Portable Ultra-Low Field MRI Deep-Learning Algorithms for White Matter Lesion Segmentation Improve Accuracy and Reflect Clinical Disability in Multiple Sclerosis

Technical Summary: Portable Ultra-Low Field MRI Deep-Learning Algorithms for White Matter Lesion Segmentation

Problem Statement
White matter lesion (WML) identification via magnetic resonance imaging (MRI) is fundamental for diagnosing and monitoring multiple sclerosis (MS). While high-field (HF) MRI (>1.5T) is the clinical standard, it is costly, immobile, and inaccessible in low-resource settings or for patients with specific contraindications. Portable ultra-low-field (pULF) MRI at 64 millitesla (mT) offers a cost-effective, mobile alternative capable of visualizing lesions larger than 4 mm. However, pULF images suffer from lower signal-to-noise ratios, heterogeneous magnetic fields, and artifacts that complicate lesion detection. Existing segmentation tools designed for HF MRI (e.g., MIMoSA) or general deep-learning models (e.g., WMH-SynthSeg) have not been rigorously validated against pULF-specific ground truths, often relying on HF reference masks that may not account for pULF's inherent detectability limits. There is a critical need for automated segmentation algorithms specifically trained and validated on pULF data to provide accurate, quantitative WML burden estimates that correlate with clinical disability.

Methodology
The study utilized a cohort of 104 adults with known or suspected MS who underwent same-day paired pULF (64mT) and HF (3T) MRI scans. The dataset was split into a training set (n=20) and a testing set (n=84).

1. Reference Standard Creation:

   • HF Reference: Generated using the C-DEF algorithm and manually refined by a neuroimaging researcher.
   • pULF Reference: Manually annotated by two raters on native pULF T2-FLAIR images. Crucially, these annotations were cross-referenced with rigidly registered HF T2-FLAIR images to confirm true-positive lesions, establishing a pULF-specific ground truth that accounts for pULF's resolution limits.

2. Automated Segmentation Models:
   Six automated methods were applied to the pULF testing cohort:

   • MIMoSA: A voxel-based regression method trained on HF data, applied without retraining.
   • WMH-SynthSeg: A convolutional neural network (CNN) trained on synthesized data, applied without retraining.
   • nnU-Net: Two models trained on the pULF training set using 5-fold cross-validation. One used only T2-FLAIR (nnU-Net-FL), and the other used T2-FLAIR plus registered T1w (nnU-Net-FL/T1).
   • PLAn (Pseudo-Label Assisted nnU-Net): A transfer learning framework. Models were pre-trained on HF reference masks (downsampled) and then fine-tuned on the pULF training set. Similar to nnU-Net, two variants were created: PLAn-FL (T2-FLAIR only) and PLAn-FL/T1 (T2-FLAIR + T1w).

3. Evaluation Metrics:
   Performance was assessed against the pULF reference masks using:

   • Volumetric Agreement: Spearman correlations and Bland-Altman analysis.
   • Spatial Overlap: Dice Similarity Coefficient (DSC), lesion-wise precision, and sensitivity.
   • Clinical Correlation: Linear regression models testing associations between automated WML volumes and clinical disability scores (Expanded Disability Status Scale [EDSS] and Scripps Neurologic Rating Scale [SNRS]), with and without age adjustment.

Key Contributions and Results

• pULF Detectability Limits: The study established that pULF reference segmentations capture approximately 78% of HF-confirmed lesions larger than 50 µL (approx. 4.5 mm diameter) but only 7% of lesions smaller than 50 µL. pULF volumes were slightly higher than HF volumes due to partial volume effects, yet they remained strongly correlated (ρ = 0.96).
• Superiority of Transfer Learning: The PLAn models (both FL and FL/T1 variants) demonstrated the highest performance.
  • Dice Similarity Coefficient (DSC): PLAn-FL achieved a mean DSC of 0.50 ± 0.24, significantly outperforming MIMoSA (0.24), WMH-SynthSeg (0.30), and nnU-Net models (0.41).
  • Balance of Metrics: PLAn models achieved a balanced trade-off between sensitivity (median ~0.60) and precision (median ~0.83–0.85), with low false-positive lesion counts. In contrast, nnU-Net models were highly conservative (high precision, low sensitivity), while MIMoSA and WMH-SynthSeg suffered from extensive oversegmentation (high false positives).
• Impact of Multi-Contrast Input: The addition of T1w images to the input (FL/T1 variants) did not significantly improve performance over T2-FLAIR alone for either nnU-Net or PLAn models. This suggests that T2-FLAIR alone provides sufficient information for pULF lesion segmentation, potentially reducing acquisition time.
• Clinical Relevance: Automated WML volumes derived from PLAn, nnU-Net, and WMH-SynthSeg were significantly associated with EDSS and SNRS scores (p < 0.05), mirroring the associations found with HF reference volumes. MIMoSA volumes showed no significant association with clinical disability.

Significance and Claims
The paper claims that deep-learning algorithms, specifically those utilizing transfer learning (PLAn) and trained on pULF-specific reference data, can accurately segment white matter lesions on portable 64mT MRI. The study demonstrates that:

1. Accuracy: These models provide quantitative WML volume estimates that align with pULF-specific ground truths and outperform existing HF-trained or generic models.
2. Clinical Utility: The automated volume estimates reflect clinical disability, validating their use as outcome measures.
3. Feasibility: The findings support the use of pULF-MRI in clinical trials and settings where HF-MRI is logistically constrained, particularly for patients with mobility issues or advanced disability. The ability to use T2-FLAIR alone further streamlines the workflow.

The authors conclude that while pULF misses smaller lesions, the accurate quantification of larger lesions by these deep-learning tools is sufficient to capture clinically relevant disease burden, offering a pathway to expand MRI access for MS research and care.