Deep learning technology has been widely applied in the field of medical image segmentation; however, precise segmentation of bladder images remains challenging. This study addresses the issue of bladder MRI image segmentation by proposing an improved algorithm based on feature fusion and attention mechanisms (MIV-Net). The algorithm introduces a Multi-scale Core Feature Fusionizer (MCF) to enhance the model's image feature extraction capability, and a Vision Transformer (VIT) Network to optimize global information processing. Additionally, the utilization of contextual information is enhanced during the decoding process through the Multilevel Attention Fusion (MAF) module. Experimental results demonstrate that the MIV-Net model exhibits superior performance in the segmentation of the bladder wall and bladder tumors. Key evaluation metrics such as IoU, Precision, and Dice coefficients significantly surpass current mainstream methods, achieving 83.16%, 91.66%, and 90.81% for bladder wall segmentation, and 82.49%, 90.50%, and 90.40% for bladder tumor segmentation, respectively. This model not only shows technical innovation but also holds promise for significantly improving the accuracy of bladder cancer diagnosis in practical applications.