Abstract
Precise segmentation of anatomical brain barriers is essential for defining the Clinical Target Volume (CTV) in radiation therapy for gliomas. This paper presents an ensembled 3D ResUnet framework developed for the Anatomical Brain Barriers to Cancer Spread (ABCs) challenge. By utilizing an asymmetric architecture, Tversky loss to mitigate class imbalance, and a pseudo-labeling ensemble strategy, the proposed method achieves state-of-the-art performance, outperforming standard baseline models in both segmentation accuracy and anatomical alignment.
Introduction
In the treatment of brain tumors, clinicians must accurately define the CTV, a boundary largely determined by anatomical structures that act as natural barriers to cancer cell migration. Manually segmenting these structures across multi-modal imaging (CT and MRI) is a labor-intensive process prone to inter-observer variability due to complex geometries and varying tissue appearances. The ABCs challenge aims to automate this process, requiring robust solutions that can handle significant class imbalances and integrate multi-modal data efficiently.
Methodology
The proposed framework is built upon the nnU-Net foundation, incorporating several architectural and training innovations to enhance performance on medical segmentation tasks.
Asymmetric ResUnet Architecture
The core of the pipeline is a U-shaped network designed with an asymmetric structure. The encoder is "deep," utilizing residual blocks to extract complex, high-level features from the multi-modal input. Conversely, the decoder is "shallow," which serves to balance reconstruction accuracy with inference efficiency while acting as a form of regularization to prevent overfitting.
Loss Function and Supervision
To address the extreme class imbalance between small anatomical structures and the large background, the authors employed Tversky loss. By setting $\alpha=0.3$ and $\beta=0.7$, the loss function prioritizes the penalization of false negatives, ensuring that critical small barriers are not overlooked. Additionally, deep supervision was implemented by adding auxiliary output blocks at various decoder levels, facilitating faster convergence and multi-scale feature learning.
Ensemble and Pseudo-training
The final model performance was boosted through a multi-stage ensemble strategy. Predictions from multiple top-performing models were averaged to generate "pseudo-labels" for test volumes. The model was then retrained using a combination of the original annotated training data and these high-confidence pseudo-labels, a technique that significantly improved generalization.
Experimental Setup and Datasets
The study utilized the ABCs Challenge dataset, comprising 45 training/validation cases and 30 test cases. Each case included co-registered CT scans, contrast-enhanced T1-weighted MRI, and T2-weighted FLAIR MRI, all resampled to a 1.2mm isotropic resolution. The multi-modal inputs were concatenated as separate channels for the network.
Results and Discussion
The proposed framework was evaluated against the standard nnU-Net baseline. The ensembled pseudo-training approach showed consistent improvements across both Dice Score (DSC) and Surface Dice (SDSC) metrics.
| Method | Task 1 DSC | Task 1 SDSC | Task 2 DSC | Task 2 SDSC |
|---|---|---|---|---|
| nnU-Net | 0.872 | 0.974 | 0.777 | 0.926 |
| Ensembled Pseudo-training | 0.876 | 0.974 | 0.782 | 0.929 |
Qualitative analysis confirms that the model produces segmentations that are highly congruent with the ground truth annotations, effectively capturing the intricate boundaries of brain barriers.
Key Insights
- Spatial Sensitivity: A notable finding was that horizontal flipping as data augmentation negatively impacted Task 2. This is attributed to the brain's paired but non-identical structures (e.g., eyes, cochlea), where x-axis flipping disrupts the model's spatial context.
- Decoder Depth: The success of the shallow decoder suggests that for medical segmentation, excessive depth in the reconstruction path may not be necessary and can lead to poorer generalization compared to more constrained architectures.
Conclusion
The ensembled 3D ResUnet framework provides a robust solution for the automated segmentation of anatomical brain barriers. By combining asymmetric architecture with specialized loss functions and pseudo-labeling, the method achieves high accuracy suitable for clinical target volume definition. Future work will focus on improving the computational efficiency of the ensemble process to facilitate real-time clinical application.
Summary updated