June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Automatic detection of drusen-like deposits on OCT using EfficientNet models
Author Affiliations & Notes
  • Luís Mendes
    Associacao para a Investigacao Biomedica e Inovacao em Luz e Imagem, Coimbra, Coimbra, Portugal
  • Thiago Martins
    Universidade Federal de Sao Paulo, Sao Paulo, São Paulo, Brazil
  • Paulo Schor
    Universidade Federal de Sao Paulo, Sao Paulo, São Paulo, Brazil
  • Jose G Cunha-Vaz
    Associacao para a Investigacao Biomedica e Inovacao em Luz e Imagem, Coimbra, Coimbra, Portugal
    Universidade de Coimbra, Coimbra, Coimbra, Portugal
  • Rufino Silva
    Associacao para a Investigacao Biomedica e Inovacao em Luz e Imagem, Coimbra, Coimbra, Portugal
    University of Coimbra, Coimbra Institute for Clinical and Biomedical Research, Faculty of Medicine (iCBR- FMUC), Coimbra, Portugal
  • Footnotes
    Commercial Relationships   Luís Mendes None; Thiago Martins None; Paulo Schor None; Jose Cunha-Vaz Carl Zeiss Meditec, Code C (Consultant/Contractor), Alimera Sciences, Code C (Consultant/Contractor), Allergan, Code C (Consultant/Contractor), Bayer, Code C (Consultant/Contractor), Gene Signal, Code C (Consultant/Contractor), Novartis, Code C (Consultant/Contractor), Pfizer, Code C (Consultant/Contractor), Roche, Code C (Consultant/Contractor); Rufino Silva Allergan, Code C (Consultant/Contractor), Alimera Sciences, Code C (Consultant/Contractor), Bayer, Code C (Consultant/Contractor), Novartis, Code C (Consultant/Contractor), Roche, Code C (Consultant/Contractor), Novus Nordisk, Code C (Consultant/Contractor), Thea Pharmaceuticals, Code C (Consultant/Contractor)
  • Footnotes
    Support  Fundação para a Ciência e a Tecnologia ( 02/SAICT/2017 – 032412 )
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 2091 – F0080. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Luís Mendes, Thiago Martins, Paulo Schor, Jose G Cunha-Vaz, Rufino Silva; Automatic detection of drusen-like deposits on OCT using EfficientNet models. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2091 – F0080.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Drusen-like deposits are hallmarks of age-related macular degeneration (AMD) that can be found between the retinal pigment epithelium and Bruch’s membrane. On OCT they can present different sizes, shapes, and localizations. We evaluate the performance of models based on EfficientNet B0 architectures to automatically detect B-scans with drusen-like deposits (DLD).

Methods : OCT(A) data acquired in the scope of a DR screening program that is running in Coimbra, Portugal were used for training the models. The imaging data were acquired by the ZEISS AngioPlex OCT Angiography (ZEISS, Dublin. CA) using the Angio 6x6 mm protocol. The dataset includes 960 B-scans collected from 67 healthy eyes (71±4.8 years) and 842 Bscans collected from eyes having drusen-like deposits (DLD) from 123 eyes (68±3.5 years). The DLD were identified by an ophthalmologist that classified the pathological scans into two categories: with the presence of DLD with a size equal or superior to 63 µm (28%), and with the presence of DLD of size less 63 µm (72%). Deep learning models based on EfficientNets B0 architectures were trained and evaluated using a 5-fold cross-validation approach to detect B-scans having DLS. The specificity and the sensitivity metrics were used to evaluate the performance of the models.

Results : An overall sensibility equals to 84.4±10.5 % and a specificity equals to 80±11.9 % were measured on the testing folders. When were only considered DLD equals or largers than 63 µm the value of the sensibility was equal to 92±7% and the and specificity to 80.0±11.6%.

Conclusions : Our results show that EfficientNet B0 architectures can detect DLD even when trained with a small dataset. Is expected that the addition of more data will improve the performance of the models. Automatic detection of DLD is especially important in the scope of screening programs.

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.

×