Research map / Architectures and efficiency / RNNs and normalization
MultiMorph: On-demand Atlas Construction
Read the paperCheck your idea against 108,967 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 MultiMorph, a feedforward UNet-based model with a GroupBlock mean-feature-sharing layer and centrality layer that builds unbiased 3D brain atlases in a single forward pass for arbitrary groups and modalities.
Problem
Atlas construction for anatomical populations is computationally slow (days to weeks), so researchers reuse mismatched precomputed atlases; learning-based alternatives need retraining per population or modality.
Approach
A convolutional network with a GroupBlock layer that concatenates each image's features with the group mean at every layer (linear-complexity feature sharing), plus a centrality layer that subtracts the group-mean velocity field, lets a single trained model produce group-specific, unbiased atlases at test time without fine-tuning, aided by synthetic training data for unseen modalities.
Key results
- On unseen IXI T1-w data, MultiMorph reached Dice 0.913 versus 0.894 for AtlasMorph and 0.863 for ANTs, with 10.5 minutes construction time versus 1141.5 for AtlasMorph and 4345.2 for ANTs.
- On OASIS-3 T1-w (212 subjects), MultiMorph achieved Dice 0.910 in 5.9 minutes versus 2858 minutes for ANTs.
- Ablation: the centrality layer reduced the centrality measure by about 1000x, and the Dice loss improved Dice by over 2 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MultiMorph instead of baseline atlas construction methods | Atlas construction · IXI T1-w, T2-w, and PD-w | Dice: 0.913 ± 0.006 vs 0.894 ± 0.015 (T1-w); 0.906 ± 0.007 vs 0.882 ± 0.018 (T2-w); 0.900 ± 0.009 vs 0.884 ± 0.018 (PD-w) | Helped |
| Use MultiMorph instead of baseline atlas construction methods | Atlas construction · IXI T1-w, T2-w, and PD-w | Construction time (min.): 10.50 vs 325.20 (T1-w); 10.40 vs 261.00 (T2-w); 7.80 vs 163.80 (PD-w) | Helped |
| Use MultiMorph instead of baseline atlas construction methods | Atlas construction · IXI subgroups | Dice Transfer: 0.904 ± 0.002 vs 0.893 ± 0.005 | Helped |
| Use MultiMorph instead of baseline atlas construction methods | Atlas construction · OASIS-3 T1-w test | Dice Transfer: 0.910 ± 0.014 vs 0.886 ± 0.017 | Helped |
| Add the centrality layer (ablation) | MultiMorph · OASIS-1 test | Centrality: 16125 ± 11494 -> 12.0 ± 39.48 | Helped |
| Add the GroupBlock mechanism (ablation) | MultiMorph · OASIS-1 test | Dice Transfer: 0.870 ± 0.021 -> 0.884 ± 0.020 | Helped |
| Replace mean GroupBlock summary with variance (ablation) | MultiMorph · OASIS-1 test | Dice Transfer: 0.884 ± 0.020 -> 0.883 ± 0.020 | No clear effect |
| Replace mean GroupBlock summary with max (ablation) | MultiMorph · OASIS-1 test | Dice Transfer: 0.884 ± 0.020 -> 0.880 ± 0.019 | No clear effect |
| Add the Dice loss (ablation) | MultiMorph · OASIS-1 test | Dice Transfer: 0.884 ± 0.020 -> 0.919 ± 0.011 | Helped |
| Train with synthetic data (ablation) | MultiMorph · IXI | Dice: up to 1.8 Dice points | Helped |
Surprising findings
- Aladdin yields zero folds but much higher centrality error, and the centrality layer costs about 1 Dice point in ablation.
- MultiMorph generalizes to the PD-w modality never seen in training.
Limitations
- Assumes diffeomorphic transformations, so it cannot handle topology-changing pathologies.
- Stores all activations in memory at inference, limiting group size for large 3D volumes.
- Currently trained only on neuroimages (brain MRI).
Experimental setup
Datasets: OASIS-1 OASIS-3 Mind Brain Body (MBB) IXI SynthSeg synthetic images
Benchmarks: IXI OASIS-3 OASIS-1
Metrics: Dice run time (min) folds (Jacobian determinant < 0) centrality (mean displacement norm) segmentation transfer Dice
Baselines: ANTs SyGN AtlasMorph Aladdin TAG
Builds on: UNet SynthMorph AtlasMorph SyGN Aladdin
Closest papers by meaning
- Point2SSM: Learning Morphological Variations of Anatomies from Point Clouds (ICLR 2024 · similarity 0.84)
- DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases (NeurIPS 2025 · similarity 0.84)
- MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI (NeurIPS 2026 · similarity 0.84)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.82)
- Representing Part-Whole Hierarchies in Foundation Models by Learning Localizability Composability and Decomposability from Anatomy via Self Supervision (CVPR 2024 · similarity 0.82)
- AnaDiffusion: Anatomically Compositional Latent Diffusion for Controllable 3D Brain MRI Generation (NeurIPS 2026 · similarity 0.82)
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.