@article{ReinhardHelmerichBorasetal.2022,
  author    = {Reinhard, Sebastian and Helmerich, Dominic A. and Boras, Dominik and Sauer, Markus and Kollmannsberger, Philip},
  title     = {ReCSAI: recursive compressed sensing artificial intelligence for confocal lifetime localization microscopy},
  series = {BMC Bioinformatics},
  volume    = {23},
  journal   = {BMC Bioinformatics},
  number    = {1},
  doi       = {10.1186/s12859-022-05071-5},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-299768},
  year      = {2022},
  abstract  = {Background Localization-based super-resolution microscopy resolves macromolecular structures down to a few nanometers by computationally reconstructing fluorescent emitter coordinates from diffraction-limited spots. The most commonly used algorithms are based on fitting parametric models of the point spread function (PSF) to a measured photon distribution. These algorithms make assumptions about the symmetry of the PSF and thus, do not work well with irregular, non-linear PSFs that occur for example in confocal lifetime imaging, where a laser is scanned across the sample. An alternative method for reconstructing sparse emitter sets from noisy, diffraction-limited images is compressed sensing, but due to its high computational cost it has not yet been widely adopted. Deep neural network fitters have recently emerged as a new competitive method for localization microscopy. They can learn to fit arbitrary PSFs, but require extensive simulated training data and do not generalize well. A method to efficiently fit the irregular PSFs from confocal lifetime localization microscopy combining the advantages of deep learning and compressed sensing would greatly improve the acquisition speed and throughput of this method. Results Here we introduce ReCSAI, a compressed sensing neural network to reconstruct localizations for confocal dSTORM, together with a simulation tool to generate training data. We implemented and compared different artificial network architectures, aiming to combine the advantages of compressed sensing and deep learning. We found that a U-Net with a recursive structure inspired by iterative compressed sensing showed the best results on realistic simulated datasets with noise, as well as on real experimentally measured confocal lifetime scanning data. Adding a trainable wavelet denoising layer as prior step further improved the reconstruction quality. Conclusions Our deep learning approach can reach a similar reconstruction accuracy for confocal dSTORM as frame binning with traditional fitting without requiring the acquisition of multiple frames. In addition, our work offers generic insights on the reconstruction of sparse measurements from noisy experimental data by combining compressed sensing and deep learning. We provide the trained networks, the code for network training and inference as well as the simulation tool as python code and Jupyter notebooks for easy reproducibility.},
  language  = {en}
}
@article{BerberichKurzReinhardetal.2021,
  author    = {Berberich, Andreas and Kurz, Andreas and Reinhard, Sebastian and Paul, Torsten Johann and Burd, Paul Ray and Sauer, Markus and Kollmannsberger, Philip},
  title     = {Fourier Ring Correlation and anisotropic kernel density estimation improve deep learning based SMLM reconstruction of microtubules},
  series = {Frontiers in Bioinformatics},
  volume    = {1},
  journal   = {Frontiers in Bioinformatics},
  doi       = {10.3389/fbinf.2021.752788},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-261686},
  year      = {2021},
  abstract  = {Single-molecule super-resolution microscopy (SMLM) techniques like dSTORM can reveal biological structures down to the nanometer scale. The achievable resolution is not only defined by the localization precision of individual fluorescent molecules, but also by their density, which becomes a limiting factor e.g., in expansion microscopy. Artificial deep neural networks can learn to reconstruct dense super-resolved structures such as microtubules from a sparse, noisy set of data points. This approach requires a robust method to assess the quality of a predicted density image and to quantitatively compare it to a ground truth image. Such a quality measure needs to be differentiable to be applied as loss function in deep learning. We developed a new trainable quality measure based on Fourier Ring Correlation (FRC) and used it to train deep neural networks to map a small number of sampling points to an underlying density. Smooth ground truth images of microtubules were generated from localization coordinates using an anisotropic Gaussian kernel density estimator. We show that the FRC criterion ideally complements the existing state-of-the-art multiscale structural similarity index, since both are interpretable and there is no trade-off between them during optimization. The TensorFlow implementation of our FRC metric can easily be integrated into existing deep learning workflows.},
  language  = {en}
}