Segment nerve section to compute axon density

[hermancollin]

Our collaborator wants to compute the axon density in his images. We have a shared OneDrive folder with him containing hundreds of images; segmentations and morphometrics are already computed. He wants to compute the axon density. To do so, we will train a model to segment the total nerve area per image. For example, the img below contains 5 fascicles. The mask we want will delimit the inner area inside the perineurium

After this, we will be able to do 2 things:

• estimate axon density per image
• filter false positives located outside the nerve area

We will thus be able to compute a new metric for this data and improve the current segmentation quality. Here is a somewhat exhaustive TODO list:

1. select ~30 images and start gathering annotations
2. send to Osvaldo and wait for feedback
3. train model (probably nnunet)
4. inference on all his data
5. approximate axon density: script to compute nerve area from the masks and read morphometrics CSVs

[hermancollin]

For future reference, this project is almost complete. More details on this repo: www.github.com/axondeepseg/model_seg_nerve

[hermancollin]

Update: our collaborator provided a lot of manual measurements so we will be able to accurately evaluate our method.