Label free prediction fnet
fnet is a deep-learning method that can be used to predict fluorescent labels from brightfield data. This page contains information to help you train Fnet networks in google colab using your own images.
Fnet was created by the Johnson laboratory at the Allen Institute and you can access the original paper here:
fnet code and documentation from the original publication (used here) is freely available in GitHub.
Please also cite these original papers when using fnet with our notebook.
To train an Fnet network you need matching 3D images of cells in a bright-field and in a fluorescent channel (see example below).
Training fnet in the provided colab notebook requires 3D stacks in two channels, e.g. brightfield and fluorescent which can be acquired for example on confocal microscopes. To prepare a training set, the stacks need to be split into individual folders by channel. The signal files (brightfield) and their respective targets (fluorescence) must be in the same order in their respective folders. It is therefore advisable to number source-target pairs or to give the files the same names.
The dataset provides as an example with our notebooks was generated to predict mitochondrial structure (labelled for the outer membrane translocase TOM20) from transmitted light images recorded in parallel on a Leica SP8 confocal microscope. For this purpose, HeLa ATCC cells were seeded on fibronectin-coated 8-well chamber slides (Sarstedt, Germany, 1.5 x 104 cells/well). Cells were grown for 16h at 37°C and 5% CO2 in Dulbecco's modified Eagle's medium containing 4.5 g/l glucose, 10% FBS and 1% GlutaMAX (all acquired from Thermo Fisher, USA).
Chemical fixation:
For fixation, we employed a protocol shown to preserve the cytoskeleton and organelles (adapted from here). In brief, the culture medium was directly removed with PHEM buffer containing 3% methanol-free formaldehyde (Thermo Fisher, USA) and 0.2% EM-grade glutaraldehyde (Electron Microscopy Sciences, USA) and the samples were incubated for 1 h at room temperature. Cells were washed thrice with PBS, quenched with 0.2% sodium borohydride in PBS for 7 min and washed again thrice with PBS.
Immunofluorescence staining:
Fixed specimens were permeabilised and blocked using 0.25% TX-100 (Sigma Aldrich, Germany) and 3% IgG-free BSA (Carl Roth, Germany) in PBS for 1.5 h. Cells were labelled for TOM20 using 5 µg/ml rabbit anti-TOM20 primary antibody (#sc-11415, Santa Cruz, USA) and 10 µg/ml donkey anti-rabbit-secondary antibody (Alexa Fluor 594 conjugated, #A32754, Thermo Fisher, USA) in PBS containing 0.1% TX-100 and 1% BSA for 1.5 h each. Samples were rinsed twice and washed 3x with PBS (5 min) after each incubation step. After the staining procedure, the samples were post-fixed for 10 min using 4% methanol-free formaldehyde in PBS and washed thrice with PBS subsequently.
Image acquisition:
Image stacks were acquired on a Leica SP8 confocal microscope (Leica Microsystems, Germany) bearing a 63x/1.40NA oil objective (Leica HC PL APO CS2). The pixel size was set to 90 nm in xy-dimensions and 150 nm in z (32 slices) and fluorescence image stacks were recorded using 561 nm laser excitation (4x line average, 762 ns pixel dwell time) and an emission window of 570 – 630 nm collected by a PMT. The corresponding transmitted light image stack was recorded in parallel using a transmitted light PMT.
| Network | Link to example training and test dataset | Direct link to notebook in Colab |
|---|---|---|
| Label-free prediction (fnet) | here |  |
or:
To train fnet in Google Colab:
-
Download our streamlined ZeroCostDL4Mic notebooks
-
Open Google Colab
-
Once the notebook is open, follow the instructions.