Abstract
This study proposes a ViT network for classification and localization to aid in detecting abnormal PGs in parathyroid scintigraphy. The network analyzes MIBI-early, MIBI-late, and Tc04 images to classify patients as normal or abnormal. The Grad-CAM++ algorithm is then used to visualize the critical image regions where the network made its predictions, enabling abnormal PG localization. This method is also utilized for evaluating the ViT model’s explanations. The study’s ViT network achieved high accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score in identifying abnormal cases and cases with no abnormal findings in 648 participants with pHPT who underwent parathyroid scintigraphy. Of the 450 scans with at least one abnormal finding, the model accurately identified most cases. However, when evaluating at the PG level using Grad-CAM++, the model had a sensitivity of 0.9385 and a specificity of 0.7333. The positive predictive value was 0.8808, and the negative predictive value was 0.8502. The model had a higher rate of false positives and false negatives at the PG level.
