Self-supervised Advanced Deep Learning for Characterization of Brain Tumor Aggressiveness and Prognosis Analysis Through Multimodality MRI Imaging

Identificador:  TDX:4289

Handle: https://hdl.handle.net/20.500.11797/TDX4289

Autores:  Mazher, Moona

Resumen:
Early detection, automatic delineation, and volume estimation are vital tasks for survival prediction and treatment planning of brain tumor patients. However, gliomas are often difficult to localize and delineate with conventional manual segmentation due to their high variation of shape, location, and appearance. Moreover, manual mark delineation is laborious and time-consuming work for a neurosurgeon. In addition, it is difficult to replicate the segmentation results due to certain practical operation factors. In recent years, convolution neural networks (CNNs) are widely used for the automated classification and segmentation of medical images. Therefore, the focus of the present thesis is to develop a system for automating brain tumor analysis (such as brain tumor segmentation, and survival prediction), using deep learning techniques and applying them to the MRI images for segmenting the brain tumor classes (Enhancing Tumor, Non-enhancing Tumor, and Peritumoral Edema) and estimating the survival days of the patients for prognosis analysis. In this study, various 2D and 3D based deep learning models were designed and tested for the multi-class brain tumor segmentation and survival prediction of the brain tumor patients. We proposed a 2D CNN model (BrainSeg-DCANet) then we proposed a 2D multiview (axial, sagittal, and coronal) deep inception residual network for brain tumor segmentation. Thereafter, a 3D CNN based Two-stage Self-supervised Contrastive Learning using Parallel Multiview Multiscale Attention-based CNN Transformers for 3D brain tumor volumetric segmentation was introduced. Finally, a 3D MR image-based survival prediction was performed. Multiple feature extraction techniques are used to extract the features from the 3D volumetric MRI image and then different regression techniques are applied to the extracted features. The thesis’s findings showed that the proposed techniques can produce a clinically helpful computer-aided tool for brain tumor segmentation and survival prediction by MRI Images.