Recently, a collaboration between researchers at the University of Tromsø and the Indian Institute of Technology (Indian School of Mines) has used Deep Learning to improve the quality in terms of noise of nanoscopy images. In the general case, this can be done by training a DL-based algorithm to learn the mapping between noisy images to their clean version (also called ground-truth) by having a large set of training pairs. However, in nanoscopy this is not possible as ground-truth data is challenging to obtain in the first place.
In their work published recently1, the team went through the route of simulating fluorescence pairs microscopy images of 3D structures for a wide variety of conditions such as emission wavelength and numerical aperture among others, one with noise and another without. These images were then processed using MUSICAL2, a super-resolution algorithm for obtaining super-resolution.
As the noise will undoubtedly impact the reconstruction quality, the problem was trying to obtain a MUSICAL reconstruction from the noisy image as close as possible to its noise-free counterpart. After the system was trained, it was tested for real microscopy images, showing that the system was able of generalise its simulated-based knowledge to real experimental data.
The work, which you can read here, contains several experimental results and a detailed description of the network used.
References:
- Suyog Jadhav, Sebastian Acuña, Ida S. Opstad, Balpreet Singh Ahluwalia, Krishna Agarwal, and Dilip K. Prasad, "Artefact removal in ground truth deficient fluctuations-based nanoscopy images using deep learning," Biomed. Opt. Express 12, 191-210 (2021)
- K. Agarwal and R. Macháň, “Multiple signal classification algorithm for super-resolution fluorescence microscopy,” Nat. Commun. 7(1), 13752 (2016).