Image Enhancement and Segmentation for Deep Tissue Microscopy

描述（由申请人提供）：当前的2D和3D反褶积和分析技术不太适合深度组织显微镜。大多数假设点扩展函数（psf）是对称的或空间不变的，这与厚样本的情况不同。此外，许多使用的先验概率分布不是数据的良好模型，会使增强结果产生偏差。最后，许多技术都是计算密集型的。长期目标是开发适用于深层组织多光子显微镜的自动和无偏3D增强（反卷积），分割和定量工具，并将这些工具实现到多平台图像分析软件中，以实现多光子图像数据的高效，交互式定量。当前应用的目标是开发有效的二维增强和分割技术，可用于多光子数据的分析。该方法将基于(1)估计真实但未知的图像的未知概率分布（即先验分布）和(2)使用估计的先验分布在分割之前增强图像，而不是直接分割它们。未知概率分布的估计将通过假设噪声的泊松模型和施加一些深部组织成像特征的物理限制（如低光子计数），从获得的数据中得出。该研究的基本原理是，对psf和先验分布的不准确假设降低了增强的质量，从而降低了分割技术的准确性和灵敏度。该目标将通过以下具体目标来实现：(1)估计先验分布，并使用估计的先验通过最大后验（MAP）估计来增强图像。增强后的图像将通过跟踪物体边界的技术进行分割。(2)利用数学形态学对增强图像中已知形状的物体进行分割。这种方法是创新的，因为它通过施加更适合深部组织成像的约束，直接从数据中导出未知数据概率分布。这将导致独特的图像分析工具的发展，更适合于深组织荧光图像的不寻常的特点。本研究对提高多光子显微镜的定量化能力具有重要意义。

DESCRIPTION (provided by applicant): Current 2D and 3D deconvolution and analysis techniques are not well suited for deep tissue microscopy. Most assume the point spread functions (PSFs) to be symmetric or spatially invariant, which is not the case for thick samples. In addition, many utilize prior probability distributions that are not good models of the data that bias the enhancement results. Finally, many techniques are computationally intensive. The long term goal is to develop automatic and unbiased 3D enhancement (deconvolution), segmentation, and quantification tools suitable for deep-tissue multi-photon microscopy, and to implement these tools into multiplatform image analysis software for efficient, interactive quantitation of multi-photon image data. The objective of the current application is to develop efficient 2D enhancement and segmentation techniques that can be used in the analysis of multi-photon data. The approach will be based on (1) estimating the unknown probability distribution of the true but unknown images (i.e. the prior distribution) and (2) using the estimated prior distribution to enhance images prior to segmentation, rather than segmenting them directly. The estimate of the unknown probability distribution will be derived from the acquired data by assuming a Poisson model for the noise and by imposing some physical constraints that are characteristic of deep tissue imaging, such as low photon count. The rationale for the proposed research is that inaccurate assumptions about PSFs and the prior distributions lessen the quality of enhancement and consequently decrease the accuracy and sensitivity of segmentation techniques. The objective will be achieved through the following specific aims: (1) Estimating the prior distributions and using the estimated priors to enhance images via maximum a posteriori (MAP) estimation. The enhanced images will then be segmented via techniques that track object boundaries. (2) Utilizing mathematical morphology to segment objects of known shape from the enhanced images. This approach is innovative because it derives the unknown data probability distributions directly from the data by imposing constraints that are better suited for deep tissue imaging. This will result in the development of unique image analysis tools better suited to the unusual characteristics of deep-tissue fluorescence images. The proposed research is significant, because it will significantly enhance the quantitative capabilities of multi-photon microscopy. Public Health Relevance Statement: The proposed studies will address an under investigated area of deep tissue microscopy. The proposed research will enhance the ability to quantitatively analyze large microscopy image volumes, providing the final link in the effective implementation of multi-photon microscopy as a quantitative method in biomedical research.