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

Toqa Khaled, Ahmad Al-Kabbany|arXiv (Cornell University)|Mar 8, 2026

Intracranial Aneurysms: Treatment and Complications0 citations

TL;DR

The paper proposes a 3D Soft Concept Bottleneck Model (Soft-CBM) for intracranial aneurysm classification using CTA, integrating 3D morphological and CFD-derived hemodynamic concepts, achieving high accuracy with interpretable clinical bottlenecks.

ABSTRACT

We are concerned with the challenge of reliably classifying and assessing intracranial aneurysms using deep learning without compromising clinical transparency. While traditional black-box models achieve high predictive accuracy, their lack of inherent interpretability remains a significant barrier to clinical adoption and regulatory approval. Explainability is paramount in medical modeling to ensure that AI-driven diagnoses align with established neurosurgical principles. Unlike traditional eXplainable AI (XAI) methods -- such as saliency maps, which often provide post-hoc, non-causal visual correlations -- Concept Bottleneck Models (CBMs) offer a robust alternative by constraining the model's internal logic to human-understandable clinical indices. In this article, we propose an end-to-end 3D Concept Bottleneck framework that maps high-dimensional neuroimaging features to a discrete set of morphological and hemodynamic concepts for aneurysm identification. We implemented this pipeline using a pre-trained 3D ResNet-34 backbone and a 3D DenseNet-121 to extract features from CTA volumes, which were subsequently processed through a soft bottleneck layer representing human-interpretable clinical concepts. The model was optimized using a joint-loss function to balance diagnostic focal loss and concept mean squared error (MSE), validated via stratified five-fold cross-validation. Our results demonstrate a peak task classification accuracy of 93.33% +/- 4.5% for the ResNet-34 architecture and 91.43% +/- 5.8% for the DenseNet-121 model. Furthermore, the implementation of 8-pass Test-Time Augmentation (TTA) yielded a robust mean accuracy of 88.31%, ensuring diagnostic stability during inference. By maintaining an accuracy-generalization gap of less than 0.04, this framework proves that high predictive performance can be achieved without sacrificing interpretability.

Motivation & Objective

• Bridge high predictive performance with clinical interpretability in intracranial aneurysm classification.
• Map 3D imaging and hemodynamic features to a human-interpretable set of clinical concepts.
• Reduce data leakage and improve generalization through staged training and multi-level data augmentation.

Proposed method

• Use 3D backbones (ResNet-34 pre-trained on MedicalNet and DenseNet-121 trained from scratch) to extract volumetric features from CTA.
• Replace final layer with an identity embedding and add a soft bottleneck predicting 26 clinical concepts.
• Concatenate latent visual embedding with predicted concepts for final diagnosis (z ⊕ c).
• Optimize with joint loss: Ltotal = β Ltask + α Lconcept, with β=1.0 and α=0.01.
• Apply staged fine-tuning with encoder frozen then unfreeze, and use 5-fold stratified CV.
• Employ multi-level 3D augmentation and 8-pass Test-Time Augmentation (TTA) for robustness.

Experimental results

Research questions

• RQ1 Can a 3D Soft Concept Bottleneck Model achieve high accuracy on aneurysm classification while providing interpretable clinical reasoning?
• RQ2 What are the effects of using pre-trained versus from-scratch 3D backbones on accuracy and generalization in this CBM framework?
• RQ3 Does multi-level augmentation and TTA improve inference stability without sacrificing interpretability?
• RQ4 What is the impact of constraining concepts to 26 clinically meaningful indices on performance and leakage mitigation?

Key findings

• Peak task accuracy: 93.33% ± 4.5% for ResNet-34 with merging strategy.
• Peak task accuracy: 91.43% ± 5.8% for DenseNet-121 trained from scratch.
• 8-pass TTA yields mean accuracy 88.31% with robust inference.
• Accuracy-generalization gap ≤ 0.04 across trials, indicating low overfitting.
• ROC AUC mean 0.960 ± 0.032 for ResNet-34 merged configuration; 0.899 for TTA, and 0.899 for DenseNet-121.
• Sensitivity peaks around 97.8% in standard inference for ResNet-34 merged; TTA improves specificity.