Gsoc-2026: TotalSegmentator v2 Integration for Multi-Organ Segmentation and Surface Texture Pipeline

[Kaipapurandeswarreddy]

Hi @paulojamorim @rmatsuda @tfmoraes,

While working on the Surface Texture integration, I realized that
the quality of surface texture mapping depends heavily on having
clean, isolated organ surfaces. This led me to explore a natural
pipeline extension for InVesalius.

Proposed: TotalSegmentator v2 Integration

InVesalius currently supports brain segmentation (FastSurfer),
trachea, and mandible. However, there is no support for chest,
abdominal, or cardiovascular organ segmentation.

I am proposing to integrate TotalSegmentator v2 which segments
104 anatomical structures (lungs, heart, liver, kidneys, spine,
vessels etc.) from CT and MRI.

I am a machine learning student with hands-on deep learning
experience. I have previously built a brain tumor segmentation
project using deep learning (segResNet)which you can check here(Iam capable or not)
https://github.com/Kaipapurandeswarreddy/BrainTumorseg
This gives me confidence that I can implement the TotalSegmentator integration correctly and efficiently.

Why this pairs perfectly with Surface Texture:
CT Scan → TotalSegmentator → Clean organ mask
→ Surface generation → Surface Texture (Rust)
→ 3D organ → VRML/OBJ export
Key insight — why segmentation improves texture quality:

Threshold-based surfaces (Marching Cubes on raw CT) include mixed
tissues — bone, air, soft tissue all together. When raycasting shoots
rays along surface normals, it samples across mixed tissue boundaries,
producing noisy textures.

With TotalSegmentator segmentation first:

• Organ boundaries are precisely defined
• Surface normals point exactly into the correct tissue region
• Raycasting rays stay within the organ's actual coordinate space
• Result: clean, anatomically accurate texture with no cross-tissue noise

Technical approach:

• ROI subset strategy — CPU friendly, no GPU required
• Lazy model download — weights downloaded only when needed
• Task-based UI (Bones, Organs, Cardiovascular)
• Direct DICOM/NIfTI input — no conversion needed

Would this direction align with InVesalius's ?
Happy to start implementing if this sounds promising.

[paulojamorim]

Hi @Kaipapurandeswarreddy

I think that's the way to go.

Have you done anything with Total Segmentor yet? Focus on the texture now.

I already answered that in the other thread: #1255

[Kaipapurandeswarreddy]

Thank you @paulojamorim! I will focus on completing the Surface Texture integration first. I have already gathered the technical information about TotalSegmentator v2 (I studied the similiar applications like 3D slicer and related documents how they manage the TotalSegmentor integration )the ROI subset strategy, lazy model loading, and the pipeline connections. I will implement TotalSegmentator after the texture work is complete, as you suggested. I will follow your guidance from #1255 and update progress there.

Meanwhile, if you agree on the TotalSegmentator integration into InVesalius could you share any specific integration strategy or preferred approach you have in mind? For example:
Should it follow the same pattern as the existing FastSurfer integration?
Any preference on how model weights should be managed (lazy download vs bundled)?
Should the UI follow the same dialog pattern as Brain Subpart Segmentation?
Any guidance before I start would help me align with the project's architecture from the beginning.

[paulojamorim]

Should it follow the same pattern as the existing FastSurfer integration?

Yes! Any additional dependencies should be avoided. Pre-processing and post-processing should be implemented using NumPy, SciPy, and scikit-image, which are already dependencies of InVesalius.

Any preference on how model weights should be managed (lazy download vs bundled)?

The weights will be added to this repository (https://github.com/invesalius/weights
). They should be converted to JIT and ONNX to avoid including architecture code and to make loading faster. If the user does not have the weights in the directory, they should be downloaded automatically.

Should the UI follow the same dialog pattern as Brain Subpart Segmentation?
Any guidance before I start would help me align with the project's architecture from the beginning.

Yes! It would be good to have a window where the user can select which organs will be segmented. Please propose an interface.

Remember, you can submit up to two different proposals.

[Kaipapurandeswarreddy]

Hi @paulojamorim, thank you for the detailed guidance on TotalSegmentator integration! I wanted to confirm since TotalSegmentator is not currently listed in the GSoC 2026 ideas list, can I still submit it as a second GSoC proposal as a new idea? Or would it need to be added to the ideas list first?
Thank you!

[paulojamorim]

Yes, you can submit the idea for the totalsegmentator.

[Kaipapurandeswarreddy]

Hi @paulojamorim

As the GSoC submission deadline is approaching, I have prepared my proposal for the TotalSegmentator v2 integration.

I would really appreciate it if you could take a look and share any feedback to ensure my approach aligns well with the project expectations here is the link of my proposal:
https://docs.google.com/document/d/1bWPHBaBHit6YFOT0orx4vtrZc8DDFQrWRCXURYZv-ZI/edit?usp=sharing
Also, could you please let me know whether this idea would be considered small, medium, or large scope for submission in Gsoc?

Thank you!

[paulojamorim]

This is a medium-sized project (12 weeks). You can submit it.

Thank you.