Digitale Histopathologie

RWTH Publications (RWTH Aachen)

Abstract

Digital pathology is an emerging field which requires dedicated algorithms for the automated analysis of digital histopathological images as presented in this thesis. Histopathological workflows analyze biomedical tissue samples on macroscopic and microscopic level, which requires e.g. cell nuclei counting, morphological inspection of cells and cell nuclei, or the analysis of the tissue composition. These analysis tasks are necessary for both the diagnosis of cancer and research related to a better understanding of tumor characteristics, growth and immune system interaction. Previously applied analog histopathological workflows may suffer from a subjective rating resulting from different levels of experience of the pathologist as well as fatigue (of the visual system) during the analysis of large amounts of brightfield microscopy images. Facilitated by the availability of whole-slide scanning systems, digitized microscopy images can easily be obtained and utilized to automate tasks. Current challenges arising from whole-slide images (WSI) are the massive data size which are usually in the range of gigapixels, color variations due to the staining procedure, and the inherently strong biological variance of biomedical tissue. The methods presented in this work address these challenges in order to provide scalable and reliable solutions for an algorithmic analysis of tissues and immune responses. Specifically, this thesis addresses the questions how to handle the data and file sizes, how to appropriately deal with inherent sources of variation, how to achieve fast and reproducible analysis results, and how to communicate the reliability of the algorithmic results to the medical expert. Regarding the massive data file sizes, this thesis presents an adaptation of a new compression scheme that achieves superior compression ratios compared to the standard used in current WSI file formats. For accelerating the processing, a fast and reliable foreground segmentation is introduced, which prevents brightfield microscopic background patches from entering computationally complex processing steps. Moreover, a stain normalization algorithm is developed that incorporates deep features from a pretrained neural network, which results in excellent normalization performance. However, the simultaneous development of stain augmentation raises the question if normalization as a preprocessing step is still required in conjunction with deep learning. As this work heavily builds on modern deep learning algorithms, challenges regarding the amount of labeled data needed for training have to be addressed. To this end, a method for texture augmentation via image quilting is introduced that performs an unsupervised realistic rearrangement of image patches and ground truth labels. It is shown in a classification scenario that this augmentation can lead to an improvement of tissue classification accuracy for underrepresented classes. Further contributions are workflows for tissue classification and cell nuclei detection. In both cases classical approaches and deep learning methods are considered and evaluated. At the price of high training times and expensive hardware, the deep learning approaches clearly outperform classical methods. On large and diverse tissue datasets, the prediction of tissue maps with deep learning achieves an F1-score of 84% exceeding the classical baseline method by nearly 28%. Besides a superior performance, the presented deep learning architecture offers probabilistic predictions via runtime dropout that can be utilized as a quality control for the result. This partially alleviates the black-box decision making that neural networks are often criticized for, as it improves the indication of reliable and unreliable predictions. In the cell nuclei detection and localization task, F1-scores of up to 88% are achieved, with a less pronounced, but still notable, difference of 8% in favor of the deep learning approaches. In this task, the main contribution is a novel augmentation that enables the transformation of a very basic ellipse model for cell nuclei simulation into realistic cell nuclei images. Herein, the challenge to preserve the classes “standard cell nuclei” and “immune cell nuclei” is crucial to train networks for immune response quantification. This is achieved by a novel loss function which provides a trade-off between enforcing consistency of the transform and the freedom to adapt the nucleus model locally. Furthermore, a broader view on cell nuclei localization is provided, where an attempt to generalize the concept to other microscopy modalities and stains is made by defining network layers and architectures that are capable of adapting to the respective modality or WSI stain. However, it is shown that these additional capabilities do not further improve the results of cell nuclei analysis in histopathology, while tasks in other image processing domains benefit from the approach. Together with the considerations regarding stain normalization and augmentation, this indicates a necessary shift in the way tasks have to be approached when working with deep neural networks. Finally, this thesis incorporates the different methods into a digital pathology workflow and exemplifies their intended use in a single-mouse-trial study. This use case defines a potential application and showcases a setup comprising preprocessing, probabilistic classification, visualization and quality assurance, the definition of meta features and methods for the graphical and quantitative assessment of tumor characteristics. This finally illustrates the benefits of the proposed approaches in a real-world preclinical research application.

Authors 1

1. Daniel Bug corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen