Magnetic resonance imaging-based radiomics for early detection of primary Sjögren's disease using ZOOMit-DWI: a non-invasive diagnostic approach.

OBJECTIVES

This study established a non-invasive, artificial intelligence-assisted diagnostic model aimed at detecting early damage to the parotid gland in patients diagnosed with primary Sjögren's disease (pSjD) using MRI-based radiomics analysis, representing one of the first efforts to integrate machine learning with ZOOMit-DWI MRI sequences for pSjD diagnosis.

METHODS

We retrospectively enrolled 15 early pSjD patients and 40 healthy controls. Radiomics features were extracted from magnetic resonance imaging-derived parotid gland volumes of interest (VOIs). Five machine learning classifiers were combined with four feature selection methods to identify the optimal model. Model performance was evaluated using ROC analysis, with interpretability enhanced by SHAP analysis.

RESULTS

The models demonstrated good diagnostic performance, with AUC values ranging from 0.91 to 0.96 (training) and 0.73 to 0.92 (test). The random forest-based early pSjD predictor (RFEP), utilising mutual information (MI) for feature selection, achieved the highest performance, with an AUC of 0.96 (training) and 0.92 (test). Notably, RFEP correctly identified 13 out of 15 pSjD cases (9/10 and 4/5, respectively), demonstrating its potential for early pSjD screening.

CONCLUSIONS

The RFEP model, leveraging parotid ZOOMit DWI radiomics, shows high diagnostic accuracy and interpretability, offering a non-invasive tool for early parotid gland injury in pSjD. This study pioneers the application of AI-driven radiomics in pSjD diagnosis, paving the way for future research and clinical translation.