Research map / Point clouds / Shape and surface completion
Point2SSM: Learning Morphological Variations of Anatomies from Point Clouds
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 Point2SSM, an unsupervised DGCNN-encoder plus attention-module network that learns correspondence-based statistical shape models directly from raw point clouds, and benchmarks SOTA point cloud networks on anatomical SSM.
Problem
Constructing correspondence-based statistical shape models (SSMs) of anatomy normally needs complete, noise-free meshes or volumes and slow cohort-wide optimization; existing point cloud networks are untested for SSM under clinical challenges like noise, sparsity, partiality, and limited data.
Approach
An encoder-attention network that predicts correspondence points as attention-weighted combinations of input points, trained self-supervised with Chamfer distance plus a pairwise mapping-error regularizer over minibatches, enabling fast inference on new unseen point clouds.
Key results
- On the pancreas dataset, Point2SSM reaches CD 2.72 mm and P2F 0.283 mm versus lower-accuracy point baselines, with the best compactness on the left atrium dataset.
- On the pancreas tumor size classification downstream task, Point2SSM achieves 89.28% accuracy versus 85.71% for PSM and most baselines.
- On the L4 vertebrae test set, Point2SSM gives the best CD (2.61 mm), EMD (1.48) and P2F (0.304), outperforming the other point-based methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Point2SSM for pancreas tumor classification instead of the strongest reported baseline | Pancreas SSM classifier · Pancreas tumor classification | accuracy: 85.71% -> 89.28% | Helped |
| Use Point2SSM instead of the strongest point-based baseline | L4 vertebrae SSM · L4 vertebrae test set | CD: 4.66 -> 2.61 | Helped |
| Replace PointNet with a DGCNN encoder (ablation) | Point2SSM · Pancreas Point2SSM ablation | CD: 3.00 -> 2.87 | Helped |
| Replace the MLP-based attention module with the attention-based SFA module (ablation) | Point2SSM · Pancreas Point2SSM ablation | CD: 3.40 -> 2.87 | Helped |
| Add ME loss (alpha 0.1) (ablation) | Point2SSM · Pancreas Point2SSM ablation | compactness: 26 -> 24 | Helped |
| Change batch size from 8 to 2, 4, 6, 10, or 12 (ablation) | Point2SSM · Pancreas Point2SSM ablation | CD: 2.72; 2.74; 2.61; 2.64; 2.67; 2.71 | No clear effect |
| Train on spleen, pancreas, and left atrium together instead of only the tested anatomy (ablation) | Point2SSM · Spleen, pancreas, and left atrium | CD: 3.42 -> 3.33; 2.72 -> 3.10; 2.03 -> 2.15 | Mixed |
Surprising findings
- The ablation shows that replacing the attention module with an MLP leaves a model that still outperforms the comparison methods, suggesting gains are not solely due to capacity.
- Multi-anatomy training neither helps nor hinders Point2SSM accuracy.
Limitations
- Deterministic framework produces overconfident estimates, especially with noisy, partial, or sparse input.
- Requires roughly aligned input point clouds and is designed for a single anatomy.
- Performance declines given very sparse input (N=128).
Experimental setup
Datasets: spleen pancreas left atrium VerSe L4 vertebrae
Benchmarks: spleen pancreas left atrium L4 vertebrae
Metrics: Chamfer distance earth mover's distance point-to-face distance compactness generalization specificity mapping error tumor classification accuracy
Baselines: PSM PN-AE DG-AE CPAE ISR DPC
Builds on: DGCNN PointNet particle-based shape modeling canonical point autoencoder deep point correspondence
Closest papers by meaning
- MultiMorph: On-demand Atlas Construction (CVPR 2025 · similarity 0.84)
- SPEAL: Skeletal Prior Embedded Attention Learning for Cross-Source Point Cloud Registration (AAAI 2024 · similarity 0.84)
- MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI (NeurIPS 2026 · similarity 0.84)
- Unraveling Normal Anatomy via Fluid-Driven Anomaly Randomization (CVPR 2025 · similarity 0.83)
- DiffCorr: Conditional Diffusion Model with Reliable Pseudo-Label Guidance for Unsupervised Point Cloud Shape Correspondence (AAAI 2025 · similarity 0.83)
- Towards a Comprehensive, Efficient and Promptable Anatomic Structure Segmentation Model Using 3D Whole-Body CT Scans (AAAI 2025 · similarity 0.83)
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.