The browser you are using is not supported by this website. All versions of Internet Explorer are no longer supported, either by us or Microsoft (read more here: https://www.microsoft.com/en-us/microsoft-365/windows/end-of-ie-support).

Please use a modern browser to fully experience our website, such as the newest versions of Edge, Chrome, Firefox or Safari etc.

Default user image.

Pablo Villanueva Perez

Senior lecturer

Default user image.

Reusability report : Unpaired deep-learning approaches for holographic image

Author

  • Yuhe Zhang
  • Tobias Ritschel
  • Pablo Villanueva Perez

Summary, in English

Deep-learning methods using unpaired datasets hold great potential for image reconstruction, especially in biomedical imaging where obtaining paired datasets is often difficult due to practical concerns. A recent study by Lee et al. (Nature Machine Intelligence 2023) has introduced a parameterized physical model (referred to as FMGAN) using the unpaired approach for adaptive holographic imaging, which replaces the forward generator network with a physical model parameterized on the propagation distance of the probing light. FMGAN has demonstrated its capability to reconstruct the complex phase and amplitude of objects, as well as the propagation distance, even in scenarios where the object-to-sensor distance exceeds the range of the training data. We performed additional experiments to comprehensively assess FMGAN’s capabilities and limitations. As in the original paper, we compared FMGAN to two state-of-the-art unpaired methods, CycleGAN and PhaseGAN, and evaluated their robustness and adaptability under diverse conditions. Our findings highlight FMGAN’s reproducibility and generalizability when dealing with both in-distribution and out-of-distribution data, corroborating the results reported by the original authors. We also extended FMGAN with explicit forward models describing the response of specific optical systems, which improved performance when dealing with non-perfect systems. However, we observed that FMGAN encounters difficulties when explicit forward models are unavailable. In such scenarios, PhaseGAN outperformed FMGAN.

Department/s

  • Synchrotron Radiation Research
  • NanoLund: Centre for Nanoscience
  • LTH Profile Area: Nanoscience and Semiconductor Technology
  • LU Profile Area: Light and Materials
  • LTH Profile Area: Photon Science and Technology

Publishing year

2024-02-15

Language

English

Pages

284-290

Publication/Series

Nature Machine Intelligence

Volume

6

Issue

3

Document type

Journal article

Publisher

Springer Nature

Topic

  • Atom and Molecular Physics and Optics
  • Computer graphics and computer vision

Status

Published

ISBN/ISSN/Other

  • ISSN: 2522-5839