Development of an algorithm for intraoperative autofluorescence assessment of parathyroid glands in primary...

Previous work showed that normal and abnormal parathyroid glands exhibit different patterns of autofluorescence, with the former appearing brighter and more homogenous. However, an objective algorithm based on quantified measurements was not provided. The aim of this study is to develop objective algorithms for intraoperative autofluorescence assessment of parathyroid glands in primary hyperparathyroidism using artificial intelligence.

The utility of near-infrared fluorescence imaging in parathyroidectomy procedures was evaluated in a study approved by the institutional review board. Autofluorescence patterns of parathyroid glands were measured intraoperatively. Comparisons were performed between normal and abnormal glands, as well as between different pathologies. Using machine learning, decision trees were created.

Normal parathyroid glands were brighter (higher normalized autofluorescence pixel intensity) and more homogenous (lower heterogeneity index) compared to abnormal glands. Optimal cutoffs to differentiate normal from abnormal parathyroid glands were >2.0 for normalized autofluorescence intensity (sensitivity 73%, specificity 70%, area under the curve 0.756) and <0.12 for parathyroid heterogeneity index (sensitivity 75%, specificity 81%, area under the curve 0.839). Decision trees created by machine learning using normalized autofluorescence intensity, heterogeneity index, and gland volume were 95% accurate in predicting normal versus abnormal glands and 84% accurate in predicting subclasses of parathyroid pathologies.

To our knowledge, this is the first study to date reporting objective algorithms using quantified autofluorescence data to intraoperatively assess parathyroid glands in primary hyperparathyroidism. These results suggest that objective data can be obtained from autofluorescence signals to help differentiate abnormal parathyroid glands from normal glands.