Early Experience Utilizing 4D-CT Radiomic Features for Differentiation of Parathyroid Adenomas From Lymph Nodes and Thyroid Nodules

Objective: Minimally invasive parathyroidectomy (MIP) requires high-fidelity localization of parathyroid adenomas through preoperative imaging, commonly 4-dimensional computed tomography (4D-CT). Texture analysis extracts high-order mathematical features from an image and may be applied to 4D-CT for quantitative differentiation of lymph nodes and thyroid nodules from parathyroid adenomas. Methods: This is a retrospective cohort study of 51 patients diagnosed with PHPT and known parathyroid adenoma and/or thyroid nodule who have undergone preoperative 4D-CT imaging before parathyroidectomy. Three anatomic structures (parathyroid adenoma, lymph node, and thyroid nodule) were manually segmented on 25-second arterial phase axial sections of the 4D-CT scans. Radiomic data were extracted for shape, first-order, and second-order classes (107 total features) for each of the structures in each patient. A series of t tests were conducted to assess for radiomic features with statistically significant differences in lymph nodes or thyroid nodules when compared with parathyroid adenomas. A multivariable logistic regression model for discrimination of parathyroid adenomas was trained on a subset of the data set and assessed on a hold-out test subset. Results: When comparing parathyroid adenomas and lymph nodes, 14/18 first-order features and 44/75 second-order features were statistically significantly different ( P <0.05), of which 13/18 first-order features and 16/75 second-order features were potent discriminators ( P <0.0001). No features were significantly different between parathyroid adenomas and thyroid nodules. A multivariable logistic regression model for discrimination of parathyroid adenomas from lymph nodes achieved strong predictive performance (AUC: 0.95, 95% CI: 0.86–1). Conclusions: Parathyroid adenomas and lymph nodes have statistically distinct radiomic textural signatures on arterial phase 4D-CT, with the most significant differences found in first-order textural features. These findings may facilitate the development of future machine learning models for automated differentiation of parathyroid adenomas, further enhancing uptake of MIP and improving clinical outcomes.