Histopathology is a cornerstone of disease diagnosis and treatment, traditionally relying on manually assessing tissue specimens under a microscope. However, the advent of slide scanners to produce digital tissue representations, so-called whole-slide images (WSI), has enabled computational pathology to perform quantitative and automated tissue analysis. Current developments in Artificial Intelligence, particularly Deep Learning, have accelerated the progress in this field. This thesis proposes a comprehensive Deep Learning pipeline for quantitative histopathological image analysis, integrating WSI preprocessing, algorithm development for tissue and cell-level segmentation, and clinical application in an end-to-end workflow. The approach not only improves the quantitative evaluation of WSI but also extracts diagnostic and prognostic markers while automatically characterizing tissue dynamics through morphological tissue features. Segmenting entire tissue sections into classes like tumorous or non-tumorous requires the consideration of global tissue patterns as well as local cell morphologies. Following this, we introduce the Memory Attention Framework that can be incorporated into any encoder-decoder segmentation architecture. This framework enables the adaptive incorporation of tissue context during fine-grained local segmentation. The method was evaluated on two public datasets (breast, liver) and an internal kidney cancer dataset, demonstrating superiority over non-context and multiscale segmentation approaches. Notably, the approach reduced the number of false-positive tumor regions. Building on this, we applied the framework to a pancreatic cancer cohort consisting of 400 internal and 182 external patients to quantify the tumor microenvironment and correlate it with patient outcomes. In doing so, we were able to stratify patients into two risk groups based on tissue composition and spatial tumor-stroma distribution, which showed significant (p < 0.05) differences in their survival probabilities. Next to tissue analysis, segmentation on the cellular level is crucial to uncover the cellular composition of tissue samples. While convolutional neural networks have been extensively used for this task, we evaluate the capabilities of Transformer-based networks and incorporate so-called foundation models to improve accuracy compared to existing solutions. The proposed CellViT and CellViT++ models have proven to achieve State-of-the-Art results on several benchmark datasets, covering a broad spectrum of tissue types and cell classes, bringing cell segmentation solutions closer to clinical practice. The models require minimal data for fine-tuning and exhibit remarkable zero-shot cell segmentation quality. This capability allows for a considerably faster adaptation to new research hypotheses without the need for extensive development time. In summary, this work presents Deep Learning techniques for quantifying tissue at both the macro and micro levels, enhancing diagnostic workflows, and identifying prognostic markers.
Fabian Horst (Wed,) studied this question.