In digital histopathological image analysis, two conflicting objectives are often pursued: closeness to the original tissue and high classification performance. The former objective tries to recover images (stains) that are as close as possible to the ones obtained by staining the tissue with a single dye. The latter objective requires images that allow the extraction of better features for an improved classification, even if their appearance is not close to single stained tissues. In this paper we propose a framework that achieves both objectives depending on the number of stains used to mathematically decompose the scanned image. The proposed framework uses a total variation prior for each stain together with the similarity to a given reference color-vector matrix. Variational inference and an evidence lower bound are utilized to automatically estimate all the latent variables and model parameters. The proposed methodology is tested on real images and compared to classical and state-of-the-art methods for histopathological blind image color deconvolution and prostate cancer classification.

在数字组织病理学图像分析中，经常追求两个相互矛盾的目标：与原始组织的接近性和高分类性能。前一个目标试图恢复图像（污点），该图像尽可能接近用单一染料对组织进行染色获得的图像（污点）。后一个目的要求图像能够提取更好的特征以改善分类，即使它们的外观与单个染色的组织不相近。在本文中，我们提出了一个框架，该框架可以实现两个目标，具体取决于用于数学分解扫描图像的污点数量。所提出的框架使用每个变体的先验总变化以及与给定参考颜色矢量矩阵的相似性。利用变分推断和证据下界自动估计所有潜在变量和模型参数。所提出的方法在真实图像上进行了测试，并与用于组织病理学盲图像颜色反卷积和前列腺癌分类的经典方法和最新方法进行了比较。