Read on Twitter
Imaging of live cells and clinical tissue can be limited by lack of labels. Why not measure their intrinsic density and anisotropy? Our work on label-free analysis of biological architecture @czbiohub is published. Congrats @symguo, @LiHao_Yeh , @JennyFolkesson, and all authors! https://twitter.com/eLife/status/1296190956606193666
Quantitative label-free imaging with phase and polarization (QLIPP for short) works across biological scales. This neat experiment by @ieivanov1 shows density and anisotropy at the scale of organelles
Above movie shows a time-lapse at a single plane of a volume. 3D pseudocolor rendering of density and anisotropy begins to identify specific organelles. Reconstruction and rendering by @LiHao_Yeh.
QLIPP enables sensitive and high-resolution imaging of tissue slices as well. We can identify specific regions of brain tissue, detect changes in myelination, and measure orientation of axons - all without label. Figure by @LiHao_Yeh and @symguo.
We are excited for the impact this method can have on study of archival brain tissue. Data in collaboration with @LabNowakowski
https://twitter.com/LabNowakowski/status/1296229354398076928?s=20
https://twitter.com/LabNowakowski/status/1296229354398076928?s=20
QLIPP is modular and can be added to any microscope. The software is open-source: https://github.com/mehta-lab/reconstruct-order
Analyzing specific structures from these information-dense images is hard. We developed deep learning models to predict specific structures, e.g., myelination in human brain tissue. These models are inspired by Greg Johnson's work @AllenInstitute.
The prediction accuracy with our 2.5D U-Net architecture approaches the accuracy of 3D U-Net architecture, but requires significantly less memory to train. Results by @symguo
and @JennyFolkesson.
and @JennyFolkesson.
The data used for training the models ( https://www.ebi.ac.uk/biostudies/BioImages/studies/S-BIAD25) and the code to train the models ( https://github.com/czbiohub/microDL) are available.
We hope above methods are broadly useful. Let us know if you need inputs to implement the computational imaging method or image translation models. (End)
