July 2018
Volume 59, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2018
The Utilization of Artificial Intelligence for Corneal Nerve Analyses of In Vivo Confocal Microscopy Images for the Diagnosis of Neuropathic Corneal Pain
Author Affiliations & Notes
  • Neslihan Dilruba Koseoglu
    Ophthalmology, Tufts Medical Center, Boston, Massachusetts, United States
  • Andrew Beam
    Biomedical Informatics, Harvard University , Boston , Massachusetts, United States
  • Pedram Hamrah
    Ophthalmology, Tufts Medical Center, Boston, Massachusetts, United States
  • Footnotes
    Commercial Relationships   Neslihan Dilruba Koseoglu, None; Andrew Beam, None; Pedram Hamrah, Allergan (C), Dompe (C), Heidelberg (C), Tissue Tech (C)
  • Footnotes
    Support  Tufts Medical Center Institutional Support
Investigative Ophthalmology & Visual Science July 2018, Vol.59, 3440. doi:https://doi.org/
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Neslihan Dilruba Koseoglu, Andrew Beam, Pedram Hamrah; The Utilization of Artificial Intelligence for Corneal Nerve Analyses of In Vivo Confocal Microscopy Images for the Diagnosis of Neuropathic Corneal Pain. Invest. Ophthalmol. Vis. Sci. 2018;59(9):3440. doi: https://doi.org/.

      Download citation file:

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

  • Supplements

Purpose : Neuropathic corneal pain (NCP) is one of the most underdiagnosed ocular surface diseases due to lack of clinical signs explaining patients’ symptoms. The presence of micro-neuromas in the subbasal nerve plexus visualized by in vivo confocal microscopy (IVCM) can be used as an adjunct to making the differential diagnosis possible. We hypothesize that a form of artificial intelligence known as deep neural networks can be utilized in the automated analyses of corneal subbasal nerve alterations and identifying micro-neuromas associated with NCP.

Methods : All images from IVCM sequences from 15 DED and NCP patients (total 12,212 images) were analyzed. The presence or absence of a micro-neuroma was identified and confirmed by at least 2 ophthalmologists. A deep neural network with over 24 million parameters (ResNet-50, pre-trained on Imagenet) was trained to predict to presence or absence of micro-neuromas in each image. We used a standard cross-validation strategy to assess the performance of the algorithm. We trained the network using the data from 12 patients and assessed the AUC, sensitivity, and specificity of the algorithm with respect to the labels provided by the ophthalmologists on the 3 patients who were withheld. We repeated this procedure 5 times and computed confidence intervals for each metric. The model was trained using 8 Titan X graphics processing units over a 24-hour period.

Results : 407 images were found to contain a micro-neuroma while 11,805 did not contain a micro-neuroma. The sensitivity for detection of micro-neuromas was calculated as 0.71 with a 95% confidence interval (CI) range lying between 0.54 – 0.87. The specificity of our method was 0.93 (95% CI 0.91 – 0.95). The area under the receiver-operator curve (AUC) was calculated as 0.91 (95% CI 0.85 – 0.97), signifying high accuracy.

Conclusions : The AI system had a very high AUC of 0.91 for detecting micro-neuromas, and the inclusion of additional patients will likely increase this performance further. The deep neural network shows great promise in identifying micro-neuromas associated with NCP, potentially allowing for the standardization of IVCM image analysis. Our study suggests that artificial intelligence can rapidly evaluate IVCM images, while maintaining a high degree of accuracy.

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.