SURVIVAL PREDICTION IN GLIOBLASTOMA BRAIN TUMOR USING SEGMENTATION AND DETECTION WITH ADVANCE COMPUTATIONAL TECHNIQUES

Abstract

In recent times, a wealth of evidence has emerged, indicating a notable increase in brain tumor cases, solidifying its status as the 10th most prevalent type of tumor, affecting both children and adults. Glioma tumors, assessed pathologically, are divided into the formidable glioblastoma (GBM/HGG) and the less aggressive lower grade glioma (LGG). Glioblastoma, among various brain tumors, stands out as the most lethal and aggressive. Within gliomas, diverse histological subfields include peritumoral edema, a necrotic core, and enhancing/non-enhancing tumor cores. Radiology, specifically magnetic resonance imaging (MRI), plays a vital role in unraveling the phenotypic intricacies and intrinsic heterogeneity of gliomas. Utilizing multimodal MRI scans, such as T1-weighted, contrast-enhanced T1-weighted (T1GD), T2-weighted, and fluid attenuation inversion recovery (FLAIR) images, provides a holistic understanding of different glioma subfields. The need for precise predictions in overall survival, diagnosis, and treatment planning for glioma patients is met through automated algorithms embedded in a brain tumor segmentation and detection framework. These algorithms leverage fragmented tumor subfields and radiometric characteristics from multimodal MRI scans. The thesis introduces a model framework encompassing tumor classification, detection, and localization, integrating advanced Deep Learning algorithms as a foundational layer with localization techniques. The first method employs cutting-edge algorithms like Inception-V3, InceptionResNet-V2, MobileNet, NASNetMobile, ResNet-101, Xception, DenseNet, ResNet-50, and EfficientNetV7 for classification. This involves adding an extra layer of activation, normalization, and density. The power of repeated blocks is utilized in a multi-branch network topology, clearly defining the head (prediction), body (data processing), and stem (data intake). This design pattern, initiated with the first two or three convolutions processing the object in the stem, persists in contemporary deep networks. The second strategy applies the RESUNET model to segment and localize brain tumors, utilizing TCGA MRI data. This comprehensive approach delves into the intricate landscape of glioblastoma, merging advanced computational techniques with medical imaging to push the boundaries of survival prediction and treatment guidance. According to the BraTS classification, a patient's survival outcome falls into three groups: short-term survivors, mid-term survivors, and long-term survivors. The thesis emphasizes effective feature selection from MRI images and advocates for the use of a deep learning-inspired replicator neural network for the task of Overall Survival prediction.