Abstract
This paper introduces a fully automated pipeline for segmenting gliomas in Fluid-Attenuated Inversion Recovery (FLAIR) MRI by classifying superpixels rather than individual voxels. By combining Extremely Randomized Trees (ERT) with a lean set of fractal and textural features, the authors achieved a Dice overlap score of 0.91, matching expert performance while significantly reducing computational overhead. The method demonstrates robustness across various tumour grades (II through IV) and provides a more efficient alternative to traditional voxel-wise classification.
Introduction
Delineating brain tumours, specifically gliomas, is a critical step in clinical treatment planning and monitoring. However, manual segmentation remains a significant bottleneck; it is time-consuming, requiring approximately 12 minutes per patient, and is highly subjective, leading to inter-observer variability.
Automated methods face two primary challenges: the irregular and indistinct boundaries of gliomas (particularly low-grade ones) and the high computational cost of traditional voxel-based machine learning models that treat every pixel as an independent data point. This study addresses these hurdles by proposing a superpixel-based approach that is both faster and more robust across different tumour grades.
Methods
The proposed pipeline consists of four distinct stages designed to move from raw MRI data to a final tumour mask.
1. Pre-processing
Initial steps include standard skull stripping and histogram matching. These ensure intensity consistency across different scans, which is vital for the performance of subsequent machine learning models.
2. Superpixel Segmentation
Instead of analyzing individual voxels, the image is partitioned into "superpixels" using the Simple Linear Iterative Clustering (SLIC) algorithm. Superpixels are irregular patches of pixels with similar intensities that respect the natural boundaries of brain tissue. This grouping reduces the number of entities to classify by orders of magnitude and ensures cleaner segmentation boundaries.
3. Feature Extraction and Selection
For each superpixel, 28 features were calculated, including intensity statistics (mean, variance, skewness), textons (via Gabor filter banks), fractal analysis (using Segmentation-based Fractal Texture Analysis), and curvature. To optimize performance, the Minimum Redundancy Maximum Relevance (mRMR) algorithm was used to select the five most impactful features.
4. Classification
The study employed Extremely Randomized Trees (ERT), an ensemble method similar to Random Forests. ERT introduces more randomness in node splitting, which helps prevent overfitting and improves accuracy on imbalanced datasets where healthy tissue significantly outnumbers tumour tissue.
Results
The model was validated using an internal clinical dataset (19 FLAIR images, Grades II-IV) and the BRATS 2012 benchmark dataset (30 FLAIR images).
- Internal Dataset: Achieved an average Dice overlap of 0.91 and a detection sensitivity of 89.48%.
- BRATS 2012: Achieved a Dice overlap of 0.88 and a sensitivity of 88.09%.
- Robustness: The ERT method maintained a high Dice score (~0.91) across all tumour grades. In contrast, a Support Vector Machine (SVM) baseline struggled with Grade II tumours, achieving a mean Dice of only 0.81.
Discussion and Conclusion
The transition from voxel-wise to superpixel-based classification offers significant advantages. Superpixels adhere to image boundaries, resulting in segmentations that are naturally more aligned with anatomical structures. Furthermore, the inclusion of fractal analysis proved crucial, as tumour tissue is structurally more "chaotic" than healthy tissue, a characteristic effectively captured by fractal dimensions.
While the current study relies solely on FLAIR MRI sequences, future work should incorporate multi-modal data (e.g., T1-weighted with contrast, T2-weighted) to better distinguish between specific tissue types like necrosis. Additionally, automating the selection of optimal superpixel sizes via optimization algorithms could further enhance the pipeline's efficiency and accuracy.
Summary updated