Research map / 3D vision / 3D generation
REFORM-3D: A Representation-Centric Evaluation Framework for 3D Medical Vision Foundation Models
Maregu Assefa, Muhammad Muzammal Naseer, Divya Velayudhan, Gedamu Kumie, Naoufel Werghi
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 REFORM-3D, a representation-centric evaluation framework that assesses 3D medical vision foundation models separately from downstream adaptation choices, motivated by spacing and anisotropy effects on image statistics.
Problem
Evaluation of 3D medical foundation models relies on downstream task scores that entangle representation quality with adaptation choices, and spacing/anisotropy can change image statistics without changing anatomy.
Approach
Shifts evaluation toward intrinsic representation quality of 3D medical vision foundation models, disentangling it from downstream fine-tuning or adaptation decisions, with explicit attention to spacing and anisotropy effects.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use voxel-relative sampling instead of isotropic normalization in pretraining (ablation) | Med3DINO (CLS) · Controlled same-scan spacing, AbdomenAtlas | cosine drift: 19.00 -> 10.90 | Improved |
| use spacing-aware resampling instead of isotropic normalization in pretraining (ablation) | Med3DINO (CLS) · Controlled same-scan spacing, AbdomenAtlas | semantic transfer: 73.60 -> 81.00 | Improved |
| use spacing-aware resampling instead of voxel-relative sampling (ablation) | Med3DINO (CLS) · Controlled same-scan spacing, ImageCAS | semantic transfer: 54.60 (REL) vs 54.20 (SA) | No clear effect |
| use spacing-aware resampling instead of voxel-relative sampling (ablation) | Med3DINO (CLS) · Controlled same-scan spacing, AbdomenAtlas | cosine drift: 10.90 -> 18.40 | Hurt |
| use spacing-aware resampling instead of isotropic normalization (ablation) | Med3DINO (GAP) · CT-MR retrieval, TotalSegmentator | mAP: 16.20 -> 27.20 | Improved |
| use spacing-aware resampling instead of isotropic normalization (ablation) | Med3DINO (GAP) · CT-MR retrieval, AMOS22 | mAP: 21.60 -> 44.80 | Improved |
| use GAP features instead of CLS features for SA retrieval (ablation) | Med3DINO-SA · CT-MR retrieval, TotalSegmentator | mAP: 19.90 -> 27.20 | Improved |
| use spacing-aware resampling instead of isotropic normalization (ablation) | Med3DINO (GAP) · Anatomical transfer, TotalSegmentator CT-MR | probe balanced accuracy: 44.30 -> 58.80 | Improved |
| use spacing-aware resampling instead of voxel-relative sampling (ablation) | Med3DINO (GAP) · Anatomical transfer, MMWHS | centroid balanced accuracy: 75.10 (REL) vs 74.50 (SA) | No clear effect |
| use VoCo features rather than Med3DINO-SA under spacing change | VoCo · Controlled same-scan spacing, AbdomenAtlas | drift vs CKA: drift 0.80, CKA 86.70, transfer 74.60 | Mixed |
| switch from Med3DINO-ISO to Med3DINO-SA pretraining (ablation) | Med3DINO segmentation · BTCV | Dice: 64.71 -> 75.69 | Improved |
| use Med3DINO-SA instead of 3DINO | segmentation · BTCV | Dice: 82.10 -> 75.69 | Hurt |
Closest papers by meaning
- Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.91)
- Autoregressive Sequence Modeling for 3D Medical Image Representation (AAAI 2025 · similarity 0.88)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.87)
- 3DMedAgent: Unified Perception-to-Understanding for 3D Medical Analysis (ICML 2026 · similarity 0.87)
- Learning General-purpose Biomedical Volume Representations using Randomized Synthesis (ICLR 2025 · similarity 0.87)
- How Well Do Supervised 3D Models Transfer to Medical Imaging Tasks? (ICLR 2024 · similarity 0.87)
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.
