Pupil localization for ophthalmic diagnosis using anchor ellipse regression

Recent developments of deep neural networks, such as Mask R-CNN, have shown significant advances in simultaneous object detection and segmentation. We thus apply deep learning to pupil localization for ophthalmic diagnosis and propose a novel anchor ellipse regression approach based on region proposal network and Mask R-CNN for detecting pupils, estimating pupil shape parameters, and segmenting pupil regions at the same time in infrared images. This new extension of anchor ellipse regression for Mask R-CNN is demonstrated to be effective in size and rotation estimations of elliptical objects, as well as in object detections and segmentations, by experiments. Temporal pupil size estimations by using the proposed approach for normal and abnormal subjects give meaningful indices of pupil size changes for ophthalmic diagnosis.