Interconnected Pathology: A Graph-Based Framework for Analyzing Multimodal Histopathology Samples
Diagnosing cancer from tissue samples still relies on pathologists visually inspecting stained slides under time and workload pressure, interpreting complex patterns of cell shape, spatial arrangement, and staining. Most computational approaches process these images as independent small patches, missing the long-range spatial structure, mixed cell populations, and artifacts that characterize whole-slide images. This thesis instead represents tissue as a graph of interacting cells and regions, using graph neural networks to model that structure directly, aiming for models that better reflect how cancer and its surrounding microenvironment are actually organized.
The first contribution builds a graph-based framework for characterizing melanoma cells using multiplexed MELC imaging data, representing individual cells as graph nodes connected by spatial proximity or feature similarity, with high-dimensional protein expression profiles compressed down to serve as node features. On a dataset of suspected melanoma cases, graph neural networks using spatial neighborhood information classified melanoma cells more accurately than standard tabular baselines that ignore spatial context, improving accuracy and weighted F1 score by roughly 4 to 8 percentage points, showing that explicitly modeling local tissue structure adds real diagnostic value.
The second contribution scales this up to a multi-level graph spanning both patient-level and cell-level structure across five different cancer types: melanoma, non-small cell lung cancer, triple-negative breast cancer, head and neck squamous cell carcinoma, and gastric cancer. Whole-slide immunohistochemistry images are converted into heterogeneous graphs with disease, patient, and cell nodes, using pathomic and radiomic descriptors as node features, and the resulting graph neural networks handle both patient-level disease classification and cell-level prediction. This shared representation consistently outperformed tabular and patch-based baselines by about 3 to 7 percentage points across tasks, suggesting that linking patients and cells across different diseases lets the model pick up on patterns that get lost when each slide is analyzed in isolation.
The third contribution asks a more practical question: which cell-level features actually work best for graph-based models under real clinical constraints. Two studies compare high-dimensional multi-stain or gene expression profiles against simpler radiomic features extracted from the same images, first in melanoma and then in a broader dermatological setting covering eczema, lymphoma, and healthy controls. Radiomic features matched, and in several cases exceeded, the performance of the more complex molecular descriptors, improving accuracy and weighted F1 score by up to 5 to 10 percentage points, while being far easier to compute and better suited to large-scale deployment. This suggests radiomics offers a practical, informative way to describe cell appearance and texture for graph-based digital pathology.
Across all three contributions, spatial context, graph structure, and feature choice consistently emerge as the key factors driving better classification performance and more biologically sensible models. The thesis points toward several directions for future work, including adaptive graph construction, integrating molecular and clinical variables, validating across institutions, and improving interpretability, all aimed at building reliable graph-based decision support for precision oncology.