ARCHIVES

This issue remains challenging due to the absence of annotated image data, as well as the demand for incorporating both long-distance spatial dependencies and finer details. In this paper, we propose a novel deep learning architecture, which employs a staged transfer learning along with probabilistic fusion strategy to effectively integrate transformer architecture with CNN models. Specifically, a Swin Transformer Tiny model is progressively fine-tuned to learn hierarchical global relationships, while the ResNet50 is used as a backbone to extract effective local features. Herein, we introduce a sequential fine-tuning approach alongside score-level ensemble, unlike existing hybrid approaches.A brain tumor MRI dataset that contains the multi- class classification of glioma, meningioma, pituitary, and non- tumor classes is used for evaluating the proposed framework. Many experiments demonstrate that the optimal Swin Trans- former achieves an accuracy rate of 98.52% during the validation process, whereas the proposed hybrid ensemble outperforms individual classifiers in reaching the maximum testing accuracy rate of 99.53%.From the results, it is clear that using global attention by means of transformers together with CNN-based local features has made classification more robust, thus making the proposed system an excellent choice for medical image analysis.