SlideGraph+: Whole Slide Image Level Graphs to Predict HER2Status in Breast Cancer (2110.06042v1)

Abstract: Human epidermal growth factor receptor 2 (HER2) is an important prognostic and predictive factor which is overexpressed in 15-20% of breast cancer (BCa). The determination of its status is a key clinical decision making step for selection of treatment regimen and prognostication. HER2 status is evaluated using transcroptomics or immunohistochemistry (IHC) through situ hybridisation (ISH) which require additional costs and tissue burden in addition to analytical variabilities in terms of manual observational biases in scoring. In this study, we propose a novel graph neural network (GNN) based model (termed SlideGraph+) to predict HER2 status directly from whole-slide images of routine Haematoxylin and Eosin (H&E) slides. The network was trained and tested on slides from The Cancer Genome Atlas (TCGA) in addition to two independent test datasets. We demonstrate that the proposed model outperforms the state-of-the-art methods with area under the ROC curve (AUC) values > 0.75 on TCGA and 0.8 on independent test sets. Our experiments show that the proposed approach can be utilised for case triaging as well as pre-ordering diagnostic tests in a diagnostic setting. It can also be used for other weakly supervised prediction problems in computational pathology. The SlideGraph+ code is available at https://github.com/wenqi006/SlideGraph.

• The paper introduces SlideGraph+, a graph neural network framework that models whole slide images as graphs to predict HER2 status.
• It applies agglomerative clustering to capture tissue architecture, moving beyond traditional patch-level analysis to a holistic view.
• Evaluations on TCGA-BRCA and independent datasets show AUROC values over 0.75, highlighting its potential to streamline breast cancer diagnostics.

Overview of "SlideGraph$^{+}$: Whole Slide Image Level Graphs to Predict HER2 Status in Breast Cancer"

The paper presents a novel methodology utilizing graph neural networks (GNN) named SlideGraph$^{+}$, which aims to predict the HER2 status in breast cancer directly from whole slide images (WSIs) of routine Haematoxylin and Eosin (H&E) stained tissue sections. HER2 status is crucial for determining appropriate treatment strategies in breast cancer, yet current diagnostic methods, such as immunohistochemistry (IHC) and in situ hybridisation (ISH), are costly and subject to variability due to manual assessments. SlideGraph$^{+}$ offers a computational pathology approach that mitigates these limitations by analyzing the WSIs in a more holistic manner compared to patch-based methods.

Key Methodology

SlideGraph$^{+}$ models a WSI as a graph, where each node represents clusters of similar histological patterns, determined through an agglomerative clustering algorithm based on spatial and feature similarities of image patches. This process results in a representative graph structure capturing the tissue's cellular architecture. The graph is then processed by a GNN to predict HER2 status, overcoming the constraints of traditional pixel or patch-level analyses which fail to leverage the broader histological context.

The GNN employed in SlideGraph$^{+}$ adopts a message-passing framework, allowing the integration of node-level information to produce an aggregate prediction for the WSI. This method allows nuanced consideration of spatial relationships and morphological patterns across the entire slide.

Results and Comparisons

The performance of SlideGraph$^{+}$ was evaluated on the TCGA-BRCA dataset and validated on independent datasets (HER2C and an internal Nottingham University Hospital dataset). The model showed superior performance with AUROC values over 0.75, outperforming existing methods that rely on patch-level analysis. Particularly notable is the model's utilization of DAB density estimates—a novel predictive feature derived from paired H&E and corresponding IHC images—demonstrating significant predictive power.

Implications and Future Directions

SlideGraph$^{+}$ provides an innovative framework for HER2 status prediction, illustrating the power of integrating graph-based methodologies in digital pathology. Its ability to render WSI-level predictions efficiently may facilitate pathologists' workflow, possibly reducing dependency on more laborious and costly diagnostic tests like IHC and ISH. The potential to integrate similar techniques for other biomarkers or histological analyses makes SlideGraph$^{+}$ a versatile tool for computational pathology. Future advancements might further refine the GNN architecture, improve graph construction methodologies, or extend to other cancer types or medical imaging applications, thus broadening the utility within the field of digital pathology.