This retrospective study analyzed 180 parotid sialo-CBCT scans obtained between 2018 and 2024 from patients presenting with symptoms of xerostomia. Maxillofacial experts classified all scans according to their arborization pattern into two groups: normal-appearing glands (90 scans) and ductopenic glands (90 scans). The technical implementation involved a sophisticated data preprocessing pipeline, utilizing a Convolutional Neural Network (CNN) backbone for volume of interest (VOI) definition and gland classification. To enhance the robustness of the AI model, extensive data augmentation was performed through the generation of multidirectional maximum intensity projection (MIP) images, creating 343 unique projections for each scan. This augmentation resulted in a comprehensive dataset of 61,740 images. The dataset was then randomally divided into three subsets: a training set comprising 120 scans (41,160 MIP images), a validation set of 20 scans (6,860 images), and a testing set containing 40 scans (13,720 images). This distribution ensured proper model training while maintaining adequate independent testing data.