Updated: Apr 17, 2020
This poster will be presented by Luciano Lucas, PhD on Tuesday December 11th from 1:30 - 3:00 PM. Poster number P2709.
The technology presented in this poster is a part of our latest release, Aivia Cloud. Request a demo to try it out: https://www.drvtechnologies.com/demo
Long term imaging of dynamic sub-cellular phenomena in living cells is limited by phototoxicity and photobleaching. This problem is exacerbated in microscopes with higher spatiotemporal resolution, particularly super-resolution imaging. The machine learning approach presented here aims to drastically reduce the amount to light required to image intracellular protein dynamics with structured illumination microscopy (SIM). As we show, for selected protein distributions our approach allows acquisition of over 10,000 images (>700 imaging volumes) without significant photobleaching, without appreciably compromising signal to noise ratio (SNR) or spatial resolution, and without making any modifications to the underlying microscope hardware.
For this application the deep learning (DL) model is trained using pairs of high and low quality images for a given field of view. During training the DL model learns to statistically relate the intrinsic features of the low quality input images to the high quality output images. The DL model used in this study is a 3D fully convolutional network based on Deep Residual Channel Attention Networks (RCAN). Training was done using thirty six 3D image pairs depicting cells with fluorescently labelled actin. Images were acquired on an instant structured illumination microscope (iSIM) and were composed of 1.9Kx1.5Kx14 voxels. The input images were captured using 3.4, 6.7, 9, 13, 29 and 49 W/cm2 and the ground truth data was acquired using 0.37 kW/cm2. Routine live cell iSIM imaging is done at ~100 W/cm2 – which significantly limits the length of recordings. The trained model was applied to 3D images, akin to the input images used for training, that had been acquired at arrange of low excitation power conditions.
We found that images acquired using as little as 13 W/cm2 could be restored with resolution similar to ground truth images (acquired at 0.37 kW/cm2). The trained model was also applied to very long (>700 volumetric time points) recordings of cells with labelled mitochondria using low excitation power (20 W/cm2). While the input images were unusable for quantitative or qualitative analysis, the DL restored images are of a similar quality to the images acquired using 10x more excitation power.This shows that our trained DL model has some general applicability as it was trained using actin labelled cells and was successful at improving the SNR and resolution of images captured using low excitation power containing either fluorescently labelled actin or mitochondria.
The present machine learning enabled approach provides microscopists with a new solution for the problems of phototoxicity and photobleaching.
Authors: H.Sasaki, J.Chen, Y.Su, C.Huang, J.S.Lee, H.Shroff, L.A.Lucas
J.Chen and H.Shroff are a part of Advanced Imaging and Microscopy Resource, National Institutes o fHealth, Bethesda, MD
H.Shroff is also a part of Section on High Resolution Optical Imaging, National Institute of Biomedical Imaging and Bioengineering, Bethesda, MD
Other authors are part of DRVision Technologies LLC, Bellevue, WA