cryoSENSE: Compressive Sensing Enables High-throughput Microscopy with Sparse and Generative Priors on the Protein Cryo-EM Image Manifold

Conference: CVPR 2026 arXiv: 2511.12931 Code: https://cryosense.github.io Area: Medical Image Analysis / Cryo-EM Keywords: Cryo-EM, Compressive Sensing, Diffusion Models, Sparse Priors, High-throughput Microscopy

TL;DR

This paper proposes cryoSENSE, the first computational framework for compressed cryo-EM imaging, demonstrating that protein cryo-EM images can be faithfully reconstructed from undersampled measurements under both sparse priors (DCT/Wavelet/TV) and generative priors (diffusion models), achieving up to 2.5× throughput gain while preserving 3D reconstruction resolution.

Background & Motivation

Background: Cryo-EM is a cornerstone technique in structural biology, yet modern direct electron detectors generate several gigabytes of data per second, far exceeding storage and transmission bandwidth. Current mitigation strategies include: (1) sub-frame summation, (2) shortening acquisition time followed by idle data transfer, and (3) post-acquisition compression — none of which resolves the real-time bandwidth bottleneck.

Limitations of Prior Work: The data deluge constrains practical throughput — instruments spend the majority of time waiting for data transfer rather than acquiring. Sub-frame summation sacrifices temporal resolution, while post-acquisition compression does not alleviate real-time bandwidth demands.

Key Challenge: Raw cryo-EM image data is highly structured (protein images reside on a low-dimensional manifold), yet existing workflows acquire and transmit data at full resolution, failing to exploit inherent redundancy.

Goal: Can compressive sensing be applied at the acquisition stage to reconstruct high-fidelity 2D particle images from undersampled measurements, thereby preserving 3D reconstruction resolution?

Key Insight: The method exploits two forms of low-dimensional structure in cryo-EM images — (1) sparsity under predefined bases, and (2) residence on a low-dimensional manifold learnable by diffusion models — to design two complementary reconstruction strategies.

Core Idea: Sparse priors + generative priors = complementary operating regimes for compressed cryo-EM imaging.

Method

Overall Architecture

cryoSENSE addresses the inverse problem of recovering \(\mathbf{x}^*\) from \(\mathbf{y} = \mathcal{A}(\mathbf{x}^*) + \boldsymbol{\eta}\), where \(\mathcal{A}\) is a known linear projection (pixel-domain or Fourier-domain masking). The framework supports two sampling schemes — pixel space and Fourier space — as well as two reconstruction strategies: sparse priors and generative priors.

Key Designs

1. Pixel-Domain and Fourier-Domain Masking Strategies:

   • Pixel-domain masking: Physically realizable via coded apertures or nanofabricated patterns.
   • Fourier-domain masking: Realizable via back-focal-plane modulation (phase plates, holographic gratings); supports uniform subsampling, annular, and radial spoke patterns.
   • Design Motivation: Each domain offers distinct advantages — Fourier-domain masking is more compatible with sparse priors, while pixel-domain masking is better suited to generative priors.

2. Sparse Prior Reconstruction (Proximal Gradient Descent):

   • Function: Solves the convex optimization problem \(\hat{\mathbf{x}} = \arg\min_{\mathbf{x}} \|\mathcal{A}(\mathbf{x}) - \mathbf{y}\|_2^2 + \lambda \Psi(\mathbf{x})\)
   • Three regularization options: DCT-basis sparsity, wavelet (WT) basis sparsity, and total variation (TV).
   • Alternates gradient steps with proximal operators (soft thresholding) until convergence.
   • Design Motivation: Sparse priors are universal and training-free, making them well-suited for moderate compression rates and Fourier-domain sampling.

3. Generative Prior Reconstruction (DDPM Posterior Sampling):

   • Function: A DDPM is trained on EMPIAR data to learn the cryo-EM image manifold; posterior sampling is performed via the Tweedie formula with modified reverse-diffusion guided sampling: \(\nabla_{\mathbf{x}_t} \log p(\mathbf{y}|\mathbf{x}_t) \simeq -\frac{1}{\sigma^2} \nabla_{\mathbf{x}_t} \|\mathcal{A}(\hat{\mathbf{x}}_0) - \mathbf{y}\|_2^2\)
   • Nesterov accelerated gradients are used to improve sampling efficiency.
   • Design Motivation: Generative priors leverage data-driven manifold structure, imposing weaker assumptions than sparse priors, and perform better at higher compression rates and under pixel-domain sampling.

Loss & Training

• Sparse reconstruction: training-free; purely optimization-based.
• DDPM training: standard score matching on EMPIAR cryo-EM data.
• Posterior sampling: combines unconditional score with measurement-consistency gradients.

Key Experimental Results

Main Results — 2D Reconstruction Quality

Pixel-Domain Masking (K=4, C≈2):

Prior	LPIPS↓	SSIM↑
Sparse-DCT	0.11	0.59
Sparse-WT	0.13	0.59
Sparse-TV	0.20	0.64
Gen-DDPM	0.12	0.50

Fourier-Domain Masking (Radial spoke, C≈2.5):

Prior	LPIPS↓	SSIM↑
Sparse-DCT	0.12	0.72
Sparse-WT	0.11	0.71
Sparse-TV	0.30	0.37
Gen-DDPM	0.11	0.63

3D Volumetric Reconstruction

Compression Factor	Best Prior (Pixel-Domain)	Best Prior (Fourier-Domain)	3D FSC Resolution Retention
1.5×	Gen-DDPM	Sparse-DCT	Near-perfect
2.5×	—	Sparse-DCT	Maintained
>2.5×	Degraded	Degraded	Reduced

Ablation Study / Key Comparisons

Property	Sparse Priors	Generative Priors
Preferred Sampling Domain	Fourier-domain	Pixel-domain
Optimal Compression Range	Moderate (≤2.5×)	Higher (suitable for extreme undersampling)
Requires Training	No	Yes
Biological Signal Preservation	✓	✓

Key Findings

• Core Finding: Sparse priors favor Fourier-domain sampling at moderate compression rates, while generative priors favor pixel-domain sampling at higher compression rates — the two strategies are complementary.
• Fourier-domain sparse reconstruction maintains near-perfect FSC resolution at a 2.5× compression factor.
• CryoDRGN conformational heterogeneity analysis achieves 80–88% clustering consistency on reconstructed images.
• ModelAngelo atomic model building yields backbone RMSD of only 2.1–2.3 Å on reconstructed images.

Highlights & Insights

• Hardware-Software Co-design: Rather than post-acquisition compression, cryoSENSE applies compressive sensing at the point of data generation, addressing the bandwidth bottleneck at its source.
• Complementary Prior Framework: The work systematically evaluates two major classes of priors under two sampling schemes, providing clear operational guidelines for practitioners.
• Biological Downstream Validation: Beyond 2D reconstruction quality, the framework is validated on core biological tasks including 3D reconstruction, conformational analysis, and atomic model building.
• Physical Realizability: Fourier-domain masking is achievable with existing phase plate technology, and pixel-domain binning is already a standard feature of modern detectors.

Limitations & Future Work

• The current work constitutes computational validation rather than physical hardware experiments.
• DDPM training requires existing cryo-EM datasets and may not generalize well to entirely novel protein classes.
• All methods degrade at extreme compression rates (>2.5×).
• Adaptive sampling strategies — dynamically adjusting masking patterns based on image content — remain unexplored.

Related Work & Insights

• Compressive sensing is well-established in MRI (CS-MRI); this work extends the paradigm to cryo-EM.
• Prior compressive sensing work in 4D-STEM provides a precedent within the electron microscopy community.
• Posterior sampling frameworks for diffusion models (DPS, DDRM) are effectively adapted to the cryo-EM setting.

Rating

• Novelty: ⭐⭐⭐⭐⭐ First compressive sensing framework for cryo-EM, opening an entirely new research direction.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Exceptionally comprehensive — multiple priors × multiple sampling schemes × multiple compression rates × downstream biological validation.
• Writing Quality: ⭐⭐⭐⭐⭐ Theoretical derivations are clear; experimental design is systematic.
• Value: ⭐⭐⭐⭐⭐ Transformative potential for high-throughput cryo-EM imaging.

Related Papers

• [NeurIPS 2025] Multiscale Guidance of Protein Structure Prediction with Heterogeneous Cryo-EM Data
• [CVPR 2026] Solving a Nonlinear Blind Inverse Problem for Tagged MRI with Physics and Deep Generative Priors
• [ICCV 2025] CryoFastAR: Fast Cryo-EM Ab initio Reconstruction Made Easy
• [CVPR 2026] X-WIN: Building Chest Radiograph World Model via Predictive Sensing
• [ICLR 2026] CryoNet.Refine: A One-step Diffusion Model for Rapid Refinement of Structural Models with Cryo-EM Density Map Restraints