Temporal unmixing Optoacoustics – Machine learning to enable routine whole animal Optoacoustic imaging of genetically encoded photo-modulatable labels.

光声成像（OA）或光声成像是一种新兴的成像方式，它提供了无与伦比的高分辨率穿透深度。它提供了更全面的时间分辨率的3D活体图像，深度远远超出了光学方法的范围。此外，由于OA只需要一个光激发源和超声探测器，在基础设施方面，它与光学显微镜相比，成本高的放射学方法更具可比性。OA在生命科学中的应用是新的，现有的少数遗传可编码标签（可靶向体内成像的先决条件）缺乏足够高的OA信号进行全动物研究。这防止了从组织中血红蛋白的主要信号分离。这一挑战可以通过使用基于遗传编码的光着色蛋白（以下称为可逆可切换OA蛋白，rsOAP）的标记来克服。这些蛋白质的信号可以通过光进行调制，从而可以将调制标签信号从非调制背景中分离出来。rsoap在整个生物体水平上研究细胞动力学的潜力最近已得到证实。标记的细胞在体内可见，深度可达1厘米，数量较少（约500）。到目前为止，基于不同的调制特性，多达三个标签的并行可视化（多路复用）是可能的。然而，这些研究使用了专用的高端OA设备，并表明对调制动力学的解释要求很高，而且容易产生伪影。这阻碍了常规应用，特别是关于组织深处非常小的细胞数量或不同标签的纠缠群体，这是许多生物学研究的典型情况。最近，我们展示了通过机器学习（ML）分析的众多特征（例如信号强度、深度、动力学、背景噪声）的解释，极大地提高了使用现成仪器进行的基于rsoap的测量分析。然而，在该方法成为生命科学中OA全动物成像的常规应用之前，还需要进一步提高准确性和灵敏度。在这个项目中，我们将解决机器学习方法的挑战，并对其进行优化，重点是对少量不同标记的细胞进行多路检测。建议的工作将有助于在OA中充分实现rsOAP成像的常规使用。OA作为一种标准的生命科学成像方式将允许其更广泛的使用，并为研究人员提供一种不可缺少的工具来可视化整个生物体体内小细胞群的相互作用。此外，多路复用将允许在更大范围内解释细胞的动态相互作用。这种观察对免疫学、发育和肿瘤生物学至关重要，可以深入了解癌症等疾病机制背后的动态相互作用。

Optoacoustic (OA) or Photoacoustic imaging is an emerging imaging modality that provides unsurpassed penetration depth at high resolution. It delivers more comprehensive time-resolved 3D in vivo images at depths far beyond the reach of optical methods. Moreover, with OA only requiring an optical excitation source and ultrasound detector, in terms of infrastructure, it is more comparable to optical microscopy than to cost intensive radiological methods. The application of OA in life sciences is new and the few existing genetically encodable labels (a prerequisite for targetable in vivo imaging) lack sufficiently high OA signal for whole animal studies. This prevents the separation from the predominant signal of blood hemoglobin in tissue. This challenge can be overcome by employing labels based on genetically encoded photochromatic proteins (hereafter called reversibly switchable OA proteins, rsOAP). These proteins’ signal can be modulated by light which allows a clean separation of modulating label signal from non-modulating background.The potential of rsOAPs for studying the dynamics of cells at the level of the whole organism has recently been demonstrated. Labeled cells were visualized in vivo at a depth of up to one centimeter and at low numbers (~500). So far, parallel visualization (multiplexing) of up to three labels is possible based on different modulation characteristics. However, those studies used dedicated high-end OA setups and showed that the interpretation of the modulation kinetics is demanding and prone to artifacts. This hampers the routine application, especially with regard to very small cell numbers deep in the tissue or entangled populations of different labels, which are typical situations for many biological studies. Recently, we showed that the interpretation of a multitude of features (e.g. signal strength, depth, kinetics, background-noise) that were analyzed via machine learning (ML) strongly improves the analysis of rsOAP-based measurements conducted with of-the-shelf instrumentation. Yet, further improvement in accuracy and sensitivity are necessary before this approach can become a routine application enabling OA whole animal imaging in the life sciences.In this project, we will address the challenges of the ML approach and optimize it with a focus on multiplexed detection of small numbers of differently labeled cells. The proposed work will help to fully enable the routine use of rsOAP imaging in OA. OA as a standard life-science imaging modality will allow its wider use and provide researchers with an indispensable tool to visualize the interactions of small cell populations in vivo in whole organisms. Moreover, multiplexing will allow interpretation of dynamic interactions of cells on the larger scale. Observations of this kind are crucial for immunology, developmental and tumor biology, allowing insights into the dynamic interplay that underlies disease mechanisms such as in cancer.