Introduction
Cerebral aneurysms—pathological bulges in brain blood vessels—pose a significant risk of life-threatening rupture. Precise segmentation from 3D CT images is essential for clinical risk assessment and surgical planning. However, automation faces two primary hurdles: extreme class imbalance, as aneurysms are minute compared to total brain volume, and visual ambiguity, where surrounding tissues mimic aneurysm appearance.
While manual segmentation is the current gold standard, it is labor-intensive and suffers from significant inter-observer variability. This paper presents a high-performing approach based on the 3D nnU-Net framework, demonstrating that for the difficult task of segmenting tiny cerebral aneurysms, "bigger is better" when it comes to image patches. By configuring a 3D U-Net with a significantly larger patch size to capture more anatomical context, the authors secured second place in the MICCAI 2020 CADA challenge.
Methodology
The authors utilized the 3D nnU-Net, a popular self-configuring framework for medical image segmentation. Their primary innovation was a specific configuration choice: maximizing the patch size.
Model Configuration and Training
The model was configured with a patch size of $192 \times 224 \times 192$ and a batch size of 2, representing the hardware limits of an NVIDIA TITAN V100 GPU. The architecture utilized a 3D U-Net with six resolution levels. Images were resampled to an iso-voxel spacing of $0.5429 mm^3$ and normalized. Training was conducted using Stochastic Gradient Descent (SGD) with a combined Dice and Cross-Entropy loss function. To prevent overfitting, they employed heavy data augmentation, including rotation, scaling, and Gaussian noise. Final inference utilized an ensemble of five models derived from five-fold cross-validation.
Results and Discussion
The method proved highly effective, ranking second in the CADA challenge with a Jaccard score of 0.7593 on the hidden test set.
Quantitative Performance
As shown in the final evaluation, the model achieved a Jaccard score of 0.759 and a Volume Pearson R of 0.998, indicating an extremely high correlation between predicted and actual aneurysm volumes. During internal cross-validation, the model maintained an average Jaccard of 0.8112 and a Dice score of 0.8861.
Qualitative Analysis
Visual inspections confirm that the model effectively captures the complex shapes of aneurysms. However, the punishing nature of Jaccard and Dice metrics for small objects means that even minor boundary discrepancies result in significant score reductions. In some challenging cases, the model failed to detect the aneurysm entirely, resulting in a Jaccard score of zero.
Conclusion
This study underscores that "bigger is better" regarding patch size for small-object segmentation in medical imaging. By providing more anatomical context, the model successfully navigated the challenges of class imbalance and visual ambiguity. Future work will focus on addressing "zero-detection" cases, improving inference speed for clinical deployment, and exploring full-volume processing as hardware capabilities evolve.
Summary updated