Deep Learning Dominated Method for Assessment of Toric Intraocular Lens Rotation

Abstract

In this study a deep learning dominated method is proposed for automatic assessment of the postoperative rotation of toric intraocular lens (IOL). This is designed to diagnose a postoperative astigmatism like disorder, which can arise several weeks or month after a cataract surgery implanting a toric IOL for treating both cataract and astigmatism together. Additionally the method is also intended to be used for researching the cause of postoperative IOL rotation. The proposed method works on pairs of retroillumination slit lamp microscope images. Using an automated assessment of the postoperative rotation of toric IOL exempts the medical personnel from a huge workload when using the state of the art semi-automated or manual methods. The method also utilizes numerous visual computing algorithms for preprocessing and extracting some feature components enabling the application of reduced complexity convolutional neural network (CNN) model. The core functionality is realized by standard CNN model, the ResNet18. The proposed method is evaluated on a labelled medical data set and the test results are justified on an enhanced data set including more image pairs with pathological cases. A detailed investigation is provided on the absolute error of the assessed IOL rotation angle and its components caused by the proposed method. The results show that the proposed method achieves 85 % classification accuracy for the task of separating the non-pathological and pathological cases by setting the threshold of the predicted IOL rotation angle properly.