ML Deep Dive: nnU-Net for Segmentation of Medical Data

One of the best tools to build an automatic segmentation model for 2D and 3D medical data is nnU-Net. Here’s why.

nnU-Net is not just a model architecture, it’s in fact a whole process for building the right training pipeline that fits your dataset.

Let me explain.

nnU-Net is a semantic segmentation method that automatically adapts to a given dataset.

It will analyze the provided training cases and automatically configure a matching U-Net-based segmentation pipeline.

Given a new dataset, nnU-Net will systematically analyze the provided training cases and create a ‘dataset fingerprint’.

nnU-Net then creates several U-Net configurations for each dataset:

• 2d: a 2D U-Net (for 2D and 3D datasets).
• 3d_fullres: a 3D U-Net that operates on a high image resolution (for 3D datasets only).
• 3d_lowres → 3d_cascade_fullres: a 3D U-Net cascade where first a 3D U-Net operates on low resolution images and then a second high-resolution 3D U-Net refined the predictions of the former (for 3D datasets with large image sizes only).

Upon release, nnU-Net was evaluated on 23 datasets belonging to competitions from the biomedical domain.

Despite competing with handcrafted solutions for each respective dataset, nnU-Net’s fully automated pipeline scored several first places on open leaderboards!

Btw, nnU-Net is the main tool behind TotalSegmentator!

You can find out more about nnU-Net in its original paper. You can train your own nnU-Net model using their github repo.

​

Cost Effective Deployment Method for Deep Learning Models

Lately I’ve been experimenting with Serverless GPUs by building deep learning models endpoints. Specifically, I wanted to try a service that offered per-use billing.

The product that I used is RunPod.

With RunPod you can easily deploy deep learning models that require heavy computation on the GPUs.

You can then turn your deep learning model into an API endpoint that you can run synchronously or asynchronously.

Aaaand you pay by the second!

Meaning that if your model takes 1 second to do the computation, and you call your model 100 times per day, then you only pay for 100 seconds!

This is an extremely efficient way to do computation on the cloud and it’s very cost effective. Which would translate into lots of costs saving. Especially for startups and small companies with a tight budget.

Here’s a high level overview of how you deploy your deep learning model on a RunPod endpoint:

• You build a docker image that uses your model to do inference. It needs to contain a RunPod handler file inside.
• You push your docker image to a container registry. It can be public or private registry. An example would be Docker Hub.
• You create what is called a “Template” on RunPod platform. This will specify several parameters such as the name of the docker image to pull from the container registry.
• You create an endpoint that makes use of the previously created Template and you specify the type of GPU you would like to use, as well as the number of workers you want to run and some other parameters.

And Voilà! You now have your model deployed as an API that you can call from anywhere and you’re only billed for the time that the API was actually used.

I believe that most deep learning models in the future will be deployed in this fashion. The value is just so clear!

​

Tweet (or X) of the Day

​

Allied Market Research @allied_market AI in Medical Imaging Market: CAGR Forecast! 🇦🇺Australia 40% 🇨🇦Canada 39.8% 🇮🇳India 38.7% 🇰🇷South Korea 37.3% 🇪🇸Spain 36.5% 🇮🇹Italy 35.7% 🇿🇦France 35.2% 🇫🇷Japan 32.6% 🇨🇳China 31.7% 🇺🇸U.S. 29.2% 🇩🇪Germany 28.8% 🇬🇧U.K. 28% 📈Find out what CAGR other countries will have:… https://twitter.com/i/web/status/1709078426756538600 7:30 AM • Oct 3, 2023 3 Retweets 7 Likes

​

​

Meme of the Day 😂