Quantification of optical absorption from preclinical photoacoustic images with semisupervised deep learning (Conference Presentation)

Photoacoustic imaging promises functional characterisation of biological tissues, but challenges in quantifying derived biomarkers hinder progress. Additionally, the validation of photoacoustic data processing algorithms is complex due to a simulation gap between simulated and experimental data. In this work, we present a two-stage semi-supervised deep learning approach to tackling these problems. In the first stage, we use time series-domain data to close the aforementioned simulation gap using a semi-supervised training approach. In the second stage, we train a supervised network for the quantification of the optical absorption coefficient from a hybrid dataset of experimental and simulated spatial-domain images. Our results show that we can effectively close the simulation gap and transform the qualitative photoacoustic images into a quantitative domain that is highly correlated with the optical absorption coefficient.