Pg-GAT: A Complete Graph Model for Cancer Detection and Subtyping in Whole Slide Images Analysis

Susu Hu, Nithya Bhasker, Guido Montufar, Jakob Nikolas Kather, Lena Seifert, Stefanie Speidel

Keywords: graph neural network, whole slide images, cancer research

TL;DR: We propose a lightweight spatial- and context-aware complete graph model for cancer detection and subtyping in whole slide images analysis, outperforming state-of-the-art methods.

Abstract:

Whole-Slide-Images (WSIs) have generated significant interests in cancer research community, owing to their availability and the rich information that they provide. Previous Multiple Instance Learning (MIL) methods often neglect the topological structure of tissues which is closely related to tumor evolution. Some attempts with transformer-based MIL methods take spatial relation into account with a trade-off of computational complexity. We propose Projection-gated Graph Attention Network (Pg-GAT), a lightweight model that effectively leverages graph neural network to provide structural prior, learns spatial and contextual relations through graph attention, and mitigates tissue morphology redundancy with differentiable projection-gated pooling, maintaining a data-adaptive decision boundary. In addition, Pg-GAT outputs region-of-interest (ROI) with respect to the graph-level prediction with post-hoc graph explainer, offering tumor localization and model interpretability. We evaluate our method on lymph node metastasis datasets (CAMELYON16 and CAMELYON17) and non-small cell lung cancer (TCGA-NSCLC), achieving AUCs of 97.6% and 95.6% and 99.6% respectively, outperforming state-of-the-art methods.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2384