MRI-Based Classification of Neuropsychiatric Systemic Lupus Erythematosus Patients With Self-Supervised Contrastive Learning

Francesca Inglese, Min‐Seon Kim, Gerda M. Steup‐Beekman, T. Huizinga, Mark A. van Buchem, Jeroen de Bresser, Dae‐Shik Kim, Itamar Ronen

TL;DR

Researchers have developed a self-supervised contrastive learning pipeline to classify Neuropsychiatric Systemic Lupus Erythematosus (NPSLE) using 3D MRI data. By training on a small but well-defined cohort, the model achieved up to 83% accuracy, identifying specific structural brain patterns—like ventricular enlargement and periventricular white matter changes—that distinguish NPSLE from other causes of neuropsychiatric symptoms.

What Problem Does This Paper Address?

Systemic Lupus Erythematosus (SLE) is a complex autoimmune disease that can attack almost any organ, including the brain. When it affects the central nervous system, it’s called Neuropsychiatric SLE (NPSLE). However, diagnosing it is a nightmare for clinicians.

Patients often present with "non-specific" symptoms like headaches, seizures, or anxiety. While MRI scans frequently show brain abnormalities, these don't always correlate with the symptoms—a phenomenon known as the clinical-radiological paradox. Furthermore, many SLE patients have neuropsychiatric symptoms caused by other factors (like medication side effects or infections), making it incredibly difficult to distinguish true NPSLE from "non-NPSLE."

Traditional supervised deep learning requires thousands of labeled images, which is impossible for rare diseases like NPSLE. This paper explores whether self-supervised contrastive learning can bridge this gap by learning meaningful brain features from a limited dataset without needing massive amounts of human-labeled data.

Methods

The authors adapted the SimCLR framework (a popular contrastive learning approach) for 3D medical imaging.

  1. The Architecture: The pipeline uses a convolutional base encoder (three layers with batch normalization and max pooling) and a two-layer multi-layer perceptron (MLP) projector.
  2. Contrastive Learning: Instead of learning from labels (e.g., "this is NPSLE"), the model learns by comparing images. It takes an MRI, creates two different "views" of it using random 3D transformations (cropping, flipping, and rotation), and trains the network to recognize that these two views represent the same patient while distinguishing them from other patients.
  3. Loss Function: They used the Normalized Temperature-scaled Cross Entropy (NT-xent) loss, which maximizes the similarity between different versions of the same image in a latent space.
  4. Fine-tuning: After the self-supervised "pre-training," they replaced the projector with a single linear layer and fine-tuned it on a small labeled set to perform the final classification.
  5. Interpretability: To understand why the model made its decisions, they used Grad-CAM (Gradient-weighted Class Activation Mapping) to generate heatmaps of the brain regions the model found most significant.

Datasets

The study used a retrospective dataset from the Leiden University Medical Center (LUMC), a national referral center in the Netherlands.

  • Total Patients: 163 SLE patients with neuropsychiatric symptoms.
  • Breakdown: 34 patients were confirmed NPSLE; the rest were non-NPSLE.
  • Imaging: 3T T1-weighted MRI scans, all registered to a standard brain template (MNI152).
  • Validation: They used a Monte Carlo cross-validation approach with 6 repeated trials to ensure the results were robust despite the small sample size.

Key Insights and Ideas

  • Overcoming Data Scarcity: The most significant "aha" moment is that self-supervised learning actually works for rare diseases. By letting the model "explore" the data structure through augmentations before asking it to classify, it becomes much more efficient than standard supervised learning.
  • Commonality vs. Heterogeneity: The Grad-CAM results revealed a fascinating finding: the model found common structural features across the NPSLE group, but found no such communality in the non-NPSLE group. This suggests that while NPSLE has a distinct (though subtle) structural "signature," the non-NPSLE group is a grab-bag of various unrelated conditions.
  • The "Signature" of NPSLE: The model consistently flagged the lateral and third ventricles, periventricular white matter, and basal cisterns. This points toward tissue loss and local atrophy as the primary drivers of the classification, rather than just isolated lesions.

Key Results

  • Classification Accuracy: Across 6 trials, the model correctly classified patients with 79% to 83% accuracy.
  • Recall vs. Precision: The model showed high recall (~83%), meaning it was very good at catching true NPSLE cases. However, precision was lower (~51%), indicating a higher rate of "false positives" (non-NPSLE patients being flagged as NPSLE).
  • Robustness: The results were largely independent of the specific data augmentation strategy used (crop vs. flip vs. rotation), suggesting the learned features were stable.

Limitations and Future Work

  • The "Gold Standard" Problem: There is no perfect diagnostic test for NPSLE; the "labels" used for training were based on expert consensus, which is the best available but still subjective.
  • Single Modality: The study only used T1-weighted images. The authors suggest that adding other modalities, like FLAIR (which is better at showing white matter hyperintensities) or Diffusion Tensor Imaging (DTI), could significantly improve precision.
  • Sample Size: While 163 is a strong cohort for this specific disease, it is still small for deep learning. Future work would benefit from multi-center studies to see if these "signatures" hold up across different scanners and populations.
  • Interpretability: While Grad-CAM provides a "where," it doesn't provide a "what." Further research is needed to correlate these highlighted regions with specific pathological changes like gliosis or myelin loss.

Summary updated

Abstract

The self-supervised contrastive learning model was effective in capturing common brain MRI features from a limited but well-defined cohort of SLE patients with NP symptoms. The interpretation of the Grad-CAM results is not straightforward, but indicates involvement of the lateral and third ventricles, periventricular white matter and basal cisterns. We believe that the common features found in the NPSLE population in this study indicate a combination of tissue loss, local atrophy and to some extent that of periventricular white matter lesions, which are commonly found in NPSLE patients and appear hypointense on T<sub>1</sub>-weighted images.
BibTeX
@inproceedings{inglese2022mri,
  author = {Francesca Inglese and Min‐Seon Kim and Gerda M. Steup‐Beekman and T. Huizinga and Mark A. van Buchem and Jeroen de Bresser and Dae‐Shik Kim and Itamar Ronen},
  title = {{MRI-Based Classification of Neuropsychiatric Systemic Lupus Erythematosus Patients With Self-Supervised Contrastive Learning}},
  booktitle = {Frontiers in Neuroscience},
  year = {2022},
  doi = {10.3389/fnins.2022.695888},
}
Figure