Brain Tumor Classification, Detection, and Segmentation in MRI using Deep Learning Approaches: A Survey

Abstract

Automated classification, detection, and segmentation of brain tumors from magnetic resonance imaging (MRI) have become central applications of deep learning in neuro-oncology, driven by the time-intensive nature of manual radiological assessment and the inter-observer variability associated with it. This paper first surveys recent deep learning approaches for glioma, meningioma, and pituitary tumor analysis, grouping them by methodological strategy: transfer learning with pretrained convolutional backbones, hybrid CNN-Vision Transformer ensembles, activation-function optimization, attention-enhanced object detectors, encoder-decoder segmentation networks, and explainability-oriented ensemble classifiers. Across the reviewed literature, transfer learning and ensembling consistently outperform single models trained from scratch, attention mechanisms are increasingly embedded in both classification and detection pipelines, and Grad-CAM-based visual explanation is emerging as a standard component of trustworthy systems. However, the majority of studies rely on a narrow set of overlapping Kaggle datasets, apply single train-test splits rather than systematic cross-validation, and depend on colour or intensity-based feature maps that provide only implicit boundary information, resulting in persistently poor delineation of diffuse, infiltrative tumor margins such as those seen in glioma. Motivated by this gap, the paper proposes EdgeFocus U-Net, a boundary-aware segmentation architecture that combines a ResNet-based encoder, a multi-scale feature pyramid, an explicit Sobel-derived edge detection branch, and an Edge-Guided Attention Module that fuses edge and semantic features before decoding. The network is trained under a dual-supervision scheme with a joint segmentation head and edge-prediction head, optimized using a combined Dice, binary cross-entropy, and edge loss. This design is intended to sharpen boundary localization while remaining lighter than transformer- or Mamba-based alternatives. The architecture, its constituent modules, and the training objective are described in detail; a discussion of the reviewed literature is used to contextualize the proposed design choices and to identify directions for its future experimental validation.