June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Automated Machine learning models applied to Optical Coherence Tomography Angiography for detection and classification of Diabetic Retinopathy in Diabetes Mellitus type 1
Author Affiliations & Notes
  • Maximiliano Olivera
    Medical Retina Research, Moorfields Eye Hospital NHS Foundation Trust, London, London, United Kingdom
    Vitreo Retina, Hospital Universitario Insular de Gran Canaria, Las Palmas de Gran Canaria, Canarias, Spain
  • Carolina Bernal-Morales
    Medical Retina Research, Moorfields Eye Hospital NHS Foundation Trust, London, London, United Kingdom
  • Anibal Alé-Chilet
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Marina Barraso
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Sara Marin
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Silvia Feu
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Josep Rosinés
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Cristian Oliva
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Teresa Hernandez
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Irene Vila
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Robbert Struyven
    Medical Retina Research, Moorfields Eye Hospital NHS Foundation Trust, London, London, United Kingdom
  • Mariana Batista-Gonçalves
    Medical Retina Research, Moorfields Eye Hospital NHS Foundation Trust, London, London, United Kingdom
  • Siegfried Wagner
    Medical Retina Research, Moorfields Eye Hospital NHS Foundation Trust, London, London, United Kingdom
  • Pearse Andrew Keane
    Medical Retina Research, Moorfields Eye Hospital NHS Foundation Trust, London, London, United Kingdom
  • Javier Zarranz-Ventura
    Hospital Clinic de Barcelona, Barcelona, Catalunya, Spain
  • Footnotes
    Commercial Relationships   Maximiliano Olivera None; Carolina Bernal-Morales None; Anibal Alé-Chilet None; Marina Barraso None; Sara Marin None; Silvia Feu None; Josep Rosinés None; Cristian Oliva None; Teresa Hernandez None; Irene Vila None; Robbert Struyven None; Mariana Batista-Gonçalves None; Siegfried Wagner None; Pearse Keane None; Javier Zarranz-Ventura None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 2326. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Maximiliano Olivera, Carolina Bernal-Morales, Anibal Alé-Chilet, Marina Barraso, Sara Marin, Silvia Feu, Josep Rosinés, Cristian Oliva, Teresa Hernandez, Irene Vila, Robbert Struyven, Mariana Batista-Gonçalves, Siegfried Wagner, Pearse Andrew Keane, Javier Zarranz-Ventura; Automated Machine learning models applied to Optical Coherence Tomography Angiography for detection and classification of Diabetic Retinopathy in Diabetes Mellitus type 1. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2326.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To evaluate the diagnostic capacity of Automated Machine Learning (AML) algorithms to classify the degree of diabetic retinopathy (DR) from optical coherence tomography (OCT) and OCT angiography (OCTA) images.

Methods : Cross-sectional study on a pseudonymised dataset of images collected during a previous prospective clinical trial (NCT03422965), corresponding to a cohort of 485 patients with type 1 Diabetes Mellitus and 115 healthy controls. The retinal images were labeled with the corresponding demographic and clinical data collected in the clinical trial. The classification models based on Deep Learning and computer vision algorithms were trained hosted in a cloud-based system, under the Vertex AI platform of Google Cloud Platform, which allows the training of algorithms without the need automated fine-tuning of hyperparameters (AML). The performance was compared for superficial and deep capillary plexuses (SCP and DCP, respectively) on 3x3 mm and 6x6 mm captures, following 3 labeling strategies: No DR vs no DR; No DR vs. Mild DR vs. Referable DR, and No RD vs. Non-Proliferative RD vs. Proliferative RD.

Results : The performance of the classification models was evaluated through precision, sensitivity, specificity and f1-score. Overall, despite having good average precision values (ranging between 0.6 to 0.9 for structural OCT and 0.59 to 0.94 for OCTA), the models trained on structural OCT images had consistently a significative lower power of detection on the greater DR grade on each strategy (ranging from 0% to 33%) compared to 30% to 88% for the models trained on OCTA images. Models performance was consistently superior in those trained using SCP images and was greatest using 3x3 mm scanning protocol.

Conclusions : AML algorithms are able to detect the DR status when trained in OCTA images of DM patients, in particular using OCTA images from the SCP and high definition 3x3mm scan field. Unbalanced datasets may give rise to non-significative good performance metrics if not adequatelly analyzed. These preliminary results raise the interesting hypothesis of using OCTA images to investigate the potential ability of AML algorithms to detect systemic microvascular damage elsewhere in the body, by means of exploring existing relationships between OCTA images and quantitative systemic variables through Oculomics.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

×
×

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.

×