Abstract
This paper presents a robust 3D segmentation framework for brain aneurysms in Computed Tomography Angiography (CTA). By integrating implicit deformable models with non-parametric statistical information derived from a k-nearest neighbor (kNN) approach, the method overcomes the limitations of traditional Gaussian distribution assumptions. The approach achieves sub-voxel accuracy, providing objective measurements that rival or exceed the precision of manual clinical assessments.
Introduction
Precise measurement of brain aneurysm dimensions—specifically the neck, width, and depth—is critical for "coiling" procedures, where platinum coils are used to induce clotting and prevent rupture. Traditionally, these measurements are derived from 2D Maximum Intensity Projection (MIP) images. However, 2D views are inherently subjective and dependent on the radiologist's chosen angle. While 3D automated segmentation offers a more objective alternative, CTA images present significant challenges due to the intensity overlap between blood vessels and adjacent bone structures, which often leads to segmentation "leakage" or failure in standard algorithms.
Methodology
The proposed method utilizes a hybrid approach combining non-parametric probability estimation with deformable models to ensure robust segmentation.
Non-Parametric Probability Estimation
Unlike conventional models that assume vessel and bone intensities follow a Gaussian distribution, this method employs a k-Nearest Neighbor (kNN) rule. The system was trained on 1,830 manually labeled points using feature vectors that include local intensity, gradient magnitude, and Hessian matrix eigenvalues to capture local shape and curvature.
MAP Classification and Deformable Models
A Maximum A Posteriori (MAP) classifier, refined by anisotropic smoothing, generates an initial partition of the image into vessel, bone, and background. This partition serves as the starting point for the Geodesic Active Regions (GAR) model. The GAR model treats the segmentation boundary as an evolving front. It leverages the kNN-derived probabilities to attract the boundary toward vessel-like regions while repelling it from bone and background.
Key Insights
- Flexible Statistical Modeling: The failure of Gaussian models in this context is attributed to the non-unimodal and overlapping intensity distributions of vessels and bone. The kNN approach effectively "learns" these complex distributions.
- The Peaking Phenomenon: The study found that increasing the dimensionality of the feature space (e.g., through multiscale analysis) without a proportional increase in training data actually degraded performance, a known issue in classifier design.
- Sub-voxel Precision: By utilizing level sets within the implicit deformable model, the algorithm achieves precision beyond the discrete voxel grid, allowing for highly accurate boundary detection.
Results and Evaluation
The method was evaluated on a clinical database of 39 brain aneurysm cases using Bland-Altman analysis against a "gold standard" of manual measurements from two experts.
- Accuracy: The average error (bias) was consistently less than 1 mm, with worst-case errors roughly equivalent to a single voxel size (~0.69 mm).
- Reliability: The agreement between the automated method and human observers was higher than the inter-observer agreement between two human experts.
- Robustness: The system successfully handled complex cases, including giant aneurysms and those in challenging anatomical locations like the Circle of Willis.
Conclusion
The integration of non-parametric region information with implicit deformable models provides a highly accurate and objective tool for aneurysm measurement in CTA. Future work will focus on refining feature selection to mitigate the peaking phenomenon and exploring more automated initialization techniques to streamline clinical workflows.
Summary updated