Cross-Dataset UNet: A UNet Method with Cross-Dataset Aware Module for Multi-center Medical Image Segmentation

One of the key challenges in the field of medical image segmentation is the dataset offset problem caused by multi-center data differences. This paper proposes an innovative Cross-Dataset UNet, which combines Dataset-Aware Blocks (DAB) with an efficient residual connection to address this challenge. DAB consists of center-shared convolutional layers and center-specific batch normalization. We further introduce residual connections in the encoder to form a residual dataset-aware block (RDAB) structure, thereby enhancing gradient flow and feature expression capabilities. Experimental results on the public head and neck tumor segmentation dataset (HECKTOR2021) show that compared with the standard U-Net using direct mixed data training, our method significantly improves the average Dice coefficient on multi-center data from 0.751 to 0.776. This study proposes a framework for segmenting medical images in multi-center PET/CT research scenarios.