Explainable Deep Learning for localising abnormal Parathyroid Glands in parathyroid scintigraphy

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 and are time-consuming. Artificial Intelligence methods can offer an assisting solution and can be embedded at the edge of Medical Decision Support Systems and hospital computer stations. This retrospective study enrolled 632 patients, who underwent parathyroid scintigraphy with double-phase and thyroid subtraction techniques. The study proposes a three-path approach, employing the state-of-the-art Convolutional Neural Network called VGG19. 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 visualised suggested areas of interest produced by the Grad-CAM algorithm are examined to evaluate the PG-level agreement between the model and the experts. Medical experts identified 545 abnormal glands in 452 patients. On a patient basis, the Deep Learning (DL) model attained an accuracy of 94.8% (sensitivity 93.8%; specificity 97.2%) in distinguishing normal from abnormal scintigraphic images. On a PG basis and in achieving identical positioning of the findings with the experts, the model correctly identified and localised 453/545 glands (83.1%) and yielded 101 false focal results (false positive rate 18.23%). Concerning surgical findings, the expert's sensitivity was 89.68% on patients and 77.6% on a PG basis, while that of the model reached 84.5% and 67.6%, respectively. Deep Learning in parathyroid scintigraphy can potentially assist medical experts in identifying abnormal findings. Despite the time- consuming training procedure and the high computational demands, these limitations apply only to the training procedure. Once the model is trained, it can deliver its predictions swiftly, due to its lightweight nature. This fact enables the utilisation of such models in a plethora of systems, even wearable devices