Abstract
Purpose :
Measurement of macular pigment optical density (MPOD) by autofluorescence technique is underestimated by cataract due to its interference to excitation and emission lights. We reported a correction method using regression equation with age, grade of nuclear cataract, and imaging quality index as independent variables, but some errors remained (Obana A, IOVS 2018). In this study, we applied artificial intelligence (AI) to increase accuracy of correction.
Methods :
MPOD was measured with Spectralis (Heidelberg Engineering, GmbH) before and after cataract surgery for 112 patients under Ethics Committee approval. The correction factor (CF) to estimate a true MPOD value (= postoperative value) from the preoperative value based on the preoperative image was obtained. Three types of images, autofluorescence images by blue and green light, and subtraction images of these two, were input to VGG16, a type of convolutional neural network (CNN). In order to compensate for the limited amount of data, the data was augmented by random cropping and left / right inversion. The accuracy of estimation was improved by fine tuning using a model pre-trained in the ImageNet database. This method was evaluated by 37-fold cross-validation. The CNN parameters were optimized by back propagation + Adam using the mean square error as a loss function.
Results :
The error between corrected value (= preoperative value times CF) and true value was calculated for local MPOD values at 0.25°, 0.5°, 1°, and 2° eccentricities and total volume of macular pigment (MPOV) in the area within 9° eccentricities. The mean value of the errors of MPOD at four eccentricities was 32% without any correction, 14% with correction using the previous regression equation, and 11% with the present correction using AI. The error decreased with increasing eccentricity. The error of MPOV was 22% without any correction, 15% with correction using the regression equation, and 8% with correction using AI.
Conclusions :
The error between corrected MPOD value and true value was reduced by using CFs that were extracted from estimation of the fluorescence image by AI. The AI correction method was considered useful for the measurement of MPOD in aging eyes with cataracts.
This is a 2020 ARVO Annual Meeting abstract.