Research map / Vision-language and multimodal / Cross-modal learning
SphereFlow: Missing Modality Imputation via Geometric Transport on Hypersphere
Huihan Wang, Juncheng Wang, Yuxiang Feng, Wenlong Hou, Guoqi Yu, Chao Xu, Yang Liu, Baigui Sun, Shujun Wang Emma
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SphereFlow, which imputes missing modalities by transporting latents along geometry on a shared hyperspherical space obtained from a frozen self-supervised encoder.
Problem
Missing modality imputation: synthesizing unobserved modalities from observed ones. Existing methods treat each modality as on its own latent manifold, forcing the generator to learn a long unstructured jump between modalities.
Approach
Projects all modalities into a shared hyperspherical latent space via a frozen self-supervised encoder so cross-modal imputation becomes geometric transport on the hypersphere rather than an unstructured jump between per-modality manifolds.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Gaussian noise source with observed-modality latent and add manifold-expanded decoder; 1-to-1 BraTS | SphereFlow · BraTS 1-to-1 | rFID: 29.99 -> 19.93 | Improved |
| SphereFlow vs best baseline on downstream segmentation | SphereFlow · BraTS segmentation avg | Dice: 0.5930 -> 0.7473 | Improved |
| use DINOv2 encoder instead of MedMAE (ablation) | SphereFlow · Stage-1 T1c reconstruction | iFID: 144.45 -> 40.64 | Improved |
| use DINOv2 encoder instead of MedMAE (ablation) | SphereFlow · Stage-1 T1c reconstruction | PSNR: 31.23 -> 30.57 | Hurt |
| add Step 2 manifold expansion to decoder (ablation) | SphereFlow RAE · on-manifold | PSNR: 33.14 -> 33.83 | Improved |
| add Step 2 manifold expansion to decoder (ablation) | SphereFlow RAE · on-manifold | rFID: 28.29 -> 30.92 | Hurt |
| add Step 3 shared-encoder alignment (ablation) | SphereFlow RAE · on-manifold | PSNR: 33.83 -> 44.06 | Improved |
| SphereFlow vs Zhang on CT to MRI | SphereFlow · CT to MRI | PSNR: 25.51 -> 24.70 | Hurt |
| SphereFlow vs Zhang on CT to MRI | SphereFlow · CT to MRI | rFID: 30.91 -> 13.65 | Improved |
| average latents of multiple source modalities (N-to-N) | SphereFlow · BraTS N-to-N | PSNR: 35.13 -> 35.88 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CodeBrain — CodeBrain: Towards Decoupled Interpretability and Multi-Scale Architecture for EEG Foundation Model (ICLR 2026) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Builds on: flow matching
Closest papers by meaning
- Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness (NeurIPS 2026 · similarity 0.85)
- UniMRSeg: Unified Modality-Relax Segmentation via Hierarchical Self-Supervised Compensation (NeurIPS 2025 · similarity 0.85)
- Riemannian Flow Matching for Brain Connectivity Matrices via Pullback Geometry (NeurIPS 2025 · similarity 0.84)
- KMD: Koopman Multi-modality Decomposition for Generalized Brain Tumor Segmentation under Incomplete Modalities (CVPR 2025 · similarity 0.84)
- MAGNET: Manifold-Aware Graph Diffusion Network for Connectome Generation (NeurIPS 2026 · similarity 0.84)
- Semantic-guided Masked Mutual Learning for Multi-modal Brain Tumor Segmentation with Arbitrary Missing Modalities (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
