This repo is a quick and dirty pass at getting a U-net up and running using tensorflow and a first pass at using a U-net to segment our microscope images of DRG neurons cultured in vitro. Importantly, it is configured to run on an AMD GPU. If you're running an NVIDIA card, you'll probably need to tweak a few things to get it running.
Lots of the model training code was written by Gemini. By no means is the currrent architecture the best for the job, just what the robot and I came up with.
The ones to pay attention to are:
- ./amd_nd2_model_train.py: trains the model (note: models not included due to file size constraints)
- ./6 image AO train/*.nd2: microscope images used to train the model we're currently using in Nikon
- Each includes raw fluorescent intensity data for DAPI and TUJI (channels 2 and 3, zero-indexed), as well as 3 binary layers (neuronal soma, neuronal nuclei, and neurites)
- ./test/*.nd2: microscope image not in train dataset with DAPI and TUJI present (channels 0 and 1), no binary layers
- ./test_amd_nd2_model.py: visualizes predictions made by the trained model