Automated Brain Tumor Segmentation in MRI using Superpixel Over-segmentation and Classification

Brain tumor segmentation is a challenging task due to the strong fluctuation in intensity and shape. It has attracted the attention of medical imaging community for several years. This work introduces a fully automated brain tumor segmentation approach from multimodal MRI images. Segmentation in three different MRI modalities; T1 (gadolinium-enhanced), T2, and Fluid-Attenuated Inversion-Recovery (FLAIR) are compared to choose the best one. The proposed approach utilizes a super-pixel over-segmentation technique and applying a classification for each super-pixel which leads to more smooth segmentation. Several features including statistical, fractal, and texture features are calculated from each super-pixel of the normalized (T1, T2, and flair) images to ensure a robust classification. Additionally, the class imbalance problem is tackled to allow the algorithm to accurately segment abnormal tissue. The Random Forest (RF) classification algorithm is utilized for final segmentation. The RF classifier is being chosen in the proposed approach because it provides a better performance according to the confusion matrix results. The proposed approach has been trained using 10 Low-Grade and 20 High-Grade cases and evaluated using different 5 Low-Grade and 5 High-Grade cases from BRATS 2013 dataset. Dice, average precision, sensitivity, and F1-score metrics are used for segmentation accuracy evaluation. The average precision, sensitivity, fl-score and dice overlap for tumor segmentation are 92%, 95%, 96% and 94% for flair images, 89%, 92%, 90% and 93% for T2 and 89%, 90%, 89% and 90% for T1. Finally, the voting strategy is being used to get the best segmentation between these different modalities.