Multi-stain cancer detection in histological whole-slide-images of breast cancer resection specimen from female primary breast cancer patients

Sammanfattning

Breast cancer continues to be a major cause of mortality among women. In recent years, machine learning has emerged as a potential tool in detecting and grading cancer. Using machine learning techniques in computational pathology has the potential to improve precision medicine, enabling more personalized and more accurate treatment plans. The machine learning models can even detect structures that cannot be seen with human eyes. The first step is often to identify tissue areas with cancerous cells using machine learning models. Those models often rely solely on Haematoxylin and Eosin slides for training due to the time-consuming and costly nature of annotations by pathologists. Because of that, valuable information for training might be lost since some cancerous cells are more visible in the immunohistochemistry slides. In this thesis, Haematoxylin and Eosin slide annotations are registered to immunohistochemistry slides for training singlestain and multi-stain models. The registration of the annotations is not straightforward since the tissue of the slides is not necessarily from consecutive cuts, and they are sometimes applied to the slide at different angles. An algorithm evaluated during the ACROBAT challenge was used to register the slides. Using the transferred annotations, individual models are trained for each stain (K167, HER2, PGR, ER). Of the single-stain model, the HER2 stain model is showing the most promising results. As a second step, a multistain model is trained using all stains. The multi-stain model performs equally well as the single-stain models specializing in individual stains. This shows that there is no need to train specialized single-stain models. Thus being able to train one model for four different stains makes it possible to detect cancer in whole slide images stained with one of those four stains without the need to train a specialized model and only needing annotations in one stain. While the multi-stain model is a nice addition this thesis shows that it is possible to reuse annotations, which reduces the amount of manual labour from pathologists and allows for training models on immunohistochemistry slides with only having annotations from one stain.