Insights from GNN and Cellular Explanations: A Multistage Framework to Interpret Automatic Diagnosis in Histopathology

Catarina Palmeirão, Rita Pereira, Maria Rita Fonseca Verdelho, Catarina Barata, Carlos Santiago

Keywords: Whole Slide Imaging, Digital Pathology, Graph Neural Networks, Explainable Artificial Intelligence, GNNExplainer, HoVer-Net, Multiscale Interpretability

Abstract: Whole Slide Imaging has transformed digital pathology by capturing tissue architecture at cellular resolution, yet its gigapixel scale and complex spatial organization challenge automated analysis. Traditional deep learning methods often overlook these spatial dependencies, limiting their diagnostic reliability. Graph-based learning addresses this limitation by representing tissue as interconnected cellular and structural entities, preserving spatial and morphological context essential for accurate cancer diagnosis. This paper investigates how Graph Neural Networks and Explainable Artificial Intelligence can jointly enhance the performance and interpretability of histopathological diagnosis. By modeling both cellular and tissue level relationships, GNN capture biologically meaningful patterns, while methods such as GNNExplainer reveal the rationale behind predictions. The integration with HoVer-Net further enables multiscale interpretability, reflecting the hierarchical reasoning process of pathologists. Extensive experiments show that attention-based GNN architectures outperform standard graph convolutional models while remaining efficient and interpretable. Beyond accuracy, the results demonstrates that graph-based learning combined with XAI provides a robust, biologically grounded foundation for reliable diagnostic systems in computational pathology.

Primary Subject Area: Interpretability and Explainable AI

Secondary Subject Area: Application: Histopathology

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 237