Abstract: Thoracic and abdominal medical image segmentation plays an irreplaceable role in the screening and diagnosis of early lesions in clinical medicine. In view of the fact that the thoracic and abdominal medical images usually have factors such as artifact interference and blurred boundaries, and existing models are prone to segmentation errors, a semantic segmentation network MQUNet is proposed by constructing a dual-branch feature interaction mechanism and combining cross-sample feature fusion. Firstly, the dual-branch architecture based on CNN-Transformer is adopted to extract the local and global features of the current image, capturing the local details and spatial correlations between organs as comprehensively as possible. It also integrates the two features by adding a negative attention branch based on an improved activation function, while taking into account both similarities and differences between features, avoiding the loss of semantic information. In addition, a memory queue is established based on the category labels of the training samples to store anatomical prior features. Through cross sample feature retrieval and weighted fusion, stable anatomical prior features from the training samples are dynamically added to the current segmentation process, effectively correcting feature shifts caused by artifacts and blurring. The experimental results on the publicly available cardiac cancer (ACDC) and abdominal multi-organ (Synapse) datasets show that the Dice coefficient index reaches 92.53% and 85.17% respectively, which is superior to the segmentation methods compared in the paper. MQUNet demonstrates significant advantages in the task of thoracic and abdominal medical image segmentation.