Automatic Brain Tumor Segmentation In Multimodal MRI Using Deep Neural Networks

Brain cancer is one of the leading causes of cancer death,so that early screening is the best strategy to diagnose and treat brain tumors.MRI is one of the most prominent methods commonly utilized to visualize brain tissue in medical imaging.Brain tumors are usually diffuse and poorly contrasted.Its borders are frequently fuzzy and difficult to differentiate from healthy tissue(white,gray,and CSF),making them difficult to segment.Gliomas are very difficult to recognize with handcrafted segmentation due to differences in brain tumor scale,shape,and function.It is also time intensive and tedious to segment manually.Automated segmentation can lead to a more accurate and more straightforward diagnosis and treatment.Automatic image segmentation approaches using deep learning methods have recently made significant strides.The main research work of the dissertation can be divided into the following sections:Firstly,we proposed Hybrid Two-Track U-Net(HTTU-Net)architecture to address the challenges of brain tumor segmentation.HTTU-Net extracts more semantic information and gives more consideration to the information of small-scale brain tumors,which improves the segmentation of brain tumors.HTTU-Net is based on the excellent achievement of U-Net based architectures.It also updates the U-Net network by adding batch normalization at the end of each block to reduce the mean and variance problems and stable the layers.The first track focuses on the tumor’s form and size,while the second track captures the contextual information.Each track consists of a different number of convolution blocks and uses a different kernel size to handle the different tumor sizes.We have introduced a new hybrid loss feature to mitigate the class imbalance,combining focal loss and generalized dice loss functions.We validate the performance on Brats 2018 dataset.The experimental results show that our proposed model performs comprehensively better than the existing state-of-the-art methods and baseline.Secondly,we proposed a novel Multi Inception Residual Attention U-Net model(MIRAU-Net)that extracts substantially more features,besides,to gain and restore information about the locations of tumors of the brain,which enhances segmentation efficiency.Encoder and decoder sub-networks are linked in MIRAU-Net by Inception-Res paths to deeper and extend the proposed network.The re-modeled skip paths of the architecture with gate signal are forwarded to the attention gate attempt to improve the capacity of expression and feature extraction and to decrease the gap between the encoder and decoder sub-networks.A new multi-loss function is introduced,combining weight loss,generalized dice loss,and focal Tversky loss to mitigate the class imbalance.The evaluation performance of MIRAU-Net was checked with Brats’ 2019 dataset online.Experimental results reveal that the suggested MIRAU-Net beats its baselines,attention gate u-net,and provides excellent performance,especially for segmentation small-scale brain tumors compared to recent techniques for the segmentation of brain tumors by a large margin.Thirdly,we proposed a novel inception residual dense nested U-Net(IRDNUNet)for solving the insufficient precision of small-scale tumors with fewer numbers of parameters.It can extract more representative features from brain tumors,enhancing segmentation accuracy.In IRDNU-Net,the standard convolutional layers used in the U-Net architecture are replaced by carefully designed residual inception modules to widen the network structure.The IRDNU-Net encoder and decoder subnetworks are linked by many nested dense paths to increase the network’s depth.The proposed segmentation architecture was evaluated on two large brain tumor segmentation benchmark datasets.Experimental results illuminate that IRDNU-Net outperforms u-net and inception-residual methods.Moreover,IRDNU-Net is capable of achieving comparable segmentation accuracy with fewer parameters.Lastly,we proposed an encoder-decoder network with a depth-wise atrous spatial pyramid pooling network(EDD-Net)for brain tumor segmentation to extract multi-context information.In the encoder and decoder module,the Dilated-Res Net block with Squeeze-Excitation module is introduced to improve segmentation accuracy by training deeper networks with fewer parameters.We integrate the DASPP block with the backbone encoder-decoder structure and the final output layer.We propose an up-sampling module based on the residual model and a downsampling module that can better integrate low and high-level feature information.On the Brats 2019 datasets,our proposed model’s high accuracy and robustness were verified.Experimental results showed a significant improvement and robustness in handling small tumor regions with fewer parameters.In conclusion,this research aims to introduce novel and efficient approaches for segmenting the abnormal tissues associated with brain tumors from the multimodal MRI images,which will aid radiologists in the diagnosis and treatment planning of patients with cancer also enhancing the segmentation accuracy,especially for small tumors,at the same time less memory and computation consumption.Therefore,this dissertation has great practical significance.