Abstract
Background: Preoperative imaging methods for the localization of abnormal parathyroid glands are widely used to facilitate ensuing surgery. Parathyroid scintigraphy with 99mTc-sestamibi (MIBI) is an established technique. However, little has been investigated regarding MIBI scan and Deep Learning (DL) algorithms for parathyroid gland identification. Materials and Methods: This retrospective study enrolled 418
patients, 397 with primary and 21 with secondary or tertiary hyperparathyroidism, who underwent parathyroid scintigraphy with double-phase and thyroid subtraction techniques. Data were collected from the archive of our laboratory. The study proposes a three-path network approach, employing the state-of-the-art Convolutional Neural Network called VGG-19. Image input to the model involved a set of three scintigraphic images in each case: MIBI early phase, MIBI late phase and 99mTcO4 thyroid scan. A medical expert’s diagnosis provided the ground truth for positive/negative results. Moreover, the image produced by the model was compared with the original scintigraphic images to examine the exact localization of findings. Results: Medical experts identified 391 abnormal glands in 296 patients. On a patient basis, the DL model attained an accuracy of 95.0% (sensitivity 94.6%; specificity 96.2%) in distinguishing normal from abnormal scintigraphic images. On a parathyroid gland basis and in achieving identical positioning of the findings with experts, the model correctly identified and localized 324/391 glands (82.9%) and yielded 74 false focal results (false positive rate 18.6%). These numbers correspond to an 81.4% positive and a 60.4% negative predictive value on a parathyroid gland level. Conclusion: Deep Learning in parathyroid scintigraphy can potentially assist medical experts in identifying abnormal findings. Future research should be directed mainly towards false positive reduction methods.
