Wilm F (2025)
Publication Language: English
Publication Type: Thesis
Publication year: 2025
URI: https://open.fau.de/handle/openfau/35170
The visual assessment of microscopic samples by pathologists constitutes an essential component of cancer diagnostics. Traditional pathology workflows were based on the visual assessment of samples under the microscope. The development of designated slide scanners has facilitated the digitization of microscopy samples which not only allowed for digital archiving and remote expert consultancy but also facilitated the use of machine learning-based image analysis algorithms for computer-aided diagnosis. Meanwhile, a wide range of computer-aided systems has been developed in the field of histopathology, often matching the performance of trained pathologists. Previous work has shown that machine learning-based image analysis algorithms, and especially convolutional neural networks, can be very susceptible to changes in the visual appearance of images. In pathology, these \emph{domain shifts} can be caused when applying trained algorithms to different morphologies, or samples prepared at a different pathology lab. The preparation of histologic samples follows routine stages, including tissue fixation, dehydration, paraffin embedding, and microtome sectioning. Subsequently, a sample is usually stained with a specific dye and digitized with a designated scanning system. The visual manifestation of these sample preparation steps can be very unique for the respective pathology lab. This thesis investigates the impact of different domain shifts on the performance of convolutional neural networks in histopathology. For these experiments, three routine tasks in cancer diagnostics were considered: cross-scanner mitotic figure detection, cross-domain tumor segmentation, and pan-tumor T-lymphocyte detection on immunohistochemistry samples. For the task of cross-scanner mitotic figure detection, domain adversarial training was employed. Evaluations of the learned embeddings demonstrated the successful extraction of scanner-agnostic features. For the task of cross-domain tumor segmentation, representation learning and, in particular, self-supervised learning was explored as a pre-training strategy to align feature embeddings across domains and thereby enhance the domain agnosticity for the downstream task. The results provide insights into the applicability of self-supervised learning in the context of histopathology. To date, this technique has mostly been employed in the field of natural images. In a project addressing the detection of tumor-infiltrating lymphocytes in immunohistochemistry samples, fine-tuning was leveraged to bridge the domain gap between different tumor indications. Initial experiments exhibited degraded performance on out-of-distribution samples. By exploiting fine-tuning on a limited number of target domain samples, this degradation was effectively mitigated. The experiments allowed for recommendations on the development of robust algorithms for the detection of lymphocytes across different tumor morphologies. In the course of the thesis, several cross-domain datasets were curated, focusing on different sources of domain shift. This includes a fully annotated dataset of 350 whole slide images covering seven canine cutaneous tumor subtypes, which constitutes one of the most comprehensive open histopathology segmentation datasets to date. A high annotation quality of each published dataset was ensured through extensive multi-rater experiments on selected subsets of the data. By making these datasets publicly available, future work on the cross-domain generalization for histopathology was facilitated.
APA:
Wilm, F. (2025). Cross-Domain Generalization of Deep Learning-Based Image Analysis Algorithms in Histopathology (Dissertation).
MLA:
Wilm, Frauke. Cross-Domain Generalization of Deep Learning-Based Image Analysis Algorithms in Histopathology. Dissertation, 2025.
BibTeX: Download