Automatic choroid plexus assessment in SLE: a deep learning-enabled study

This study developed a deep learning model for automated choroid plexus (ChP) segmentation and examined its relationship with systemic inflammation and processing speed and attention deficits (PSAD) in SLE patients without major neuropsychiatric syndromes.

In this multicenter retrospective study, 137 SLE patients without major neuropsychiatric syndromes and 159 healthy controls (HCs) were enrolled. The Swin-UNETR model was trained for ChP segmentation on 3D T1-weighted MR images. SLE patients were classified as with processing speed and attention deficits (SLE-PSAD, n = 43) or intact processing speed and attention (SLE-IPSA, n = 94). Clinical, laboratory, and imaging data were compared among groups. Correlation, mediation, and LASSO regression analyses were performed.

Swin-UNETR achieved high segmentation accuracy (median DSC = 0.89 internal, 0.82 external, P < 0.001). ChP volume was significantly greater in SLE-PSAD patients than in SLE-IPSA patients and HCs (P < 0.001) and positively correlated with systemic inflammation index (SII, r = 0.34, P < 0.001). Bayesian logistic regression identified increased ChP volume (aOR = 2.57), elevated SII (aOR = 2.47), and low complement component 3 (C3, aOR = 0.47) as independent PSAD risk factors. ChP volume mediated 39.2% of the SII-PSAD relationship (P < 0.001). LASSO regression identified a minimal three-biomarker model (ChP volume + C3 + SII) with excellent discriminative ability (AUC = 0.79 training, 0.77 validation).

Enlarged ChP volume is a biomarker and mediator of PSAD in SLE patients. The Swin-UNETR model enables accurate ChP quantification, and the three-biomarker panel provides a practical tool for SLE-PSAD risk stratification.