Photonics West 2022 in San Francisco
Starts Saturday
>
Conference 11960 > Paper 11960-165
Paper 11960-165

Network-generated synthetic training data for photoacoustic sO2-estimation

Abstract

Data-driven approaches to estimating sO2 from photoacoustic images have been successfully implemented in several in silico experiments. However, networks trained on simulated datasets may not produce accurate estimates from real tissues due to differences in the data domains describing real and simulated images. Unfortunately, obtaining a training set of images of real tissues and their accompanying sO2 distributions is non-trivial. However, differences between the data domains can be minimised by improving the simulated data. Here we describe two methods: (1) using unsupervised domain adaptation via Cycle Consistent Adversarial Networks, and (2) generating training data using Ambient Generative Adversarial Networks.

Presenter

Ciaran Bench
Univ. College London (United Kingdom)
Presenter/Author
Ciaran Bench
Univ. College London (United Kingdom)
Author
Univ. College London (United Kingdom)