Detection and Localisation of Abnormal Parathyroid Glands: An Explainable Deep Learning Approach

Parathyroid scintigraphy with 99mTc-sestamibi (MIBI) is an established technique for localising abnormal parathyroid glands (PGs). However, the identification and localisation of PGs require much attention from medical experts are time-consuming. Artificial intelligence methods can offer assisting solution. This retrospective study enrolled 632 patients who underwent double-phase thyroid subtraction techniques. The proposes a three-path approach, employing state-of-the-art convolutional neural network called VGG19. Images input to model involved set three scintigraphic images in each case: MIBI early phase, late 99mTcO4 scan. A expert’s diagnosis provided ground truth positive/negative results. Moreover, visualised suggested areas interest produced by Grad-CAM algorithm examined evaluate PG-level agreement between experts. Medical identified 545 452 patients. On patient basis, deep learning (DL) attained accuracy 94.8% (sensitivity 93.8%; specificity 97.2%) distinguishing normal images. PG basis achieving identical positioning findings experts, correctly localised 453/545 (83.1%) yielded 101 false focal results (false positive rate 18.23%). Concerning surgical findings, sensitivity was 89.68% on 77.6% while that reached 84.5% 67.6%, respectively. Deep potentially assist identifying findings.