Team OCT-opus
A Deep Learning Algorithm to Enhance Retinal Images
2020 University of Waterloo Software Engineering capstone design project
Catch us in California! 🌴
We are presenting our paper at
the Applications of Machine Learning
conference of the SPIE Optics + Photonics event.
It'll be at the San Diego Convention Center on Wednesday
26 August 2020.
Our goal:
Develop an algorithm that can be used to infer blood flow from OCT
cross-section images of human retinas, from only 1 cross-section per spot, captured by
Prof. Bizheva's lab.
The sequence of inferred cross-sections can then be combined to form a vascular map of the retina, which
empowers clinicians to perform diagnosis and monitoring of diseases affecting the eye.
Our solution:
A deep learning based image-to-image translation algorithm using the pix2pix cGAN
architecture. The end result is that our predicted images result in vascular maps that resemble
ground truth images that were created using the optical microangiography (OMAG) algorithm
on multiple acquisitions of each spot. This visual success is mirrored by the quantitative comparison
with the ground truth using k-folds cross-validation (k=5, 65 eyes of data) on sedated rat images,
revealing a structural similarity index (SSIM) of 0.62.
Structural cross-section (B-scan)
An example of an input structural image.
Predicted functional cross-section
Our algorithm's corresponding output enhancing visibility of vasculature.
Functional cross-section from OMAG algorithm
The ground truth image, obtained from the optical microangiography (OMAG) algorithm, which performs temporal decorrelation of 4 adjacent B-scan images from repeated acquisitions of the same spot.
En-face vascular map from structural cross-sections (B-scans)
En-face vascular map from predicted functional cross-sections
En-face vascular map from OMAG images
Want to hear more?
Take a look at our
capstone symposium
presentation slides
and keep an eye on the 2020 SPIE O+P conference proceedings.