Deep Learning Approaches for Brain Tumor Detection in MRI Images: Recent Advances, Challenges, and Future Research Directions

Brain tumors are serious disorders of the brain and require early and precise diagnosis for proper treatment. MRI is the imaging modality of choice as it provides excellent soft tissue contrast, and doesn't involve any type of radiation. In recent years, deep learning techniques such as CNN, U-Net, Vision Transformers, hybrid architectures and attention-based models have contributed to the advancement of automated brain tumor detection, segmentation, and classification. This review gives an overview of the recent deep learning methods, including image pre-processing, feature extraction methods, segmentation methods, classification methods, benchmark datasets, and evaluation metrics. It also covers the main problems encountered in clinical use, including a lack of annotated datasets, class imbalance, lack of model generalizability, computational cost, interpretability, and deployment. Lastly, future research trends are emphasized such as self-supervised learning, explainable AI, federated learning, multimodal learning, lightweight architectures, and foundation models. This review will be valuable for suggesting directions for future research and development that could lead to more accurate, efficient, and clinically usable deep learning applications in the diagnosis of brain tumors.