Advancing Medical Image Analysis in Digital Pathology: Diffusion Models and Efficient Annotation for HER2 Analysis in Breast Cancer

Medical image analysis plays an important role in modern healthcare, supporting precise diagnoses, guiding therapeutic decisions, and improving patient outcomes. Despite advances in deep learning, the development of reliable algorithms remains constrained by the scarcity of high-quality, annotated data, leading to challenges such as non-representative datasets and single-rater annotations. These challenges are particularly evident in digital pathology, where whole-slide imaging of stained tissue enables quantitative analyses but introduces substantial complexity. A clinically important application is the evaluation of Human Epidermal Growth Factor Receptor 2 (HER2) expression in breast cancer, where accurate scoring of invasive tumor tissue is critical for therapy selection. Manual assessment is time-consuming and subject to inter-rater variability, motivating the development of algorithmic methods that are both efficient and interpretable. This thesis advances medical image analysis with efficient annotation strategies, an interpretable HER2 scoring pipeline, two data-centric diffusion model approaches, and a task-specific adaptation of diffusion models for segmentation – using HER2 tissue analysis as a shared use case. First, annotation effort is reduced through superpixel pre-segmentations, which provide editable region proposals aligned with pathologist annotations and facilitates the creation of larger, more representative datasets. Second, a segmentation-driven HER2 scoring pipeline is developed that directly implements clinical guidelines, distinguishes invasive from non-invasive tumor tissue, and aggregates subtype predictions into interpretable patient-level scores with calibrated adjustments. Third, the challenge of subtype imbalance in tumor segmentation is addressed by employing generative networks, in particular diffusion models, to synthesize subtype-balanced training datasets. This improves segmentation accuracy and reduces variance across HER2 subtypes without requiring additional manual labels. Fourth, Style-Extracting Diffusion Models (STEDMs) are introduced to leverage unannotated whole-slide images. By extracting style cues from unlabeled data and combining them with known content masks, STEDMs synthesize training images that enable semi-supervised segmentation, particularly beneficial in low-data regimes. Finally, to improve probability calibration under single-rater supervision, this work investigates the behavior of diffusion models with segmentation targets and introduces Weighted Noise-Range Schedules. These task-specific schedules enhance both segmentation accuracy and probability calibration, enabling more reliable probabilistic modeling from single-rater data. Together, these contributions mitigate annotation effort, address subtype imbalance, integrate unannotated data, and adapt diffusion segmentation to single-rater supervision. The proposed methods not only establish practical pathways for developing interpretable and robust HER2 analysis pipelines but are also broadly applicable to other challenges in medical image analysis.