Research map / Label-efficient and robust learning / Self-supervised and contrastive
VoCo: A Simple-yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis
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 VoCo, a self-supervised volume contrastive pretext task that predicts which region of a 3D medical volume a random sub-volume comes from by contrasting it against base crops, with a decorrelation regularizer on the bases.
Problem
Self-supervised pre-training for 3D medical image analysis lacks high-level semantics, which limits downstream performance on tasks needing semantic understanding.
Approach
Exploits consistent contextual position priors (geometric relations between organs) by cropping non-overlapping base volumes as prototype-like class assignments and predicting each random sub-volume's overlap proportions with them via cosine-similarity volume contrast, plus a loss that makes bases linearly independent; this replaces position-classification linear heads and online clustering.
Key results
- On BTCV, VoCo reaches 83.85% average Dice, versus 82.01% for the best prior method GL-MAE and 80.53% for scratch Swin-UNETR.
- On LiTS liver segmentation, Swin-UNETR with VoCo pre-training reaches 96.52% Dice, up from 93.42% from scratch.
- On CC-CCII COVID-19 classification, VoCo reaches 90.83% accuracy versus 89.61% for SwinMM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pre-train with VoCo rather than GL-MAE | Swin-UNETR on BTCV · BTCV | Average Dice Score (%): 82.01 -> 83.85 | Helped |
| pre-train with VoCo rather than Rubik++ | Swin-UNETR on LiTs · LiTs | Dice Score (%): 95.72 -> 96.52 | Helped |
| pre-train with VoCo rather than GVSL | Swin-UNETR on MSD Spleen · MSD Spleen | Dice Score (%): 95.47 -> 96.34 | Helped |
| pre-train with VoCo rather than GVSL | Swin-UNETR on MM-WHS · MM-WHS | Dice Score (%): 88.27 -> 90.54 | Helped |
| pre-train with VoCo rather than the strongest listed baseline | Swin-UNETR on BraTS 21 · BraTS 21 | AVG Dice Score (%): 77.85 -> 78.53 | Helped |
| pre-train with VoCo rather than SwinMM | Swin-UNETR on CC-CCII · CC-CCII | Accuracy (%): 89.61 -> 90.83 | Helped |
| add the position-prediction loss Lpred (ablation) | VoCo · BTCV | Average Dice Score (%): 80.53 -> 82.96 | Helped |
| add the position-prediction loss Lpred (ablation) | VoCo · MM-WHS | Average Dice Score (%): 86.11 -> 88.82 | Helped |
| add the regularization loss Lreg (ablation) | VoCo with Lpred · BTCV | Average Dice Score (%): 82.96 -> 83.85 | Helped |
| add the regularization loss Lreg (ablation) | VoCo with Lpred · MM-WHS | Average Dice Score (%): 88.82 -> 90.54 | Helped |
| increase the number of bases from 2×2×1 to 3×3×1 (ablation) | VoCo · BTCV | Average Dice Score (%): 81.56 -> 82.77 | Helped |
| increase the number of bases from 3×3×1 to 4×4×1 (ablation) | VoCo · MM-WHS | Average Dice Score (%): 89.31 -> 90.54 | Helped |
Surprising findings
- General SSL methods such as MAE, SimCLR, and MoCo v3 underperform most medical-specific SSL methods on BTCV, which the authors attribute to large-batch negatives being impractical for 3D volumes.
Limitations
- Number of base crops is limited by computation; cropping along the Z direction did not help.
- Pre-training uses only about 1.6k CT scans and was not scaled up.
- Evaluated mainly on segmentation with Swin-UNETR and a few classification and MRI transfer tasks.
Experimental setup
Datasets: BTCV TCIA Covid19 LUNA LiTS MSD Spleen MM-WHS BraTS 21 CC-CCII
Benchmarks: BTCV LiTS MSD Spleen MM-WHS BraTS 21 CC-CCII
Metrics: Dice score accuracy
Baselines: MAE3D SimCLR SimMIM MoCo v3 Jigsaw PositionLabel MG ROT PCRLv1 PCRLv2 Rubik++ Swin-UNETR
Builds on: MoCo v3 Swin-UNETR PositionLabel prototype-level contrastive learning
Closest papers by meaning
- Autoregressive Sequence Modeling for 3D Medical Image Representation (AAAI 2025 · similarity 0.86)
- Learning General-purpose Biomedical Volume Representations using Randomized Synthesis (ICLR 2025 · similarity 0.86)
- 3DMedAgent: Unified Perception-to-Understanding for 3D Medical Analysis (ICML 2026 · similarity 0.84)
- Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.84)
- MESC-3D:Mining Effective Semantic Cues for 3D Reconstruction from a Single Image (CVPR 2025 · similarity 0.84)
- Revisiting MAE Pre-training for 3D Medical Image Segmentation (CVPR 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.