Welcome to the spine segmentation with nnUNet course 👊

This course consists of training a machine learning model for spine segmentation (multiclass). The model that will be used is a UNet, but not just a normal UNet, we will use the framework nnUNet to train the model as it is a state of the art model for medical image segmentation.

Check out our blog post about nnUNetv2 here.

The model has been built on 1089 CT scans taken from the TotalSegmentator dataset. The dataset used ofor the course is hosted in Kaggle and can be accessed here.

Materials

This repository is hosting some materials that helps you prepare your dataset, train the model and even some demos of how you can use or deploy your model with different technologies such as QT, Streamlit or Trame.

• Prepare your dataset with the different modules available in dataset.py file.
• Life time access to the dataset in Kaggle.
• The model trained during the course.
• Sample STLs for 3D visualization.
• Multiple demos where you can deploy your model after training.

In the figure below you can see the training graph:

And in the table below you can find some useful information:

N. train data	N. valid data	N. eval data	N. epochs	Train dice	Valid dice	Eval dice	Model ckpt	Valid samples
608	152	217	250	91.44%	88%	85%	ckpts	samples

To get an idea about the whole process steps, you can check this guidelines doc.

Our Medical Imaging Resources

• Free medical imaging ebook.
• Weekly medical imaging newsletter.
• > 50 medical imaging blog posts.
• 30:30 medical imaging notebooks.