Training Model Head/Neck CTs with groundtruth segmentation and volume .nrrd files

[danielsh3n]

I am hoping to train a 3D U-Net model on lymphatic/fatty tissue in the head and neck. Regarding my dataset, I have 18 cases along with their manual segmentations (done in 3DSlicer) in .nrrd files. Is it possible for me to use MONAILabel to train and test on this dataset?

@hw-ju

[diazandr3s]

Hi @danielsh3n,

This is a good question. Certainly, you can train a MONAI Label model for this.

Please follow these instructions

They are simple.

Once you create the Python env, install monai label and download the Radiology app (https://github.com/Project-MONAI/MONAILabel/tree/main#getting-started-with-monai-label).

• Update the label names here with the consistent label indexes for the segment you already have
• As training from scratch, disable or set to false this flag: https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/lib/configs/segmentation.py#L72
• Organize the dataset in the MONAI Label structure
• Train

let us know how that goes

[diazandr3s]

Hi @danielsh3n,

It is great! You already have most of the things done.

To trigger the training, please follow these instructions: https://youtu.be/3HTh2dqZqew?si=fQVJ_9fSuMgpoi2w&t=194

As you already have the labels, you don't need to submit anything in Slicer. Just connect to the server and trigger the training with the default configuration.

Let us know how that goes,

[danielsh3n]

Hi @diazandr3s ,

Thanks for the quick response! Glad to hear that I'm almost there! Unfortunately, when I tried starting the server with the new path to my dataset, I get these errors:

[diazandr3s]

Hi @danielsh3n,

It seems an error from the Python env. Can you please create a fresh Python environment and try again?
BTW, no need to use double quotes for the studies argument.

Let us know,

[danielsh3n]

Hi @diazandr3s ,

I just set up with a new environment and connected to the server in Slicer and have these outputs into Slicer and in terminal (WSL2 on Ubuntu 22.04). Thoughts? Also, after training the model, how would I go about testing the model's predictions on a new head/neck CT?

(monailabel) [email protected]

Thanks!

[diazandr3s]

Hi @danielsh3n,

Sorry for my late reponse.
I checked the logs and it seems there is an issue with the default transforms or the labels.
Are there any empty labels? Did you make sure the label (CTV) is represented by 1 in all the label volumes?
Do you know the Hounsfield units range in which these regions are more visible?

Let us know,