Abstract:
Due to the shortage of experienced endoscopists, Computer-Aided Diagnosis (CAD) systems for colonoscopy have recently attracted many research interests. There exist sever...Show MoreMetadata
Abstract:
Due to the shortage of experienced endoscopists, Computer-Aided Diagnosis (CAD) systems for colonoscopy have recently attracted many research interests. There exist several public polyp segmentation datasets, giving way to the adoptions of domain adaptation methods to address the shift in distributions. Current domain adaptation frameworks often comprise (i) a domain discriminator trained with an adversarial loss and (ii) an image-translation network. Due to the complexity of image-translation networks, such methods are generally hard to train to achieve satisfactory results. Hence, we propose a domain adaptation method that leverages Fourier transform as a simple alternative to the image-translation network. We introduce an adversarial contrastive training strategy to jointly learn an embedding space that considers both style and content of the sample. Our method demonstrated consistent gains over state-of-the-arts on polyp semantic segmentation task for four public datasets. The code is available at: https://github.com/tadeephuy/CoFo
Date of Conference: 28-31 March 2022
Date Added to IEEE Xplore: 26 April 2022
ISBN Information: