Julien Cohen-Adad

Use squeletonization to close myelin

2

As pointed out in #397, there is a need to manually correct myelin masks, including closing them so that the flood fill algo works. There might be ways to "close"...

Test other software package offering "easy label" and segmentation of images

2

### Context A couple of recent publications and software packages offer the possibility to upload an image in a GUI (e.g. ImageJ, FIJI, etc.), perform an easy labeling, train a...

Train model on synthetic data, agnostic to contrast

2

The idea would be to generate synthetic data from real data + ground truth, and use approaches such as GANs. The goal would be to generate images with infinite amount...

Browser or cloud solutions for processing large datasets

2

### Context Wish list: - user uploads image (ideally drag and drop) - processing is done on the cloud or on local station - user can download results ### Methodology...

**Describe the problem** It would be nice to test integrity, ie. is the prediction correct? **Proposed solution** Add an assertion below this code: https://github.com/neuropoly/axondeepseg/blob/e32dcfc582ab9cb0b886af57aa9075a95cf958ee/AxonDeepSeg/integrity_test.py#L61 Every time we update the model...

Create Docker-based workflow to generate model

12

The current problem is that models are not easy to reproduce given that it depends on what data the user downloads, what version of ADS, etc. To mitigate the issue,...

Update contribution guideline

11

we need to establish conventions on PR titles and description. We could get inspiration from: https://github.com/neuropoly/spinalcordtoolbox/wiki/Contribution-Guidelines#submit-a-pull-request

Create unit test data of dummy axons

2

Currently, images of synthetic axons are [generated by hand on a software like paint, and then put on OSF.](https://github.com/neuropoly/axondeepseg/pull/399#discussion_r541314003) We shoudl avoid that. Instead, we should have python code that...

Implement new method to separate myelins

1

Instead of the watershed which is fairly unstable and inaccurate, we could think of other (or complimentary) ways to separate myelin sheaths that touch. Similarly to what has been implemented...

For an application on nerve segmentation, there is a working branch that implements transfer learning: https://github.com/neuropoly/axondeepseg/tree/vs/transfer-nerve-seg This branch is deemed to disappear at some point once ivadomed becomes a dependency...