Related Experiment Video
Updated: Jan 3, 2026
12:22
Multimodal Volumetric Retinal Imaging by Oblique Scanning Laser Ophthalmoscopy oSLO and Optical Coherence Tomography OCT
Published on: August 4, 2018
8.9K
Deep spectral learning for label-free optical imaging oximetry with uncertainty quantification.
Rongrong Liu1, Shiyi Cheng2, Lei Tian2
11Department of Biomedical Engineering, Northwestern University, Evanston, IL 60208 USA.
Light, Science & Applications
|November 23, 2019
Summary
A new deep spectral learning method offers robust, non-invasive blood oxygen saturation (sO2) measurement. This approach provides uncertainty quantification, improving accuracy over traditional methods for optical imaging oximetry.
Area of Science:
- Biomedical Optics
- Medical Imaging
- Oximetry
Background:
- Label-free optical imaging oximetry measures blood oxygen saturation (sO2) using sO2-dependent spectral contrast from hemoglobin.
- Traditional methods rely on analytical models, but suffer from uncertainties due to biological variability, tissue properties, and experimental conditions.
Purpose of the Study:
- To introduce a data-driven approach, deep spectral learning (DSL), for robust in vivo oximetry with uncertainty quantification.
- To demonstrate the generalizability and accuracy of DSL compared to traditional methods.
Main Methods:
- Developed and tested two neural network-based DSL models using visible light optical coherence tomography (vis-OCT) data from rat retinas.
- Compared DSL predictions against the traditional least-squares fitting (LSF) method.
Main Results:
- DSL demonstrated high adaptability to experimental variabilities and depth-dependent spectral properties.
- DSL-predicted sO2 showed significantly lower mean-square errors than LSF.
- Generated the first en face maps of retinal oximetry with pixel-wise confidence assessment.
Conclusions:
- Deep spectral learning provides a more flexible, robust, and reliable method for in vivo non-invasive label-free optical oximetry.
- DSL overcomes limitations of traditional analytical approaches, offering improved accuracy and uncertainty quantification.
