June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Automated machine learning detection of transcellular pores in Schlemm’s canal endothelial cells exposed to stretch
Author Affiliations & Notes
  • Justino Rodrigues
    Bioengineering, Imperial College London, London, London, United Kingdom
  • Anil Bharath
    Bioengineering, Imperial College London, London, London, United Kingdom
  • Darryl R Overby
    Bioengineering, Imperial College London, London, London, United Kingdom
  • Footnotes
    Commercial Relationships   Justino Rodrigues, None; Anil Bharath, None; Darryl Overby, None
  • Footnotes
    Support  NIH EY022359
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 483. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Justino Rodrigues, Anil Bharath, Darryl R Overby; Automated machine learning detection of transcellular pores in Schlemm’s canal endothelial cells exposed to stretch. Invest. Ophthalmol. Vis. Sci. 2021;62(8):483.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To drain from the eye, aqueous humor (AH) passes through micron-sized pores in the inner wall endothelium of Schlemm’s canal (SC). SC pores are reduced in glaucoma, which may contribute to increased AH outflow resistance and elevated IOP. To investigate pore formation mechanisms, we previously developed a fluorescent assay to label pores in SC cells (Braakman et al. ARVO, 2014). The cells are stretched, which triggers pore formation, and exposed to fluorescent avidin that crosses the cell body at pore locations to create a fluorescent “hotspot” on the substrate that can be imaged by confocal microscopy. This technique produces high volumes of image data that must be examined by a trained human observer to identify individual pores. Because cultured SC cells form immature intercellular junctions, pores are often difficult to discern from noisy junctional labelling, and pores themselves are relatively sparse. These difficulties make pore labeling time consuming and labor-intensive. To increase efficiency, we developed an automated machine learning algorithm to detect SC pores in fluorescent images.

Methods : The algorithm consisted of a custom pre-processing block for interest point detection, a Histogram of Oriented Gradients feature extractor and a RUSBoost classifier (Seiffert et al. IEEE, 2008). This architecture was trained and tested on separate datasets derived from ~1,500 manually labelled pores and ~65,000 non-pore objects, obtained from 300 previously analyzed images of two SC cell strains, one normal (SC67) & one glaucomatous (SC57g).

Results : Our trained algorithm significantly accelerates SC pore detection relative to identification by a human observer (0.2 s vs. ~8 min per 0.2mm2 image containing ~60 cells). Based on a testing set of 100 pores and 20,000 non-pores, our algorithm achieves 80% recall, 16% precision and 4% false positive rate (FPR) on SC pores.

Conclusions : Machine learning algorithms maximize the value of fluorescent pore detection in SC cells. The high recall indicates that the algorithm can successfully identify most pores. The low FPR indicates that the algorithm has high discriminative capacity, but the low precision prevents the algorithm from being used as a stand-alone pore detection tool. However, our algorithm can be used to rapidly identify candidate pores to pass to a human observer, accelerating the process of pore detection in SC cells.

This is a 2021 ARVO Annual Meeting abstract.

×
×

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.

×