Source-linked AI summary
FORESEE-MS: Toward Prospective Spatial Forecasting of Future Multiple Sclerosis Lesions from Baseline FLAIR MRI
Maryam Akhavan Aghdam, Sudip Mittal, Noorbakhsh Amiri Golilarz, Shahram Rahimi
TL;DR
Prior MS image analysis has largely focused on detecting lesions already present, leaving open whether baseline FLAIR can forecast future lesion locations beyond population-level spatial patterns. FORESEE-MS evaluates a fold-specific atlas and five deep learning models using pooled discrimination and within-patient localization. The atlas led pooled AUROC and AUPRC, while two models had higher mean within-patient recall, but the small cohort leaves the value of patient-specific forecasting uncertain.
Problem
Prior MS image analysis has largely addressed retrospective lesion segmentation, leaving open whether baseline FLAIR can identify regions at risk for future lesions beyond population-level spatial patterns.
Source: §1 · p. 1Method
FORESEE-MS compares a fold-specific population-prior atlas with five deep learning models using pooled AUROC/AUPRC and within-patient Recall@Top-K% on baseline imaging.
Source: §1 · p. 1, §2 · pp. 1–2Results
The atlas achieved the highest pooled AUROC and AUPRC, while the U-Net and Attn+ResBlock+VAE had higher mean within-patient voxel recall at all evaluated thresholds, with most paired confidence intervals including zero.
Source: §4 · p. 4, §3.4 · p. 4Takeaways & Limitations
Baseline FLAIR may contain patient-specific predictive information, but whether it improves lesion forecasting beyond the population-prior atlas remains uncertain.
Source: §4 · p. 4The small cohort of 39 subjects, including 28 with new lesions, limits statistical precision and contributes to wide confidence intervals.
Source: §3.4 · p. 4
Abstract
from arXiv · showhide
Traditional multiple sclerosis (MS) image analysis has primarily focused on retrospectively detecting and segmenting lesions. This study investigates whether baseline FLAIR MRI can identify regions at risk for future lesion formation beyond population-level spatial patterns. We introduce FORESEE-MS, an evaluation protocol that uses a fold-specific population-prior atlas as a baseline and assesses predictions using pooled voxel-level discrimination (AUROC and AUPRC) and within-patient lesion localization (Recall@Top-K%), with patient-level bootstrap confidence intervals. We evaluated five deep learning models (3D U-Net, Attn+ResBlock+VAE, ViT, and two SwinUNETR variants) using 5-fold cross-validation on 39 subjects from the MSSEG-2 dataset. The atlas achieved the highest pooled voxel-level AUROC and AUPRC, with no deep learning model reliably outperforming it on either metric. However, the 3D U-Net and Attn+ResBlock+VAE achieved higher mean voxel-level within-patient Recall@Top-K% than the atlas at all evaluated thresholds, although most paired confidence intervals for the differences included zero. These findings suggest that baseline FLAIR may contain patient-specific predictive information, but in this small cohort, it remains uncertain whether this information improves lesion forecasting beyond a population-prior atlas.
1. INTRODUCTION
Prior MS image analysis has largely focused on segmenting existing lesions, leaving open whether baseline FLAIR can forecast where new lesions will emerge. FORESEE-MS addresses this question by evaluating whole-brain voxel-wise risk maps against population-level spatial patterns.
1. INTRODUCTION
- Prior MS image analysis mainly segmented lesions already present, rather than forecasting new lesion locations from a baseline scan.The paper also notes prior work predicting patient- or region-level future activity and a treatment-aware voxel-level model.Source: §1 · p. 1
- Earlier evidence reports elevated T2 signal and reduced magnetization transfer ratio in normal-appearing white matter before it becomes lesional; FLAIR is closely related to T2-weighted contrast.Source: §1 · p. 1
- FORESEE-MS evaluates dense whole-brain voxel-wise risk maps from baseline FLAIR and derived spatial information without treatment or clinical data.The protocol compares a fold-specific population-prior atlas with five deep learning models using pooled AUROC/AUPRC and within-patient Recall@Top-K%, with patient-level bootstrap confidence intervals.Source: §1 · p. 1
2. METHOD
FORESEE-MS forecasts future lesion risk from baseline imaging and compares deep learning predictions with a fold-specific population-prior atlas. Held-out follow-up lesion masks are reserved for evaluation.
2. METHOD
- Each model receives baseline FLAIR and a distance-to-brain-boundary map, while held-out follow-up masks are used only for evaluation.Training-fold follow-up masks supervise the models and contribute to the atlas.Source: §2 · p. 1
- The atlas and each model independently produce voxel-wise future-lesion risk maps, which are evaluated using pooled discrimination and within-patient localization.The workflow evaluates the risk maps using AUROC, AUPRC, and Recall@Top-K%; held-out follow-up masks are evaluation-only.Source: §2 · p. 1, Figure 1 · p. 2
- The atlas assigns each template voxel the fraction of training subjects with a new lesion there, smooths the frequencies, and is rebuilt from training subjects in each fold.It is warped into held-out subjects’ native space using the inverse template registration; held-out lesion data do not contribute.Source: §2 · pp. 1–2
- The five evaluated models are a 3D U-Net, Attn+ResBlock+VAE, ViT, and frozen and fine-tuned SwinUNETR variants.The SwinUNETR variants use pretrained weights; their input-facing layers and output layer are re-initialized for the study’s two-channel input and single-channel output.Source: §2 · p. 2
3. EXPERIMENTAL RESULTS AND DISCUSSION
In 39-subject cross-validation, the atlas led pooled voxel-level discrimination, while two models had higher mean within-patient voxel recall than the atlas at every evaluated threshold. Most paired recall confidence intervals included zero, and the small cohort limits precision.
3.1. Experimental Setup
- After excluding one subject with 27.6% lesion-mask overlap with the brain mask, the study analyzed 39 subjects; new lesions comprised approximately 0.035% of brain voxels (58,292 of 1.65 × 108).Source: §3.1 · p. 2
- Patient-level bootstrap intervals were unadjusted for multiple testing, so intervals excluding zero were treated as nominal, exploratory findings.Subjects’ voxels and lesion objects were resampled together, with the same resamples used for paired method comparisons.Source: §3.1 · p. 3
3.2. Quantitative Results
- The atlas achieved the highest pooled AUROC (0.747) and AUPRC (1.24 × 10−3); no deep learning model reliably outperformed it on either metric.All five models had lower AUROC point estimates (0.643–0.682), and the paired AUROC confidence intervals included zero.Source: §3.2 · p. 3, Table 1 · p. 3
- At K = 20%, mean voxel-level recall was 0.671 for Attn+ResBlock+VAE and 0.660 for the U-Net, versus 0.572 for the atlas.Both models also had higher mean recall than the atlas at K = 1%, 5%, and 10%; most paired recall confidence intervals included zero.Source: §3.2 · p. 3, Figure 2 · p. 4
3.3. Qualitative Results
- Qualitative examples showed a hit, a partial hit, and a miss; the partial hit highlighted only one of two lesions.The risk maps represent relative lesion risk, not calibrated probabilities.Source: §3.3 · p. 3
3.4. Discussion
- The higher mean recall of the U-Net and Attn+ResBlock+VAE may partly reflect visible pre-existing lesions; the 39-subject cohort limits precision and leaves improvement beyond the atlas uncertain.The atlas was built from approximately 31 subjects per fold and could benefit from a larger reference cohort.Source: §3.4 · pp. 3–4
4. CONCLUSION
FORESEE-MS evaluates whether baseline FLAIR can forecast future lesion locations beyond population-level spatial patterns. The atlas led on pooled voxel-level discrimination, while two deep learning models had higher mean within-patient recall, with uncertainty around most differences.
4. CONCLUSION
- Figure 3 illustrates a hit, a partial hit, and a miss across three subjects, with maps showing relative risk within the brain.Source: Figure 3 · p. 4
- The atlas achieved the highest pooled voxel-level AUROC and AUPRC, with no deep learning model reliably outperforming it.Source: §4 · p. 4
- At K ∈{1, 5, 10, 20}, 3D U-Net and Attn+ResBlock+VAE had higher mean voxel-level within-patient Recall@Top-K% than the atlas, although most paired confidence intervals included zero.Source: §4 · p. 4
- Future work will investigate baseline lesion masks and atlas-informed models using larger multi-site cohorts.Source: §4 · p. 5
5. COMPLIANCE WITH ETHICAL STANDARDS
The study used retrospective, de-identified MSSEG-2 MRI data under a Data Use Agreement and required no additional ethical approval for this secondary analysis.
5. COMPLIANCE WITH ETHICAL STANDARDS
- Retrospective, de-identified MSSEG-2 MRI data were distributed through Shanoir under a Data Use Agreement; no additional ethical approval was required.Source: §5 · p. 5