An Accurate Brain Tumor Segmentation using Deep Learning

Gliomas, which can appear in different sizes, and locations, and with scattered boundaries, pose a significant challenge as an aggressive type of tumor. Convolutional Neural Networks (CNNs), one of the most effective deep learning approaches for image analysis problems, have been utilized to develop an automatic deep learning-based 2D-CNN model for brain tumor segmentation in this study. The architecture of the model was designed to be deeper by using small convolution filters (3x3). Additionally, the number of convolutional layers was increased to 16 for the HGG model and 13 for the LGG model to facilitate efficient learning of complex features from large datasets and achieve better results. Fine tuning among the dataset and hyperparameters was employed to obtain the results. The pre-processing for this model includes the generation of a brain pipeline, intensity normalization, bias correction, and data augmentation. The proposed method’s performance was evaluated using the BRATS-2015 dataset and the Dice Similarity Coefficient (DSC). Our method achieved DSC scores of 0.79, 0.77, and 0.78 for complete, core, and enhanced tumor regions, respectively. These results are comparable to other 2D CNN architecture-based methods.