Advancing Medical Image Analysis in Digital Pathology: Diffusion Models and Efficient Annotation for HER2 Analysis in Breast Cancer
Medical image analysis increasingly supports precise diagnosis, treatment decisions, and better patient outcomes, but building reliable deep learning algorithms is still held back by a shortage of high-quality annotated data, often leading to non-representative datasets and annotations from only a single rater. This is especially true in digital pathology, where whole-slide imaging enables detailed quantitative analysis but adds a lot of complexity. One clinically important example is scoring Human Epidermal Growth Factor Receptor 2 (HER2) expression in breast cancer, where accurately assessing invasive tumor tissue directly determines therapy choice. Manual scoring is slow and varies between raters, motivating this thesis’s push toward efficient, interpretable algorithmic alternatives, using HER2 tissue analysis as a running example throughout.
The thesis makes four contributions. First, it reduces annotation effort using superpixel pre-segmentation, which generates editable region proposals that align well with pathologists’ own annotations, making it easier to build larger, more representative datasets. Second, it develops a segmentation-driven HER2 scoring pipeline that implements clinical guidelines directly, separates invasive from non-invasive tumor tissue, and combines subtype predictions into interpretable, calibrated patient-level scores. Third, it tackles subtype imbalance in tumor segmentation by using diffusion models to generate subtype-balanced synthetic training data, improving segmentation accuracy and reducing variance across HER2 subtypes without any extra manual labeling. Fourth, it introduces Style-Extracting Diffusion Models (STEDMs), which pull style information from unannotated whole-slide images and combine it with known content masks to synthesize training images for semi-supervised segmentation, particularly useful when labeled data is scarce.
Finally, to improve how well predicted probabilities reflect reality when training on single-rater annotations, the thesis studies how diffusion models behave on segmentation tasks and introduces Weighted Noise-Range Schedules, a task-specific adjustment that improves both segmentation accuracy and probability calibration, enabling more trustworthy probabilistic predictions even from single-rater data.
Together, these four contributions cut down annotation effort, correct for subtype imbalance, make use of unannotated data, and adapt diffusion-based segmentation to work well with single-rater supervision. Beyond HER2 analysis specifically, the methods developed here offer a practical foundation for building interpretable, robust pipelines that should generalize to other problems in medical image analysis.