Precision oncology relies on advancing tools for diagnosing and treating cancer patients, aiming for a more personalized medicine. A novel frontier in this pursuit is illustrated by Digital Pathology (DP), which facilitates the discovery of predictive biomarkers. Utilizing high-definition scanners, DP enables the acquisition and storage of histological slides, establishing precious resources for research. Another key aspect of this emerging field is the development of machine learning models, particularly those automating the identification of diverse cell types within the tumor microenvironment. However, to ensure a high level of accuracy in cell identification, these models depend heavily on the creation of extensive annotated datasets, which typically take a lot of time and expertise to generate.
This master’s thesis introduces an innovative digital pathology project called HE2mIF which employs a dual-staining approach combining Hematoxylin and Eosin (H&E) with Multiplex Immunofluorescence (mIF) on the same pathology slides. After staining the slides with H&E and scanning them, a discoloration phase is initiated, followed by re-staining with mIF and subsequent scanning. This approach generates Whole Slide Images (WSIs) that are then analyzed using an image processing platform called IFQuant, resulting in phenotype coordinates. Through an optimized pipeline, these phenotypes are directly superposed onto their H&E counterparts, creating a comprehensive annotated dataset. The primary objective of this dataset is to train predictive algorithms capable of accurately identifying melanoma cells, lymphocytes, and myeloid cells (macrophages and dendritic cells) within H&E whole slide images. HE2mIF represents a notable advancement in the field of digital pathology, contributing to the generation of training datasets for deep learning models with adaptability and applicability in pathology and clinical routine.