Quantized feature alignment for unsupervised domain adaptation in abdominal and prostate segmentation

Unsupervised Domain Adaptation (UDA) plays a crucial role in medical image segmentation, especially when annotated target domain data is unavailable. Traditional continuous feature alignment methods face challenges due to mini-batch limitations and often fail to capture the full domain distribution. To address these issues, we propose QFASeg-Net, a novel UDA strategy that utilizes quantized feature alignment instead of conventional continuous approaches. Our grouped quantization technique transforms continuous features into discrete representations, constructing a codebook that progressively learns to capture the entire feature distribution of the domain, rather than adapting to the limited distributions within mini-batches. Moreover, we incorporate multi-scale high-order statistical alignment to refine the alignment of quantized features across different domain spaces, enhancing cross-domain feature consistency. Experimental results on abdominal and prostate segmentation tasks demonstrate that QFASeg-Net outperforms existing methods, validating the effectiveness of quantized feature alignment for cross-modality medical image segmentation.