add instructions on using contrast-agnostic weights with nnunet

quick_start_guides/use_contrast_agnostic_weights_with_nnUNet.md

@@ -0,0 +1,43 @@
+## Using contrast agnostic weights with nnUNet
+
+This guide provides instructions on how to train nnUNet models with the contrast-agnostic pre-trained weights. Ideal use cases include using the pretrained weights for training/finetuning on any spinal-cord-related segmentation task (e.g. lesions, rootlets, etc.). 
+
+### Step 1: Download the pretrained weights
+
+Download the pretrained weights from the [latest release](https://github.com/sct-pipeline/contrast-agnostic-softseg-spinalcord/releases) of the contrast-agnostic model. This refers to the `.zip` file of the format `model_contrast_agnostic_<date>_nnunet_compatible.zip`. 
+
+> [!WARNING]  
+> Only download the model with the `nnunet_compatible` suffix. If a release does not have this suffix, then that model weights are not directly compatible with nnUNet.
+
+
+### Step 2: Modify the plans file
+
+In your `$nnUNet_preprocessed/<dataset_name_or_id>/nnUNetPlans.json`, change the following keys for the `3d_fullres` configuration: 
+
+- `strides`: 
+
+```
+[[1, 1, 1], [2, 2, 2], [2, 2, 2], [2, 2, 2], [2, 2, 2], [1, 2, 2]]
+```
+
+- `patch_size`:
+
+```
+[64, 192, 320]
+```
+
+### Step 3: Train/Finetune the nnUNet model on your task
+
+Provide the path to the downloaded pretrained weights:
+
+```bash
+
+nnUNetv2_train <dataset_name_or_id> <configuration> <fold> -pretrained_weights <path_to_pretrained_weights> -tr <nnUNetTrainer_Xepochs>
+
+```
+
+> [!IMPORTANT]  
+> * Training/finetuning with contrast-agnostic weights only works when for 3D nnUNet models. 
+> * Ensure that all images are in the RPI orientation before running `nnUNetv2_plan_and_preprocess`. This is because the updated `patch_size` refers to patches in RPI orientation (if images are in different orientation, then the patch size might be sub-optimal).
+> * Ensure that `X` in `nnUNetTrainer_Xepochs` is set to a lower value than 1000. The idea is the finetuning does not require as many epochs as training from scratch because the contrast-agnostic model has already been trained on a lot of spinal cord images (so it might not require 1000 epochs to converge).
+