EndoUDA: A Modality Independent Segmentation Approach for Endoscopy Imaging

Gastrointestinal (GI) cancer precursors require frequent monitoring for risk stratification of patients. Automated segmentation methods can help to assess areas more accurately, and assist in therapeutic procedures or even removal. In clinical practice, addition the conventional white-light imaging (WLI), complimentary modalities such as narrow-band (NBI) fluorescence are used. While, today most approaches supervised only concentrated on a single modality dataset, this work exploits use target-independent unsupervised domain adaptation (UDA) technique that is capable generalize an unseen target modality. context, we propose novel UDA-based method couples variational autoencoder U-Net with common EfficientNet-B4 backbone, uses joint loss latent-space optimization samples. We show our model NBI (target) when trained using WLI (source) Our experiments both upper lower GI endoscopy data effectiveness approach compared naive state-of-the-art UDA methods.