Abstract
Systemic lupus erythematosus (SLE) is a chronic auto-immune disease with a broad spectrum of clinical presentations, including heterogeneous and uncommon neuropsychiatric (NP) syndromes. Accurate diagnosis of neuropsychiatric SLE (NPSLE) is challenging due to a lack of clinically useful biomarkers. This study applies self-supervised contrastive learning on T1-weighted images from a well-defined cohort of 163 SLE patients to distinguish between NPSLE and non-NPSLE cases. Using a 3D convolutional encoder and a linear projector trained with the NT-xent loss, the model achieved 79-83% accuracy. Grad-CAM analysis identified common features in the NPSLE cohort involving the ventricles and periventricular white matter, suggesting a combination of tissue loss and local atrophy.
Introduction: The Clinical-Radiological Paradox
Systemic Lupus Erythematosus (SLE) is a complex autoimmune disease that can affect nearly any organ, including the central nervous system—a condition known as Neuropsychiatric SLE (NPSLE). Diagnosing NPSLE remains a significant clinical challenge. Patients often present with non-specific symptoms such as headaches, seizures, or anxiety. While MRI scans frequently reveal brain abnormalities, these do not always correlate directly with clinical symptoms, a phenomenon termed the "clinical-radiological paradox." Furthermore, distinguishing true NPSLE from "non-NPSLE" (where symptoms arise from other factors like infections or medication) is difficult due to the lack of robust biomarkers.
Traditional supervised deep learning models typically require thousands of labeled images, which are unavailable for rare conditions like NPSLE. This research explores whether self-supervised contrastive learning can learn meaningful brain features from a limited dataset without requiring massive amounts of human-labeled data.
Methodology
The researchers implemented a self-supervised contrastive learning pipeline based on the SimCLR framework, adapted for 3D medical imaging.
Model Architecture and Training
The pipeline consists of a convolutional base encoder and a multi-layer perceptron (MLP) projector. The encoder comprises three layers, each featuring 3D convolution, batch normalization, and max pooling. During the training phase, the model learns by comparing different "views" of the same MRI scan. These views are generated through random 3D transformations, such as cropping, flipping, and rotation.
The network is trained using the Normalized Temperature-scaled Cross Entropy (NT-xent) loss function, which maximizes the similarity between different augmented versions of the same patient's image in a latent space while distinguishing them from others.
Dataset and Validation
The study utilized a retrospective dataset of 163 SLE patients from the Leiden University Medical Center (LUMC). The cohort included 34 confirmed NPSLE cases and 129 non-NPSLE cases. All 3T T1-weighted MRI scans were registered to the MNI152 standard template. To ensure robust results with a small sample size, the team employed Monte Carlo cross-validation with six repeated random sub-samplings.
Results and Key Insights
The self-supervised model demonstrated strong performance in classifying NPSLE vs. non-NPSLE:
- Accuracy: Across the six trials, the model achieved a classification accuracy between 79% and 83%.
- Recall and Precision: The model showed high recall (~83%), effectively identifying true NPSLE cases. However, precision was lower (~51%), indicating a higher rate of false positives in the non-NPSLE group.
- Feature Commonality: A critical finding from the Grad-CAM (Gradient-weighted Class Activation Mapping) analysis was that the model identified common structural features across the NPSLE group. In contrast, no such commonality was found for the non-NPSLE group, suggesting that NPSLE has a distinct structural "signature" while non-NPSLE symptoms stem from a heterogeneous mix of causes.
Discussion and Conclusion
The model consistently highlighted regions including the lateral and third ventricles, periventricular white matter, and basal cisterns. These findings suggest that the "signature" of NPSLE involves a combination of tissue loss, local atrophy, and periventricular white matter lesions, which appear hypointense on T1-weighted images.
While promising, the study has limitations. The "gold standard" labels were based on expert consensus, which remains somewhat subjective. Additionally, the use of a single modality (T1-weighted MRI) may limit precision. Future work could incorporate other modalities like FLAIR or Diffusion Tensor Imaging (DTI) to better capture white matter hyperintensities and improve diagnostic accuracy. Overall, this study demonstrates that self-supervised contrastive learning is a powerful tool for capturing diagnostic features in rare diseases with limited datasets.
Summary updated