Straight from the Heart: Teaching AI to Sculpt Blood Vessels from Shadows
Mapping the Vascular Maze
Medical imaging is, at its core, an attempt to see the invisible. Among the most challenging of bodily landscapes to map are our blood vessels — a dense network of tubes curving, branching, and narrowing through organs and tissue. From the large, meandering aorta to delicate brain capillaries, these vessels hold the keys to diagnosing and treating conditions like aneurysms, strokes, and heart failure. But accurately capturing their shape in 3D for simulation, diagnosis, or surgical planning is an extraordinarily difficult task.
The 2025 arXiv paper “Parametric shape models for vessels learned from segmentations via differentiable voxelization” by Alina F. Dima and colleagues marks a significant leap forward in how artificial intelligence can not only observe but understand and replicate these anatomical structures. By integrating deep learning with a mathematical tool called differentiable voxelization, the researchers have enabled a machine to learn vessel shapes from imaging data with unprecedented precision and flexibility — without even needing a ground truth of the shape to begin with.
Why This Research Inspires Awe
What’s striking about this work is that it tackles three of medical image analysis’s toughest challenges at once: reducing data annotation burdens, improving geometric accuracy, and providing representations that are both interpretable and manipulable. This isn’t just a better segmentation algorithm. It’s a comprehensive system that can translate the fuzzy outlines of blood vessels in medical scans into clean, manipulable 3D models, complete with smooth centerlines and radii, as if the AI were a skilled sculptor carving from fog.
Most crucially, this method does not merely generate vessel segmentations — it produces parametric models. That means the output is a compact mathematical representation that can be adjusted, analyzed, and simulated in other clinical applications. This is the leap from mere image processing to an anatomical language machines can speak fluently and use creatively.
How It Works: Teaching the Machine to Shape Vessels
The technical brilliance of the approach lies in three key ideas:
Parametric Representation with B-Splines: The vessel’s centerline is modeled using smooth mathematical curves called cubic B-splines. These allow the vessel path and radius to be represented continuously and sparsely, capturing complex shapes with just a few control points. Think of it like fitting a flexible backbone through a cloud of dots, where that backbone can stretch and bend smoothly.
Mesh and Volume Conversion: From these splines, a 3D mesh (a network of tiny triangles forming a surface) is generated. This mesh is then “voxelized” — converted into a 3D grid of cubes, the format used in medical scans. But unlike traditional voxelization, which is a one-way street, the team introduces a differentiable version. This means the process is reversible and can be optimized using gradient descent — a backbone of deep learning.
Multi-Stage Training Pipeline: Fitting a vessel’s shape involves several moving parts: the centerline, the radius, and fine adjustments to match irregular shapes. The researchers trained the system in four stages: first fitting the centerline, then the radius, correcting the centerline, and finally adjusting each radial direction. This staged training addresses the notorious issue of conflicting gradients, which can derail training when too many variables are tuned simultaneously.
The system is trained on CT and MRI scans, using vessel segmentations as reference. However, crucially, it does notrequire ground truth centerlines or radii — which are rarely available. Instead, it learns these features by minimizing a novel loss function based on voxel-level differences between the predicted shape and the reference segmentation.
The Team Behind the Innovation
This groundbreaking research was made possible by an international and interdisciplinary collaboration involving:
Technical University of Munich (TUM) and TUM University Hospital, Germany
University of Zurich, Switzerland
Stanford University, USA
ZHAW — Zurich University of Applied Sciences, Switzerland
Imperial College London, UK
German Cancer Consortium (DKTK) and the Munich Center for Machine Learning (MCML)
The lead author, Alina F. Dima, worked alongside a talented group including Suprosanna Shit, Huaqi Qiu, Robbie Holland, Tamara T. Mueller, Fabio Musio, Kaiyuan Yang, Bjoern Menze, Rickmer Braren, Marcus Makowski, and Daniel Rueckert — spanning AI, medical imaging, and clinical radiology expertise.
Real-World Results: Aorta to Aneurysms
The team tested their framework on three diverse datasets: Aorta24 (human CT scans of the aorta), TopCoW (brain vessels from the Circle of Willis), and MouseAneurysm (MRI scans of mouse aortas with aneurysms).
On the Aorta24 dataset, the model achieved a Dice score of 94.66%, with just 673 vertices and 1326 mesh faces, compared to 76,803 vertices and 153,600 faces from traditional segmentation-plus-marching-cubes pipelines. This means their method produced equally accurate, but far more efficient and sparse representations.
Even on more challenging datasets, like the MouseAneurysm images with irregular shapes and lower imaging quality, the model still managed 86.43% Dice accuracy. What’s more, when annotations were reduced to just 5% of the original slices, performance only dropped by about 1%. This suggests a dramatic potential reduction in annotation workload for future medical studies.
A Future of Fewer Clicks and Smarter Insights
Why is this important? Because vessel models aren’t just visual — they’re the basis for simulating blood flow, detecting anomalies, planning surgeries, and building patient-specific computational models. Currently, this often requires hand-cleaned 3D segmentations and meshes, painstakingly prepared by clinicians and engineers. With this new pipeline, AI can take over that burden.
A cardiologist could one day adjust a patient’s vessel model in real-time to simulate how a stent would affect blood flow. A radiologist could rapidly screen for aneurysms by comparing parametrically smoothed vessels across time points. A neurosurgeon might get an instantly deformed model reflecting planned interventions.
And researchers could model entire vascular systems with fewer annotations — something that might dramatically accelerate rare disease modeling or drug development.
Open for All
In a welcome move toward open science, the authors provide a public codebase for their implementation at github.com/alinafdima/paravess. This includes the differentiable voxelization module, the training pipeline, and details for each dataset. While the current voxelization process is computationally expensive (about one minute per 3D scan), the team notes that GPU-accelerated implementations could soon bring it closer to clinical timelines.
A New Anatomical Alphabet
This paper doesn’t just provide a new tool — it introduces a new language for describing vascular structures. It unifies three powerful representations: splines, meshes, and voxels — bridging the gap between how clinicians see, how computers learn, and how simulations work.
The phrase “bringing order to chaos” is often used lightly. Here, it’s apt. By translating noisy segmentations into smooth, flexible, anatomically faithful shapes, this research might one day power a new generation of AI-native clinical tools. It’s the kind of foundational advance that makes other breakthroughs possible — because when machines can finally understand the structure of a vessel, everything else — from diagnosis to treatment — flows better.
This blog post is based on this 2025 ArXiv paper.
If you liked this blog post, I recommend having a look at our free deep learning resources or my YouTube Channel.
Text and images of this article are licensed under Creative Commons License 4.0 Attribution. Feel free to reuse and share any part of this work.