Exploiting standardised tissue-mimicking phantoms to enable deep learning-based estimation of optical tissue properties on experimental photoacoustic data

点击翻译按钮获取中文摘要

Photoacoustic imaging (PAI) is an emerging imaging modality that offers the capability of measuring optical tissue properties non-invasively and in real-time. One of the most promising applications of the technique is the estimation of functional tissue properties, such as blood oxygenation. Such analysis requires the accurate and spatially resolved measurement of optical absorption at several wavelengths. However, PAI measurements are not only dependent on the optical absorption, but also the fluence, the distribution of light in tissue, which makes the problem of ill-posed inverse nature. As a result, a key computational challenge of PAI is the recovery of quantitative values of the underlying absorption and scattering properties of the tissue from reconstructed photoacoustic images. As it is practically impossible to obain ground truth measurements of the underlying optical tissue properties in vivo, methods that tackle this challenge have to rely on simulated data. While several promising approaches that tackle this problem have already been presented, current methods suffer from a systematic gap between numerical models and experimental data.The central hypothesis of this project is that the problem can be tackled by leveraging a combination of novel data-driven approaches and advanced tissue-mimicking phantoms. To this end, state-of-the-art physical forward models will be partnered with a recently developed formulation for standardised tissue-mimicking phantoms to create pairs of simulated and experimental photoacoustic measurements. The acquired dataset can be used to test the central hypothesis by: (1) examining and quantifying the gap between simulated and experimental measurements; (2) training data-driven models on experimental data in a supervised manner due to the availability of ground truth optical property information; and (3) investigating the feasibility of applying the data-driven inversion algorithm trained on tissue-mimicking phantoms to different in vitro and in vivo experimental data.