Accurate classification of brain tumors from magnetic resonance images is essential for supporting clinical assessment, although variations in tumor morphology, anatomical location, image orientation, and visual similarity among diagnostic categories can limit the effectiveness of conventional deep learning models. This study proposes a hybrid learning framework that integrates local structural information, global contextual relationships, and frequency domain characteristics for four class brain tumor classification. The model combines an ImageNet pretrained ConvNeXt Tiny branch for local feature extraction, a Swin Transformer Tiny branch for hierarchical contextual representation, and a wavelet convolutional neural network branch based on two level Haar decomposition. The resulting branch features are aligned within a shared representation space and integrated through bidirectional cross attention and adaptive gated fusion before classification. A dataset containing 10,464 brain magnetic resonance images was evaluated at the individual image level, including 2,726 glioma, 2,737 meningioma, 2,300 no tumor, and 2,701 pituitary tumor images. The dataset was divided into 9,417 training images, 653 validation images, and 394 independent testing images. Brain contour cropping, contrast limited adaptive histogram equalization, normalization, and controlled augmentation were applied during preprocessing. On the independent test set, the proposed model correctly classified 389 of 394 images, achieving an accuracy of 98.73% with a 95% confidence interval from 97.06% to 99.46%. It further obtained a macro precision of 98.84%, macro recall of 98.82%, macro F1 score of 98.82%, macro specificity of 99.56%, Matthews correlation coefficient of 0.9830, and macro area under the receiver operating characteristic curve of 0.9995. The proposed framework outperformed the strongest individual backbone, Swin Transformer Tiny, which achieved 96.45% accuracy and a macro F1 score of 96.63%. These findings indicate that coordinated integration of spatial, contextual, and frequency information can improve image level brain tumor classification. However, the absence of patient identifiers and magnetic resonance sequence metadata limits patient level interpretation and sequence specific analysis.
Copyrights © 2024