August 2021
Volume 62, Issue 11
Open Access
ARVO Imaging in the Eye Conference Abstract  |   August 2021
Using image similarity score to select images for training deep learning models to optimize annotation and computational resources
Author Affiliations & Notes
  • Niranchana Manivannan
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Gary Lee
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Lars Omlor
    Corporate Research and Technology, Carl Zeiss Inc., Pleasanton, California, United States
  • Hugang Ren
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Sophia Yu
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Rishi P. Singh
    Center for Ophthalmic Bioinformatics, Cole Eye Institute, Cleveland Clinic, Cleveland, Ohio, United States
  • Katherine Talcott
    Center for Ophthalmic Bioinformatics, Cole Eye Institute, Cleveland Clinic, Cleveland, Ohio, United States
  • Alline G.R. Melo
    Center for Ophthalmic Bioinformatics, Cole Eye Institute, Cleveland Clinic, Cleveland, Ohio, United States
  • Tyler E. Greenlee
    Center for Ophthalmic Bioinformatics, Cole Eye Institute, Cleveland Clinic, Cleveland, Ohio, United States
  • Eric R. Chen
    Center for Ophthalmic Bioinformatics, Cole Eye Institute, Cleveland Clinic, Cleveland, Ohio, United States
  • Thais F. Conti
    Center for Ophthalmic Bioinformatics, Cole Eye Institute, Cleveland Clinic, Cleveland, Ohio, United States
  • Mary Kathryn Durbin
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Footnotes
    Commercial Relationships   Niranchana Manivannan, Carl Zeiss Meditec, Inc. (E); Gary Lee, Carl Zeiss Meditec, Inc. (E); Lars Omlor, Carl Zeiss Inc. (E); Hugang Ren, Carl Zeiss Meditec, Inc. (E); Sophia Yu, Carl Zeiss Meditec, Inc. (E); Rishi Singh, Aerie (F), Alcon (C), Apellis (F), Bausch and Lomb (C), Genentech (C), Graybug (F), Gyroscope (C), Novartis (C), Regeneron (C); Katherine Talcott, Carl Zeiss Meditec, Inc. (F), Genentech (C), Roche (C); Alline Melo, None; Tyler Greenlee, None; Eric Chen, None; Thais Conti, None; Mary Durbin, Carl Zeiss Meditec, Inc. (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science August 2021, Vol.62, 61. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Niranchana Manivannan, Gary Lee, Lars Omlor, Hugang Ren, Sophia Yu, Rishi P. Singh, Katherine Talcott, Alline G.R. Melo, Tyler E. Greenlee, Eric R. Chen, Thais F. Conti, Mary Kathryn Durbin; Using image similarity score to select images for training deep learning models to optimize annotation and computational resources. Invest. Ophthalmol. Vis. Sci. 2021;62(11):61.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Deep learning networks (DLN) have been shown to provide good performance for classification and segmentation tasks in optical coherence tomography (OCT) imaging. Training a DLN requires large data, annotation, and computational resources. In this research we explore: 1) Whether we can use selected B-scans from an OCT cube to train the model and get comparable performance to using all the B-scans, 2) Whether we can use image similarity metrics to select which B-scans need to be annotated to optimize the performance.

Methods : This retrospective study uses 76,544 B-scans from 512 x 128 macular cubes of 598 subjects (one eye per subject) acquired using CIRRUS™ 5000 HD-OCT (ZEISS, Dublin, CA). Training and testing set contains B-scans from 478 and 120 OCT cubes respectively. Each B-scan is annotated for retinal pathologies by two retina specialists1.

The baseline algorithm (BSOI-128) is trained using all 128 B-scans from 478 OCT cubes. ResNet-50 based DLN is retrained using randomly selected 96 (BSOI-R96), 64 and 32 B-scans from each cube. For each B-scan in an OCT cube, square differences (SqD) and cross correlation (CC) with other scans from the cube are calculated. SqD and CC image similarity scores are computed by normalized summations. DLN is retrained using 96, 64 and 32 B-scans selected with lowest similarity scores from each cube. Accuracy and time taken to train the DLN are measured. Carbon footprints are estimated using ML CO2 impact calculator2.

Results : Figure 1 shows the accuracy, training time in Intel® Xeon® Processor E5-1650 v2 @3.50GHz 48.0 GB RAM with NVIDIA GTX 1080 GPU, carbon footprint and annotation time for various models. Model trained using CC similarity score achieved higher accuracy than SqD and randomly selected B-scans.

Conclusions : The findings in this study suggest image similarity score can be used to select smaller number of images to annotate and train a DLN with low impact to performance when compared with the DLN trained with large data. This will reduce the computational and annotation time as well as reducing the GPU’s carbon emission footprint. Future studies may include exploration of other similarity metrics and additional use cases.

References:
1Yu et al. IOVS 2020: Abstract PB0085
2Lacoste et al. arXiv:1910.09700v2 [cs.CY]

This is a 2021 Imaging in the Eye Conference abstract.

 

Figure 1. Performance metrics for various DLN models

Figure 1. Performance metrics for various DLN models

×
×

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.

×