The segmentation of malignant breast tissue from histological images represents a crucial task for the diagnosis of breast cancer (BC). This is a time-consuming process that could be alleviated with the help of computerized segmentation methods, leading to elevated precision and reproducibility results. However, this automated segmentation poses a challenge due to the large size of histological whole-slide images and the significant variability, heterogeneity and complexity of features in them.

In this research, we propose a processing pipeline for the automatic segmentation of stained BC images presenting different types of histopathological patterns. To deal with the gigantic size of whole-slide images, the digital preparations were processed in a tile-wise manner: a large part of the image is split into patches. Then, the segmentation of each tile was accomplished by applying a deep convolutional neural network (DCNN) along with an encoder-decoder with separable atrous convolution architecture, which, once successfully validated, has revealed to be a promising method to segment pathological image patches. Next, in order to combine the local segmentation results (segmented tiles), while avoiding discontinuities and inconsistencies, an improved merging strategy based on an efficient fully connected Conditional Random Field (CRF) was applied.

Experimental results on a collection of patches of breast cancer images demonstrate how the designed processing pipeline performs properly regardless the size, texture or any other colour-shape features typical of the malignant carcinomas considered in this study. The estimated segmentation accuracy and frequency weighted intersection over union (FWIoU) were 95.62%, 92.52%, respectively. Additionally, in order to facilitate the collaboration between pathologists and researchers to extract the specialist knowledge in form of training datasets that allows the training of new algorithms, a web-based platform which includes a slide-viewer and an annotation tool was developed. The automatic segmentation method proposed in this work was integrated into this platform and currently, it is being used as a decision support tool by pathologists.

从组织学图像中对恶性乳腺组织进行分割代表了诊断乳腺癌（BC）的关键任务。这是一个耗时的过程，可以借助计算机化的分割方法来缓解，从而提高了精度和重现性。然而，由于组织学全幻灯片图像的大尺寸及其特征的显着可变性，异质性和复杂性，这种自动分割带来了挑战。

在这项研究中，我们提出了一种用于自动分割呈现不同类型组织病理学模式的染色BC图像的处理管线。为了处理整个幻灯片图像的巨大尺寸，对数字准备进行了平铺处理：将图像的很大一部分分成小块。然后，通过应用深度卷积神经网络（DCNN）以及具有可分离的无规卷积架构的编码器/解码器来完成每个图块的分割，一旦成功验证，该方法已被证明是分割病理图像斑块的一种有前途的方法。接下来，为了合并局部分割结果（分段的图块），同时避免出现不连续和不一致的情况，

一系列乳腺癌图像的实验结果表明，无论本研究中考虑的恶性肿瘤的大小，质地或任何其他颜色形状特征如何，设计的处理管道如何正常运行。估计的分割精度和频率加权并集（FWIoU）分别为95.62％和92.52％。此外，为了促进病理学家和研究人员之间的协作，以训​​练数据集的形式提取专业知识，从而可以训练新算法，开发了一个基于Web的平台，其中包括幻灯片查看器和注释工具。这项工作中提出的自动分割方法已集成到该平台中，目前，病理学家正在将其用作决策支持工具。