TL;DR
This paper introduces GAMER-MRI, a deep learning approach that uses attention mechanisms to identify the most sensitive diffusion MRI measures for detecting focal damage in multiple sclerosis (MS) lesions. By combining these key measures, the study demonstrates a stronger correlation with clinical disability (EDSS) and neuroaxonal damage (sNfL) than individual measures or conventional MRI metrics.
What Problem Does This Paper Address?
Multiple Sclerosis (MS) is a complex neurological disease characterized by focal brain lesions. While conventional MRI (cMRI) is crucial for diagnosis and monitoring, it struggles to capture the subtle microstructural changes within the brain tissue and the heterogeneity of MS lesions. Multishell diffusion-weighted imaging (mDWI) offers a more detailed view by probing water diffusion within different cellular compartments (like axons and free water). However, a vast array of diffusion models and derived measures exist, making it challenging to determine which ones are most informative for MS pathology and how to best combine them for clinical relevance. This paper aims to address this by developing a method to systematically select and validate the most discriminative diffusion MRI measures for MS lesion characterization.
Methods
The core of this work is the GAMER-MRI (Gated-Attention Mechanism for Enhanced Ranking of MRI) framework, a modified attention-based Convolutional Neural Network (CNN). The network takes multiple diffusion MRI contrasts as input and uses an attention mechanism to learn the relative importance of each measure for classifying MS lesions versus perilesional tissue.
Key methodological aspects include:
- Feature Extraction: Each diffusion contrast is processed through convolutional blocks to extract spatial features.
- Gated Attention Mechanism: This mechanism learns to assign attention weights to different diffusion measures, highlighting those most relevant for the classification task. A modification was introduced to enhance the differentiation of attention weights for highly correlated diffusion measures.
- Selection Procedure: Instead of relying solely on attention weights, a novel iterative selection process was developed to identify a subset of measures whose combined attention weights exceeded a threshold (0.5). This helps mitigate the instability of attention weights when dealing with highly correlated inputs.
- Clinico-Biological Validation: The selected diffusion measures, and their combinations, were correlated with clinical disability (Expanded Disability Status Scale - EDSS) and a biomarker of neuroaxonal damage (serum neurofilament light chain - sNfL) using Spearman's correlation.
Datasets
The study utilized data from 123 MS patients. This dataset included:
- Conventional MRI sequences (FLAIR, MP2RAGE).
- Multishell Diffusion-Weighted Imaging (mDWI).
The dataset was split into:
- A training/validation set (84 patients) used for 5-fold cross-validation.
- A pure test set (39 patients) for independent evaluation.
White matter lesions were segmented using FLAIR and MP2RAGE, and perilesional tissue was defined within a 3-voxel radius. Patches of 5x5x5 voxels were sampled from lesions and perilesional areas.
Key Insights and Ideas
- Attention as a Feature Selector: GAMER-MRI effectively uses attention weights not just for classification but as a powerful tool to rank and select the most informative diffusion MRI measures for a specific pathology.
- Handling Correlated Data: The modification to the attention mechanism and the iterative selection process are crucial for dealing with the inherent high correlation between different diffusion MRI measures, preventing instability in feature selection.
- Beyond Individual Measures: The study strongly suggests that combining specific diffusion measures can provide a more comprehensive picture of tissue damage than any single measure alone, leading to stronger correlations with clinical outcomes.
- Axonal Integrity is Key: The selected measures primarily focused on the intra-axonal compartment, indicating that axonal damage is a dominant microstructural feature in MS lesions that correlates with clinical severity.
Key Results
- Lesion Classification Performance: GAMER-MRI achieved good performance in classifying MS lesions and perilesional tissue, with AUCs around 90-91% and balanced accuracies around 81-83% on both validation and test datasets.
- Most Discriminating Measures: The study identified four key diffusion measures with high attention weights:
- Neurite Density Index (NDI) from NODDI (0.12)
- Intra-axonal compartment from Microstructural Bayesian approach (Intra-MB) (0.12)
- Isotropic compartment from Microstructural Bayesian approach (Iso-MB) (0.15)
- Intra-axonal compartment from Spherical Mean Technique NODDI (Intra-SMT) (0.13)
- Stronger Clinical Correlations: Combinations of these selected measures, particularly the combination of Intra-MB, Iso-MB, and NDI, showed significantly stronger correlations with EDSS (e.g., ρ = -0.45, p = 0.004) and sNfL (e.g., ρ = -0.51, p = 0.004) compared to individual measures or conventional lesion load metrics.
Limitations and Future Work
- Normalization Impact: The authors acknowledge that the subject-wise normalization of diffusion maps makes it challenging to interpret the pathological meaning of the combined metrics.
- Cross-Validation Strategy: Dividing cross-validation folds by patient rather than by patch led to varying distributions of lesion and perilesional patches, potentially causing fluctuations in validation results. However, this strategy prevented information leakage.
- Future Directions: The paper suggests future work on directly finding the best combinations without relying on statistical tests and addressing the multiple comparison problem. Further development of combined diffusion-based microstructural measures is encouraged for improved interpretability and clinical application.
Summary updated