Investigation of image processing and computer-assisted diagnosis system for automatic video vision development assessment.

摘要

Amblyopia (lazy eye) affects about 2-5% of general population. It is a leading cause of single-eye blindness in adults. Video-based photoscreenirig, pioneered by Dr. Gerhard W. Cibis, is a technique effective for screening very young children for amblyopia-causing factors, which is essential to the prevention. This dissertation presents the first ever work to automate this screening technique. The main challenge in algorithm development in this study arises from many dynamic factors that can not be controlled, including the patient's direction of sight, focusing, and physical motion. These factors make the observed imagery highly variable even for the same patient. This dissertation provides detailed descriptions of the various algorithms that combine to reach screening decisions from the video data. The algorithms are divided into two main groups, image processing and computer-assisted diagnosis.; Combining new image features and sequence processing, we have developed robust algorithms for locating irises, pupils, and Hirschberg points, which are useful for diagnosing misalignment between eyes. The new technical contributions include the use of sclera in iris boundary location and the use of possibilistic shell clustering with additional constraints to simultaneously locating a pair of pupils.; We also investigated computer-assisted diagnosis of various amblyogenic vision disorders, including high refractive errors (hyperopia and myopia), refractive error difference between the two eyes (anisometropia), astigmatism, and misalignment between the two eyes, (strabismus). Diagnoses of these factors are then combined to yield the final screening decision.; We evaluate the performance of our algorithm by comparing our screening decisions to those given by Dr. Cibis (the ground truth) over the same data set, which include video data of 182 different children. Our results indicate that the screening decisions made by the algorithms are 89% correct when compared with those made by Dr. Cibis. Further investigation points to several areas for future improvement, including better analysis of cases with misalignment and myopia.