Rectification of Cornea Induced Distortions in Microscopic Images for Assisted Ophthalmic Surgery

Image distortions induced by the high refractive power of the eye's optical components challenge the accuracy of geometric information derived from intra-operative sensor data in ophthalmic surgery. Correcting these distortions is vital for advancing surgical assistance systems that rely on geometric scene comprehension. In this work, we focus on cornea induced distortions (CIDs) in surgical microscope images of the anterior eye.

We employ a convolutional neural network (CNN) with stereo fusion layers to predict distortion distribution maps (DDMs) to correct CIDs in stereo images. To enable supervised learning, we introduce CIDCAT, a synthetic surgical microscope dataset generated through a rendering pipeline using a digital eye model. We address the domain gap between the synthetic training data and the unlabeled target domain of real surgical images by employing an auxiliary task of semantic segmentation to regularizes the feature encoder.

Our rectification model reduces the cornea induced pupil radius error from 8.56% to 0.72% and improves the structural similarity by over 9% for synthetic CIDCAT images. Our semantic segmentation driven domain regularization technique enables the translation to real surgical images.

The CIDCAT dataset enables the investigation of CIDs and the implementation of a CID rectification model. Our proposed CID rectification model demonstrate successful minimization of CIDs while preserving image integrity.

The work represents a significant advancement in research on computer-assistance and robotic solutions for ophthalmic surgery that rely on a distortion-free, three-dimensional understanding of the patient's eye.