Analysis of Macula on Color Fundus Images Using Heightmap Reconstruction through Deep Learning

Peyman Tahghighi, MS1 *, Reza A. Zoroofi, PhD1 , Sare Safi, PhD2 , Alireza Ramezani, MD3 , Hamid Ahmadieh, MD4 , Seyed Farzad Mohammadi, MD5

  1. School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran
  2. Ophthalmic Epidemiology Research Center, Research Institute for Ophthalmology and Vision Science Shahid Beheshti University of Medical Sciences, Tehran, Iran
  3. Ophthalmic Epidemiology Research Center, Research Institute for Ophthalmology and Vision Science, Shahid Beheshti University of Medical Sciences, Tehran, Iran; Negah Aref Ophthalmic Research Center, Shahid Beheshti University of Medical Sciences, Tehran, Iran
  4. Ophthalmic Research Center, Research Institute for Ophthalmology and Vision Science, Shahid Beheshti University of Medical Sciences, Tehran, Iran
  5. Translational Ophthalmology Research Center, Tehran University of Medical Sciences, Tehran, Iran

Abstract: Color fundus photography is a routine imaging modality. However, it lacks height information due to the two-dimensional nature of imaging and transparency of neuroretina by visible light. Hence, we aim to model the height information of the macula area using the state-of-the-art artificial intelligence methods based on Optical Coherence Tomography (OCT) outputs.

Methods: Totally, 3407 three-dimensional macular OCT outputs and two-dimensional fundus images were extracted from the DRIOCT Triton (Topcon, Tokyo, Japan) at Negah Eye Hospital, Tehran, Iran. The macula area on the fundus images and the heightmap information from the same region on OCT output were targeted. Then, the deep learning models were developed on 80% of the data to automatically generate a heightmap information estimation from the color fundus image. Finally, two retina specialists rated 50 generated heightmaps that were randomly selected from the untouched 20% dataset. The scores were defined as zero to three. Zero means that the generated images did not add any information in comparison to fundus images which were classified as negative. Scores of one to three were classified as positive which means that the generated images can add more information in comparison to color fundus images.

Results: All images were classified as positive by the first retina specialist, while 46 (92%) images were classified in the positive class by the second retina specialist. Additionally, both retina specialists rated most of them as score two (42% and 32%). The mean selected score by the first and second retina specialists were 1.94±0.76 and 2.0±0.81, respectively.

Conclusion: This study showed that heightmaps generated from two dimensional macula images through machine learning may provide additional information.





اخبــار



برگزار کنندگان کنگره


حامیان کنگره