Abstract
Brain tumor segmentation in multimodal MRI is significantly challenged by extreme class imbalance and structural heterogeneity. This paper presents a cascaded 3D convolutional neural network (CNN) framework utilizing a two-stage V-Net architecture. By first localizing the whole tumor and subsequently refining sub-region delineation within a constrained Region of Interest (ROI), the approach effectively focuses learning on tumorous tissue, mitigating bias toward healthy brain volume. The method demonstrates competitive performance on the BraTS 2017 dataset, particularly in whole tumor localization.
Introduction
In medical imaging, segmenting brain tumors—particularly gliomas—is often described as a "needle in a haystack" problem. Tumors typically occupy only 3–5% of the total MRI volume. Standard machine learning models trained on the entire volume often converge toward a trivial solution, predicting no tumor presence due to the overwhelming prevalence of healthy tissue.
Furthermore, gliomas are highly heterogeneous, comprising sub-regions such as the necrotic core, enhancing tumor, and peritumoral edema. Each region exhibits distinct intensity profiles across MRI modalities (T1, T1c, T2, and FLAIR). Accurate delineation is critical for surgical planning and treatment monitoring, yet fuzzy boundaries and high inter-patient variability remain significant hurdles for automated systems.
Methodology
The proposed architecture, Cascaded V-Net, extends the 3D U-Net for volumetric data with several key refinements. To improve efficiency and gradient flow, the authors replaced $5 \times 5 \times 5$ filters with $3 \times 3 \times 3$ filters and reformulated short residual connections using identity mappings. Additionally, Batch Normalization (BN) was integrated before ReLU activations to handle internal covariate shift.
Two-Stage Cascade
The core strategy involves a two-stage pipeline designed to handle class imbalance:
- Net-1 (Localization): A binary segmentation network that identifies the "Whole Tumor" (WT) region across all four MRI modalities.
- Net-2 (Delineation): This network focuses exclusively on the area identified by Net-1. It uses an ROI mask to constrain training and inference to the vicinity of the tumor, effectively ignoring healthy brain tissue.
A key innovation is the use of the ROI mask during training to zero out loss and gradients for background voxels. This "dense training" approach on the whole ROI allows the network to capture global spatial context while forcing its "attention" entirely onto the tumor sub-structures.
Results and Discussion
The model was evaluated on the BraTS 2017 dataset, consisting of 285 training cases and 46 validation cases. Performance was measured using the Dice Similarity Coefficient (DSC), where 1.0 represents a perfect match.
The system achieved a DSC of 0.877 for the Whole Tumor (WT) on the validation set. However, smaller, more complex sub-regions proved more challenging: the Enhancing Tumor (ET) and Tumor Core (TC) reached DSCs of 0.714 and 0.637, respectively.
As shown in Figure 6, the model maintains high specificity (>0.99) but exhibits lower sensitivity in smaller sub-structures. Notably, the "Tumor Core" metric (Fig 6b) shows signs of overfitting as training progresses, even as the overall loss improves. This suggests that internal class imbalance within the tumor itself—such as a small enhancing core relative to a large edema—remains a significant hurdle for the network.
Conclusion
The cascaded V-Net with ROI masking effectively addresses the global class imbalance inherent in brain tumor segmentation. While the localization of the whole tumor is robust, future work will focus on addressing the internal imbalance of tumor sub-regions, potentially through weighted loss functions to improve sensitivity in critical areas like the tumor core.
Summary updated