Review on Hybrid Deep Learning Approaches for Early Detection and Segmentation of Brain Tumors from MRI Images Using the BraTS Dataset

Abstract

Brain tumors require accurate and timely detection and segmentation to support effective diagnosis and treatment planning. Magnetic Resonance Imaging (MRI) provides valuable multimodal information, but manual analysis is time-consuming and challenging because of tumor heterogeneity, irregular morphology, and unclear tumor boundaries. This review aims to examine hybrid deep learning approaches for automated brain tumor detection and segmentation, with particular emphasis on the BraTS dataset. The reviewed studies are analyzed according to their architectural strategies, datasets, evaluation metrics, and reported performance, covering CNN-based, U-Net-based, CNN–Transformer, attention-based, 3D, optimization-based, and multimodal approaches. The findings indicate that hybrid architectures can effectively combine local spatial, global contextual, volumetric, and multimodal features to improve tumor detection and segmentation. Reported studies achieved promising performance, including 98.4% accuracy and 96.8% sensitivity, Dice scores of up to 0.975, and classification accuracy of 98.78%. However, important challenges remain, including limited annotated data, class imbalance, computational complexity, poor cross-dataset generalization, limited interpretability, and insufficient clinical validation. Future research should focus on integrated detection and segmentation, robust multimodal fusion, efficient CNN–Transformer models, semi-supervised learning, explainable AI, and multi-center clinical validation. Overall, hybrid deep learning demonstrates strong potential for accurate, robust, and clinically applicable MRI-based brain tumor analysis.