Neural Graph Modelling of Whole Slide Images for Survival Ranking

Callum Christopher Mackenzie, Muhammad Dawood, Simon Graham, Mark Eastwood, Fayyaz ul Amir Afsar Minhas
Proceedings of the First Learning on Graphs Conference, PMLR 198:48:1-48:10, 2022.

Abstract

Evaluation of a cancer patient’s prognostic outlook is an essential step in the clinical decision-making process, involving the assessment of complex tissue structures in multi-gigapixel whole slide images (WSIs). Effective risk stratification of patients from WSIs has proven challenging despite several approaches across the literature due to their large size and inability of existing approaches to effectively model inter-relationships between different tissue components. We propose a graph neural network (GNN) model that performs pairwise ranking of graph representations of WSIs based on survival scores. The proposed approach translates spatially-localised deep features along with their spatial context to a graph neural network to produce survival scores. Analysis over breast cancer patients from The Cancer Genome Atlas (TCGA) shows that the proposed GNN approach is able to rank patients with respect to their disease-specific survival times with a concordance index of \textdollar 0.672 \pm 0.058\textdollar . This is a significant improvement over existing state of the art and paves the way for neural graph modelling of WSI data for survival prediction for other cancer types.

Cite this Paper

BibTeX
@InProceedings{pmlr-v198-mackenzie22a, title = {Neural Graph Modelling of Whole Slide Images for Survival Ranking}, author = {Mackenzie, Callum Christopher and Dawood, Muhammad and Graham, Simon and Eastwood, Mark and ul Amir Afsar Minhas, Fayyaz}, booktitle = {Proceedings of the First Learning on Graphs Conference}, pages = {48:1--48:10}, year = {2022}, editor = {Rieck, Bastian and Pascanu, Razvan}, volume = {198}, series = {Proceedings of Machine Learning Research}, month = {09--12 Dec}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v198/mackenzie22a/mackenzie22a.pdf}, url = {https://proceedings.mlr.press/v198/mackenzie22a.html}, abstract = {Evaluation of a cancer patient’s prognostic outlook is an essential step in the clinical decision-making process, involving the assessment of complex tissue structures in multi-gigapixel whole slide images (WSIs). Effective risk stratification of patients from WSIs has proven challenging despite several approaches across the literature due to their large size and inability of existing approaches to effectively model inter-relationships between different tissue components. We propose a graph neural network (GNN) model that performs pairwise ranking of graph representations of WSIs based on survival scores. The proposed approach translates spatially-localised deep features along with their spatial context to a graph neural network to produce survival scores. Analysis over breast cancer patients from The Cancer Genome Atlas (TCGA) shows that the proposed GNN approach is able to rank patients with respect to their disease-specific survival times with a concordance index of \textdollar 0.672 \pm 0.058\textdollar . This is a significant improvement over existing state of the art and paves the way for neural graph modelling of WSI data for survival prediction for other cancer types. } }
Endnote
%0 Conference Paper %T Neural Graph Modelling of Whole Slide Images for Survival Ranking %A Callum Christopher Mackenzie %A Muhammad Dawood %A Simon Graham %A Mark Eastwood %A Fayyaz ul Amir Afsar Minhas %B Proceedings of the First Learning on Graphs Conference %C Proceedings of Machine Learning Research %D 2022 %E Bastian Rieck %E Razvan Pascanu %F pmlr-v198-mackenzie22a %I PMLR %P 48:1--48:10 %U https://proceedings.mlr.press/v198/mackenzie22a.html %V 198 %X Evaluation of a cancer patient’s prognostic outlook is an essential step in the clinical decision-making process, involving the assessment of complex tissue structures in multi-gigapixel whole slide images (WSIs). Effective risk stratification of patients from WSIs has proven challenging despite several approaches across the literature due to their large size and inability of existing approaches to effectively model inter-relationships between different tissue components. We propose a graph neural network (GNN) model that performs pairwise ranking of graph representations of WSIs based on survival scores. The proposed approach translates spatially-localised deep features along with their spatial context to a graph neural network to produce survival scores. Analysis over breast cancer patients from The Cancer Genome Atlas (TCGA) shows that the proposed GNN approach is able to rank patients with respect to their disease-specific survival times with a concordance index of \textdollar 0.672 \pm 0.058\textdollar . This is a significant improvement over existing state of the art and paves the way for neural graph modelling of WSI data for survival prediction for other cancer types.
APA
Mackenzie, C.C., Dawood, M., Graham, S., Eastwood, M. & ul Amir Afsar Minhas, F.. (2022). Neural Graph Modelling of Whole Slide Images for Survival Ranking. Proceedings of the First Learning on Graphs Conference, in Proceedings of Machine Learning Research 198:48:1-48:10 Available from https://proceedings.mlr.press/v198/mackenzie22a.html.

Related Material