August 2021
Volume 62, Issue 11
Open Access
ARVO Imaging in the Eye Conference Abstract  |   August 2021
Synthesizing degree-of-polarization-uniformity from non-polarization sensitive OCT by convolutional neural network
Author Affiliations & Notes
  • Kensuke Oikawa
    Computational Optics Group, Tsukuba Daigaku, Tsukuba, Ibaraki, Japan
  • Shuichi Makita
    Computational Optics Group, Tsukuba Daigaku, Tsukuba, Ibaraki, Japan
  • Masahiro Miura
    Department of Ophthalmology, Tokyo Ika Daigaku Ibaraki Iryo Center, Inashiki-gun, Ibaraki, Japan
  • Takuya Iwasaki
    Department of Ophthalmology, Tokyo Ika Daigaku Ibaraki Iryo Center, Inashiki-gun, Ibaraki, Japan
  • Toshihiro Mino
    Kabushiki Kaisha Topcon, Itabashi-ku, Tokyo, Japan
  • Tatsuo Yamaguchi
    Kabushiki Kaisha Topcon, Itabashi-ku, Tokyo, Japan
  • Yoshiaki Yasuno
    Computational Optics Group, Tsukuba Daigaku, Tsukuba, Ibaraki, Japan
  • Footnotes
    Commercial Relationships   Kensuke Oikawa, Kao (F), Nikon (F), Sky Technology (F), TOPCON (F), Yokogawa (F); Shuichi Makita, Kao (F), Nikon (F), Sky Technology (F), Tomey (P), TOPCON (F), Yokogawa (F); Masahiro Miura, Alcon (F), Novartis (F), Santen (F); Takuya Iwasaki, None; Toshihiro Mino, TOPCON (E); Tatsuo Yamaguchi, TOPCON (E); Yoshiaki Yasuno, Kao (F), Nikon (F), Sky Technology (F), Tomey (P), TOPCON (F), Yokogawa (F)
  • Footnotes
    Support  JSPS grant 18K09460 and 18H01893, JST grant JPMJMI18G8
Investigative Ophthalmology & Visual Science August 2021, Vol.62, 41. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Kensuke Oikawa, Shuichi Makita, Masahiro Miura, Takuya Iwasaki, Toshihiro Mino, Tatsuo Yamaguchi, Yoshiaki Yasuno; Synthesizing degree-of-polarization-uniformity from non-polarization sensitive OCT by convolutional neural network. Invest. Ophthalmol. Vis. Sci. 2021;62(11):41.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Degree of polarization uniformity (DOPU) highlights the retinal pigment epithelium (RPE) pathology. However, polarization sensitive (PS-) OCT is mandatory to obtain DOPU. This presentation demonstrates a convolutional-neural-network based method generating a DOPU image from a non-PS OCT image, so that it will enables DOPU imaging by a conventional OCT device.

Methods : A simplified PS-OCT was used, it provides DOPU and non-PS-OCT images by a single measurement.
117 pathological eyes of 96 subjects and 4 normal eyes of 4 subjects were examined.
A neural network model inspired by a U-Net was trained to generate a DOPU image from a non-PS OCT image. Both input and output image patch sizes are 64 × 64 pixels.
105 of the 117 pathologic eyes (84 subjects) were used for training. 8,000 image patches were used to the training itself, while 2,000 patches were for validation.
12 pathological eyes were used to evaluate the generated pseudo DOPU (pDOPU) images. An expert ophthalmologist selected the region of pathology by using non-PS-OCT, and 5 B-scans were selected from the pathologic region of each eye. The DOPU and pDOPU images of the B-scan were presented to another expert grader. And the grader evaluated the visibility of RPE defect, RPE thickening, RPE elevation and hyper-refractive foci (HRF). For normal eyes, the grader evaluated the soundness of RPE appearance in equally spaced 5 B-scans of DOPU and pDOPU for each eye.

Results : Fig. 1 shows an example of OCT, DOPU and pDOPU of a pathologic eye. Although RPE elevation is seen in OCT, DOPU does not show this elevation. It indicates the lack of melanin. pDOPU shows consistent appearance. In addition, both DOPU and pDOPU images exhibit an RPE defect (arrow).
The agreements of clinical features, which is defined as the number of features found both in DOPU and pDOPU over the total number of features, are 35.7% for RPE defect, 75.0% for RPE thickening, 83.3% for RPE elevation, and 12.5% for HRF (see Fig. 2 for details). For the normal eyes, 3 positions of the 20 DOPU images showed RPE-defect artifacts, while all RPEs appeared as normal in pDOPU.

Conclusions : pDOPU successfully highlighted RPE thickening and elevations, while more improvements are needed for RPE defect and HRF. For normal RPE, pDOPU was less prone to RPE defect artifacts.

This is a 2021 Imaging in the Eye Conference abstract.

 

An example of OCT, DOPU and pDOPU.

An example of OCT, DOPU and pDOPU.

 

The summary of clinical findings in DOPU and pDOPU.

The summary of clinical findings in DOPU and pDOPU.

×
×

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.

×