Joint Denoising of Cryo-EM Projection Images using Polar Transformers

Joakim Andén, Justus Sagemüller · 2025

Read Joint Denoising of Cryo-EM Projection Images using Polar Transformers by Joakim Andén, Justus Sagemüller (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-11283.txt.

Introduction

Joakim Andén and Justus Sagemüller introduce a neural network design for cleaning up cryo-electron microscopy images in a 2025 paper. The method images tiny particles from many angles, but each shot is very noisy. Combining shots is key, yet most tools clean only one step of the pipeline at a time.

Their polar transformer works in polar coordinates and honors the rotational symmetry of the imaging process. It learns features that let it group, align, and clean many particle images at once. On simulated data with very low signal, it cut error by up to half compared with prior methods. The authors see promise for other imaging tasks that rebuild objects from noisy rotated views.

Catalog record

CREATORJoakim Andén, Justus Sagemüller
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE57 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-11283

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Joint Denoising of Cryo-EM Projection Images using Polar Transformers Joakim Andén, Justus Sagemüller Abstract Deep neural networks (DNNs) have proven powerful for denoising, but they are ultimately of limited use in high-noise settings, such as for cryogenic electron microscopy (cryo-EM) projection images. In this setting, however, datasets contain a large number of projections of the same molecule, each taken from a different viewing direction. This redundancy of information is useful in traditional denoising techniques known as class averaging methods, where images are clustered, aligned, and then averaged to reduce the noise level. We present a neural network architecture based on transformers that extends these class averaging methods by simultaneously clustering, aligning, and denoising cryo-EM images. Results on synthetic data show accurate denoising performance using this architecture, reducing the relative mean squared error (MSE) single-image DNNs by 45 % 45\% at a signal-to-noise (SNR) of 0.03 0.03 . 1 Introduction Many imaging problems amount to reconstructing some unknown quantity (often two-dimensional images or three-dimensional density maps) from several measurements obtained through a lossy, noisy forward process. One such problem is single-particle cryogenic electron microscopy (cryo-EM) ( Dubochet et al. 1982 ) , a molecular imaging technique that uses transmission electron microscopes to analyze the structure of molecules, typically proteins of interest for biological or medical research. This involves freezing a sample containing millions of copies of a single molecule in a thin layer of vitreous ice, imaging the frozen sample in an electron microscope, and then reconstructing the 3D structure of the molecule from the resulting 2D projection images, known as micrographs . The reconstruction process involves several steps, such as first extracting images of individual particles from the micrographs, denoising these images, constructing a low-resolution initial reference map, and then refining this into the final 3D density map ( Frank 1996 ) . An important challenge of cryo-EM imaging is the high noise level. Indeed, since the electron dose is kept low to avoid radiation damage, each projection contains only a weak signal and a large amount of shot noise ( Bendory et al. 2020 ) . Denoising is therefore a central task in any cryo-EM pipeline. For example, it is needed for visual inspection ( Singer & Sigworth 2020 ) , detecting contamination (outliers) ( Bhamre et al. 2016 ) , generating templates for particle picking ( Singer & Sigworth 2020 ) , identifying high-quality projection images ( Scheres 2015 ) , and as preprocessing for certain ab-initio reconstruction methods ( Singer & Shkolnisky 2011 ) . Traditional denoising methods can be divided into two categories. The first is based on applying a Wiener filter to the images ( Frank et al. 1996 ; Tang et al. 2007 ; Sindelar & Grigorieff 2011 ; Bhamre et al. 2016 ) . While this approach exploits prior information on the content of the images and is optimal in the class of linear estimators, it often results in a high level of blurring for the relevant noise levels. The second group of conventional denoising methods is known collectively as class averaging methods. These rely on the geometry of the problem described above – given the large number of images, they can be divided into clusters having approximately the same viewing direction up to in-plane rotation. Class averaging methods therefore first classify the images, where each cluster consists of images that are similar up to some in-plane rotation, then align those images rotationally and average them ( van Heel 1984 ; Park & Chirikjian 2014 ; Zhao & Singer 2014a ; Scheres 2015 ) . This average has less noise, no bias introduced by a filter, and can then be used for subsequent reconstruction tasks. However, the high noise level makes both classification and alignment very challenging ( Jensen 2001 ) . Like many other disciplines, cryo-EM processing has been improved in several ways by the introduction of methods from deep learning. This ranges from particle picking ( Wang et al. 2016 ; Wagner et al. 2019 ) and postprocessing of 3D densities ( Sanchez-Garcia et al. 2021 ) to implicit neural representations for 2D images ( Bibas et al. 2021 ; Nasiri & Bepler 2022 ; Kwon et al. 2023 ) and 3D structures ( Zhong et al. 2021 ; Nashed et al. 2021 ; Levy et al. 2022 ; Schwab et al. 2024 ) . DNNs have also been used for denoising, but in the setting of entire micrographs ( Bepler et al. 2020 ; Buchholz et al. 2019 ; Palovcak et al. 2020 ) . This problem is different from the denoising of projection images in several respects, such only being able to leverage information within a single micrograph. These methods also do not explicitly incorporate rotational symmetries of the data distribution, limiting their performance. This work proposes to solve the denoising problem by simultaneously denoising multiple images in a manner that extends traditional class averaging by leveraging prior information on the content of the projection images. The novel architecture rotationally aligns and integrates information from multiple cryo-EM projection images using a new convolutional attention mechanism that also guarantees the overall rotational equivariance of the system. This network, known as polar transformer , is shown to successfully cluster, align, and denoise sets of cryo-EM images on simulated datasets, yielding a reduction of 45\% in relative mean squared error (MSE) compared to deep neural networks operating on single projections of signal-to-noise ratio (SNR) as low as 0.03 . The rest of this paper is structured as follows. Section 2 describes the three main denoising tasks: single-image denoising, directional set denoising, and general set denoising. Then Section 3 reviews some of the existing literature on the subject of cryo-EM denoising. Sections 4 and 5 describe the construction of the polar transformer and some of its properties. Finally, Section 6 gives details of concrete models and how we trained and tested them. 2 Problem setup The denoising problem in cryo-EM aims at recovering a clean image x\in X from a noisy version y\in X . We represent all images through their pixel values on an L\times L raster, i.e., X=\mathbb{R}^{L\times L} . The clean images are assumed to be tomographic projections of a 3D potential density from a biological macromolecule. These are calculated using line integrals of that density along some chosen viewing direction. Following this, the images are filtered using a contrast transfer function (CTF) to simulate the optical properties of the microscope, and shifted randomly. We assume that these are individual particle projetions (i.e., we are past the particle-picking stage). While simple, the above model has proven to be relatively accurate ( Frank 1996 ; Vulović et al. 2013 ) and is used in a range of cryo-EM reconstruction methods ( Barnett et al. 2017 ; Scheres 2012 ; Punjani et al. 2017 ; Zhong et al. 2021 ) . To simulate the noisy projections y from the clean image x , we then use additive Gaussian noise, a common assumption in many reconstruction methods ( Punjani et al. 2017 ) . We will now consider three distinct denoising tasks. Single images The simplest setting is to consider each projection separately, potentially combining it with information from other projections later on.

Read the complete text