Research map / Label-efficient and robust learning / Self-supervised and contrastive

Revisiting MAE Pre-training for 3D Medical Image Segmentation

CVPR 2025 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Self-supervised neuroimaging representations · 19 citations (OpenAlex)

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.

Revisits masked autoencoder pre-training for 3D CNNs on a 39k brain MRI dataset using a ResEnc U-Net in nnU-Net, producing Spark3D, which beats SSL baselines and the nnU-Net from-scratch baseline across 11 brain MRI segmentation datasets.

Problem

SSL pre-training for 3D medical image segmentation has been hampered by small pre-training datasets, outdated backbones, and weak evaluation, so it is unclear whether it beats strong from-scratch nnU-Net baselines.

Approach

Applies a carefully configured masked autoencoder to a residual encoder U-Net (CNN) with sparse convolutions, mask tokens, densification convs, dynamic masking ratio, and a tuned fine-tuning schedule, pre-trained on 39k brain MRIs and validated on separate development and test datasets.

Key results

• Spark3D achieves average DSC 72.37 across 11 test datasets vs 70.40 for nnU-Net fixed-config from scratch and 71.83 for Models Genesis.
• Spark3D outperforms the from-scratch nnU-Net (fixed) on 10 of 11 test datasets, with about +2 DSC on average.
• With only 40 labeled training images, the pre-trained model nearly matches the from-scratch model trained on the full dataset.

Measured effects reported in the paper

Surprising findings

• Models Genesis, a 2019 method, outperforms more recent VoCo and Volume Fusion when transferred to the same CNN backbone.
• Longer pre-training beyond 250k steps degraded downstream benefits, and larger batch sizes did not improve results.

Limitations

• Pre-training data is proprietary and restricted to head and neck/brain MRI.
• Scaling pre-training dataset size or model parameters was not shown to help and remains untested.
• Generalization to unseen MRI modalities (e.g., TOF angiography) remains weaker than Models Genesis.

Experimental setup

Datasets: proprietary brain MRI pre-training dataset (39,168 volumes) MS FLAIR Brain Mets Hippocampus (MSD) ATLAS v2.0 CrossModa Cosmos22 HaNSeg ISLES22 HNTS-MRG24 BraTS Africa T2 Aneurysms

Benchmarks: MS FLAIR Brain Mets Hippocampus Atlas22 CrossModa Cosmos22 ISLES22 HaNSeg HNTS-MRG24 BraTS Africa T2 Aneurysms

Metrics: Dice similarity coefficient normalized surface distance average rank

Baselines: nnU-Net (dynamic) nnU-Net (fixed) VoCo Volume Fusion Models Genesis

Builds on: masked autoencoder nnU-Net Models Genesis VoCo SparK

Closest papers by meaning

• 3D Self-Supervised Methods for Medical Imaging (NeurIPS 2020 · similarity 0.91)
• Learning Generalizable 3D Medical Image Representations from Mask-Guided Self-Supervision (CVPR 2026 · similarity 0.90)
• Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis (CVPR 2022 · similarity 0.89)
• Autoregressive Sequence Modeling for 3D Medical Image Representation (AAAI 2025 · similarity 0.87)
• Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.86)
• Enhancing vision-language models for medical imaging: bridging the 3D gap with innovative slice selection (NeurIPS 2024 · similarity 0.85)