Abstract
Predicting disease progression in Multiple Sclerosis (MS) is a critical clinical challenge due to the sparse and heterogeneous nature of brain lesions. Traditional deep learning approaches, such as 3D Convolutional Neural Networks (CNNs), often struggle to capture global disease patterns from localized pathology. This paper presents a novel two-stage framework that represents MS lesions as nodes in a graph and utilizes a self-pruning Graph Neural Network (GNN) to predict future Inflammatory Disease Activity (IDA). By focusing on lesion-specific features and their spatial relationships, the model achieves superior predictive performance and provides a mechanism for clinical explainability through its self-pruning module.
Introduction
Multiple Sclerosis (MS) is a chronic inflammatory disease of the central nervous system characterized by the formation of lesions throughout the brain. A primary goal in clinical management is predicting Inflammatory Disease Activity (IDA)—defined as the appearance of new or enlarging lesions—to guide treatment decisions. However, predicting IDA from baseline MRI scans is a "needle in a haystack" problem. Lesions are sparse, often occupying less than 1% of the total brain volume, and vary significantly in size, shape, and location across patients.
Standard machine learning approaches, particularly 3D CNNs, process the entire brain volume. This often results in a low signal-to-noise ratio as the model struggles to distinguish critical pathological features from the vast amount of healthy tissue. Furthermore, these "black-box" models offer little insight into which specific lesions drive a prediction, limiting their utility in clinical settings.
Methodology
To address these challenges, we propose a two-stage pipeline that shifts the focus from voxel-level analysis to a lesion-centric graph representation.
Stage 1: Detection and Feature Extraction
The first stage involves identifying and characterizing individual lesions. We employ a state-of-the-art nn-Unet for the initial segmentation of lesions from FLAIR and T1-weighted MR images. For each detected lesion, a 3D patch is extracted and processed through a Transformer-based masked autoencoder. This self-supervised learning component generates a 768-dimensional feature vector for each lesion, capturing intricate details of its texture and morphology that might be missed by broader filters.
Stage 2: Graph Construction and Self-Pruning
In the second stage, the patient’s pathology is modeled as a graph where each lesion is a node.
- Graph Construction: Edges are established between lesions using a k-Nearest Neighbor (k-NN) approach based on their spatial coordinates. This allows the model to incorporate the spatial context and proximity of lesions, which are known to be relevant in MS progression.
- Graph Convolutional Network (GCN): A GCN aggregates information across the graph, enabling the model to learn how the distribution and characteristics of multiple lesions interact to signal future activity.
- Self-Pruning Module (SPM): The core innovation is the SPM, which assigns an importance score to each node. The module automatically prunes the least relevant 50% of lesions, retaining only those most indicative of disease activity. This not only refines the predictive signal but also serves as an inherent attention mechanism.
- Classification: The features from the remaining nodes are aggregated to predict IDA at 1-year and 2-year intervals.
Experimental Results
The framework was evaluated on a cohort of 430 patients with relapsing-remitting MS. Ground truth labels were established by three neuro-radiologists using longitudinal subtraction imaging to identify new or enlarged lesions over time.
Predictive Performance
The GNN-based approach consistently outperformed traditional 3D CNN architectures:
- One-Year Prediction: The proposed model achieved an AUC of 0.67, significantly higher than the multi-resolution CNN (0.61) and 3D ResNet (0.59).
- Two-Year Prediction: The model maintained an AUC of 0.66, compared to 0.60 for the CNN baseline.
The Impact of Self-Pruning
Ablation studies confirmed that the Self-Pruning Module improved performance across various graph architectures, including GAT, EdgeConv, and GraphSage. By discarding non-predictive "noise" lesions, the SPM helped the model focus on the most relevant pathology, enhancing both accuracy and the potential for clinical interpretation.
Discussion and Insights
This research highlights several key advantages of graph-based modeling in neuroimaging:
- Handling Variability: Unlike CNNs, which require fixed-size inputs, GNNs naturally accommodate a varying number of lesions per patient, making them highly flexible for clinical data.
- Spatial Biomarkers: The superior performance of the GNN over models that ignore spatial distance (such as standard MLPs) suggests that the relative positioning of lesions is a vital biomarker for MS activity.
- Explainability: The pruning mechanism allows clinicians to visualize which lesions the AI identifies as high-risk, bridging the gap between automated prediction and clinical intuition.
Conclusion and Future Work
While the self-pruning GNN represents a significant step forward, predicting MS activity remains a complex task, as evidenced by the performance ceiling in the high 0.60s. Future research will focus on integrating additional clinical variables—such as patient age, gender, and disability scores (EDSS)—into the node features to further enhance predictive power. Additionally, we aim to explore the model's utility in monitoring long-term treatment response, potentially providing a more robust tool for personalized MS care.
Summary updated