July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Automatic tracing of corneal sub-basal nerves using deep learning
Author Affiliations & Notes
  • Alessia Colonna
    University of Padova, Italy
  • Fabio Scarpa
    University of Padova, Italy
  • Alfredo Ruggeri
    University of Padova, Italy
  • Footnotes
    Commercial Relationships   Alessia Colonna, None; Fabio Scarpa, None; Alfredo Ruggeri, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 4237. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Alessia Colonna, Fabio Scarpa, Alfredo Ruggeri; Automatic tracing of corneal sub-basal nerves using deep learning. Invest. Ophthalmol. Vis. Sci. 2019;60(9):4237.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : In-vivo confocal microscopy is a rapid and noninvasive technique that provides information on the health state of the cornea. Several studies have shown a correlation between corneal nerves morphometry and a wide range of diseases. Despite the interest in those parameters, the use of this technique in clinical practice is limited due to the time cost and the difficulty in obtaining an accurate tracing of the nerves. This study provides automated and robust corneal descriptors in order to allow corneal nerves tracing through Convolutional Neural Network (CNN), a deep learning technique.

Methods : We used 8909 confocal images of sub-basal corneal nerve plexus from different subjects, covering a field of 400x400 μm (384x384 pixels), acquired using the Heidelberg Retina Tomograph (HRT-II) with the Rostock Cornea Module (Heidelberg Engineering GmbH, Heidelberg, Germany). Images were acquired in different clinical centers and anonymized. With them, we trained a CNN based on U-net. This architecture uses a contracting-encoder path (5 blocks of 2 convolutional units and max pooling) to provide the features that represent the nerves and an expanding-decoder path (5 blocks of 2 convolutional units concatenated with the corresponding blocks in the previous path and up-sampled convolution) to create a probability map (Fig. 1) and trace the nerves. Since very few public datasets are available with the manual tracking, we used as ground truth for training the tracing obtained with a previous algorithm proposed in literature (Guimarães et al., Transl Vis Sci Technol, 2016).

Results : We evaluated the CNN’s performance on 70 test images, not included in the training set and manually segmented by an expert. The True Positive Rate, i.e., the percentage of nerves correctly traced by the CNN is 98% and the False Discovery Rate, i.e., the percentage of wrongly traced nerves, is 14%. We compared the CNN’s results also with the previous technique, used here to produce ground truth (Fig. 2).

Conclusions : The proposed CNN appears capable of extracting features that correctly describe the nerves, since it was able to provide a very accurate nerve tracing. The FDR is relatively high probably because the CNN does not include any post-processing. The CNN appears capable of tracing nerves not traced by the previous technique (Fig. 2).

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

 

Convolutional Neural Network architecture.

Convolutional Neural Network architecture.

 

Comparison between proposed and literature (Guimarães et al.) method.

Comparison between proposed and literature (Guimarães et al.) method.

×
×

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.

×