search

sct-pipeline / contrast-agnostic-softseg-spinalcord

Contrast-agnostic spinal cord segmentation project with softseg
MIT License
4 stars 3 forks source link

Re-train `contrast-agnostic` model with `EPI` data #83

Open rohanbanerjee opened 9 months ago

rohanbanerjee commented 9 months ago

Here, we discuss the re-training of the contrast-agnostic model using the entire data that was used originally and, in addition, the EPI data.

Method: fine-tuning or training from scratch.

Data: all the data included so far on the contrast-agnostic model + the EPI data.

Note: The EPI data need to have soft GT https://github.com/sct-pipeline/fmri-segmentation/issues/24

Location of the trained contrast-agnostic model checkpoint: duke/temp/muena/contrast-agnostic/final_monai_model/nnunet_nf=32_DS=1_opt=adam_lr=0.001_AdapW_CCrop_bs=2_64x192x320_20230918-2253

Currently the main.py script under the monai folder in the repository does not have the functionality of loading the weights from checkpoint and loads the runs from the wands run.

I will keep updating this issue with further details.

valosekj commented 9 months ago

Tagging @plbenveniste -- we applied the contrast-agnostic model on canproco PSIR/STIR images (context here).

EDIT by naga: These GT for PSIR/STIR were also manually corrected -- hence they can be used to fine-tune the contrast-agnostic model on these additional contrasts.