PhD Defense by Mohamed Alsheakhali

Speaker: Mohamed Alsheakhali Date: Tuesday, June 27th, 2017 Time: 02:00 PM Location: FMI-Building, Room 01.06.020

Machine Learning for Medical Instrument Detection and Pose Estimation in Retinal Microsurgery

Abstract:

Instrument detection and pose estimation has attracted great interest in retinal microsurgery. Automatic detection of the instrument parts and estimating the instrument pose promote many applications to guide the surgeon in the operation room. One important application is the automatic positioning of Optical Coherence Tomography (OCT) scans to estimate the distance between the detected instrument tip and the retina to minimize the damage during the surgery. There are many other applications that employ the pose of the instrument such as activity recognition and surgical workflow analysis. This work addresses the problem of detecting the instrument parts (tips and/or the joint point). The first contribution employs instrument’s color and geometric information, while the second exploits deep learning to localize the instrument tip and the shaft’s orientation. In our third contribution, we formulate the problem as a regression task, where a new Hough Forest pipeline is proposed to predict the locations of the instrument parts within the 2D image space. Finally, the dependencies among the instrument parts are employed in our fourth contribution using a Conditional Random Field (CRF) model to estimate the instrument’s orientation in addition to the parts localization. Therefore, this work introduces different techniques to assist surgeons in minimally invasive procedures. The techniques are not limited to retinal microsurgery but also can be applied to laparoscopic surgery.