The Pitfall of Pixel-Level Domain Adaptation for Cross-Device Polyp Segmentation

Cross-device domain adaptation is essential for deploying polyp segmentation models across different endoscope manufacturers. Pixel-level domain adaptation methods such as Contrastive Unpaired Translation (CUT) have been widely adopted for style transfer; however, their effectiveness in gastrointestinal polyp segmentation remains controversial. This paper presents a systematic study that reveals a critical, overlooked factor determining CUT's performance: the input normalization strategy of the segmentation model. Evaluated on four mainstream models (PolypPVT, MSNet, PraNet, and TransResUNet) across two target domains, our results show that CUT consistently harms ImageNet-pretrained models, with average external mIoU dropping by up to 14.7%. In contrast, models employing simple /255.0 normalization benefit significantly from CUT, gaining 17.9% in external performance. We identify the root cause: ImageNet normalization inherently 'neutralizes' pixel-level style adjustments, rendering them ineffective. Furthermore, we highlight a critical clinical risk where CUT degrades small polyp detection recall by up to 19.3% in ImageNet-normalized models. These findings challenge the prevalent assumption that pixel-level domain adaptation universally benefits medical image segmentation, and strongly advocate for a normalization compatibility assessment prior to applying such techniques. To our knowledge, this is the first systematic study quantifying the normalization-style transfer incompatibility specifically in the context of cross-device endoscopic polyp deployment, bridging a critical gap between general computer vision theory and clinical AI practice.