A visual deep learning model to predict abnormal versus normal parathyroid glands using intraoperative autofluorescence signals

Background: Previous work demonstrated that abnormal versus normal parathyroid glands (PGs) exhibit different patterns of autofluorescence, with former appearing darker and more heterogenous. Our objective was to develop a visual artificial intelligence model using intraoperative autofluorescence signals to predict whether a PG is abnormal (hypersecreting and/or hypercellular) or normal before excision during surgical exploration for primary hyperparathyroidism.

Methods: A total of 906 intraoperative parathyroid autofluorescence images of 303 patients undergoing parathyroidectomy/thyroidectomy were used to develop model. Autofluorescence image of each PG was uploaded into the visual artificial intelligence platform as abnormal or normal. For deep learning, randomly chosen 80% of data was used for training, 10% for testing, 10% for validation. The area under the receiver operating characteristic (AUROC), area under the precision-recall curve (AUPRC), recall (sensitivity), and precision (positive predictive value) of the model were calculated.

Results: AUROC and AUPRC of the model to predict normal and abnormal PGs were 0.90 and 0.93, respectively. Recall and precision of the model were 89% each.

Conclusion: Visual artificial intelligence platforms may be used to compare the autofluorescence signal of a given parathyroid gland against a large database. This may be a new adjunctive tool for intraoperative assessment of parathyroid glands during surgical exploration for primary hyperparathyroidism.