Deep Learning-Based Localization of Preoperative Parathyroid Glands in Secondary Hyperparathyroidism Using Dual-Modality Data from Unenhanced CT and Contrast-Enhanced CT

Abstract

Abstract Background Accurate preoperative localization of parathyroid glands (PGs) is crucial in patients with secondary hyperparathyroidism (sHPT) who are scheduled for parathyroidectomy (PTx) surgery. Nonetheless, despite its importance, this remains a challenging task. The existing medical imaging techniques used for preoperative PG localization have varying levels of sensitivity and accessibility. This study aims to construct a deep-learning model based on a multimodal framework for identifying PGs drawing on a dual-modality dataset consisting of plain CT and enhanced CT, we validate the model’s sensitivity in clinical performance. Methods A retrospective study was conducted using a dataset of 94 CT images from 47 patients. For each patient there is a plain CT and an enhanced CT scanned image. The data were randomly partitioned into a training set (38 cases, 76 CT images) and a test sets (9 cases, 18 CT images). A U-Net model was trained on the training set then validated on the test set. In our analysis, the sensitivity of recognizing PGs with imaging information of various modalities was compared between the developed model and clinical physicians. An error analysis and an inter-modal imaging complementarity analysis were performed to provide references for subsequent model enhancement and application. Results The identification of parathyroid glands (PGs) using dual-modality CT has shown a diagnostic sensitivity of 94.44%. This result is significantly higher than those obtained by clinicians using ultrasound (61.11%, P = 0.0013) and CT (72.22%, P = 0.0238). Additionally, the sensitivity achieved by dual-modality CT is comparable to that of Tc-MIBI SPECT/CT (86.11%, P = 0.429). We also found that combining the predictions from this model with other imaging modalities could further improve the detection rate of PGs. Conclusions To the best of our knowledge, this study is the first to use artificial intelligence techniques with CT bimodal data for preoperative localization of parathyroid glands. The findings of the study suggest that using a deep learning model with plain and enhanced CT data can improve the ability to identify parathyroid glands prior to thyroidectomy or parathyroidectomy.