Purchase this article with an account.
Ke-Wei Chen, Perera Chandrashan, Hsu-kuang Chiu, Andrea Lora Kossler, Myung David, Chiou-Shann Fuh, Cherng-Ru Hsu, Ming-Chun Tseng, Shu-Lang Liao, Ju Yi Hung; Automatic Blepharoptosis Measurement by Iris Edge Detection with a Deep Learning Model. Invest. Ophthalmol. Vis. Sci. 2022;63(7):172 – F0019.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Our study proposes a method for margin reflex distances-1 (MRD1) measurement through facial photographs.
A deep learning U-net based model was trained with 100 external ocular photography to automate segmentation of the iris image. Images were acquired from pre-operative photos of patients prior to ptosis surgery. For each patient’s photo, the segmentation result was used to extract the iris edge. A circle of best fit was generated based on the visible iris edges. The margin reflex distance 1 (MRD1) was calculated as the distance between the center of this circle to the margin of the upper eyelid. A fixed size reference marker on the patient’s forehead was used to convert pixel measurements to millimeters. The proposed deep-learning-based MRD measurement algorithm was then evaluated with 500 single eye photos by comparing the measurement results from the model with measurements taken by a physician using the same photos. The physician was provided with custom software to enable measurement (available at https://gosienna.github.io/MRD_measurement/)
A total of 500 photos, with MRD’s ranging from -3mm to 6.15mm were used to evaluate the performance of the model. Using the physician measurements as the ground truth, we found that 95 % of the model measurements were within 1mm of the ground truth. The Pearson’s correlation between model measurements and the ground truth was 0.95. The P-value for non-correlation testing was less than 0.001.The Bland-Altman plot shows good agreement between automatic and manual MRD measurements (95% limits of agreement). Prediction errors occur in images without distinct upper eyelid margins. For example, long eyelashes, severe ptosis or large pterygium can result in difficulty for the model to recognize the iris edge.
Using a deep learning segmentation model, the MRD can be measured from clinical photos with a high degree of accuracy.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
A sample result from the study. (Left upper figure) Iris segmentation AI model's output result. (Right upper figure) Edge with circle fitting. (Left lower figure) Manual measurement is shown with the green circle. Automatic result is shown with the red dot and the vertical line. (Right lower figure) Marker used to translate pixel measurements to millimeters
Bland-Altman plot for MRD measurements from automatic and manual. Dashed light blue line indicate values of two standard deviations.
This PDF is available to Subscribers Only