Image-Based Localization and Its Application in Eye Laser Surgery Automation

• Main Authors: WU, JHENG-HAN, 吳政翰
• Other Authors: LIN, HORNG-HORNG
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/58027517281507060044

碩士 === 南臺科技大學 === 資訊工程系 === 105 === Eye detection applications have been brought into our daily lives, such as driver fatigue monitoring, consumer attention tracking, and amyotrophic lateral sclerosis patient assistance. In this thesis, we explore the medical applications of eye detection to the automations of eye surgery using laser. In particular, we have developed image-based detection modules for eye localization and for automatic laser positioning to simplify the conventional operation procedures manipulated by surgeons and to improve the surgical precision of refractive eye surgery. For the application of surgery automation with excimer laser, we integrate two auxiliary functions of target eye localization and of laser head positioning to corneal surface into real surgical equipment. Specifically, we propose a new eye detection method using template matching based on image gradient histograms as features for automatic localizations of a target eye. Upon the eye localization, image-guided alignment of positioning markers, i.e., a read line and a green dot of visible light, is then performed to measure the distance between the corneal surface and the excimer laser head to automatically set laser head position. Moreover, the same technique of image-guided alignment can also be applied to femto laser surgery to monitor the surface changes of cornea in laser docking process and to measure the depth from the corneal surface to the anterior chamber for laser cutting. In the experiments, the proposed eye detection method using template matching is compared with the SVMs and the ResNets of deep convolutional neural network. Using the real human eye images captured from the surgery instrument as test samples, the proposed eye detection method achieves 97% detection accuracy, which outperforms the SVM classifier, and gives comparable performance to the computationally costly ResNet detector. We also validate the effectiveness of the proposed auxiliary surgical automation by conducting system-level localization tests in real surgical equipment with long runs. Regarding the experiment of laser positioning for femto laser, the proposed method for monitoring the corneal surface changes during laser docking and for measuring the depth from the corneal surface to the anterior chamber is implemented in femto laser equipment as well. Preliminary experimental results using pig eyes show the feasibility of the proposed method for automatic surgical parameter measurements of real eyes.