This study explores the application of computer image processing technology in brain tumor segmentation. Based on the BraTS 2021 dataset, the performance of traditional methods, machine learning methods, and deep learning methods were compared. Experimental results show that deep learning methods, especially the improved SegFormer model, performed the best on multimodal data. By using hybrid methods and ensemble learning strategies, the Dice coefficient was increased to 87.2%, and metrics such as accuracy, sensitivity, and specificity also showed excellent performance. The study further analyzed the computational efficiency of different models, providing references for practical applications. Through the comprehensive application of various advanced technologies, this research made significant progress in both the accuracy and efficiency of brain tumor segmentation, offering reliable technical support for clinical applications.