I Trained a Model that Segments the Aorta

Lately I wanted to experiment with building a machine learning model that can segment parts of the human body that are not easy to identify.

Usually, I see a lot of examples online about segmenting different organs of the body such as the liver, the lungs or the heart.

But these organs typically have a wide and more or less clear structure.

I wanted to segment a part of the body that looks challenging. So I chose the aorta.

The aorta is the main artery that distributes oxygenated blood to all parts of the body.

It has a strange looking shape when visualized in 3D. I joked about this in a LinkedIn post where I compared the aorta to the worms from "Men in Black" movie 😂

To carry this experiment, I chose nnUNet tool. I have shared before how this tool is designed to make it easy to build a segmentation model for any medical imaging dataset. You will mostly need to configure some folders structure and you'll be able to run the training without much configuration from your side.

I also wanted to use a dataset that had a format that I haven't seen being used that often with nnUNet. So I opted for a dataset that had all patients volumes as well as the labels (segmentation masks) in NRRD format.

You can find this dataset inside this Google Drive folder.

Below you can see a sample output of my nnUNet model on a test case.

I am using 3D Slicer software here to visualize the segmentation output next to the patient case.

As you can see, even though the aorta shape is difficult to get, the model is still able to segment it very well.

There are 2 reasons why I was able to get such results:

1. The dataset is actually very well refined. The people who did the segmentation definitely did a great job here. It should also be noted that this dataset is actually pretty small with a little over 50 patients cases. But since the annotations are well refined, the model can learn the right features to be able to segment this part of the human body.
2. nnUNet is a powerful segmentation tool.

One final note about nnUNet. Although, this tool is powerful and easy to use to train a segmentation model, it can quickly become confusing because of all the available options.

Also, the tool was very well developed for training. But when it comes to inference and deployment, you'll find yourself spending a ton of time going through the documentation, open and closed issues on their github repo, as well as the actual code.

This is where PYCAD comes into the picture. We can help you accelerate the usage of this tool to actually build a model and deploy it into a development or production environment. Details in the next section.

X Post of the Day

​

freeCodeCamp.org @freeCodeCamp There are many helpful ways you can use artificial intelligence. For example, doctors can use it to detect diseases like cancer in patients. In this course a physician + developer teaches you how to apply AI models to medical imaging using TensorFlow. https://www.freecodecamp.org/news/medical-ai-models-with-tensorflow-tutorial/ freecodecamp.org Medical AI Models with TensorFlow – Tutorial Machine learning is transforming many industries, including healthcare. Artifici... 7:2 PM • Nov 26, 2023 71 Retweets 353 Likes Read 5 replies

​