Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Self-supervised prior learning improves structured illumination microscopy resolution

Ze-Hao Wang, Tong-Tian Weng, Long-Kun Shan, Xiang-Dong Chen, Guang-Can Guo, Fang-Wen Sun, Tian-Long Chen

TL;DR

Results show that learned priors can substantially extend SIM resolution and robustness, enabling rapid, large-scale imaging with STORM-level detail and a self-distilled variant that improves noise robustness while preserving high resolution at extremely low photon counts.

Abstract

Structured illumination microscopy (SIM) is a wide-field super-resolution technique normally limited to roughly twice the diffraction-limited resolution ($\approx 100$--$200$~nm). Surpassing this bound is a classic ill-posed inverse problem: recovering high-frequency structure from band-limited raw data. We introduce SIMFormer, a fully blind SIM reconstruction framework that learns a powerful, data-driven prior directly from raw images via self-supervision. This learned prior regularizes the solution and enables reliable extrapolation beyond the optical transfer function cutoff, yielding an effective resolution of approximately 45~nm. We validate SIMFormer on synthetic data and the BioSR dataset, where it resolves features such as flattened endoplasmic reticulum lipid bilayers previously reported to require STORM-level resolution. A self-distilled variant, SIMFormer+, further improves noise robustness while preserving high resolution at extremely low photon counts. These results show that learned priors can substantially extend SIM resolution and robustness, enabling rapid, large-scale imaging with STORM-level detail.

Self-supervised prior learning improves structured illumination microscopy resolution

TL;DR

Results show that learned priors can substantially extend SIM resolution and robustness, enabling rapid, large-scale imaging with STORM-level detail and a self-distilled variant that improves noise robustness while preserving high resolution at extremely low photon counts.

Abstract

Structured illumination microscopy (SIM) is a wide-field super-resolution technique normally limited to roughly twice the diffraction-limited resolution (--~nm). Surpassing this bound is a classic ill-posed inverse problem: recovering high-frequency structure from band-limited raw data. We introduce SIMFormer, a fully blind SIM reconstruction framework that learns a powerful, data-driven prior directly from raw images via self-supervision. This learned prior regularizes the solution and enables reliable extrapolation beyond the optical transfer function cutoff, yielding an effective resolution of approximately 45~nm. We validate SIMFormer on synthetic data and the BioSR dataset, where it resolves features such as flattened endoplasmic reticulum lipid bilayers previously reported to require STORM-level resolution. A self-distilled variant, SIMFormer+, further improves noise robustness while preserving high resolution at extremely low photon counts. These results show that learned priors can substantially extend SIM resolution and robustness, enabling rapid, large-scale imaging with STORM-level detail.

Paper Structure

This paper contains 28 sections, 2 equations, 15 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Discussion
  4. Materials and Methods
  5. Acknowledgments
  6. Figures
  7. Extended Data Figures

Figures (15)

  • Figure 1: Fig. 1. Overview of SIMFormer. (A) The image encoder extracts representations from SIM raw images and is updated through self-supervised training. (B) SIMFormer is a self-supervised learning architecture based on masked autoencoders. (C) The representation is mapped to emitters using the decoder and Softplus activation. (D) The representation is mapped to light patterns using the decoder and low-rank coding (LRC). (E) The representation is mapped to the background using the decoder and Gaussian blur. (F) The PSF is generated from fixed noise using a convolutional network.
  • Figure 2: Fig. 2. Evaluation of estimated emitters for synthetic data. (A--C) Visual comparison of the super-resolution reconstruction of conventional SIM, Sparse-SIM, Blind-SIM, SIMFormer, and SIMFormer+ on the raw SIM images. (D--F) Quantitative analysis results using NRMSE and MS-SSIM. (G--I) Resolution obtained from the decorrelation analysis. (A) Visualization results of emitters with synthetic filament structures. (B) Visualization results of emitters with synthetic microtubule structures. (C) Visualization results for emitters with synthetic ring structures. Scale bar: 0.5 µ m. (D) NRMSE and MS-SSIM results of synthetic filament emitters. (E) NRMSE and MS-SSIM results of synthetic microtubular emitters. (F) NRMSE and MS-SSIM results of synthetic ring structure emitters. (G) Resolution of estimated emitters with synthetic filament structures. (H) Resolution of estimated emitters with synthetic microtubule structures. (I) Resolution of estimated emitters with synthetic ring structures. Box plots show the median (centre line), 25th--75th percentiles (box), and 1.5 $\times$ IQR whiskers; outliers are plotted as dots.
  • Figure 3: Fig. 3. Experiments on the BioSR dataset. (A--D) Visualization results of the BioSR data from wide field imaging, BioSR SIM, Blind-SIM, SIMFormer, and SIMFormer+. (E--H) Resolution obtained from decorrelation analysis. (A) Comparison results for ER. (B) Comparison results for CCPs. (C) Comparison results for MTs. (D) Comparison results for F-actin. Scale bar: 0.5 µ m. (E) Resolution for ER. (F) Resolution for CCPs. (G) Resolution for MTs. (H) Resolution for F-actin. Box plots show the median (centre line), 25th--75th percentiles (box), and 1.5 $\times$ IQR whiskers; outliers are plotted as dots.
  • Figure 4: Fig. 4. Light pattern estimation. (A--C) Comparison of estimated light patterns from synthetic data. The emitters are (A) synthetic filament, (B) synthetic microtubule, and (C) synthetic ring. (D) Quantitative evaluation of the light pattern using NRMSE. (E--H) Comparison of estimated light patterns from BioSR data. The emitters are (E) CCPs, (F) ER, (G) MTs, and (H) F-actin.
  • Figure 5: Fig. 5. Noise robustness of synthetic data. Comparison of the super-resolution reconstruction of conventional SIM, Sparse-SIM, Blind-SIM, SIMFormer, and SIMFormer+ on the raw SIM images with average photon count equal to 0.5, 1, 10, 100, 500, and 1000. (A) Comparative presentation of high and low photon-number results for synthetic filaments. (B) Quantitative evaluation using NRMSE and MS-SSIM. Scale bar: 0.5 µ m. Lines represent the mean; error bars denote the standard error of the mean.
  • ...and 10 more figures