The rising occurrence of intracranial tumors highlights the pressing demand for prompt and precise medical evaluation. Techniques based on deep neural networks for medical image partitioning play a pivotal role in streamlining diagnostic workflows and enhancing assessment efficiency. Our study introduces an enhanced U-Net++ architecture, which utilizes short skip connections in place of long connections to better capture edge information. Additionally, the convolutional blocks are replaced by residual structures to mitigate model degradation, while batch normalization is applied to the residual blocks to accelerate convergence. The enhanced U-Net++ model achieved a sensitivity of 0.913, an accuracy of 0.995, and a Dice score of 0.927 on the Tumor Imaging Archive dataset. The enhanced U-Net++ network demonstrates improved feature fusion across various depths and enables real-time weight updates, ensuring the network approaches optimal performance without degradation during training, while also improving segmentation accuracy.