BRAIN TUMOR DETECTION FROM MRI USING A HYBRID CNN–TRANSFORMER DEEP LEARNING FRAMEWORK

Abstract

Early and accurate detection of brain tumors from magnetic resonance imaging (MRI) is crucial for timely clinical intervention, however, is challenging due to the heterogeneous appearance of the tumors and to the indistinct tumor boundaries, as well as to the limitations of using conventional feature extraction. To effectively capture both fine-grained local features and long-range contextual dependencies, this study proposes a Dual-Path Adaptive CNN–Transformer Network (DPA-CTNet), a hybrid network. DPA-CTNet combines a multi-scale convolutional encoder, a lightweight windowed Transformer branch, and proposes a Cross-Attention Fusion Module (CAFM) for bidirectional alignment of convolutional feature maps and Transformer tokens. A Multi-Scale Gated Boundary Attention (MSGBA) block is used to further improve the representations of tumor–tissue boundaries. DPA-CTNet consists of a series of adaptively cross-modal fused layers, allowing complementary local and global representations to interact across the network, unlike conventional hybrid architectures that fuse CNN and Transformer features at a single stage. The proposed framework obtains 98.45% accuracy, 98.12% F1-score, and 0.996 AUC-ROC on the Figshare brain tumor benchmark, which is better than the evaluated CNN, attention-enhanced CNN, ViT, and hybrid baselines. Furthermore, DPA-CTNet only uses 21.1M parameters and 5.5 GFLOPs, which shows good accuracy–efficiency balance. Based on these results, it can be concluded that it has the potential as a computationally efficient system for automatic classification of brain tumors in resource limited clinical settings.