, or vascular structures.

We developed a machine learning–based framework to support clinical decision-making in the preoperative assessment of parathyroid disease. The workflow simulated real-world practice: (1) classification of ultrasonographic images to determine the presence or absence of parathyroid lesions; and (2) automated segmentation of anatomical structures including parathyroid glands, thyroid tissue, and cervical vessels. To enhance diagnostic accuracy, a multimodal learning model was further constructed by integrating ultrasonographic features with biochemical and clinical data.

The classification model demonstrated robust performance in delineating cervical structures with high accuracy. The segmentation model effectively distinguished parathyroid lesions from non-parathyroid findings, yielding favorable diagnostic metrics. Incorporation of multimodal data improved predictive reliability compared with imaging alone. Collectively, these results indicate that the models can provide direct, intuitive assistance to physicians interpreting ultrasonographic studies in patients with suspected parathyroid disease.

This study presents a machine learning–based modality for the classification and segmentation of parathyroid lesions on ultrasonography, with integration of clinical and biochemical information. The proposed framework has the potential to reduce operator dependency, improve reproducibility, and standardize preoperative imaging assessment in hyperparathyroidism. Given that this study was conducted at a single institution, validation in multicenter and multinational cohorts will be essential to confirm generalizability and enhance predictive performance.