CrossScale-GLIO: Topology-Preserving Vision-Language Alignment of MRI and Whole-Slide Histopathology for Diffuse Glioma

1 1. Introduction

• CrossScale-GLIO preserves each modality’s graph topology during transport, extending multimodal fusion beyond implicit patient-vector or unordered-token correspondence.Its habitat and niche graphs are aligned through feature, structural, and language-derived diagnostic signals.Source: §1 · p. 1
• MRI and histology lack reliable spatial registration, but their relational structures can support cross-scale correspondence without voxel-to-slide alignment.Relevant relations include centrality, boundary position, adjacency to necrosis, interface complexity, and neighborhood diversity.Source: §1 · p. 1

2.1 2.1 Computational pathology representation learning

• Pathology foundation encoders and cellular-neighborhood summaries distinguish patch appearance from histologic ecological organization.The pathology encoder may be frozen or lightly adapted.Source: §2.1 · pp. 1–2

2.2 2.2 Multimodal oncology

• CrossScale-GLIO makes cross-modal correspondence explicit with a coupling matrix and tests it through retrieval, deletion, and graph randomization.This contrasts with fusion approaches that leave correspondence implicit.Source: §2.2 · p. 2

2.3 2.3 Glioma radiology and molecular classification

• Structured intermediate representations retain modality-specific organization, making the fusion map inspectable and enabling transport-specific counterfactual tests.Optimal transport can align compact MRI and larger histology graphs without one-to-one node identity.Source: §2.3 · p. 2

3 3. Problem Definition

• For each patient, the model learns MRI habitat and histologic niche graphs, a coupling matrix, and outputs for subtype, markers, survival, retrieval, and localization.A pairing indicator distinguishes fully paired, MRI-only, and WSI-only cases; transport losses apply only to paired cases.Source: §3 · pp. 2–3

4.1 4.1 MRI habitat graph

• MRI habitat nodes summarize tumor-region sequence, texture, spatial, learned-feature, and uncertainty information, with edges encoding adjacency and compartment relations.Spatially constrained clustering forms contiguous regions from tumor and peritumoral voxels.Source: §4.1 · p. 3, §4.6 · p. 3

4.2 4.2 Histologic niche graph

• Histologic niche nodes summarize patch embeddings and cellular composition, with edges representing physical proximity, phenotype transitions, and shared tissue interfaces.The pipeline uses low- and high-magnification tissue analysis and retains specimen and slide identity.Source: §4.2 · p. 3, §4.6 · p. 3

4.3 4.3 Diagnostic language anchors

• Diagnostic language anchors map observable radiology and pathology concepts to a shared space, and predicted concept distributions permit inference without reports.Anchor concepts include enhancement, necrosis, infiltration, vascular proliferation, cellularity, and immune-rich regions.Source: §4.3 · p. 3, §4.6 · p. 3

4.4 4.4 Topology-aware optimal transport

• The transport module combines projected feature distance, within-graph structural discrepancy, and semantic compatibility to produce a differentiable soft coupling.Entropic regularization stabilizes the coupling; learned marginal masses are constrained against collapse, and cycle consistency preserves patient-level summaries.Source: §4.4 · p. 3, §4.6 · p. 3

4.5 4.5 Patient representation and task heads

• Mass-weighted pooling of coupled node features feeds classification, survival, retrieval, and localization heads; modality-specific heads support missing-modality inference.Localization ranks node pairs by transported mass multiplied by task gradient.Source: §4.5 · p. 3, §4.6 · p. 4

4.6 4.6 Training strategy

• The training strategy progresses from encoder pretraining and graph consistency to transport, semantic anchors, joint task optimization, and calibration.Patient-level batching keeps a patient’s slides and MRI examinations within one split.Source: §4.6 · p. 3

5.1 5.1 Cohorts and pairing

• Paired TCGA-TCIA MRI and histology cases are supplemented by MRI-rich cohorts for missing-pathology transfer and external imaging validation.Patient identity is resolved before partitioning, and the pairing audit tracks acquisition timing, specimen and slide quality, sequence availability, and target completeness.Source: §5.1 · p. 4, §5.4 · p. 5

5.2 5.2 Baselines

• Baselines include unimodal models, late and gated fusion, token cross-attention, feature-only transport, and full topology-aware transport.Encoders, augmentations, patient splits, and head capacity are matched when possible; a capacity-matched control guards against attributing gains to extra parameters.Source: §5.2 · p. 4, §5.4 · p. 5

5.3 5.3 Outcomes and statistical tests

• Evaluation uses classification, survival, retrieval, and localization metrics, with patient-level bootstrap confidence intervals and Holm-corrected paired tests.Localization measures include expert plausibility, deletion faithfulness, and stability across augmentations.Source: §5.3 · p. 4

5.4 5.4 Falsification experiments

• Falsification tests perturb graph topology and edge weights, permute semantic anchors, subsample tissue, and drop MRI sequences to assess which information supports each task.Blinded pathologists rate high-mass habitat-niche pairs for plausibility and specificity, while neuroradiologists independently rate MRI habitats.Source: §5.4 · p. 5

6.1 6.1 Molecular, subtype, and survival performance

• Subtype macro-F1 was 0.789, with IDH AUROC 0.934, 1p/19q AUROC 0.884, and MGMT AUROC 0.802; the gain over feature-only optimal transport was 0.028.The subtype gain had a 95% CI of 0.012 to 0.044 and Holm-adjusted p = 0.0019; external transfer remained stable across UPenn-GBM, UCSF-PDGM, and the independent pathology cohort.Source: §6.1 · p. 5

6.2 6.2 Retrieval and cross-scale localization

• MRI-to-WSI and WSI-to-MRI Recall@1 reached 0.286 and 0.278, while Recall@5 reached 0.621 and 0.608.Pathologists rated 81.2% of high-mass habitat-niche pairs biologically plausible; removing the highest-mass correspondence reduced correct-class probability by 0.184, compared with 0.049 under random pair deletion.Source: §6.2 · p. 6, Table 4 · p. 6

6.3 6.3 Topology falsification and robustness

• Degree-preserving rewiring reduced subtype macro-F1 by 0.034 and retrieval Recall@1 by 0.090; semantic-anchor permutation reduced subtype macro-F1 by 0.017.Performance declined monotonically under stronger edge randomization, tissue subsampling, and MRI-sequence removal; the no-transport variant had no coupling for expert review.Source: §6.3 · p. 6, Table 5 · p. 6

• The joint pattern across prediction, retrieval, localization, and falsification supports topology-preserving cross-scale alignment.Targeted deletion reduced correct-class probability nearly four times as much as random deletion, linking correspondence mass to task relevance.Source: §7 · p. 6
• Degree-preserving rewiring retained node appearance and local density while reducing subtype macro-F1 by 0.034 and retrieval Recall@1 by 0.090.The larger retrieval loss indicates strong dependence of patient-specific alignment on relational geometry.Source: §7 · p. 6
• Semantic permutation produced a smaller but significant subtype loss, indicating that structural and language anchors contribute complementary constraints.Source: §7 · p. 7
• The coupling is a probabilistic association between whole-lesion MRI and selectively sampled tissue, not physical registration or verified biopsy coordinates.Tissue fraction, specimen timing, coupling entropy, and expert plausibility delimit underdetermined cases; prospective spatial sampling and multi-region pathology are the next validation step.Source: §7 · p. 7

• The release will include split manifests, preprocessing containers, graph-construction parameters, encoder versions, training seeds, transport hyperparameters, evaluation code, and permitted de-identified coupling outputs.Slide and MRI identifiers remain governed by their original repositories.Source: §8 · p. 7
• Cross-scale maps are not verified biopsy coordinates, and the study does not claim direct clinical actionability.Documentation defines validation required before clinical use.Source: §8 · p. 7
• Figure 7 presents topology falsification under progressively stronger graph perturbations.Source: Figure 7 · p. 7

9 9. Conclusion

• CrossScale-GLIO aligns MRI habitats and histologic niches through topology-aware optimal transport and semantic anchors.External prediction, bidirectional retrieval, expert review, targeted deletion, and graph rewiring jointly support the conclusion that relational topology carries clinically relevant cross-scale information.Source: §9 · p. 7

Funding

• This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the National Program for Excellence in Software, supervised by the IITP.Source: Funding · p. 7
• The Kunsan National University Industry-University Cooperation Foundation Research Fund provided support under 2023H052.Source: Funding · p. 8

A Data Integrity and Result Reconciliation

• The locked export has 131 rows and 14 fields, no exact duplicate rows, a unique composite analysis key, and two N/A values for structurally unavailable expert plausibility.Confidence intervals and adjusted p values appear only for prespecified inferential comparisons.Source: A Data Integrity and Result… · p. 10

B Extended Numerical Results

• Appendix Tables A3 and A4 are designated for inference-ready results with confidence intervals and for ablation and robustness rows, respectively.Table A5 is designated for narrative-placeholder reconciliation.Source: Table A3 · p. 10, Table A4 · p. 11, Table A5 · p. 11
• The extended-results table header lists variant, condition, metric, and value.Source: B Extended Numerical Results · p. 11
• The paired external test reports adjusted p = 0.002 and Recall@1 values of 0.286 for M2W and 0.278 for W2M; subtype macro-F1 CI bounds are 0.012 and 0.044.The lower and upper confidence-interval bounds are reported separately in the extended results.Source: B Extended Numerical Results · p. 11

C Extended Result Figures

• Figures A1–A5 are labeled as extended data-derived diagnostics.Source: Figure A1 · p. 12, Figure A2 · p. 13, Figure A3 · p. 13
• Figures A6–A10 are labeled as extended data-derived diagnostics.Source: Figure A6 · p. 14, Figure A7 · p. 14, Figure A8 · p. 14

D Reproducibility Record

• The appendix source directory contains the locked CSV export, generated tables, all figure assets, and the executed analysis notebook.Main-paper numeric statements map to export rows through cohort, method, condition, metric, and optional placeholder key.Source: D Reproducibility Record · p. 16