IntroductionAccurate preoperative localization of parathyroid glands (PGs) is crucial for patients with secondary hyperparathyroidism scheduled for parathyroidectomy. However

despite its importance

localization remains challenging since current preoperative imaging modalities for PG localization vary in sensitivity and accessibility.Materials and methodsIn this study

we developed a deep learning model for PG identification using a dual-phase computed tomography (CT) dataset

including unenhanced CT and contrast-enhanced (CE) CT data

and validated the model’s sensitivity in clinical application. A retrospective study was conducted using 94 CT images obtained from 47 patients (one plain CT image and one CE CT image per patient). The data were randomly assigned to a training set (38 cases

76 CT images) and a test set (9 cases

18 CT images) based on per-patient splits. A three-dimensional U-Net model wa

s trained using the training set and then validated using the test set. An analysis was conducted to compare the model’s and clinicians’ sensitivity in detecting PGs based on various imaging modalities. An error analysis and an intermodal imaging complementarity analysis were performed to provide references for subsequent model enhancement and application.ResultsThe dual-phase CT model identified PGs with a diagnostic sensitivity of 94.44%. This was significantly higher than the sensitivity achieved by clinicians using ultrasonography (61.11%

P

 = 0.0013) and CT (72.22%

P

 = 0.0238). Additionally

the sensitivity achieved using the dual-phase CT model was comparable to that achieved using technetium-99m-methoxyisobutylisonitrile single-photon emission CT/CT (86.11%

P

 = 0.429). We also found that combining predictions from this model with other imaging modalities further improved the PG detection rates.ConclusionsThe study findings suggest that using a deep learning model with plain and CE CT data could improve PG identification prior to thyroidectomy or parathyroidectomy.