Abstract: Medical image segmentation is important in disease diagnosis, treatment planning and outcome assessment. However, segmentation accuracy is still challenging due to the complexity and shape differences of organ edges.In order to improve the segmentation accuracy, this paper proposes an improved TransUNet segmentation model. The model is a coding and decoding structure, the encoder consists of CNN and Transformer, CNN is responsible for extracting local features, and Transformer learns global features; the decoder gradually recovers the image features by cascading the up-sampling structure, which includes the feature fusion module, the up-sampling module, and the jump connection module. In the encoding stage, the model introduces efficient channel attention to improve the attention to image channel features; in the decoding stage, dynamic convolution is used to adapt to different feature inputs in the up-sampling process, and at the same time, dense up-sampling convolution is used to better preserve image detail information. In this paper, we validate the model on the Synapse multi-organ segmentation and automatic cardiac diagnosis challenge datasets, and the experimental results show that the performance outperforms the existing models.