Abstract
Accurate segmentation of brain tumors is an indispensable component for cancer diagnosis and treatment. In this paper, we propose a novel brain tumor segmentation method based on a multi-cascaded convolutional neural network (MCCNN) and fully connected conditional random fields (CRFs). The segmentation process mainly includes two steps. First, we design a multi-cascaded network architecture by combining the intermediate results of several connected components to take the local dependencies of labels into account and make use of multi-scale features for coarse segmentation. Second, we apply CRFs to consider spatial contextual information and eliminate spurious outputs for fine segmentation. Additionally, we use image patches obtained from axial, coronal, and sagittal views to train three segmentation models, which are then combined for the final result. Evaluated on three public databases, our method achieves competitive performance compared with state-of-the-art approaches.
Introduction
Segmenting gliomas from MRI scans is a critical yet challenging task due to the high variability in tumor shape, size, and location. Boundaries between pathological and healthy tissue are often blurred, making manual segmentation time-consuming and prone to error. While deep learning has significantly advanced automated segmentation, many patch-based models treat voxels in isolation, failing to capture local label dependencies and broader spatial context. This paper addresses these challenges by proposing a "coarse-to-fine" framework that balances high-resolution spatial information with computational efficiency.
Proposed Methodology
The proposed segmentation pipeline follows a two-stage approach: coarse segmentation via a Multi-Cascaded CNN (MCCNN) followed by fine refinement using Conditional Random Fields (CRFs).
Coarse Segmentation (MCCNN)
The MCCNN architecture evolves from the InputCascadeCNN by utilizing parallel sub-networks to process input images at different scales. By cascading the output of one layer as an additional input to the next, the model learns local dependencies—effectively teaching the network that a voxel's classification is influenced by its neighbors. This multi-scale approach allows the model to capture both global context and local details simultaneously.
Fine Segmentation (CRFs)
To address the "noisy" or spurious outputs common in CNN-based segmentation, fully connected CRFs are applied. The CRF acts as a refiner, smoothing tumor edges and ensuring that the final segmentation respects the appearance and spatial consistency of the original MRI data.
Multi-View Fusion
Rather than relying on a single 2D perspective, the framework trains three separate models on image patches from axial, coronal, and sagittal views. These perspectives are fused using a majority voting strategy, providing a "pseudo-3D" understanding of the tumor structure without the heavy memory requirements of native 3D CNNs.
Experimental Results and Discussion
The model was evaluated on three benchmark datasets from the Brain Tumor Segmentation Challenge (BRATS): 2013, 2015, and 2018.
- Accuracy: On the BRATS 2013 dataset, the fused model achieved a Dice Similarity Coefficient (DSC) of 0.89 for the complete tumor, outperforming several established methods.
- Efficiency: The system can segment an entire brain volume in 1.5 to 3 minutes on a standard GPU, significantly faster than traditional patch-based methods which can take up to 100 minutes.
- Modality Importance: Ablation studies revealed that T1c (contrast-enhanced) and FLAIR sequences are the most critical for identifying the "enhancing core" and tumor edges.
Conclusion
This research demonstrates that combining multi-cascaded CNNs with CRFs and multi-view fusion provides a robust and efficient solution for brain tumor segmentation. While the current approach relies on 2D patches, it achieves state-of-the-art accuracy with clinical-grade processing speeds. Future work will explore deeper integration of anatomical views directly within the network architecture to further improve volumetric consistency.
Summary updated