The effect of imaging parameters and computational image restoration on information retrieval from fluorescence microscopy images
- Abstract number
- 163
- Presentation Form
- Poster
- Corresponding Email
- [email protected]
- Session
- Poster Session
- Authors
- Kristine Medcalf (1), Siân Culley (1)
- Affiliations
-
1. King's College London
- Keywords
Fluorescence Microscopy, Image Quality, Image Restoration, Denoising, Segmentation
- Abstract text
A standard quantitative fluorescence microscopy experiment typically involves sample preparation, image acquisition, image processing and finally image analysis. Within each of these stages there are many parameters that can be optimised, yet there are no reliable metrics available for measuring how successful these optimisations have been. For example, during image acquisition the illumination intensity and exposure time are typically adjusted by the researcher until they perceive that the image ‘looks good’. For data that does not look good, novel methods such as deep learning-based image restoration approaches can be applied to acquired images so that they ‘look better’.
However, this visual assessment carries certain risks. In a bid to improve perceived image quality, researchers may use imaging conditions (such as high illumination intensity) that can result in artefacts such as photobleaching and, in the case of live samples, phototoxicity. Conversely, overestimating the quality of acquired data can make it impossible to analyse accurately. While application of image restoration techniques, such as denoising, will frequently improve the appearance of images, such methods may be prone to generating artificial structure, particularly in very challenging datasets.
Here, we examine the impact of image acquisition settings and application of denoising methods on the accuracy of segmenting of sfGFP-labelled organelles in the fission yeast Schizosaccharomyces pombe. By repeatedly segmenting the same cells, but with methodically adjusted image acquisition and post-processing parameters, we can start to identify which experimental regimes have the greatest impact on accurate information retrieval and what gains can be made by denoising.