Research map / Image generation / Image restoration
DRACO: A Denoising-Reconstruction Autoencoder for Cryo-EM
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 DRACO, a denoising-reconstruction masked autoencoder pre-trained on odd/even cryo-EM micrographs (Noise2Noise-style) from a curated 270k-micrograph EMPIAR dataset, serving as a denoiser and foundation model for cryo-EM downstream tasks.
Problem
Self-supervised vision foundation models such as MAE overlook the extreme noise in cryo-EM micrographs, and cryo-EM lacks large, well-curated pre-training data and clean references for denoising.
Approach
Split each movie into odd and even sub-movies treated as independent noisy observations for Noise2Noise-style denoising on visible patches, combined with MAE-style reconstruction of masked patches supervised by the full-sum micrograph, in a masked autoencoder framework.
Key results
- DRACO-L achieves F1 0.816 on Human HCN1 and 0.824 on 70S ribosome particle picking, best among compared methods, with resolutions 3.90 and 2.51 A.
- DRACO-B reaches SNR 10.13 dB-level gains over raw; DRACO-L reaches SNR 12.21 on RNA polymerase vs -4.69 raw and 8.66 for Topaz-Denoise.
- DRACO-L achieves 0.983 accuracy and 0.984 F1 on micrograph curation vs 0.836 for Miffi and 0.938 for ResNet18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DRACO-L instead of ResNet18 trained from scratch | Micrograph curation · Micrograph curation test | accuracy: 0.938 -> 0.983 | Helped |
| Use DRACO-L instead of the strongest listed baseline | Particle picking · Human HCN1 | F1 score: 0.782 -> 0.816 | Helped |
| Use DRACO-L instead of the strongest listed baseline | Particle picking · 70S ribosome | F1 score: 0.786 -> 0.824 | Helped |
| Use DRACO-L instead of the strongest listed baseline | Particle picking · LetB | F1 score: 0.682 -> 0.725 | Helped |
| Use DRACO denoising instead of the strongest listed baseline | Micrograph denoising · Human Apoferritin | SNR: -0.31 -> 2.01 | Helped |
| Use DRACO denoising instead of the strongest listed baseline | Micrograph denoising · HA Trimer | SNR: -0.83 -> 3.69 | Helped |
| Use DRACO denoising instead of the strongest listed baseline | Micrograph denoising · Phage MS2 | SNR: -6.71 -> 0.23 | Helped |
| Use DRACO denoising instead of the strongest listed baseline | Micrograph denoising · RNA polymerase | SNR: 8.66 -> 12.21 | Helped |
| Add the N2N loss (ablation) | DRACO-B · 70S ribosome | F1 score: 0.786 -> 0.810 | Helped |
| Add the N2N loss (ablation) | DRACO-B · 70S ribosome | SNR: -4.94 -> -2.86 | Helped |
| Add the reconstruction loss (ablation) | DRACO-B · 70S ribosome | F1 score: 0.761 -> 0.810 | Helped |
| Add the reconstruction loss (ablation) | DRACO-B · 70S ribosome | SNR: -4.22 -> -2.86 | Helped |
Surprising findings
- Denoising SNR on the Phage MS2 and HA trimer sets is lower for DRACO than in some other settings, and the mask-ratio ablation shows non-monotonic behavior (0.625 best on RNA polymerase, 0.75 best on HA trimer).
Limitations
- Relies heavily on the performance of motion correction algorithms.
- Dataset focuses mainly on mainstream single-particle datasets; cryo-ET data not included in pre-training.
- Supports only micrograph-level downstream tasks; particle-level tasks such as pose estimation not addressed.
Experimental setup
Datasets: EMPIAR curated pre-training set (529 datasets) particle picking annotated set (80k micrographs) Human Apoferritin (EMPIAR-10421) HA Trimer (EMPIAR-10096) Phage MS2 (EMPIAR-10075) RNA polymerase (EMPIAR-11521) HCN1 (EMPIAR-10081) LetB (EMPIAR-10350) 70S ribosome (EMPIAR-10407) HIV tilt series
Benchmarks: HCN1 particle picking 70S ribosome particle picking LetB particle picking denoising SNR on four EMPIAR datasets micrograph curation on in-house 1,194-micrograph set
Metrics: precision recall F1 score resolution (A) SNR (dB) accuracy
Baselines: Topaz crYOLO CryoTransformer Detectron2 MAE Topaz-Denoise Low-pass filtering Miffi ResNet18
Builds on: masked autoencoder Noise2Noise Topaz-Denoise Detectron2
Closest papers by meaning
- CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference (NeurIPS 2024 · similarity 0.86)
- Efficient Online Training for Zero-Shot Time-Lapse Microscopy Denoising and Super-Resolution (AAAI 2025 · similarity 0.86)
- CryoGEM: Physics-Informed Generative Cryo-Electron Microscopy (NeurIPS 2024 · similarity 0.86)
- cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM (ICLR 2025 · similarity 0.85)
- Structure-Guided Masked Autoencoders for Ultra-High Resolution Scientific Image Understanding (NeurIPS 2026 · similarity 0.85)
- Vox-UDA: Voxel-wise Unsupervised Domain Adaptation for Cryo-Electron Subtomogram Segmentation with Denoised Pseudo-Labeling (AAAI 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.