| Abstract [eng] |
To better serve cancer patients, diagnostic and digital pathology methods focus on more novel targets. One of such targets is the tumor microenvironment. Machine vision-dependent digital pathology methods are still very tumor cell-centric and largely ignore the tumor microenvironment. This work has set the aim to investigate and propose new histopathology image segmentation and classification methods by targeting tumor microenvironment-related histologic tissue components. Firstly, convolutional neural networks were identified as a group of state-of-the-art methods of sufficient capacity to handle multiple histologic object segmentation. Then, the existing tumor cell segmentation method was adapted and extended for lymphocyte segmentation and identification. Next, fibrous collagen was identified as a novel tumor microenvironment-borne target for segmentation in bright-field images of tumorous tissue. To address the collagen fiber segmentation task, a fully convolutional neural network-based approach was developed. Finally, an approach integrating knowledge gained in previous experiments was proposed enabling segmentation of lymphocytes, tumor cell nuclei, stromal cell nuclei, collagen fibers, and major tissue compartments. Additionally, by the engineering of image features, a whole slide image transformation was introduced, enabling the prediction of therapeutic biomarker status for individual breast cancer patients from complete tumor tissue whole-slide images. The proposed methods were comparable to the state-of-the-art methods while at the same time providing special additional features. |
