DEEP LEARNING APPROACHES FOR BRAIN TUMOR SEGMENTATION IN MRI: A SYSTEMATIC LITERATURE REVIEW AND EXPERIMENTAL EVALUATION

Abstract

Brain tumor segmentation from magnetic resonance imaging (MRI) is an essential step in the diagnosis, treatment planning, and follow-up of brain tumor growth. Manual tumor region annotation is time-consuming, subject to individual interpretation, and needs expert radiological assessment. Deep learning-based automatic segmentation methods have been proposed as a solution to overcome this by allowing quick, precise, and reproducible tumor delineation. This dissertation is a critical examination of deep learning approaches to brain tumor segmentation with special emphasis on designing, implementing, and assessing a new hybrid model referred to as HybridSegNet++. The model integrates convolutional neural networks (CNNs) with MobileViT blocks and gated residual skip connections to improve feature extraction, representation learning, and gradient transmission across network layers. Training and evaluation occur on the BraTS 2020 dataset using multi-modal MRI sequences (T1, T1ce, T2, and FLAIR) as inputs. The data pipeline employs a custom DataGenerator class for normalization of 2D slices to a common 192×192 resolution with one-hot encoded segmentation masks. We use a weighted categorical crossentropy and Dice loss hybrid loss function to address class imbalance for tumor subregions. Performance is quantified by the usual segmentation metrics such as per-class Dice score values (Tumor Core, Enhancing Tumor, Whole Tumor), Intersection over Union (IoU), accuracy, precision, sensitivity, specificity, and mean IoU. Extensive literature review of state-of-the-art models—ranging from U-Net architectures to Transformer-based models—is also performed in a bid to place into perspective the strengths of the model proposed. The best-performing model possesses competitive performance and is capable of generalizing across unseen test examples, thereby proving its viability for clinical use in computer-aided diagnosis systems.