An Empirical Investigation into Domain Generalization in Computational Histopathology

Overview & Motivation

While deep learning models have achieved expert-level performance in computational pathology, their clinical translation is hindered by a critical lack of generalizability. Models trained on data from one source often fail dramatically when applied to unseen data from a different domain — a different hospital, scanner, or staining protocol.

This project frames the model as a scientific instrument to rigorously investigate and quantify this "domain shift" phenomenon. The core objective is not perfect generalization, but to measure its absence and analyze the reasons for failure — providing evidence-based insights for building robust AI tools for pathology.

The Problem — Domain Gap

A model can identify cancer nuclei in one hospital's dataset but may collapse on images from another hospital or organ type. Subtle variations in tissue preparation, staining protocols, and scanner hardware create this domain gap. Quantifying this gap rigorously is the first step toward developing reliable models for clinical use.

Approach

The thesis is structured as a two-phase experiment:

• Phase 1 — Baseline validation on a specialist in-domain dataset (CoNSeP), confirming model performance under controlled conditions.
• Phase 2 — Generalization test on a cross-domain dataset (MoNuSeg), measuring the performance drop on unseen tissue types and staining conditions.

Architectures investigated include HoVer-Net and U-Net. Evaluation metrics: Panoptic Quality (PQ) for segmentation, macro F1-score for classification. A Streamlit-based visualisation tool is also developed to inspect segmented histology images and phenotype data.

Key Technologies: Python · PyTorch · Docker · HoverNet · Streamlit · ROS2 · Webots · Histopathology (CoNSeP, MoNuSeg)

Expected Outcomes & Scientific Contribution

• Quantitative Benchmark: Statistically significant drop in performance expected on out-of-domain data, providing a concrete measure of the domain gap.
• Characterization: Identifying which failure modes are most prevalent — morphology variation, stain intensity shifts, scanner artifacts.
• Actionable Insights: Analysis that guides future domain adaptation and stain normalization strategies.
• Foundation for Future Work: Basis for domain adaptation research in clinical AI, reproducible experimental pipeline.