Artificial intelligence enhances super-resolution microscopy

03 Apr 2024

New Anglo-German generative model is more efficient than established approaches.

Computational methods and algorithmic data processing are an important aspect of super-resolution microscopy, but a new approach to the task could offer a route to enhanced imaging.

A joint project between the Center for Advanced Systems Understanding (CASUS) at Germany's Helmholtz-Zentrum Dresden-Rossendorf, Imperial College London and University College London has used generative artificial intelligence (AI) to improve the quality of super-resolution images.

Generative AI is the principle behind current text- or image-creating applications such as ChatGPT or Stable Diffusion, whereby a machine learning operation does not simply perceive and classify the data it receives, but expands upon it to create further data within controlled restraints.

The CASUS project's open-source generative AI algorithm, called a Conditional Variational Diffusion Model (CVDM), "improves the quality of images by reconstructing them from randomness" according to the researchers, and is computationally less expensive than established diffusion models.

"Diffusion models have long been known as computationally expensive to train; some researchers were recently giving up on them exactly for that reason," commented Artur Yakimovich from CASUS. "But new developments like our CVDM allow minimizing of 'unproductive runs' which do not lead to the final model. By lowering the computational effort and hence power consumption, this approach may also make diffusion models more eco-friendly to train."

Super-resolution microscopy, in which imaging is possible below the notional diffraction limit, could be one valuable area of application for CVDM, since techniques such as structured illumination microscopy still face inherent hurdles relating to information loss and noise despite the strides made in performance.

Immediate applicability to medical microscopy

CVDM is intended to minimize the unproductive runs during computation, and designed so that during the training phase the model is capable of finding the optimal training for noise reduction in a particular task on its own, without programmers needing to work through overall noise reduction "schedules" first via trail and error.

In tests applying CVDM to super-resolution images and structured illumination super-resolution microscopy, the project found that it could improve resolution by 4.42 percent when compared against existing diffusion probabilistic methods, and by 26.27 percent compared to a regression-based method.

Tests using clinical images of epithelial and urinary cells showed that a CVDM approach has "immediate applicability to medical microscopy," noted the project.

"We believe our approach has some new unique properties, namely high flexibility and speed at a comparable or even better quality compared to other diffusion model approaches," said Artur Yakimovich. "In addition our CVDM provides direct hints where it is not very sure about the reconstruction; a very helpful property that suggests the path to addressing these uncertainties in new experiments and simulations."