July 2018
Volume 59, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2018
Macular hole detection and staging on optical coherence tomography images using convolutional neural network
Author Affiliations & Notes
  • Akira Ojima
    Ophthalmology, Fukushima Medical University, Fukushima, Japan
  • Tetsuju Sekiryu
    Ophthalmology, Fukushima Medical University, Fukushima, Japan
  • Ryutaro Tomita
    Ophthalmology, Fukushima Medical University, Fukushima, Japan
  • Yukinori Sugano
    Ophthalmology, Fukushima Medical University, Fukushima, Japan
  • Yutaka Kato
    Ophthalmology, Fukushima Medical University, Fukushima, Japan
  • Footnotes
    Commercial Relationships   Akira Ojima, None; Tetsuju Sekiryu, Novartis (F); Ryutaro Tomita, None; Yukinori Sugano, None; Yutaka Kato, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2018, Vol.59, 3164. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Akira Ojima, Tetsuju Sekiryu, Ryutaro Tomita, Yukinori Sugano, Yutaka Kato; Macular hole detection and staging on optical coherence tomography images using convolutional neural network. Invest. Ophthalmol. Vis. Sci. 2018;59(9):3164.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : To evaluate the accuracy of automated detection and classification of idiopathic macular hole (MH) on optical coherence tomography (OCT) images using convolutional neural network.

Methods : One thousand four hundred and six foveal-centered OCT images collected from 224 patients with MH were used to train and test a neural network; 98 males and 126 females, mean age 71 years old, 1025 images of MH (stage 1, 2, 3, and 4) collected from affected eyes and 381 normal images collected from healthy fellow eyes of the patients. All OCT images were taken at grayscale with a resolution of 768 x 496 pixels using Spectralis OCT (Heidelberg, Germany) and labeled as MH stage 1, 2, 3, 4, and normal. All images were cropped into square in the center and resized into 256 x 256 pixels. We randomly divided the data into two parts with keeping the proportion of classification; 1126 images (80%) for training and 280 images (20%) for testing. Training data was learned 50 epochs using GoogLeNet implemented by Chainer. All 280 images of testing data were verified using the trained model.

Results : In 280 images for testing, 80 images were normal and 200 images were MH. Number of images of MH stage 1, 2, 3, and 4 were 39, 42, 68, and 51, respectively. The trained model detected MH in 193 of 200 MH images and judged normal in 78 of 80 normal images. The sensitivity and specificity in detecting MH were 96.5% and 97.5%, respectively. Correct diagnosis rate of MH stage 1, 2, 3, and 4 were 76.9%, 88.1%, 95.6%, and 82.4%, respectively. Trained model tended to misdiagnose stage 1 as normal, and stage4 as stage 3. Overall correct diagnosis rate of all categories was 90.0%.

Conclusions : Convolutional neural network achieved accurate detection and classification of MH on OCT images. Trained model was applicable to clinical use in detecting MH.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×