Automatic tumor detection from brain MRI images using deep learning techniques

The detection of brain tumors from MRI images is a critical concern for human health and often requires risky surgeries for treatment. Experts can identify brain tumors using high-resolution Magnetic Resonance (MR) imaging. However, nowadays, many researchers are employing deep learning methods to detect brain tumors from MR or Computed Tomography (CT) images. Significant success has been achieved in tumor detection from MR or CT images through these studies. Nevertheless, there are still many areas in this field that need improvement, especially in accurately determining the location and size of tumors. In this thesis, a new approach has been developed using deep learning methods to automatically detect brain tumors from MRI images with a higher success rate than previously proposed methods. The brain tumor detection process in this thesis is carried out in two stages: in the first stage, segmentation is performed to determine the location and dimensions of the brain tumor, and in the second stage, classification is performed to determine the types of tumors. For the segmentation process, the UNet architecture, which is one of the deep learning networks, is used as a hybrid model with the pre-trained DenseNet121 architecture. The proposed model has been validated on the BraTS 2019 publicly available brain tumor dataset, which includes high-grade and low-grade glioma tumors. Experimental results demonstrate that our model outperforms other state-of-the-art methods in this field. For the segmentation process, Dice Similarity Coefficient (DSC) values of 95.9%, 94.3%, and 89.2% are obtained for the whole tumor, core tumor, and enhancing tumor, respectively. An 8-layer CNN-based model is designed for the classification of brain MR images. The proposed approach achieves detection accuracies of 99.64% for glioma tumors, 96.53% for meningioma tumors, and 98.39% for pituitary tumors.