Thesis (MSc) - Cyprus International University. Institute of Graduate Studies and Research Electrical and Electronics Engineering Department

Includes bibliography (sheets 56-62)

ABSTRACT
Gliomas are one the deadliest forms of brain tumor, often resulting in a short life when they reach a high grade. Early detection of glioma is critical for patient survival. Magnetic resonance images (MRI) are frequently used to evaluate brain malignancies. Segmenting tumors from Magnetic resonance images of the brain is one of the highest priorities areas of medical science. Semantic segmentation gives detection information and helps doctors know the disease's early stage. A convolutional neural network is highly effective in segmenting medical images. This study presents a new deep learning method for accurate brain tumor segmentation that can be modified and expand the residual unet architecture. It increases the network's depth while keeping an extremely high level of accuracy. This study proposes a deep learning network architecture called increased depth of 3D Residual UNET Network Architecture with an attention gate for Brain Tumors Segmentation, which contains an attention gate and advanced 3D Residual UNET. The proposed architecture has increased the depth of the normal attention residual unet from four layers to six layers. However, the network loses a corresponding amount of spatial information, lowering segmentation performance. The 3D UNet transmits contextual and spatial information from the encoder part to the decoder by using skip links. Consequently, critical spatial information lost during down sampling can be recovered more effectively. By allowing only activations from important areas on the encoder side using attention gates and creating better feature mappings at the decoder, these modifications to the network enhanced the process of learning. Furthermore, the use of a combination of dice loss and focal loss helped the model in resolving class imbalance challenges where brain tumors have a significant imbalance between foreground and background classes. Because of this, the model has improved and got a better segmentation achievement. The model outperformed baseline models such as UNet, Residual UNet, and attention gates with Residual UNet. Three separate datasets are evaluated to demonstrate that the presented model is superior to its baseline models and the existing state-of-the-art segmentation approaches. The suggested model was tested on BraTS 2020, BraTS 2019, and BraTS 2018 datasets. The model achieved the dice coefficient scores for WT, TC, and ET of 93.91%, 93.01%, and 89.21% on the BraTS 2020 dataset, 88.44%, 75.11%, and 79.87% on the BraTS 2019 dataset, and 88.36%, 83.17%, and 78.19% on the BraTS 2018 dataset, respectively.