As a significant brain disorder, brain tumors constantly threaten human health, making early diagnosis crucial. In the analysis of existing methods, deep learning, with its ability to automatically extract features through multi-level nonlinear transformations, demonstrates superior performance in diagnosing brain tumors. In this study, a deep-learning-based brain tumor diagnosis model that focuses on multi-level feature analysis is proposed. ResNext is employed to extract multi-level features. A spatial attention mechanism combining linear layers and large-kernel convolutions is designed to analyze multi-level contextual information. Moreover, the Transformer structure is integrated to dynamically fuse multi-level features, generating feature maps with high expression power for the final diagnosis. The model is trained and evaluated on the Kaggle dataset for two-class and four-class brain tumor classification tasks. Experimental results show that the model achieves an accuracy of 99.47% in distinguishing between no tumor and tumor, and an accuracy of 99.75% in distinguishing between no tumor and three types of tumors. Compared with other deep-learning models, the proposed method demonstrates superior diagnostic capabilities, enabling early diagnosis with high accuracy.