Calibration-free single-frame super-resolution fluorescence microscopy

This repository contains data and scripts required for reproducing the results presented in the paper Calibration-free single-frame super-resolution fluorescence microscopy by Anežka Dostálová, Dominik Vašinka, Robert Stárek, and Miroslav Ježek.

The paper is available on: * arXiv: https://arxiv.org/abs/2505.13293. * bioRxiv: https://www.biorxiv.org/content/10.1101/2025.05.20.655080v1

cnn-model

This folder contains the developed calibration-free convolutional neural network (CFCNN) for super-resolving image reconstruction from a single intensity frame, and a usage example.

experimental-data

In this folder, the data and script necessary to recreate Figure 2 and Table I of the paper are provided. Fig. 2 shows the experimentally acquired fluorescence microscopy images together with their ground truth, and the visual comparison of the reconstructed outputs from the Richardson-Lucy (R-L) deconvolution algorithm, multi-emitter fitting (MEF) using ThunderSTORM, and our CFCNN. These are stored in HDF5 format, respectively, with separate files for each experimental image ("Image1.h5", "Image2.h5", "Image3.h5").
Table I provides a quantitative comparison of the reconstruction quality in terms of mean absolute error and Kullback-Leibler divergence. These metrics are computed between the output of each reconstruction method and the ground truth for each experimental image.

resolution-test

This folder contains scripts for the analysis of the resolution achievable by the CFCNN. More detailed comments are included within the scripts.
"generateinputs.py" generates synthetic data for resolution testing, including input images with varying signal-to-noise ratios (SNRs) and corresponding reference images, and stores the outputs in HDF5 files.
"processcnnoutputs.ipynb" evaluates the resolving ability of the CFCNN model on the generated synthetic data.
"processinputs_cnn.py" processes the input datasets by the CFCNN and saves the results in HDF5 files.

star-test

This folder contains scripts for the recreation of Figure 3 of the paper. Synthetic data are generated and analyzed by our CFCNN, the R-L deconvolution, and MEF using ThunderSTORM for a broader and more systematic evaluation of the performance beyond the presented experimental images. More detailed comments are included within the scripts.
"generatefigure.ipynb" recreates the Fig. 3.
"generateinputs.py" generates synthetic 2D images of a star-shaped pattern with varying SNRs and corresponding ground truth images.
"processinputscnn.py" processes the input datasets by the CFCNN and saves the results in HDF5 files.
"rlmodule.py" provides functions for generating Gaussian kernels and performing Richardson-Lucy deconvolution.
"rlprocess.py" applies the R-L algorithm.