Preparing data and train model to detect/segment spinal cord

[jcohenadad]

• Contact @andreanne-lemay to understand where the code for findcord_tumor is in https://ivadomed.org
• Develop code to prepare data for training SC detection. This step will depend on the state of the code that @andreanne-lemay used
• Generate preprocessed data for training cord segmentation model.
• Package this data + version.
• Train model. Important information: in order to avoid having big blobs far away from the cord at inference (eg caused by artifacts, OOD) there is a feature in the postprocessing methods of ivadomed that can select the biggest object:

https://github.com/ivadomed/ivadomed/blob/8fd599a07e4a20e4a4ca8fdc676c31f9a41df438/ivadomed/postprocessing.py#L102

• Package model as new repository. Name: model_segment_sc_mp2rage

[uzaymacar]

Hello @charleygros, Andreanne mentioned that you might have the code for training the YOLO model for SC detection. If so, can you please refer me to the code? Thanks in advance!

[uzaymacar]

Also @andreanne-lemay mentioned that if we follow the config file for the cascaded architecture, and if we add the path to the model which is available on GH, it would work and crop around the SC. We'll just have to test this.

[jcohenadad]

@andreanne-lemay @charleygros do you have any updates for us? this project is becoming quite high priority-- many thanks!

[charleygros]

@charleygros do you have any updates for us? this project is becoming quite high priority-- many thanks!

Copy that, will try to help today

[jcohenadad]

@charleygros do you have any updates for us? this project is becoming quite high priority-- many thanks!

Copy that, will try to help today

quick reply until @uzaymacar gives more context: @uzaymacar and i had a quick chat with @andreanne-lemay who clarified a lot of things-- we've updated most of the related issues to add more context/information. So we're good for now 👍

[jcohenadad]

Following up on #8 (comment), here is the context:

• Andreanne/Charley did not use YOLO for the tumor segmentation. Instead, they used a UNet to segment the spinal cord, which was then used to create a mask for subsequent cropping.