From Stars to Molecules: AI Guided Device-Agnostic Super-Resolution Imaging

Abstract

Super-resolution imaging has revolutionized the study of systems ranging from molecular structures to distant galaxies. However, existing super-resolution methods require extensive calibration and retraining for each imaging setup, limiting their practical deployment. We introduce a device-agnostic deep-learning framework for super-resolution imaging of point-like emitters that eliminates the need for calibration data or explicit knowledge of optical system parameters. Our device-agnostic modeling utilizes diverse, numerically simulated dataset encompassing a broad range of imaging conditions, enabling generalization across different optical setups. Once trained, the model reconstructs super-resolved images directly from a single resolution-limited camera frame with superior accuracy and computational efficiency compared to state-of-the-art methods. We experimentally validate our approach using a custom microscopy setup with controllable ground-truth emitter positions. We also demonstrate its versatility on astronomy and single-molecule localization microscopy datasets, achieving unprecedented resolution without prior information. Our findings establish a pathway toward universal, calibration-free super-resolution imaging, expanding its applicability across scientific disciplines.

Figures (8)

• Figure 1: Schematic representation of the device-agnostic modeling approach and its application to intensity images of point-like emitting sources. a The model is trained using numerically simulated data pairs comprising resolution-limited noisy images alongside their super-resolved counterparts. Each training sample represents a unique combination of underlying optical parameters, such as the width of the point spread function and the signal-to-noise ratio. b Following training, this model is applied to enhance the resolution of experimental images acquired using a real-life imaging system.
• Figure 1: A detailed visualization of the optical setup with full control over the ground truth of emitters. In the illumination path, we generate an incoherent light collimated by a microscopic objective $\text{MO}_{\text{s}}$. The subsequent iris diaphragm $\text{D}$ and two achromatic doublets, $\text{L}_{\text{s1}}$ and $\text{L}_{\text{s2}}$, improve the illumination homogeneity and reduce undesired reflections from the passive DMD parts. In the preparation path, the DMD-induced ground-truth mask is re-imaged by an achromatic doublet $\text{L}_{\text{p}}$ and a high-resolution microscope objective $\text{MO}_{\text{p}}$, creating point-like emitters in the emitter plane. Lastly, this plane is projected by a low-resolution microscope objective $\text{MO}_{\text{I}}$ onto a camera with a mounted spectral filter.
• Figure 2: Performance of super-resolving methods on simulated data. The dependence of the mean absolute error on a the signal-to-noise ratio (SNR), b the width of a Gaussian point spread function (PSF), c the number of emitters in the image (concentration), and d the continuous transition between a Gaussian and an Airy PSF, respectively. The resulting averages of the Richardson-Lucy algorithm (green), its variant with total variation regularization (blue), Deep-STORM neural networks (purple), and the DAMN model (red) are accompanied by their 90% confidence intervals over the test set. Panel a is provided with a secondary horizontal axis recalculating the SNR values to the peak-to-noise ratio (PNR). Across all panels, the DAMN model consistently outperforms the alternative approaches by up to two orders of magnitude. The optical parameters not investigated in a given panel have the following values: SNR $= 500$, the average noise intensity $= 10$, the concentration $= 50$, and the Gaussian PSF with $\sigma = 2$ px.
• Figure 2: The aberration effects on the DAMN reconstruction performance. (a) The relative increase of mean absolute error (MAE) due to the aberration strength and numerical aperture values. The presented results were evaluated relative to the baseline errors obtained for the non-aberrated point spread function (PSF). (b) Visualization of the distorted point spread functions for each aberration type with $0.2 \lambda$ strength, alongside the non-aberrated baseline and 0.2 numerical aperture.
• Figure 3: Schematic illustration of the optical setup used to collect experimental data pairs. A ground-truth mask is imposed on the digital micromirror device (DMD) by configuring its mirrors. An incoherent illumination light reflected by these mirrors impinges a high-resolution preparation system. The DMD-imposed mask is re-imaged into the front sample plane of the preparation system, creating point-like emitters with the intended spatial distribution. The imaging part of the setup, comprised of a low-resolution microscope objective, images the sample-plane emitters onto a camera. The resulting camera-captured intensity image and the DMD-imposed mask represent the experimental data pairs.