Applying homology-based morphometry to longitudinal spinal cord MRI will show that topological invariants (persistent homology Betti numbers) predict MS lesion merging and splitting events over time with a higher area under the ROC curve than conventional semantic segmentation volume-change models.
Adversarial Debate Score
57% survival rate under critique
Expert panel critique
Independent views, each critiquing the hypothesis on its own — the score rewards genuine disagreement and discounts consensus.
Supporting Research Papers
- Longitudinal tracking of multiple sclerosis lesions in the spinal cord: A validation study
Longitudinal characterization of multiple sclerosis (MS) lesions remains constrained by the lack of frameworks capable of establishing consistent instance-level correspondences across time. Convention...
- Homology-based Morphometry of Brain Atrophy: Methods and Applications
Understanding the structure of the brain, and how it changes with time and disease, is a core goal of structural neuroimaging. Contemporary approaches to structural brain analysis are dominated by vox...
- How Good are Foundation Models in Longitudinal MRI Disease Progression Reasoning?
Magnetic Resonance Imaging (MRI) interpretation is fundamental to clinical decision-making, requiring radiologists to integrate multi-view anatomical planes across sequential timepoints while precisel...
- Predicting Brain Morphometry with MT-GNN: Mesh Evolution in Continuous Time with Graph-Based Metric Tensor Embeddings
Predicting how a subcortical structure's shape will evolve from a few prior scans could support prognosis and clinical-trial enrichment. Existing longitudinal mesh predictors either extrapolate shape ...
Formal Verification
Z3 checks whether the hypothesis is internally consistent, not whether it is empirically true.
This discovery has a Claude-generated validation package with a full experimental design.
Precise Hypothesis
In patients with relapsing or progressive multiple sclerosis undergoing longitudinal spinal cord MRI (minimum 3 timepoints, 6-12 month intervals), persistent homology-derived topological features (Betti-0 connected components, Betti-1 loop structures, and associated persistence landscapes computed from cubical complexes over lesion probability maps) will predict binary lesion merging/splitting events occurring between consecutive scans with an AUC-ROC that is statistically significantly higher (by at least 0.05 AUC, one-sided DeLong test p<0.05) than a conventional model using semantic segmentation volume-change features (absolute volume delta, percent volume change, lesion count change) as the sole predictors, evaluated on identical held-out patient-level folds.
Hypothesis is disproven if: (a) AUC difference between topological and volumetric models is <0.05 or favors the volumetric model across >=3 independent cohorts/cross-validation schemes; (b) DeLong test fails to reach significance (p>=0.05) after correction for multiple comparisons across cohorts; (c) topological features show high collinearity with volume features (variance inflation factor >10) such that apparent gains vanish in ablation when volume is added to the topological model; (d) bootstrap resampling (1000 iterations) shows confidence interval for AUC difference crossing zero.
Experimental Protocol
Multi-site retrospective + prospective validation (MV) design with three phases: Phase 1 (retrospective feasibility, n=80-120 patients from existing MS natural history cohorts), Phase 2 (held-out replication, n=60-100 patients from an independent site/scanner), Phase 3 (prospective validation, n=150+ patients enrolled newly with protocolized imaging). All phases use nested patient-level cross-validation (outer 5-fold, inner 3-fold for hyperparameter tuning) to prevent leakage. Models compared: (1) baseline volumetric logistic regression/gradient boosting on volume-change features; (2) topological model using persistence images/landscapes fed into gradient boosting or lightweight CNN; (3) combined model for ablation. Primary endpoint: AUC-ROC for predicting merge/split events at next timepoint.
Longitudinal spinal cord MRI datasets with lesion segmentation masks and timepoint registration: (1) existing public/consortium MS cohorts with spinal cord imaging (e.g., NAIMS consortium data, UCSF/UPenn MS longitudinal registries) targeting n>=250 patients with >=3 timepoints each; (2) expert-adjudicated merge/split event labels requiring manual review by 2+ neuroradiologists with inter-rater kappa >=0.7 reported; (3) co-registered brain+cord MRI if available for confound adjustment (EDSS, lesion load, relapse history, DMT exposure); (4) scanner/protocol metadata for harmonization (ComBat or similar) across sites.
Topological model achieves AUC-ROC >=0.75 (vs baseline <=0.70) with DeLong p<0.05 in at least 2 of 3 phases; effect replicates in independent site data; ablation shows topological features retain independent predictive value (partial R^2 or SHAP importance >15% of total) when combined with volume features; inter-rater kappa for event labels >=0.7 confirming label reliability.
AUC improvement <0.05 or non-significant in >=2 of 3 phases; topological feature importance collapses to near-zero when combined with volume in ablation; results fail to replicate across scanner vendors/sites (AUC variance >0.15 across sites); registration/segmentation errors shown to drive apparent topological signal (sensitivity analysis shows signal disappears under perturbation of registration parameters within clinically plausible noise).
850
GPU hours
270d
Time to result
$18,000
Min cost
$165,000
Full cost
ROI Projection
Moderate-to-high if validated: addressable market includes MS clinical trial sponsors (biomarker-based patient stratification/enrichment can reduce trial sample sizes and costs, estimated trial cost savings of $2-5M per Phase 2/3 trial via enrichment); potential licensing to radiology AI vendors (Icometrix, Olea Medical) for a topological-analytics add-on module; estimated standalone software/biomarker licensing revenue potential of $3-10M over 5 years if regulatory-cleared; commercial value contingent on successful external replication and regulatory engagement, which adds 2-4 years before revenue realization.
High research value independent of commercial outcome: novel methodological contribution bridging computational topology and clinical neuroimaging, likely publishable in high-impact venues (Nature Communications, Medical Image Analysis, Brain); generates reusable open-source TDA pipeline and benchmark dataset of adjudicated merge/split events, valuable to broader MS imaging research community; strengthens mathematical foundations for topological biomarkers applicable beyond MS.
🔓 If proven, this unlocks
Proving this hypothesis is a prerequisite for the following downstream discoveries and applications:
- 1Topology-based biomarker pipeline for clinical trial enrichment in MS (identify fast lesion-reorganization patients)
- 2Extension to brain MS lesion dynamics and other neuroinflammatory/neurodegenerative diseases (e.g., NMOSD, small vessel disease)
- 3Regulatory-track biomarker qualification pursuit (FDA Biomarker Qualification Program) for topological MRI metrics
- 4Commercial software module for radiology AI platforms offering topological lesion-dynamics scoring
- 5Broader research program applying persistent homology to other longitudinal medical imaging phenotyping tasks
Prerequisites
These must be validated before this hypothesis can be confirmed:
- Lesion segmentation model achieves sufficient accuracy (Dice >=0.7) on spinal cord MRI across sites
- Reliable longitudinal registration pipeline for spinal cord (given cord motion/flexion artifacts) with validated sub-voxel accuracy
- Merge/split event definition and manual adjudication protocol achieves acceptable inter-rater reliability (kappa >=0.7)
- Persistence homology computation pipeline validated on synthetic/phantom lesion data to confirm topological features are not artifacts of noise or interpolation
Implementation Sketch
PIPELINE:
- preprocess(scan) -> bias_correct, denoise, register_to_PAM50
- lesion_mask = segment(scan, model=nnUNet_cord_lesion)
- for each timepoint t: complex_t = build_cubical_complex(lesion_prob_map_t)
- PD_t = compute_persistence_diagram(complex_t, dims=[0,1]) # GUDHI/Ripser
- feat_topo_t = vectorize(PD_t, method='persistence_image', resolution=20x20)
- feat_vol_t = [volume_t, volume_delta, lesion_count_delta]
- event_label = adjudicate_merge_split(mask_t, mask_t+1, registration)
- model_topo = GradientBoosting(feat_topo_t -> event_label)
- model_vol = GradientBoosting(feat_vol_t -> event_label)
- model_combined = GradientBoosting([feat_topo_t, feat_vol_t] -> event_label)
- evaluate: AUC_topo, AUC_vol, AUC_combined via nested CV; DeLong_test(AUC_topo, AUC_vol)
- ablation: SHAP_importance(model_combined) to assess independent contribution of topo features
FAILURE MODES:
- Registration drift between timepoints creating spurious merge/split detections (mitigated via deformable registration QC and manual review)
- Persistence diagrams dominated by segmentation noise rather than true lesion topology (mitigate via filtration threshold sensitivity analysis and bootstrap stability of Betti numbers)
- Small sample size leading to overfit topological feature vectors (high-dimensional persistence images vs few hundred patients) -> apply dimensionality reduction (PCA on persistence landscapes) and strict nested CV
- Class imbalance (merge/split events rare relative to stable lesions) causing inflated AUC estimates -> use stratified sampling, report PR-AUC alongside ROC-AUC
- Confounding by lesion count/size correlating with both topology complexity and ground truth labels -> explicit ablation and partial correlation analysis
ABORT CHECKPOINTS:
- After Phase 1 (n=80-120): if AUC_topo - AUC_vol < 0.02 or inter-rater kappa <0.6 for event labels, halt and revise event definition before Phase 2
- After registration QC: if >15% of timepoint pairs fail deformable registration quality threshold, halt and improve preprocessing pipeline
- After ablation analysis: if SHAP importance of topological features <5% in combined model, abort further phases as effect likely spurious/redundant with volume
- Mid-Phase 2: if cross-site AUC variance exceeds 0.15, pause prospective Phase 3 enrollment pending harmonization investigation