June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Modular Deep Neural Network (DNN) – A new paradigm for classifying optical coherence tomography (OCT) B-scans by pathology
Author Affiliations & Notes
  • Krunalkumar Ramanbhai Patel
    Center of Application and Research in India, Carl Zeiss India (Bangalore) Pvt. Ltd., Bangalore, Karnataka, India
  • SANDIPAN CHAKROBORTY
    Center of Application and Research in India, Carl Zeiss India (Bangalore) Pvt. Ltd., Bangalore, Karnataka, India
  • Niranchana Manivannan
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Mary K Durbin
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Footnotes
    Commercial Relationships   Krunalkumar Ramanbhai Patel, Carl Zeiss India (Bangalore) Pvt. Ltd. (E); SANDIPAN CHAKROBORTY, Carl Zeiss India (Bangalore) Pvt. Ltd. (E); Niranchana Manivannan, Carl Zeiss Meditec, Inc. (E); Mary Durbin, Carl Zeiss Meditec, Inc. (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 1780. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Krunalkumar Ramanbhai Patel, SANDIPAN CHAKROBORTY, Niranchana Manivannan, Mary K Durbin; Modular Deep Neural Network (DNN) – A new paradigm for classifying optical coherence tomography (OCT) B-scans by pathology. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1780.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Macular OCT B-scans may contain various pathologies and identification of these pathologies using Artificial Intelligence-assisted technologies may help doctors diagnose macular diseases more quickly. For an accurate yet fast clinical decision support system, it is essential to develop a robust algorithm with high sensitivity that can detect the presence of multiple pathologies. This could be attempted with two different approaches such as 1) Monolithic DNN, and the proposed 2) Modular DNN. With monolithic DNN, a single DNN is involved to absorb features that are required to train all the pathologies – the usual way of handling the typical multi-label classification task. The idea of modular DNN is to dedicate a single and potentially smaller DNN for each pathology that yields individual outputs and then combines these outputs of several DNNs to arrive at conclusions.

Methods : For training and evaluating the performance of the monolithic and modular DNN based algorithms, a total of 598 CIRRUS macular OCT cubes from 598 patients were collected. The labeling was done at the B-scan level with 8 pathologies (one or more per B-scan) by two retina specialists. Ungradable B-scans were excluded from the dataset. The dataset was then divided into train and test with 80-20 split at patients’ level.
Fig. 1(a) table depicts the list of 8 pathologies and their corresponding samples used for training and evaluating both algorithms while 1(b) shows the distribution of pathologies. For both methods, inception_resnet_v2 DNNs were used – 1 for monolithic DNN and 8 for modular DNN (Fig. 1(c)).

Results : The performances of both the algorithms were evaluated on the hold-out test set. The monolithic DNN achieved maximum 75.36% AUC for RPE Elevation while modular DNN achieved maximum 98.85% AUC for Intraretinal Fluid. Fig 2 shows the achieved AUCs for modular DNN for all 8 pathologies.
Modular DNN based algorithm outperforms monolithic DNN significantly for all pathologies. The proposed approach provides a flexibility to include or exclude pathologies for designing any customized workflow.

Conclusions : In this study, we determined the modular DNN architecture-based algorithm outperformed a monolithic DNN based algorithm for the multi-pathology classification task with no loss of generality.

This is a 2021 ARVO Annual Meeting abstract.

 

Fig 1. Sample distribution & Monolithic/Modular DNN architectures

Fig 1. Sample distribution & Monolithic/Modular DNN architectures

 

Fig 2. AUCs for modular DNN architecture

Fig 2. AUCs for modular DNN 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.

×