heartchambers_highres module

[HenrikPLind]

Hi there,

I've been exploring the TotalSegmentator tool and really appreciate all the work put into it. I noticed there is a heartchambers_highres module, but I haven’t been able to find any specific documentation about it.

What I’ve tried so far:

Looked through the README and the GitHub Wiki (if available)
Searched open and closed issues here on GitHub
Checked the codebase for hints on training details (data source, hyperparameters, etc.)

Questions:

Does documentation for heartchambers_highres exist somewhere else, and I’ve just missed it?
Could you provide (or point me to) information on how this module was trained (e.g., dataset, training strategy, etc.)?
Are there any recommended settings or usage tips specific to heartchambers_highres (e.g., input image requirements, typical post-processing steps, etc.)?
Thank you in advance for any pointers or references. .

Best regards,
Henrik

[wasserth]

Hi Henrik,
there is indeed no real documentation about the training of this model.
It was trained on two datasets: the totalsegmentator dataset but in addition on a high resolution dataset of cardiac images.
It should work well on most CT images. No specific post processing is required.
There is not much documentation because in the end that does not really help you. You have to run the model on your own data and see how well it performs and if it satisfies your requirements. If I tell you it was trained on 500 or 5000 images does not really make any difference for you.

[HenrikPLind]

Thank you for your swift reply! I completely agree that visual inspection is an excellent way to validate a model's performance on specific data. However, I believe information about the training dataset can provide valuable insights into the model's generalizability.

Given that the literature on heart chamber segmentation is somewhat sparse, and many existing studies rely on datasets with fewer than 250 CT images, knowing the scale and diversity of the datasets used for this model could help users better understand its potential limitations and strengths. Additionally, details about the model architecture used for the heartchambers_highres module would be highly beneficial, as it could serve as a reference for others working on similar challenges in this domain.

Thank you again for your time and support!

Best regards,
Henrik

[tkerby]

Just to add to this, I've been using this model to reconstruct my own heart from CT scans and 3D print it with great results. The only thing that's a bit disappointing is that the superior and inferior vena cava and the left atrial appendage aren't included, as it's really only the chambers and myocardium. It would be great to have those added to the high res model.

Picture below to show the results - you'll see the obvious low resolution bits included from the whole body models