Abstract
In this paper, we present a novel and efficient method for brain tumor (and sub regions) segmentation in multimodal MR images based on a fully convolutional network (FCN) that enables end-to-end training and fast inference. Our structure consists of a downsampling path and three upsampling paths, which extract multi-level contextual information by concatenating hierarchical feature representation from each upsampling path. Meanwhile, we introduce a symmetry-driven FCN by the proposal of using symmetry difference images. The model was evaluated on Brain Tumor Image Segmentation Benchmark (BRATS) 2013 challenge dataset and achieved the state-of-the-art results while the computational cost is less than competitors.
Introduction
Segmenting gliomas—the most common primary brain tumors—is a critical task for surgical planning and treatment monitoring. Manual segmentation of 3D MRI volumes is tedious and prone to human error, yet precise boundaries are essential for clinical success. While Deep Learning has revolutionized this field, many existing methods rely on "patch-based" approaches that classify central pixels within small windows. These methods are computationally expensive and often fail to capture global context. Furthermore, deep architectures designed for natural images (like VGG-16) are often suboptimal for medical imaging, where datasets are small and small lesions can disappear during aggressive downsampling.
Methods
The proposed architecture is a streamlined Fully Convolutional Network (FCN) tailored for medical data. Unlike standard deep models, this network uses only three convolutional blocks to prevent the loss of small lesions and reduce overfitting. The network employs a single downsampling path and three upsampling paths, concatenating hierarchical features to capture both high-level context and fine-grained boundary details.
A core innovation is the symmetry-driven input. Since healthy brains are largely symmetrical, tumors represent a break in that symmetry. The authors calculate symmetry maps by locating the brain's mid-sagittal axis, computing the intensity difference between mirrored pixels, and applying a Gaussian filter. These maps are fed into the network alongside standard MRI modalities (Flair, T1, T1c, and T2).
The model was evaluated using the BRATS 2013 dataset, consisting of 20 High-Grade (HG) patient volumes for training and 10 for testing. Data augmentation, including scaling, rotating, and flipping, was used to triple the training set size to ensure robustness.
Results and Discussion
The model achieved a Dice score of 0.87 for "Complete" tumor segmentation, placing it among the top-ranking methods on the BRATS 2013 leaderboard. Incorporating symmetry maps provided a 3% improvement in Dice scores and significantly reduced false positives by helping the network distinguish tumors from similar-looking healthy tissue.
Efficiency is a major advantage of this approach. The FCN processes entire slices at once, taking approximately 2 minutes to process a full 3D volume. This is significantly faster than other high-performing methods, which can take up to 85 minutes. The authors found that a shallower 3-block architecture outperformed deeper variants in both accuracy and convergence speed, suggesting that medical images rely more on low-level texture features than high-level semantic abstractions.
Conclusion
This research demonstrates that a streamlined, symmetry-aware FCN can achieve state-of-the-art results in brain tumor segmentation with high efficiency. While 95% of current processing time is spent on CPU-based symmetry map calculation, future GPU parallelization could further enhance real-time clinical viability. Validation on larger, more diverse clinical datasets remains a priority for future work.
Summary updated