Abstract
Purpose :
Ptosis is an eye condition where the upper eyelid droops. The current diagnosis for ptosis involves cumbersome manual measurements that are time-consuming and prone to human error. This work presents a fully automated and interpretable dual model system for rapid detection of ptosis that can help save the clinics and patients valuable resources.
Methods :
The data for this study were queried from the Illinois Ophthalmic Database Atlas (I-ODA) that was developed by the Department of Ophthalmology at the University of Illinois Chicago. The dataset consisted of 820 facial images collected in a clinical setting and was augmented by 43 Flickr-Faces-HQ images. A subset of 100 images was hand-selected by an expert oculoplastic surgeon to create a held-out test set. All values reported in the results were derived from this test set.
The eye regions were extracted from the facial images and were fed into two pipelines, Deep Learning (DL) and Feature & Rule (FR). The DL used a 5-layer uni-eye convolutional neural network to learn data characteristics to predict ptosis likelihood and the FR extracted features to determine the marginal reflex distance (MRD1) and the iris visibility ratio (IR). AutoPtosis was a combination of both these pipelines, where a predictive model was trained using MRD1, IR, and DL’s likelihood to predict ptosis.
Results :
The DL performed well giving a 0.91 accuracy, 0.89 precision, 0.93 recall, and 0.95 AUC while FR gave 0.63 accuracy, 0.60 precision, 0.80 recall, and 0.77 AUC. When both pipelines were combined there was an improvement in accuracy to 0.95, precision to 0.96, recall to 0.96, and AUC to 0.98. We noticed an improvement especially in FR results when only clinical data was used. We also created class activation maps and direct feature visualization to interpret the most contributing eye regions for DL and iris, eye contours used for calculating MRD1 and IR respectively, that helps with error assessment.
Conclusions :
AutoPtosis combined the preferable aspects of both DL and FR to create a rapid and automatic system for detection of ptosis all while achieving results almost as good as an expert physician. The models performed optimally under clinical settings and can be deployed there as a helping tool to generate results which can then be verified by physicians, saving the healthcare system and patients valuable resources.
This is a 2021 ARVO Annual Meeting abstract.