Deep Learning Method Based On U-Net Model For Semantic Segmentation Of Brain Tumor In 3D-MRI Images

Automatic brain tumor segmentation is a crucial step in the diagnosis and treatment of medical image processing and has been a subject of significant research interest.However,in recent years,deep learning methods have emerged as a promising approach for brain tumor segmentation.Although many techniques for brain tumor segmentation have been developed,they still have some limitations that can affect their accuracy and reliability of segmentation.To address the issue of segmentation accuracy,this thesis proposed a novel deep learning method called "GRes U-Net," which is based on the classic U-Net structure and utilizes grouped convolutions and residual blocks aims to improve the accuracy of brain tumor segmentation.The proposed method enhances the feature maps at different stages in the encoder-decoder parts of the network to help the model learn a diverse set of low-level and high-level features and resolution information.Moreover,a recurrent block mechanism is added in the decoder part to aid in the recovery of a features lost during the down-sampling process.To address the issue of many image segmentation methods being biased toward the majority class and ignoring the minority class of small size tumors,an improved Weighted Focal Loss is proposed based on 3D U-Net to enhance the segmentation of small tumors in3 D MRI brain tumor images.The proposed loss function Weighted Focal Loss(WFL)aims to address the imbalance between classes and the imbalance between weights by giving higher weights to the minority and lower weights to the majority.After assigning these weights to different pixel values,the work is able to resolve pixel degradation,which is one of the limitations of the loss function during model training.Experimental validation of the proposed GRes U-Net method is conducted using data from multimodal brain tumor databases,such as the Bra TS 2018-2019 HGG database,and improved WFL based 3D U-Net model tested on the Bra TS 2019-2020 for high-grade glioma(HGG)and low-grade glioma(LGG).The proposed methods achieved promising results for segmenting the tumor core(TC),whole tumor(WT),and enhancing the tumor(ET).