The Accuracy of Diagnosing Salivary Gland Diseases by Artificial Intelligence: Systematic Review

Abstract

1.1 Purpose Artificial intelligence (AI) is increasingly applied in the diagnosis of salivary gland diseases, particularly Sjögren’s syndrome (SS) and salivary gland tumours (SGTs). This review aimed to evaluate the diagnostic performance of AI models in these two disease categories and identify converging patterns, limitations, and research gaps. 1.2 Method A systematic literature search was conducted in PubMed, Scopus, and Google Scholar over the past two decades (2005-2025) using predefined inclusion and exclusion criteria. Data extraction captured study design, input modality, AI model type, performance metrics (sensitivity, specificity, accuracy, AUC). Quality analysis was performed using JBI tool. Results were stratified by disease group (SS vs SGTs) and AI model type (Machine learning vs Deep learning). 1.3 Results A total of 19 studies were included from the 221 initially retrieved. Most of the included studies were assessed as moderate risk of bias, with only three low-risk and one high-risk. In SS studies , ML models showed excellent performance when applied to structured data. Logistic Regression emerged as the best-performing architecture, achieving accuracies up to 94% with AUC values ranging from 0.88 to 0.96. DL models on histopathology ranged from weak performance in baseline Residual CNNs (ResNet) (50% accuracy) to excellent outcomes with custom architectures such as CTG-PAM (100% across sensitivity, specificity, and accuracy). In SGTs, ML models on imaging inputs showed moderate ability, with Logistic Regression achieving 78–84% accuracy (AUC up to 0.91) and ultrasound reporting lower sensitivity but good specificity. DL approaches outperformed ML, particularly hybrid CNN–Transformers on MRI (85% accuracy, AUC 0.96; Liu et al., 2023) and Vision Transformers on ultrasound (87% accuracy, AUC 0.93; He et al., 2025). CNNs were more variable: Inception showed consistent results (73–85% accuracy, AUC up to 0.91), while ResNet and Densely Connected CNN (DenseNet) performance fluctuated widely even within the same input modality. 1.4 Conclusion AI demonstrates high potential in salivary gland disease diagnosis, with structured data input and custom-made models and advanced DL architectures yielding the most promising results. However, heterogeneity in input modalities and model design limits comparability, underscoring the need for standardised, multicentre validation.