June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Volumetric Choroidal Segmentation with a Novel Deep Learning Approach fusing Spatial Information
Author Affiliations & Notes
  • Dheo Cahyo
    Ocular Imaging, Singapore Eye Research Institute, Singapore, Singapore
  • Ai Ping Yow
    Nanyang Technological University, Singapore, Singapore, Singapore
  • Seang Mei Saw
    Singapore Eye Research Institute, Singapore, Singapore
  • Leopold Schmetterer
    Ocular Imaging, Singapore Eye Research Institute, Singapore, Singapore
  • Damon Wong
    Nanyang Technological University, Singapore, Singapore, Singapore
  • Footnotes
    Commercial Relationships   Dheo Cahyo, None; Ai Ping Yow, None; Seang Mei Saw, None; Leopold Schmetterer, None; Damon Wong, None
  • Footnotes
    Support  NMRC/OFLCG/004c/2018, NMRC/CG/C010A/2017_SERI, NMRC/OFIRG/0048/2017
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 2156. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Dheo Cahyo, Ai Ping Yow, Seang Mei Saw, Leopold Schmetterer, Damon Wong; Volumetric Choroidal Segmentation with a Novel Deep Learning Approach fusing Spatial Information. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2156.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To evaluate the performance of a novel deep learning-based volumetric choroidal segmentation approach incorporating spatial information on swept source optical coherence tomography (SS-OCT) images in eyes with myopia.

Methods : 126 myopia eyes were acquired using a commercial swept-source OCT (SS-OCT) system, DRI OCT Triton (Topcon Corp., Japan) at a 1050 nm wavelength, scanning speed of 100,000 A-scans/sec and 7 mm × 7 mm scanning protocol, centered at the macula. Each volumetric image contained 256 cross-sectional B-scans with dimensions of 256×128 pixels. Manual annotations of the choroid were used as the ground truth. We implemented a novel multi-task deep convolutional neural network architecture, which we named as SA-Net, that reconstructs and segments a target B-scan with the incorporation of spatial context from neighbouring B-scans. Intersection over Union (IoU) of the volumetric segmentation and Structural Similarity Index (SSIM) of the enface choroidal thickness map were used to assess the accuracy of the detected volume. Results were also compared with choroidal segmentation using U-Net.

Results : For the myopia eyes (spherical equivalent = 5.44 ± 2.11 D), the measured choroidal thickness was 0.204 ± 0.046 mm. SA-Net required a processing time of 1.08s for each volumetric choroidal segmentation, and achieved an average cross-validation segmentation IoU of 0.942 (95% CI: 0.937 to 0.946) compared to an IoU of 0.929 (95% CI: 0.923 to 0.934) with U-Net. The mean absolute difference between the ground truth choroidal volume and volumetric segmentation for SA-Net was 0.191 mm3 (95% CI: 0.164 mm3 to 0.218 mm3 ) with a mean absolute sub-foveal choroidal thickness difference of 0.010 mm (95% CI: 0.008 mm to 0.012 mm) between the ground truth and segmented. For the choroidal thickness map we obtained SSIM of 0.700 (95% CI: 0.688 to 0.711) compared to the ground truth thickness map.

Conclusions : The novel SA-Net approach showed a high accuracy in detecting the choroid from volumetric OCT cube scans in eyes with myopia. This indicates that spatial information can provide useful context for volumetric segmentation. The results are promising for the automated detection of the choroid for further analysis in myopia and other ocular diseases.

This is a 2021 ARVO Annual Meeting abstract.

 

SA-Net learns spatial information from two or more adjacent slices and infuse thespatial information to segmentation task.

SA-Net learns spatial information from two or more adjacent slices and infuse thespatial information to segmentation task.

×
×

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.

×