FoldIt: Haustral Folds Detection and Segmentation in Colonoscopy Videos

Abstract:

Haustral folds are colon wall protrusions implicated for highpolyp miss rate during optical colonoscopy procedures. If segmented ac-curately, haustral folds can allow for better estimation of missed surfaceand can also serve as valuable landmarks for registering pre-treatmentvirtual (CT) and optical colonoscopies, to guide navigation towards theanomalies found in pre-treatment scans. We present a novel generativeadversarial network, FoldIt, for feature-consistent image translation ofoptical colonoscopy videos to virtual colonoscopy renderings with haus-tral fold overlays. A new transitive loss is introduced in order to leverageground truth information between haustral fold annotations and virtualcolonoscopy renderings. We demonstrate the effectiveness of our modelon real challenging optical colonoscopy videos as well as on texturedvirtual colonoscopy videos with clinician-verified haustral fold annota-tions. All code and scripts to reproduce the experiments of this paper will be made available via our Computational Endoscopy Platform at https://github.com/nadeemlab/CEP

Model Weights for FoldIt

This was trained where domain A is optical colonoscopy, domain B is haustral fold annotation and domain C is virtual colonoscopy ground truth corresponding to the haustral fold annotations.

Testing Videos

The frames that were used for testing are included as well. Please cite the appropriate papers when using them.

Incetan, K., Celik, I.O., Obeid, A., Gokceler, G.I., Ozyoruk, K.B., Almalioglu,Y., Chen, R.J., Mahmood, F., Gilbert, H., Durr, N.J., et al.: VR-Caps: A virtualenvironment for capsule endoscopy. Medical Image Analysis. p. 101990 (2021)

Ma, R., Wang, R., Pizer, S., Rosenman, J., McGill, S.K., Frahm, J.M.: Real-time3D reconstruction of colonoscopic surfaces for determining missing regions. Inter-national Conference on Medical Image Computing and Computer-Assisted Inter-vention pp. 573–582 (2019)