June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Time-aware deep models for predicting diabetic retinopathy progression
Author Affiliations & Notes
  • Rachid zeghlache
    Universite de Bretagne Occidentale, Brest, Bretagne, France
    INSERM, LaTIM, UMR 1101, Brest, Bretagne, France
  • Pierre-henri Conze
    IMT Atlantique Bretagne-Pays de la Loire - Campus de Brest, Brest, Bretagne, France
    INSERM, LaTIM, UMR 1101, Brest, Bretagne, France
  • Mostafa EL HABIB DAHO
    INSERM, LaTIM, UMR 1101, Brest, Bretagne, France
    Universite de Bretagne Occidentale, Brest, Bretagne, France
  • YIHAO LI
    INSERM, LaTIM, UMR 1101, Brest, Bretagne, France
    Universite de Bretagne Occidentale, Brest, Bretagne, France
  • Ikram Brahim
    Universite de Bretagne Occidentale, Brest, Bretagne, France
    INSERM, LaTIM, UMR 1101, Brest, Bretagne, France
  • Hugo Le Boité
    Assistance Publique - Hopitaux de Paris, Paris, Île-de-France, France
  • Pascale Massin
    Assistance Publique - Hopitaux de Paris, Paris, Île-de-France, France
  • Ramin Tadayoni
    Assistance Publique - Hopitaux de Paris, Paris, Île-de-France, France
  • Béatrice Cochener
    Ophtalmology Department, CHRU Brest, Brest, France, France
    INSERM, LaTIM, UMR 1101, Brest, Bretagne, France
  • Gwenole Quellec
    INSERM, LaTIM, UMR 1101, Brest, Bretagne, France
  • Mathieu Lamard
    INSERM, LaTIM, UMR 1101, Brest, Bretagne, France
    Universite de Bretagne Occidentale, Brest, Bretagne, France
  • Footnotes
    Commercial Relationships   Rachid zeghlache None; Pierre-henri Conze None; Mostafa EL HABIB DAHO None; YIHAO LI None; Ikram Brahim None; Hugo Le Boité None; Pascale Massin None; Ramin Tadayoni Carl Zeiss Meditec Inc, Code C (Consultant/Contractor); Béatrice Cochener Carl Zeiss Meditec Inc, Code C (Consultant/Contractor); Gwenole Quellec Evolucare, Code C (Consultant/Contractor); Mathieu Lamard None
  • Footnotes
    Support  ANR grant ANR-18-RHU-008
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 246. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Rachid zeghlache, Pierre-henri Conze, Mostafa EL HABIB DAHO, YIHAO LI, Ikram Brahim, Hugo Le Boité, Pascale Massin, Ramin Tadayoni, Béatrice Cochener, Gwenole Quellec, Mathieu Lamard; Time-aware deep models for predicting diabetic retinopathy progression. Invest. Ophthalmol. Vis. Sci. 2023;64(8):246.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : (1) Develop a deep learning-based method to predict the progression of Diabetic Retinopathy (DR) using the sequence of previously-acquired color fundus photographs (CFP).
(2) Exploit a method that can take the time irregularity between examinations to better encode the disease progression.
(3) Evaluate whether taking into account the sequence of images can improve single-point DR prediction.

Methods : The proposed models were trained and evaluated on OPHDIAT, a large CFP database collected from the OPHDIAT network (Paris, France) and consisting of examinations acquired from 101,383 patients between 2004 and 2017. Around 30% of the patients have follow-up examinations. Each examination has at least two images for each eye. We selected two subsets of patients with longitudinal follow-up of 3 and 4 visits. This resulted in 6551 patients with an average follow-up duration of 4.35 years, and 3341 patients with an average duration of 5.7 years respectively. The database description (number of images and labels) is presented in Tab1. Time-Long short term memory (T-LSTM) consists of adding the time interval between two consecutive examinations in order to modulate both input and output gates of the LSTM. For the CNN, we used a Resnet-like architecture. We compared the three following methods: CNN (mean feature across visits), CNN+LSTM, CNN+T-LSTM (presented fig.1). The goal was to predict the severity of the last examination (according to the ICDR scale) using the follow-up examinations.

Results : For the evaluation, we compared Cohen's Kappa scores. For the dataset with 3 visits, we obtained the following Kappa value: CNN 0.804, CNN+LSTM 0.821, CNN+T-LSTM 0.844. or 4 visits: CNN 0.784, CNN+LSTM 0.793, CNN+T-LSTM 0.822. As baseline, we trained the CNN on the last image using the same parameters used for the best model. We obtained the following kappa value: 0.824 for 3 visits, 0.816 for 4 visits.

Conclusions : These results support the importance of injecting temporal information into the deep model in order to better encode the disease progression because only the CNN+T-LSTM surpasses the baseline for both experiments. Also, it indicates the possibility of using disease progression information to enhance the prediction of the DR severity. We suspect the size difference between the two dataset being the cause of the kappa drop score.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1. Proposed approach

Figure 1. Proposed approach

 

Table 1. Distribution of image, label, for the two sets during training.

Table 1. Distribution of image, label, for the two sets during training.

×
×

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.

×