Brain tumors pose a significant threat to human health, necessitating accurate segmentation for effective diagnosis and treatment. The diverse shapes and sizes of tumors, imbalance between background and foreground regions, and challenges in integrating multimodal image data complicate automated segmentation. This paper presents a multimodal brain tumor segmentation method based on a multi-encoder pyramid U-Net (MEP-UNet) designed to fully leverage multimodal MRI data and enhance segmentation accuracy. The approach employs a multi-encoder architecture with ResConvASPP convolutional blocks to extract detailed features from diverse modalities. A pyramid segmentation attention (PSA) module in the decoder facilitates attention-weighted fusion of multi-scale features, enhancing informative content. Post-processing involves a connected domain-based method to eliminate small region noise. Experimental results on the BraTS2018 dataset demonstrate the efficacy of the MEP-UNet method, the Dice similarity metric for the enhancing tumor, tumor core, and whole tumor regions was 0.769, 0.780, 0.884, while the Hausdorff distances measured 3.501, 8.145, 6.280. These results underscore the method’s robust segmentation performance.