Research map / Label-efficient and robust learning / Self-supervised and contrastive
Beyond Isolated Scans: Cross-Phase Alignment of Structure and Topology for 3D Medical Pretraining
Wenzhuo xu, YANJIE ZHOU, Yujian Hu, Hongkun Zhang, Minfeng Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spectral consistency (Lspec) to reconstruction baseline (ablation) | CAST · Parse22 | Dice: 0.7972 -> 0.8423; +4.51% | Improved |
| add geometry-aware topological consistency (Lgeo) to reconstruction baseline (ablation) | CAST · Parse22 | Dice: 0.7972 -> 0.8365; +3.93% | Improved |
| combine Lspec and Lgeo (ablation) | CAST · Parse22 | Dice: 0.8423 -> 0.8463 | Improved |
| add Lspec to reconstruction baseline (ablation) | CAST · COVID19-20 | Dice: 0.6271 -> 0.6935 | Improved |
| add Lgeo and Lspec together (ablation) | CAST · COVID19-20 | Dice: 0.6975 -> 0.7075 | Improved |
| use Diff+TopHat keypoint sampling instead of Random sampling (ablation) | CAST · Parse22 | Dice: 0.8268 -> 0.8463 | Improved |
| use Diff+TopHat sampling instead of Diff only (ablation) | CAST · Parse22 | Dice: 0.8021 -> 0.8463 | Improved |
| use 1024 keypoints instead of 512 (ablation) | CAST · Parse22 | Dice: 0.8174 -> 0.8463 | Improved |
| use 1024 keypoints instead of 2048 (ablation) | CAST · Parse22 | Dice: 0.8132 -> 0.8463 | Improved |
| use true paired NCCT-CECT pretraining instead of shuffled pairs (ablation) | CAST · COVID19-20 | Dice: 0.6364 -> 0.7075 | Improved |
| use paired cross-phase pretraining instead of NCCT-only CAST (ablation) | CAST · COVID19-20 (10% labels) | Dice: 2.15% gain at 10% labels (0.6514) | Improved |
| use 10% paired pretraining data versus full NCCT-only data (ablation) | CAST · COVID19-20 | Dice: +0.0094 Dice; full 13,850 pairs reach 0.7075 | 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.
- VoCo-v1 — VoCo: A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis (CVPR 2024) · better on 4 of 4 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.
Closest papers by meaning
- Rethinking CT Synthesis through Semantics-Structure Alignment (NeurIPS 2026 · similarity 0.87)
- REFORM-3D: A Representation-Centric Evaluation Framework for 3D Medical Vision Foundation Models (NeurIPS 2026 · similarity 0.87)
- Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.86)
- Scaling Self-Supervised and Cross-Modal Pretraining for Volumetric CT Transformers (CVPR 2026 · similarity 0.86)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.85)
- Large-scale and Fine-grained Vision-language Pre-training for Enhanced CT Image Understanding (ICLR 2025 · similarity 0.85)
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.
