An Enhanced Synthetic Cystoscopic Environment for Use in Monocular Depth Estimation

As technology advances and sensing devices improve, it is becoming more and more pertinent to ensure accurate positioning of these devices, especially within the human body. This task remains particularly difficult during manual, minimally invasive surgeries such as cystoscopies where only a monocular, endoscopic camera image is available and driven by hand. Tracking relies on optical localization methods, however, existing classical options do not function well in such a dynamic, non-rigid environment. This work builds on recent works using neural networks to learn a supervised depth estimation from synthetically generated images and, in a second training step, use adversarial training to then apply the network on real images. The improvements made to a synthetic cystoscopic environment are done in such a way to reduce the domain gap between the synthetic images and the real ones. Training with the proposed enhanced environment shows distinct improvements over previously published work when applied to real test images.