- The paper demonstrates that a stacked ensemble of three distinct nnUNetv2 configurations, integrated via an L2-regularized logistic regression meta-model, improves segmentation accuracy.
- The paper leverages diverse data augmentation and model architectures, achieving a mean Dice Coefficient of 0.9752 and a reduced Separation Distance of 0.4598 mm on test cases.
- The paper’s approach supports more reliable patient-specific biomechanical modeling while highlighting the need for broader validation on larger, heterogeneous datasets.
Stacked Ensemble Learning for AAA Segmentation: Technical Analysis
Motivation and Background
Accurate delineation of abdominal aortic aneurysms (AAA) in computed tomography angiography (CTA) is a cornerstone for patient-specific biomechanical modeling, crucial for rupture risk assessment. Manual and semi-automated approaches to AAA segmentation are inefficient and exhibit significant observer variability, restricting scalability to larger cohorts. CNN-based pipelines—particularly the nnUNet family—have established state-of-the-art performance in AAA and general medical segmentation tasks, but residual error near complex anatomical boundaries can persist. This work targets those gaps by proposing a stacked ensemble framework leveraging architectural and training diversity among three nnUNetv2 derivations. The ensemble employs a regularized logistic regression meta-model to optimally combine voxel-wise probability estimates from the base learners with the intent to further minimize residual boundary and volumetric segmentation errors.
Experimental Design and Methodology
The dataset comprises 40 anonymized, contrast-enhanced CTA scans from AAA patients, annotated using the nnInteractive extension in 3D Slicer to generate high-consistency ground truth segmentations incorporating expert prompts. Data partitioning involved 32 training and 8 test cases, with four-fold cross-validation among the training set to generate robust out-of-sample base learner predictions for meta-model training.
Three independently trained nnUNetv2 configurations served as base learners:
- Default: Baseline nnUNetv2 parameters,
- DA5: Augmented with an extended data augmentation pipeline,
- ResEncL: Utilizing a larger encoder to enhance feature extraction capacity.
Stacking is implemented via an L2-regularized logistic regression meta-model, where per-voxel base learner probability outputs constitute input features. Hyperparameter selection for regularization is optimized via LOOCV on the cross-validation set, with final training on the full dataset and evaluation on the held-out test set. Segmentation quality is quantified using Dice Coefficient (DC) for volumetric overlap and the novel Separation Distance (SD), a one-directional mean surface-to-surface distance from ground truth to prediction boundary—providing a highly sensitive measure of anatomical fidelity critical for biomechanical pipelines.
Results
All base learners achieved high standalone accuracy, with DA5 yielding the highest training DC (0.9704) and lowest SD (0.5449 mm). The ensemble systematically outperformed individual nnUNetv2 variants on the held-out set with mean DC of 0.9752 and SD of 0.4598 mm. Gains, though modest, are robust and most pronounced for SD, indicating improved average boundary alignment.
Notably, in cases with greater anatomical complexity, the ensemble demonstrated improved boundary consistency, as visually supported in the paper. The stacked meta-model capitalizes on subtle, complementary strengths of the constituent architectures and data augmentation regimes, resolving boundary ambiguities where single models diverged due to architectural bias or training regimen idiosyncrasies.
Implications and Limitations
The presented methodology evidences that stacked ensemble learning marginally but consistently enhances segmentation accuracy, even in regimes where strong single-model baselines saturate. This is especially pertinent for applications where boundary precision, not merely volumetric overlap, governs downstream computation, e.g., finite element analysis for wall stress or strain computation. Reduced SD translates directly to lower geometry-induced uncertainty in mechanical models.
Practical deployment in broader clinical scenarios, however, is contingent upon validation on larger, multi-institutional datasets encompassing greater scanner/protocol heterogeneity and anatomical variation. The current dataset’s focused field of view facilitated strong accuracy, but general settings may necessitate preliminary aortic ROI localization. The computational overhead is modest, as the ensemble meta-model involves only shallow logistic regression, but the system inherits the aggregate compute cost of multiple nnUNetv2 forward passes.
Theoretical Outlook and Future Directions
The observed complementary error profiles among diverse nnUNetv2 configurations support further exploration of meta-ensemble strategies—possibly integrating base learners from other advanced architectures (e.g., transformer-based segmenters) or leveraging non-linear meta-models. Additionally, explicit modeling of segmentation uncertainty in ensemble decision-making could improve reliability for outlier cases. Most critically, robust generalization must be established on larger-scale, federated datasets and with automatic AAA localization as a precursor.
Further research should quantify the translation of segmentation refinement (notably SD improvements) to enhanced accuracy of patient-specific biomechanical biomarkers, potentially via uncertainty propagation studies or ablation analyses within mechanical modeling workflows.
Conclusion
This study demonstrates that stacked ensembles of diverse nnUNetv2 configurations using L2-regularized logistic regression meta-models offer consistent, incremental gains in AAA segmentation accuracy in CTA, particularly for surface proximity critical in biomechanical modeling. The strategy is especially effective for minimizing residual boundary errors among high-performing learners. Broader adoption will require validation on diverse cohorts and incorporation of upstream localization, but the approach sets a new technical reference for automated AAA segmentation pipelines involved in computation-driven vascular risk stratification.