Calibration-free single-frame super-resolution fluorescence microscopy

TL;DR

The paper tackles the limitation of calibration-dependent, multi-frame super-resolution by introducing a calibration-free CNN that reconstructs SR images from a single diffraction-limited frame. Trained entirely on synthetic data, the CFCNN generalizes across PSFs, noise, NA, and aberrations, enabling plug-and-play SR on standard wide-field setups. On terrylene-in-PVA samples, it achieves up to $7\times$ enhancement beyond the Rayleigh limit and resolves separations down to $35\,\text{nm}$ at $580\,\text{nm}$ with a $150\,\text{ms}$ frame, outperforming Richardson-Lucy, ThunderSTORM, and DECODE across SNRs. This approach promises faster, less calibration-intensive SR for dense or fast-moving samples and could shorten acquisition pipelines in SMLM workflows while preserving high fidelity at the single-frame level.

Abstract

Molecular fluorescence microscopy is a leading approach to super-resolution and nanoscale imaging in life and material sciences. However, super-resolution fluorescence microscopy is often bottlenecked by system-specific calibrations and long acquisitions of sparsely blinking molecules. We present a deep-learning approach that reconstructs super-resolved images directly from a single diffraction-limited camera frame. The model is trained exclusively on synthetic data encompassing a wide range of optical and sample parameters, enabling robust generalization across microscopes and experimental conditions. Applied to dense terrylene samples with 150 ms acquisition time, our method significantly reduces reconstruction error compared to Richardson-Lucy deconvolution and ThunderSTORM multi-emitter fitting. The results confirm the ability to resolve emitters separated by 35 nm at 580 nm wavelength, corresponding to sevenfold resolution improvement beyond the Rayleigh criterion. Furthermore, we demonstrate strong generalization ability of the developed model and its resilience across a broad range of noise levels, numerical apertures, and optical aberrations. By delivering unprecedented details from a single short camera exposure without prior information and calibration, our approach enables plug-and-play super-resolution imaging of fast, dense, or light-sensitive samples on standard wide-field setups.

Figures (5)

• Figure 1: Schematic illustration of the experimental fluorescence microscope setup and the CFCNN application to the camera image. Terrylene molecules in the sample emit fluorescent light upon laser excitation. The fluorescence is collected by the microscope objective, separated from the excitation light on the dichroic mirror, and detected by an sCMOS camera. The resolution-limited camera image is directly processed by the CFCNN without requiring any prior information or calibration. The CFCNN reconstructs a super-resolved image, revealing individual molecules with significantly enhanced resolution.
• Figure 2: A visual comparison of the results from three state-of-the-art methods for super-resolution imaging and localization, and our calibration-free neural network (CFCNN). All resulting images in panels (b)--(f) are convolved with a narrow Gaussian filter for better visualisation. Panels (a) show the input intensity images of organic terrylene molecules from a fluorescence microscope acquired by an sCMOS camera, with different molecule powers and densities. (b) Results of the reconstruction by the Richardson-Lucy deconvolution algorithm provided with a specific PSF. (c) The outputs of multi-emitter fitting using ThunderSTORM with specified camera settings. (d) Resulting images by DECODE trained on a full set of parameters of the imaging setup and sample. (e) Super-resolved images reconstructed by our CFCNN using no prior information or calibration. (f) An estimation of the ground truth using temporal information beyond the single intensity image.
• Figure 3: Synthetic data with emitters arranged in the shape of a star. Images in panels (b)--(f) and (h)--(l) are convolved with a narrow Gaussian filter for better visualization. (a, g) Examples of the input image with $\textrm{SNR}=15.8$ and $\textrm{SNR}=125.9$, respectively, obtained as the ground truth (f, l) convolved with a Gaussian PSF with the width of $2$ px on the $50 \times 50$ pixel grid and the corresponding noise backgrounds added. (b, h) The image reconstructions by the Richardson-Lucy deconvolution algorithm. (c, i) The outputs of multi-emitter fitting using ThunderSTORM. (d, j) The resulting images produced by DECODE. A uniquely trained model was used for each SNR. (e, k) The reconstructed images by the CFCNN without recalibration and retraining. (m) Mean absolute error between the outputs and the ground truth as a function of SNR for R-L (orange), MEF (green), DECODE (red), and the CFCNN (blue). (n) Kullback–Leibler divergence of the outputs from the ground truth as a function of SNR. In the case of DECODE, a separate model was trained on the corresponding SNR and PSF for each data point.
• Figure 4: The level of super-resolution achieved by the CFCNN depending on the signal-to-noise ratio using a simulated dataset. The resolvable separation $\delta_\textit{R}$ between two identical emitters is expressed in terms of the Rayleigh distance. A significant spatial resolution enhancement relative to the Rayleigh criterion is achieved across the entire SNR range, exceeding a sevenfold improvement for higher SNR values and attained solely from a single camera image.
• Figure 5: Reconstruction performance of the CFCNN for varying numerical aperture values in terms of mean absolute error (a) and Kullback–Leibler divergence (c). The results demonstrate that the model maintains a consistently high reconstruction quality across the whole range. MAE (b) and KLD (d) between the CFCNN outputs of aberrated star pattern (same as in Fig. \ref{['fig:synth_data']}) and the ground truth as a function of the aberration strength $W$ show robustness to moderate optical aberrations. Astigmatism (orange), spherical aberration (red), coma (green), and defocus (blue) were studied individually.