Automatic Quantification of Pupil Dilation under Stress CADDLab @ UNC Julien Jomier, Erwann Rault, and Stephen R. Aylward Computer Aided-Diagnosis and Display Lab - University of North Carolina at Chapel Hill We seek to automate the measurement of pupil and iris areas from color digital photos so as to calculate the ratio of those areas as a measure of the amount that an individual has dark-adapted. We are particularly interested in testing the hypothesis that dark adaptation is slowed proportional to the amount of stress that an individual has experienced. Coarse Pupil and Iris Segmentation 1) Eye Localization - Sub sampling is performed to improve computation speed. - Statistics filter is applied to extract pixels that have the red component higher than the blue component (Red – Blue < ) defining two regions of interest around each eye. - Statistics of the iris are estimated on the left and right part of the pupil. - Only the radius is regions used to optimized. The center of the evaluate statistics iris is assumed to be the of the iris center of the pupil. - Calculate the optimal linear discriminant between pupil and iris color classes to compute a pupil-likelihood image (TopRight). We seek the boundary between the iris and the pupil that is emphasized by that likelihood image. - Threshold the likelihood image to form a binary image such that every pixel with a likelihood greater or equal to zero is set to one. We then use morphological operations to reduce noise (Bottom-Left). - Active contours segmentation Resulting precise segmentation of the [3] (Bottom-Right) pupil via active contours Eye localization using color statistics of the pupil (middle) from the original image (left) resulting to a definition of two regions of interest (right) 2) Coarse Pupil Segmentation The pupil is segmented in three steps. Blue Re d 0 1) The pixels that satisfy the equation are set to 1 and 0 otherwise to produce a binary image. 2) Hough Transform [1] is used to approximate the center and radius of the best fitting circle in the binary image. 3) We apply a model-to-image registration [Aylward 2001] using the 1+1 evolutionary optimizer [2]. We define the metric f of the fit of the circle with the binary image. Ar r , X Ar , X f(r,X) 2 r r Resulting segmentation of the iris Precise Pupil Segmentation Results - Training : 5 left eyes from different subjects. - Testing: 20 eyes from 10 different patients. - Comparison of the automated algorithm with handsegmentation of 5 raters shows equal accuracy. - Automatic segmentation takes less than 1 minute per image (2 times faster than manual segmentation). - Fully automatic Average Segmentation Error Raters Computer 35010 34780 1.94% 2.41% Max Error 6.76% 5.80% Average and Maximum error in % Introduction 8 7 6 5 4 3 2 1 0 Error Max Error Comparison with hand-segmentation of the pupil by 5 raters on 10 images. References Plot of the metric for a radius r=3 Resulting segmentation of the pupil 3) Coarse Iris Segmentation The segmentation of the iris is done with the same technique as the pupil. Computer-Aided Diagnosis and Display Lab Department of Radiology, UNC @ Chapel Hill [1] D. H. Ballard, “Generalized hough transform to detect arbitrary patterns,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 13, no. 2, pp. 111–122, 1981 [2] M Styner and G. Gerig, “Evaluation of 2d/3d bias correction with 1+1es-optimization,” Technical Report BIWI-TR-179 [3] Ross T. Whitaker, “Algorithms for implicit deformable models,” in Fifth International Conference on Computer Vision. IEEE, 1995, IEEE Computer Society Press [4] Insight Software Consortium, “The insight toolkit: Segmentation and Registration toolkit,” http://www.itk.org April 2004