Implicit domain adaptation with conditional generative adversarial networks for depth prediction in endoscopy

Colorectal cancer is the third most common cancer worldwide, and early therapeutic treatment of precancerous tissue during colonoscopy is crucial for better prognosis and can be curative. Navigation within the colon and comprehensive inspection of the endoluminal tissue are key to successful colonoscopy but can vary with the skill and experience of the endoscopist. Computer-assisted interventions in colonoscopy can provide better support tools for mapping the colon to ensure complete examination and for automatically detecting abnormal tissue regions.

We train the conditional generative adversarial network pix2pix, to transform monocular endoscopic images to depth, which can be a building block in a navigational pipeline or be used to measure the size of polyps during colonoscopy. To overcome the lack of labelled training data in endoscopy, we propose to use simulation environments and to additionally train the generator and discriminator of the model on unlabelled real video frames in order to adapt to real colonoscopy environments.

We report promising results on synthetic, phantom and real datasets and show that generative models outperform discriminative models when predicting depth from colonoscopy images, in terms of both accuracy and robustness towards changes in domains.

Training the discriminator and generator of the model on real images, we show that our model performs implicit domain adaptation, which is a key step towards bridging the gap between synthetic and real data. Importantly, we demonstrate the feasibility of training a single model to predict depth from both synthetic and real images without the need for explicit, unsupervised transformer networks mapping between the domains of synthetic and real data.