Abstract
Quantitative photoacoustic imaging (PAI) aims to determine the spatially varying optical absorption coefficient of a sample using the measured photoacoustic (PA) signals. When imaging tissue, this can be used to investigate the absolute concentration of the various constituent chromophores, such as oxy- and deoxyhaemoglobin. Supervised deep learning approaches have achieved promising results when trained to predict the absorption coefficient using synthetic datasets. However, models trained using synthetic data struggle to generalise to real data. Furthermore, very limited experimental data is available for this task, causing models trained using these data to overfit. The purpose of this study is to address these challenges using transfer learning. For this, convolutional neural networks (U-Nets) were pre-trained on a diverse synthetic dataset, created using 3D optical and acoustic modelling, and then fine-tuned on a publicly available experimental phantom dataset. When compared to U-Nets that were randomly initialised and trained on just the experimental dataset, the fine-tuned U-Nets achieved a ∼17% lower root mean squared error (RMSE) when predicting the optical absorption coefficient of the inclusions in the experimental phantom test dataset. This study also shows that, so long as the image formation process is the same for both training and testing data, and the training images are diverse, then U-Nets trained on synthetic data created from non-anatomical images are able to generalise to synthetic data created from an anatomically realistic mouse model.