Deep Learning-based MRI Image Segmentation Of Brain Tumors

Brain tumor examination is an active domain of study,which has received a lot of consideration from both the medical and the technical sectors in the past years.Brain tumors,especially glioma a sickness with a very low survival rate.3.6% is the relative five-year survival rate for people diagnosed with high-grade glioma.The segmentation of brain tumors into distinct groups is an important diagnostic and treatment-planning strategy for medical experts.Magnetic resonance imaging(MRI)is the imaging technique most frequently utilized in clinical work with brain malignancies.Brain tumor segmentation includes the delineation of cancerous structures,such as a whole tumor,core tumor and enhance tumor,and healthy brain tissues,commonly classified in gray matter,white matter,and cerebrospinal fluid.Timely diagnosis of brain tumors can help decrease the human death rate.The method of manually segmenting huge MRI scans is a challenging and time-consuming task.Various tumor regions remain unidentified due to their lesser size and the variation in area occupancy among tumor structures.Owing to the heterogeneous tumor regions,segmenting brain tumors has proved to be a challenging job.However,these problems can be overcome with the help of computer-aided diagnostic systems that act as a second opinion in the analysis of brain tumor segmentation.This work only focuses on brain tumor sub-regions segmentation,although several techniques have been employed to develop automated deep-learning algorithms for brain tumor segmentation.The main problem with current methods is that when brain tumor sub-regions are detected,a significant percentage of false positives are produced,and many true positive areas are missed which results in poor performance on segmentation evaluation measures.Deep learning techniques have produced excellent outcomes in the segmentation of brain tumors and many other biomedical applications.Developing deep learning-based systems is a challenging task owing to the uneven,irregular,and unstructured tumors boundary connections among substructures such as whole tumor(WT,comprising all classes of tumor structures),tumor core(TC,comprising a necrotic core,non-enhancing core),and enhancing tumor(ET).Thus,this thesis aims to develop state-of-the-art deep learning-based automatic systems for brain tumor segmentation.This thesis uses and yields motivation from the well-known U-Net architecture.The main contributions thesis is as follows:Firstly,this dissertation proposed an improved 3D U-Net encoder-decoder-based deep convolutional networks algorithm for the automatic segmentation of brain tumor sub-structures.The designed model includes Conv-Net units as the basic building blocks of U-Net architecture,followed by transition layer blocks and skip connections,which strengthens the fusion of various scales by associates at several network layers.These several skip connections support the transmission of high-resolution features from shallow layers to deeper layers.The developed model employs the transition layer block,which reduces feature maps by effectively generalizing unique local patterns.Moreover,the dice loss function helps in the suitable segmentation of trivial anomalies that are likely to be misclassified.Therefore,the designed approach performs well in segmenting brain tumors from MRI scans with diverse shapes,boundaries,and dimensions.Secondly,using a modified U-Net model,a new multi-step cascaded method was presented for segmenting brain tumor substructures from non-invasive multimodal MR data.Dense-Net Encoder blocks and Conv-Net Decoder blocks comprise the suggested U-Net structure.Both blocks contribute to the extraction of local and global relevant data from MRI slices of the brain.A modified 3D U-Net with Dense-Net Encoder and Conv-Net Decoder blocks rather than the original U-Net architecture provides a solid and more generalized performance.Finally,to reduce the cascaded network problem and increase the segmentation results on various evaluation measures,in this research,we build a U-Net system with various U-Net modules to acquire spatial data at different resolutions using artificial intelligence.To extract and fully leverage adequate features,we up sample feature maps across multiple resolutions using residual inception modules.To reduce the computational complexity of the network,we use 3D depth-wise separable convolution in the down-inception and up-inception components.Furthermore,we presented a combination of the binary cross-entropy and dice loss functions as a solution to the problem of high-class imbalance.Without the need for extensive adjustment of weight hyper-parameters,our method can produce excellent segmentation results.