Hi Reader,

I haven't sent you a newsletter email for some time now. This is because there are major events happening in my personal life.

We just had our first kid, so I'm still trying to adapt to the new routine set by this cute little creature!

I also changed my office! I used to work from home, but now I am working in a coworking space. I'm hoping that this will help me deliver more value to the newsletter subscribers as well as our clients at PYCAD.

Now, back to the newsletter! I've got some cool insights for you today.

Dealing with Thin Anatomical Structures in Medical Imaging

One issue I noticed when building medical imaging models is that they do not perform as well on thin anatomical structures as they do on large anatomical structures.
​
For example, a simple UNet3D with dice loss would perform very well or close to perfect when segmenting structures like liver, kidney or lungs.
​
But the same setup would not perform very well on thin structures like vessels or nerves.
​
Lately, a paper came out about a new type of loss called Skelton Recall Loss.
​
The purpose of this loss function is to address the segmentation of thin anatomical structures.
​
The trained model with this loss achieved state of the art performance on 5 public datasets: Roads, DRIVE, Cracks, Toothfairy and TopCoW.
​
The researchers who developed this loss function released the code on github. Which is awesome!
​
But what is even more awesome is that they integrated their loss directly into nnUNet framework! This means that you can train AI models to do segmentation of thin anatomical structures using this loss function!
​
Their github repo contains a fork from the original nnUNet repo but with the skeleton recall loss integrated within!

You can read more about this technique in the original paper.

TotalSegmentator Has a New Update!

I have covered TotalSegmentator before here. In that article, I went in-depth about how TotalSegmentator works and if it's useful to your startup/company or not!

The team behind this tool keep updating it every few months with new features. So I try to keep you updated about these new features as well!

In the latest update, they added 65 new anatomical structures, all focused on the head and neck region. Things like: "head_glands_cavities", "head_muscles", "headneck_bones_vessels", and "headneck_muscles.

​

The opportunity that I noticed this week is 3D Slicer extensions for specific areas in the medical imaging field!

Basically, you can build a 3D Slicer extension and share it with both a free and commercial licenses.

Since 3D Slicer is already widely used, you can get your extension tested and validate pretty quickly.

If the users of your extension love your extension and want to use it commercially, then they have to pay you.

An example of this is an extension for virtual surgical planning for mandibular reconstruction.