Graph Perceiver Network for Lung Tumor and Bronchial Premalignant Lesion Stratification from Histopathology

Abstract

Bronchial premalignant lesions (PMLs) precede the development of invasive lung squamous cell carcinoma (LUSC), posing a significant challenge in distinguishing those likely to advance to LUSC from those that might regress without intervention. This study followed a novel computational approach, the Graph Perceiver Network, leveraging hematoxylin and eosin-stained whole slide images to stratify endobronchial biopsies of PMLs across a spectrum from normal to tumor lung tissues. The Graph Perceiver Network outperformed existing frameworks in classification accuracy predicting LUSC, lung adenocarcinoma, and nontumor lung tissue on The Cancer Genome Atlas and Clinical Proteomic Tumor Analysis Consortium datasets containing lung resection tissues while efficiently generating pathologist-aligned, class-specific heatmaps. The network was further tested using endobronchial biopsies from two data cohorts, containing normal to carcinoma in situ histology. It demonstrated a unique capability to differentiate carcinoma in situ lung squamous PMLs based on their progression status to invasive carcinoma. The network may have utility in stratifying PMLs for chemoprevention trials or more aggressive follow-up.

Figure 1

Overview of the study. A: Digitized hematoxylin and eosin–stained whole slide images were obtained from four different datasets, including lung resection tissues [The Cancer Genome Atlas (TCGA) and Clinical Proteomic Tumor Analysis Consortium (CPTAC) datasets] with nontumor tissue adjacent to the tumor (Normal), lung adenocarcinoma (LUAD), and lung squamous cell carcinoma (LUSC) and endobronchial biopsies [University College London (UCL) and Roswell Park Comprehensive Cancer Institute (Roswell) datasets] ranging from normal to carcinoma in situ (CIS) histology. The model was trained on TCGA data and tested across TCGA, CPTAC, UCL, and Roswell datasets. B: Graph Perceiver Network overview trained to differentiate lung cancer subtypes. C: Overview of analysis conducted using the results of the model that included evaluation of model performance, calculation of clustering-based metrics using whole slide image features, and model explainability using class-specific heatmaps. Panel A was generated with BioRender.com (Toronto, ON, Canada).

Figure 2

Graph Perceiver Network (GRAPE-Net) architecture. Each whole slide image (WSI) is represented as an undirected, unweighted tissue graph, where each node is an embedding of features in an image patch. Our proposed GRAPE-Net aggregates neighborhood information via graph convolutions while preserving spatial context. To efficiently find the neighborhood interaction between the different tissue regions in the tumor microenvironment, the graph is clustered to $C$ overlapping sets by pooling the nodes using multihead cross-attention pooling (PMA) block. The fixed set embeddings are then given as input to the self-attention block (SAB), which learns morphologic interactions with its neighborhood and relevance of each set toward the specific output label of interest. The aggregated attentions are then given as input to a multilayer perceptron (MLP) for classification of the WSI as normal, lung adenocarcinoma (LUAD), or lung squamous cell carcinoma (LUSC). Here, the layer-specific feature representation of the graph is denoted as H. A classification token (CLS) is added to serve as the entire WSI representation, which is used for classification and computing the relevance of each graph node toward the prediction.

Figure 3

Classification performance. Receiver operating characteristic (ROC) and precision-recall (PR) curves showcasing the performance of Graph Perceiver Network (GRAPE-Net) toward multiclass classification on Clinical Proteomic Tumor Analysis Consortium (CPTAC) external testing dataset. The mean ± SD area under the curve (AUC) score for each label [normal, lung adenocarcinoma (LUAD), and lung squamous cell carcinoma (LUSC)] is provided for the ROC and PR curves.

Figure 4

Graph Perceiver Network identifies tissue regions that correspond with pathologic annotations. A: The network identified regions annotated as tumor and normal on lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) whole slide images (WSIs) from the Clinical Proteomic Tumor Analysis Consortium (CPTAC). CTPAC images were acquired at ×20 magnification, the LUAD image is 23.6 × 17.9 mm, and the LUSC image is 24.6 × 21.1 mm. Representative patches are 126.5 × 126.5 μm. B and C: The network was also used to classify premalignant lesions (PMLs) as one of the three labels. The LUSC-specific heat maps for PMLs predicted to be LUSC identified regions of dysplasia and carcinoma in situ (CIS) as well as other regions likely contributing to its ability to separate CIS progressors from CIS regressors. B: The University College London images were acquired at ×40 magnification, where the progressive image is 8.6 × 5.6 μm, the regressive image is 4.7 × 3.1 μm, and the representative patches are 58.2 × 58.2 μm. C: The Roswell Park Comprehensive Cancer Institute images were acquired at ×20 magnification, where the top image is 8.0 × 7.1 μm, the bottom image is 5.0 × 4.8 μm, and the representative patches are 128.8 × 128.8 μm. Z-score color bar represents the contribution of the patches towards the final prediction, with scores close to 1 providing high contributions while scores close to –1 are not relevant to the prediction.

Figure 5

Stratification of lung tumors and premalignant lesions (PMLs) relates to histologic features and outcome. A: Uniform manifold approximation and projection (UMAP) plots of Clinical Proteomic Tumor Analysis Consortium, University College London, and Roswell Park Comprehensive Cancer Institute whole slide image (WSI) features, stratified samples from normal to invasive carcinoma. B: The model separated lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) WSIs by tumor histologic patterns and carcinoma in situ (CIS) PMLs by outcome.