Cross-Domain Generalization of Deep Learning-Based Image Analysis Algorithms in Histopathology
Pathologists have long diagnosed cancer by visually assessing tissue samples under a microscope. The rise of digital slide scanners changed that, enabling not just digital archiving and remote consultation, but also the use of machine learning for computer-aided diagnosis, with many systems now matching trained pathologists’ performance. But convolutional neural networks can be surprisingly sensitive to changes in how images look, and in pathology these “domain shifts” happen easily: different tissue morphologies, or samples prepared at a different lab, each following their own routine of fixation, dehydration, embedding, staining, and scanning, which can give the resulting images a distinctly different visual signature from lab to lab.
This thesis studies how different kinds of domain shift affect neural network performance across three real diagnostic tasks: detecting mitotic figures across different scanners, segmenting tumors across different domains, and detecting T-lymphocytes across different tumor types in immunohistochemistry samples. For the cross-scanner detection task, domain adversarial training successfully extracted features that were largely independent of which scanner produced the image. For cross-domain tumor segmentation, the thesis explores self-supervised learning as a pre-training strategy to align feature representations across domains, a technique mostly used until now on natural images rather than histopathology, and the results offer new insight into how well it transfers to this setting. For the lymphocyte detection task, an initial model’s performance dropped sharply on samples from tumor types it hadn’t seen before, but fine-tuning on just a small number of samples from the new domain largely fixed this, pointing toward practical recommendations for building more robust detection algorithms across tumor types.
Alongside these experiments, the thesis also produced several new cross-domain datasets, each targeting a different source of domain shift. One of these, a fully annotated set of 350 whole slide images covering seven canine skin tumor subtypes, is now one of the most comprehensive open histopathology segmentation datasets available. Each dataset’s annotation quality was checked through extensive multi-rater experiments, and making them publicly available should help future work on cross-domain generalization in histopathology.