Evaluation of Multi-Modal MRI Images for Brain Tumor Segmentation

Brain tumors are the most threatening types of tumors and are classified into four grades. To capture the contrast variations of each grade of tumor, currently, the Magnetic Resonance Imaging (MRI) images are considered as the gold standard and are preferred due to their non-invasive nature. However, accurate segmentation of these tumors relies on computational models. Existing models use multimodal MRI images for the segmentation of each grade of a tumor which results in downgraded performance in some tumor regions. In this paper, we propose a new method that incorporates a deep learning-based model called U-Net to address the problem of brain tumor segmentation. The conceptual novelty of our method lies in modifying the U-Net model to make it suitable for multi-grade tumor segmentation task. Additionally, we use this architecture to evaluate individual MRI modalities for the segmentation of brain tumors. The proposed method is evaluated on Brain Tumor Image Segmentation (BRATS) 2015 dataset and has achieved significantly good results. It is pertinent to mention here that by using only
$\mathbf{T}_{1}$
modality, we achieved the highest score for complete sub-tumoral region. On the other hand, a better result in core sub-tumoral has been achieved by using
$\mathbf{T}_{2}$
modality, while the highest score in enhancing sub-tumoral region has been achieved with the help of
$\mathbf{T}_{\mathrm{f}}$
modality.