MeshBrush: Painting the Anatomical Mesh with Neural Stylization for Endoscopy

TL;DR

It is demonstrated that mesh stylization is a promising approach for creating realistic simulations for downstream tasks such as training networks and preoperative planning, and its components are transferable to general endoscopic and laparoscopic procedures.

Abstract

Style transfer is a promising approach to close the sim-to-real gap in medical endoscopy. Rendering synthetic endoscopic videos by traversing pre-operative scans (such as MRI or CT) can generate structurally accurate simulations as well as ground truth camera poses and depth maps. Although image-to-image (I2I) translation models such as CycleGAN can imitate realistic endoscopic images from these simulations, they are unsuitable for video-to-video synthesis due to the lack of temporal consistency, resulting in artifacts between frames. We propose MeshBrush, a neural mesh stylization method to synthesize temporally consistent videos with differentiable rendering. MeshBrush uses the underlying geometry of patient imaging data while leveraging existing I2I methods. With learned per-vertex textures, the stylized mesh guarantees consistency while producing high-fidelity outputs. We demonstrate that mesh stylization is a promising approach for creating realistic simulations for downstream tasks such as training networks and preoperative planning. Although our method is tested and designed for ureteroscopy, its components are transferable to general endoscopic and laparoscopic procedures. The code will be made public on GitHub.

Figures (4)

• Figure 1: Sample camera trajectory inside the mesh. From top to bottom: original rendered, I2I style transfers pfeiffer2019generating, and MeshBrush. Note inconsistency from frame to frame in the I2I method while MeshBrush maintains consistent features.
• Figure 2: An overview of our method. A skeletonization method skeletor generates camera views inside the anatomy. At each iteration, the model re-renders from each camera pose via differential rendering and its style-transferred image supervises the vertex textures to be updated via the view-dependent heatmap loss.
• Figure 3: Left: visualization of MeshBrush's paintings onto the mesh. Right: original renderings, heatmap (see Loss function description), ground-truth style transferred renderings, and renderings from CT using PyTorch3D (from top-left to bottom-right) where the camera is approximately placed at the green star. Corners of the images are masked identically to original endoscopy images.
• Figure 4: Sparse COLMAP reconstruction using renders from the stylized mesh. Red squares show the estimated camera trajectory