July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Healthy vs pathological classification of corneal nerves images using deep learning
Author Affiliations & Notes
  • Fabio Scarpa
    University of Padova, Padova, Italy
  • Alessia Colonna
    University of Padova, Padova, Italy
  • Alfredo Ruggeri
    University of Padova, Padova, Italy
  • Footnotes
    Commercial Relationships   Fabio Scarpa, None; Alessia Colonna, None; Alfredo Ruggeri, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 2107. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Fabio Scarpa, Alessia Colonna, Alfredo Ruggeri; Healthy vs pathological classification of corneal nerves images using deep learning. Invest. Ophthalmol. Vis. Sci. 2019;60(9):2107.

      Download citation file:

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

  • Supplements

Purpose : In-vivo confocal microscopy provides information on the corneal health state. The correlation between corneal nerves morphology and pathology has been shown several times. However, the evaluation of corneal morphology is quite difficult and often based on a tedious and subjective manual tracing of nerve fibers. We developed a completely automated method, based on a deep learning technique, able to distinguish images from healthy or diabetic with neuropathy subjects.

Methods : Images from 30 healthy and 30 diabetic subjects with neuropathy were acquired using the Heidelberg Retina Tomograph (HRT-II) with the Rostock Cornea Module (Heidelberg Engineering GmbH, Heidelberg, Germany). Image acquisition was performed in different clinical centers and each image was anonymized eliminating any patient information. Three non-overlapping images from the central part of the cornea were acquired for each subject, all covering an area of 400x400 μm (384x384 pixels).
A Convolutional Neural Network (CNN) simultaneously analyzes for each subject all three images. The CNN is composed of a feature extraction part (3 convolutional units and 3 max-pooling layers) and a classification part (2 fully connected layers and 1 softmax layer) (Fig.1). The network was trained by taking the three images in every possible order, so that the final classification is not influenced by the order in which the images of the subject are considered. In order to increase the data available to train the CNN, for each image trio we also considered its copy flipped horizontally. Thus, we considered a total of 720 blocks of three images.

Results : We trained the CNN on 59 subjects and evaluated the remaining subject, in turn for all the 60 subjects of the set (leave one out cross-validation). The CNN was able to correctly classify 57 of the 60 subjects (accuracy 95%). The CNN erroneously classified three pathological subjects as healthy.

Conclusions : We developed a method based on a deep learning technique that simultaneously analyses multiple images of one subject, automatically extracts features and finally provides an overall classification of the subject. The proposed method provides a very good classification of healthy subjects and subjects with diabetic neuropathy, demonstrating the potentiality of CNN in identifying clinically useful features.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.


Fig.1 : CNN architecture

Fig.1 : CNN architecture


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.