Interpretable Aneurysm Classification via 3D Concept Bottleneck Models: Integrating Morphological and Hemodynamic Clinical Features

Imagine you have a very smart, super-fast robot doctor. This robot can look at 3D scans of a human brain and spot a dangerous bulge in a blood vessel called an aneurysm. It's incredibly good at this; it gets it right almost every time.

But here's the problem: The robot is a "black box."

If you ask the robot, "Why did you say this patient is at risk?" it just says, "Because the computer says so." It can't explain which part of the image made it nervous. In medicine, doctors can't just trust a magic answer. They need to know why so they can double-check the work. If the robot is wrong, they need to know where it went wrong.

This paper introduces a new way to build this robot so it doesn't just give an answer, but shows its homework.

The Big Idea: The "Concept Bottleneck"

Think of the robot's brain like a factory assembly line.

The Old Way (Black Box):
Raw brain scan $\rightarrow$ Mystery Machine $\rightarrow$ "Risky" or "Safe."
The machine does all the thinking in the dark. We don't know what it looked at.

The New Way (This Paper's Solution):
Raw brain scan $\rightarrow$ Step 1: Identify Clues $\rightarrow$ Step 2: Make Decision $\rightarrow$ "Risky" or "Safe."

The authors built a "bottleneck" in the middle. Before the robot makes its final decision, it must stop and write down a list of specific, human-understandable clues it found.

The 26 Clues (The "Concepts")

Instead of looking at pixels, the robot is forced to look for 26 specific things that real neurosurgeons care about. Think of these as the robot's "checklist":

Shape: Is the bulge round or weirdly shaped? (Morphology)
Size: How big is it compared to the vessel?
Flow: Is the blood rushing too hard against the wall? (Hemodynamics)
Pressure: Is the wall being stressed in a specific way?

The robot has to say, "I see the bulge is 4mm wide, and the blood flow is hitting the wall at a sharp angle." Then, it uses those specific facts to decide if it's dangerous.

Why This is a Game-Changer

Trust: If the robot says a patient is at risk, the doctor can look at the checklist. If the robot says, "It's risky because the blood flow is too high," the doctor can verify that. If the robot says, "It's risky because the color of the scan is blue," the doctor knows the robot is confused.
No Cheating: The authors made sure the robot couldn't cheat by looking for the word "aneurysm" in the data. It has to learn the actual physics and shapes.
It's Still Super Smart: Usually, when you make a robot explain itself, it gets slower or less accurate. But this robot is still a champion! It got 93% accuracy, which is just as good as the "black box" models.

The "Test-Time" Safety Net

The paper also mentions a trick called Test-Time Augmentation (TTA). Imagine you are taking a test. To make sure you didn't just get lucky, you take the test 8 times, each time with the paper slightly rotated or the lighting changed. You then average your answers.

The robot does this too. It looks at the brain scan 8 different ways (tilted, flipped, zoomed) and averages the result. This makes the diagnosis much more stable and reliable, ensuring the robot isn't just guessing based on a weird angle.

The Bottom Line

This research is like giving the robot doctor a transparent glass wall instead of a solid steel door.

Before: The robot shouted, "Danger!" and you had to trust it blindly.
Now: The robot says, "Danger! Here is my list of 26 reasons why: The wall is thin, the flow is fast, and the shape is odd. Do you agree?"

This allows human doctors and AI to work together as a team, making surgery safer and saving more lives by ensuring the AI's logic matches human medical wisdom.

Here is a detailed technical summary of the paper "Interpretable Aneurysm Classification via 3D Concept Bottleneck Models: Integrating Morphological and Hemodynamic Clinical Features."

1. Problem Statement

The primary challenge addressed is the lack of clinical transparency in deep learning models used for intracranial aneurysm (IA) classification. While traditional "black-box" Convolutional Neural Networks (CNNs) achieve high predictive accuracy, their inability to provide causal, human-understandable reasoning hinders clinical adoption and regulatory approval.

Limitations of Current XAI: Existing Explainable AI (XAI) methods, such as saliency maps or Class Activation Mapping (CAM), offer post-hoc visual correlations that often lack causal depth and can be misleading in high-stakes neurosurgical contexts.
Clinical Need: Surgeons require diagnostic outputs anchored to established neurosurgical principles (e.g., specific vessel geometries or hemodynamic stressors) rather than opaque mathematical artifacts to validate risk assessments and plan interventions.

2. Methodology

The authors propose an end-to-end 3D Soft Concept Bottleneck Model (CBM) framework that integrates high-dimensional imaging features with human-interpretable clinical concepts.

A. Dataset and Preprocessing

Data Source: The Clinical, Morphological, and Hemodynamic Aneurysm (CMHA) dataset, containing 3D Computed Tomography Angiography (CTA) volumes, structured clinical metadata, and biomechanical indices derived from Computational Fluid Dynamics (CFD).
Cohort: 136 subjects (92 patients with aneurysms, 44 controls).
Preprocessing: Raw 3D volumes were resampled to isotropic resolution and resized to a 96 × 96 × 96 voxel grid.
Validation Strategy: 5-fold Stratified K-Fold cross-validation to maintain class distribution balance.

B. Architecture Design

The framework utilizes a dual-pathway architecture:

Backbones: Two 3D architectures were evaluated:
- 3D ResNet-34: Pre-trained on the MedicalNet dataset.
- 3D DenseNet-121: Trained from scratch.
- Feature Extraction: The final fully connected layer is replaced with an identity layer to extract latent embeddings ( $z$ ) of 512 (ResNet) or 1024 (DenseNet) dimensions.
Concept Head: A dedicated network branch predicts a vector of 26 human-interpretable clinical concepts (e.g., vessel angles, Oscillatory Shear Index, aspect ratios).
- Leakage Mitigation: A strict keyword filter removed trivial diagnostic indicators (e.g., "aneurysm," "sac") from the concept pool to prevent the model from "cheating" by identifying labels directly.
Task Head (Soft Bottleneck): The final diagnosis is derived by concatenating the latent visual embedding ( $z$ ) with the predicted concept vector ( $c$ ), denoted as $z \oplus c$ . This ensures the model retains the representational power of deep features while being constrained by clinical reasoning.

C. Training and Optimization

Loss Function: A joint-loss formulation: $L_{total} = \beta \cdot L_{task} + \alpha \cdot L_{concept}$ $L_{t o t a l} = β \cdot L_{t a s k} + α \cdot L_{co n ce pt}$ .
- $L_{task}$ : Focal Loss (addressing class imbalance).
- $L_{concept}$ : Mean Squared Error (MSE) for concept regression.
- Weights: $\beta = 1.0$ (primary), $\alpha = 0.01$ (secondary), optimized via Optuna.
Staged Fine-Tuning: For the pre-trained ResNet-34, the encoder was frozen for the first 8 epochs to stabilize the heads, followed by "unfreezing" layers 3 and 4 for joint training.
Augmentation Strategy:
- Training: Moderate transformations for real samples; aggressive regularization for synthetic controls to prevent memorization.
- Inference (TTA): An 8-pass Test-Time Augmentation (TTA) ensemble was used to average predictions across light random augmentations, ensuring diagnostic stability against orientation and noise variations.

3. Key Contributions

First 3D Soft-CBM for IAs: Introduction of the first interpretable bottleneck framework for intracranial aneurysms that unifies 3D morphological and CFD-derived hemodynamic features.
Leakage Mitigation Protocol: Implementation of a rigorous feature selection process to ensure predicted concepts are biologically relevant and free from diagnostic data leakage.
Multi-Level Augmentation: A three-tier augmentation strategy (standard training, strong regularization for synthetic controls, and 8-pass TTA) to maximize generalization in a small-sample regime.
Performance-Interpretability Trade-off: Demonstration that high predictive performance can be achieved without sacrificing clinical transparency.

4. Results

The model was evaluated using stratified 5-fold cross-validation.

Classification Accuracy:
- ResNet-34 (Pre-trained): Peak accuracy of 93.33% ± 4.5%.
- DenseNet-121 (From Scratch): Peak accuracy of 91.43% ± 5.8%.
- TTA Stability: The 8-pass TTA ensemble yielded a robust mean accuracy of 88.31%, significantly improving specificity and reducing over-diagnosis in low-risk scenarios.
Generalization: The gap between training and validation accuracy remained < 0.04 across all trials, indicating successful avoidance of overfitting despite the limited dataset size.
Discriminative Power:
- ResNet-34 achieved a mean AUC of 0.960 ± 0.032, with one fold reaching a perfect AUC of 1.000.
- Sensitivity reached 97.8% for the standard ResNet-34 model, correctly identifying 90/92 aneurysm cases.
Training Dynamics: The ResNet-34 exhibited a distinct "unfreezing shock" (loss spike) at epoch 8 followed by recovery, validating the efficacy of staged transfer learning for 3D neuroimaging.

5. Significance and Conclusion

This research represents a significant step toward trustworthy AI in neurosurgery. By embedding interpretability directly into the architecture (ante-hoc) rather than relying on post-hoc explanations, the model provides a verifiable "reasoning path" based on established clinical indices (e.g., wall shear stress, vessel geometry).

Clinical Impact: The framework bridges the gap between high-dimensional visual features and neurosurgical parameters, allowing clinicians to inspect and potentially intervene in the model's logic.
Future Directions: The authors acknowledge limitations regarding the manual curation of the concept pool (potential mathematical redundancies) and the single-center dataset. Future work will focus on concept pruning and integrating broader multimodal clinical data to further refine the bottleneck for surgical planning.

In summary, the proposed 3D Soft CBM successfully demonstrates that clinical transparency and state-of-the-art predictive performance are not mutually exclusive, offering a robust foundation for AI-assisted aneurysm risk stratification.