Author:

Keywords:

PSI_VISICS, PSI_4373

Abstract:

3D reconstruction has become an important tool in many camera-assisted medical interventions. As such, one of the goals of the EU funded project EurEyeCase is to use reconstructed 3D maps of the retina to help the surgeon during retinal interventions. To that end, a stereo microscope is used to view a small fraction of the retina at each instant. In this paper, we investigate how the reconstructed point clouds from each stereo image pair can be fused to generate a more consistent global retinal model. A global 3D model of the retina would help the surgeon to accurately localize the tools with respect to the retinal surface. We build upon previous work of stereo microscopic calibration and retinal 3D reconstruction (Havlena et al. 2016) while implementing them in a framework compatible with real-time applications. We use a coarse to fine surface registration algorithm based on 2D image features, transformation estimation using singular value decomposition (SVD) and an Iterative Closest Point (ICP) method to align each new point cloud surface to the global retinal 3D model. Fusion is achieved by filtering aligned incoming depth maps using a voxel grid. In contrast to Havlena et al., the approach can be implemented online with close to real-time performance. We also remark several encountered challenges in achieving such a task and the solutions we used to tackle them. Finally we present our results on a sequence of recorded images of a surgery performed on an open sky porcine eye.