💻 computer science

Integrating Physics-Informed Neural Networks and 3D Vascular Geometry Learning for Cerebral Aneurysm Detection and Multimodal Rupture-Risk Prediction

This study presents a modular framework that integrates 3D vascular geometry learning, physics-informed neural networks for hemodynamic simulation, and clinical variables to achieve high-performance cerebral aneurysm detection and multimodal rupture-risk prediction.

Original authors: Eshan Vipuil, Xianqi Li

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Eshan Vipuil, Xianqi Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain's blood vessels as a bustling city of tiny, winding rivers. Sometimes, a weak spot in the riverbank swells up into a balloon-like bulge called a cerebral aneurysm. Most of the time, these balloons are harmless and never pop. But occasionally, they burst, causing a dangerous event called a subarachnoid hemorrhage. The big challenge for doctors right now is figuring out which balloons are safe and which ones are about to pop, without having to pop them to find out.

Currently, doctors act like expert detectives, looking at 3D pictures of the vessels and asking, "Does this shape look scary? Does the patient's age or history make it worse?" But this study suggests we can build a super-smart, digital assistant that does this detective work even better by combining three different types of clues: the shape of the vessel, the physics of the blood flowing through it, and the patient's personal story.

The Two-Step Detective Game

The researchers built a two-stage system to solve this mystery.

Step 1: The Shape Shifter (Detection)
First, the system needs to know if a balloon even exists. The team trained a digital brain (a neural network called PointNeXt) to look at 3D "point clouds"—basically, a digital cloud of dots that forms the surface of the blood vessels. Think of this like a sculptor looking at a pile of sand and instantly recognizing if a hidden statue is buried inside.

The Result: This digital sculptor was incredibly sharp. It correctly spotted aneurysms 95.9% of the time (a score called AUROC) and was very good at not crying wolf (AUPRC of 0.859). It successfully separated the "balloon" vessels from the "normal" ones in a test set of 304 cases.

Step 2: The Risk Calculator (Rupture Prediction)
Once the system says, "Yes, there is a balloon here," the real hard work begins: Is it going to pop?
Here is where the paper gets really clever. Usually, to know how the blood flows inside a balloon, scientists have to run massive, slow computer simulations (like a wind tunnel for blood) that take forever. This study tried a shortcut using Physics-Informed Neural Networks (PINNs).

The Analogy: Imagine trying to guess how wind blows around a house. You could build a giant wind tunnel and measure it (slow and expensive). Or, you could teach a computer the laws of physics (like how air moves) and let it "dream" the wind patterns based on the house's shape. The PINN did exactly this: it learned the rules of fluid mechanics and used them to predict pressure, speed, and "wall shear stress" (how hard the blood rubs against the vessel wall) without needing a pre-made simulation for every single case.

The Magic of Mixing Clues

The team tested different ways to combine their clues to predict the risk of a rupture:

Just the Shape: Looking only at the 3D geometry of the balloon.
Just the Patient: Looking only at age, sex, and location.
Just the Flow: Looking only at the physics-based blood flow data.
The Mix: Combining them all.

The Big Surprise:
If you looked only at the shape of the aneurysm, the system was only okay at guessing the risk (AUROC of 0.611). It wasn't good enough to be trusted alone.
However, when they added the patient's history (age, sex, location), the score jumped up to 0.809.
But the champion was a "late fusion" strategy. Imagine two experts giving their opinions separately, and then a referee combining their scores.

Expert A looked at the Shape + Patient History.
Expert B looked at the Shape + Blood Flow Physics.
The Referee gave Expert A 70% of the vote and Expert B 30%.

This specific 70/30 combination was the winner, achieving a top score of 0.827 (AUROC) and 0.732 (AUPRC). This was statistically better than any other method they tried.

What Made the Difference?

When the researchers peeked under the hood to see why the model made its choices, they found some interesting patterns. The most important clues for predicting a rupture were:

OSI Distribution: This stands for "Oscillatory Shear Index." Think of it as a measure of how "confused" the blood flow is. If the blood is sloshing back and forth chaotically against the wall, that's a bad sign.
Location: Where the aneurysm sits matters a lot (e.g., near the cavernous internal carotid artery).
Radial Geometry: How round or stretched the balloon is.
TAWSS: The average force of the blood rubbing against the wall over time.

What the Paper Doesn't Say (The Fine Print)

It is crucial to understand what this study didn't do, so we don't get our hopes up too high.

No Crystal Ball: The paper explicitly states this is a "cross-sectional" study. That means it looks at a snapshot in time. It can tell you if a balloon looks like it has ruptured or not, but it cannot predict when a specific unruptured balloon will pop in the future. It's a risk score, not a fortune teller.
Simulated Physics: The blood flow numbers (pressure, velocity, etc.) were generated by the PINN based on math rules, not by measuring real blood flow in real patients or running the "gold standard" slow simulations. The authors are careful to say these are "physics-constrained descriptors," meaning they follow the laws of physics, but they haven't been fully proven against real-world measurements yet.
Internal Testing Only: The system was tested on open datasets (VMR and AneuX) using internal cross-validation. It hasn't been tested on new, independent groups of patients from different hospitals yet.

The Bottom Line

This study proposes a new, modular way to assess brain aneurysms. It suggests that by using a smart detector to find the balloon, a physics-savvy AI to guess the blood flow, and a simple mix of patient data, we can create a much better risk score than looking at the shape alone. The 70/30 late-fusion model showed the most promise, suggesting that while the patient's story is the strongest clue, the physics of the blood flow adds a vital piece of the puzzle.

However, the authors are clear: this is a powerful step forward in simulation and modeling, but it needs real-world testing on new patients and long-term follow-up data before it can be used to make life-or-death decisions in a hospital. It's a very promising prototype, not a finished product.

Technical Summary: Integrating Physics-Informed Neural Networks and 3D Vascular Geometry Learning for Cerebral Aneurysm Detection and Multimodal Rupture-Risk Prediction

Problem Statement
Cerebral aneurysms are localized intracranial artery dilations that pose a risk of rupture, leading to subarachnoid hemorrhage. Current clinical assessment relies on human interpretation of imaging (CTA, MRA, DSA) combined with clinical factors (age, sex, location). However, rupture risk is a complex function of vascular morphology, local hemodynamics, and patient physiology. Existing deep learning approaches often focus on imaging patterns or geometric shapes while omitting explicit flow-related information, or they rely on computationally expensive Computational Fluid Dynamics (CFD) simulations to derive hemodynamic features. Furthermore, many existing models lack external validation and fail to integrate vascular geometry, physics-based flow descriptors, and clinical variables into a unified quantitative framework.

Methodology
The study proposes a modular, two-stage framework:

Aneurysm Detection:
- Data: 304 vascular surface geometries (75 aneurysm-positive, 229 non-aneurysm) from the Vascular Model Repository (VMR).
- Model: A PointNeXt-based detector processes vascular point clouds (2,048 sampled points). The architecture uses hierarchical set-abstraction to aggregate local-to-global features, classifying geometries as aneurysm-positive or non-aneurysm.
- Training: Stratified five-fold cross-validation with focal loss to address class imbalance.
Rupture-Risk Prediction:
- Data: 735 aneurysm cases with documented rupture status (261 ruptured, 474 unruptured) from the AneuX cohort.
- Physics-Informed Neural Networks (PINNs): For each aneurysm geometry, a PINN generates hemodynamic descriptors without paired CFD supervision. The network solves the incompressible normalized Navier-Stokes equations under prescribed boundary conditions (pulsatile inlet, no-slip walls, zero-gauge outlet pressure).
- Descriptors: The PINN outputs dimensionless relative pressure, velocity, and derived metrics: Wall Shear Stress (WSS), Time-Averaged WSS (TAWSS), Oscillatory Shear Index (OSI), and Relative Residence Time (RRT).
- Multimodal Modeling: Several strategies were compared on the same 735 cases:
  - Geometry-Only: PointNet++, PointNeXt, Voxel CNN, and Graph Neural Networks (GNN) using 3D point clouds.
  - Clinical-Only: MLP using age, sex, location, and side.
  - Flow-Augmented: PointNeXt augmented with PINN-derived per-point flow channels.
  - Fusion Strategies: Early fusion (concatenating all embeddings) and Late fusion (a weighted probability ensemble of "Geometry + Clinical" and "Flow + Geometry" models).
- Evaluation: Stratified five-fold cross-validation with paired DeLong tests (Holm-corrected) to compare Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC).

Key Contributions

Staged Framework: A conditional workflow linking 3D point-cloud screening to multimodal rupture-risk prediction.
PINN Integration: Introduction of PINN-derived hemodynamic descriptors (pressure, velocity, WSS, TAWSS, OSI, RRT) as physics-constrained features without requiring paired CFD field labels.
Multimodal Integration: Systematic evaluation of fusing vascular geometry, PINN-derived flow, and clinical variables.
Performance Benchmarking: Quantification of the added value of each modality through matched comparisons of geometry-only, clinical-only, flow-augmented, and fusion models.

Results

Aneurysm Detection: The PointNeXt detector achieved a pooled out-of-fold AUROC of 0.959 and AUPRC of 0.859 on the VMR cohort.
PINN Optimization: The PINN successfully minimized the weighted loss (combining Navier-Stokes residuals and boundary conditions), converging to a value of ~0.0028, generating geometry-conditioned hemodynamic fields.
Rupture-Risk Prediction:
- Geometry-Only: Performance was moderate; PointNeXt achieved the best AUROC of 0.611, indicating morphology alone is insufficient for strong discrimination.
- Clinical-Only: Achieved an AUROC of 0.784, outperforming geometry-only models.
- Flow-Augmented: Adding PINN descriptors to geometry (Flow + Geometry) improved AUROC to 0.738.
- Multimodal Fusion: The fixed 70/30 late-fusion model (70% weight to Geometry + Clinical, 30% to Flow + Geometry) achieved the highest performance with a pooled AUROC of 0.827 and AUPRC of 0.732. This significantly outperformed all comparison models (adjusted $p < 0.05$ ).
Feature Importance: Analysis identified OSI distribution, aneurysm location (specifically cavernous internal carotid and anterior communicating arteries), radial geometry, and TAWSS descriptors as key contributors to rupture-risk discrimination.

Significance and Claims
The paper claims to provide a quantitative, multimodal strategy for case-specific aneurysm assessment. It demonstrates that while vascular morphology alone has limited predictive power for rupture risk, integrating clinical variables and physics-constrained hemodynamic descriptors significantly improves discrimination. The authors emphasize that the proposed framework offers a scalable basis for integrating shape, flow, and clinical data.

The authors explicitly state modesty regarding the current findings:

The models produce cross-sectional risk scores (ruptured vs. unruptured status) rather than prospective estimates of future rupture timing or probability.
The PINN-derived fields are physics-constrained descriptors but have not been validated against patient-specific CFD or in vivo flow measurements.
The study relies on internal cross-validation; external validation across different institutions and imaging protocols is required before clinical translation.
The results highlight the complementary nature of the modalities, suggesting that no single data source (geometry, flow, or clinical) is sufficient on its own for robust risk stratification.

The Two-Step Detective Game

The Magic of Mixing Clues

What Made the Difference?

What the Paper Doesn't Say (The Fine Print)

The Bottom Line

More like this