Research map / Label-efficient and robust learning / Self-supervised and contrastive
Masked-Diffusion Autoencoders for 3D Medical Vision Representation Learning
Jiachen Tu, Guanghui Qin, Theodore Zhengde Zhao, Jeya Maria Jose Valanarasu, Sheng Zhang, Tristan Naumann, Fan Lam, Sheng Wang, Hoifung Poon
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 |
|---|---|---|---|
| combine spatial masking with diffusion noise (dual corruption) (ablation) | MDAE · BraTS18 LGG vs HGG ablation | AUROC: 0.598 -> 0.658 | Improved |
| sample masking ratio from U(0.01, 0.99) instead of a fixed ratio (ablation) | MDAE · BraTS18 LGG vs HGG ablation | AUROC: 0.625 -> 0.662 | Improved |
| raise fixed masking ratio from 75% to 95% (ablation) | MDAE · BraTS18 LGG vs HGG ablation | AUROC: 0.625 -> 0.568 | Hurt |
| use block-wise 16^3 masking instead of random voxel masking (ablation) | MDAE · BraTS18 LGG vs HGG ablation | AUROC: 0.600 -> 0.623 | Improved |
| corrupt raw pixel space instead of embedding space (EMDAE) (ablation) | MDAE · BraTS18 LGG vs HGG ablation | AUROC: +4.2-10.2% AUROC over EMDAE | Improved |
| switch noise schedule among VP, VE and flow matching (ablation) | MDAE · BraTS18 LGG vs HGG ablation | AUROC: VE 0.615, Flow Matching 0.614, VP 0.613 | No clear effect |
| use ResNet encoder instead of Vision Transformer (ablation) | MDAE · BraTS18 LGG vs HGG ablation | AUROC: +12.1-13.8% AUROC over ViT | Improved |
| add FiLM time conditioning to ResNet (ablation) | MDAE · BraTS18 LGG vs HGG ablation | AUROC | No clear effect |
| add masked reconstruction loss alongside visible denoising loss (ablation) | MDAE · UCSF-PDGM segmentation (mean Dice) | Dice (%): 91.8 -> 92.6 | Improved |
| add visible-region denoising loss alongside masked reconstruction loss (ablation) | MDAE · UCSF-PDGM classification (mean) | classification mean score (%): 86.4 -> 90.7 | Improved |
| pretrain with MDAE instead of MAE | ResNet U-Net encoder · Cross-modal OOD brain MRI (6 tasks) | mean AUROC (%): 70.0 -> 78.6 | Improved |
| pretrain with MDAE instead of MAE | ResNet U-Net encoder · In-distribution brain MRI (T1/T2, 6 tasks) | mean AUROC (%): 69.5 -> 73.6 | 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 — VoCo: A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis (CVPR 2024) · better on 2 of 2 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
- LMD: Faster Image Reconstruction with Latent Masking Diffusion (AAAI 2024 · similarity 0.87)
- Multi-Modal Contrastive Masked Autoencoders: A Two-Stage Progressive Pre-training Approach for RGBD Datasets (CVPR 2025 · similarity 0.87)
- Structure-Guided Masked Autoencoders for Ultra-High Resolution Scientific Image Understanding (NeurIPS 2026 · similarity 0.87)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.87)
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.86)
- Noise-Consistent Siamese-Diffusion for Medical Image Synthesis and Segmentation (CVPR 2025 · similarity 0.86)
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.
