To address the insufficient fusion of local texture details and global semantic information in medical image segmentation, which often leads to blurred boundaries and incomplete structures, a named dual fusion branch UNet（DFB-UNet）is proposed with parallel convolution and self-attention. The encoder employs a convolutional neural network（CNN）branch and a Transformer branch in parallel to learn local and global features. Before cross-branch fusion, a deformable convolution network （DCN）is introduced to enhance the modeling of irregular object boundaries. In addition, a dual-fusion module composed of an attention fusion module（AFM）and a convolution fusion module（CFM）is constructed; CFM incorporates depthwise separable convolutions to achieve effective feature interaction and complementary aggregation with lower computational cost. The experiment results are as follows: mDice/mIoU reach 0.89/0.83 on the dataset Kvasir-SEG; 0.810/0.723 on CVC-ClinicDB. In cross-dataset testing, training on Kvasir-SEG and testing on CVC-ClinicDB yield Dice/IoU of 0.751/0.668. On ISIC2017, IoU/Dice/Acc are 0.816/0.889/0.965, and on LUNA16, Dice/IoU are 0.966 0/0.934 2. The results indicate that the proposed dual-fusion parallel encoder improves segmentation accuracy and generalization while maintaining computational efficiency. The ablation studies further verify the contribution of DCN and the dual-fusion module.