ORCGT: Ollivier-Ricci Curvature-Based Graph Model for Lung STAS Prediction

Min Cen, Zheng Wang, Zhenfeng Zhuang, Hong Zhang, Dan Su, Zhen Bao, Weiwei Wei, Baptiste Magnier, Lequan Yu, Liansheng Wang

Introduction

In lung adenocarcinoma, Tumor Spread Through Air Spaces (STAS) is a critical prognostic indicator, signaling a high risk of recurrence and poor patient outcomes. STAS is characterized by the presence of tiny tumor cell clusters in air spaces beyond the main tumor boundary. For pathologists, identifying these microscopic, sparse features in Whole Slide Images (WSIs) is a significant challenge, often described as a "needle in a haystack" problem due to the subtle nature of the signals.

Existing deep learning approaches, such as Multiple Instance Learning (MIL), frequently fail in this task because they tend to focus on the most prominent features—the large primary tumor—while ignoring the sparse, distant signals of STAS. Furthermore, standard Graph Neural Networks (GNNs) often suffer from "over-squashing," where information from distant patches is lost or distorted as it travels through the graph's bottlenecks. To overcome these limitations, the authors proposed the Ollivier-Ricci Curvature-Based Graph Model (ORCGT).

Methodology: The ORCGT Framework

ORCGT introduces a domain-inspired two-stage approach designed to mimic the focused review process of a human pathologist.

The ORCGT architecture consists of a Major Tumor Margin Extraction stage followed by a Curvature-based Graph Model to predict STAS.

Stage 1: Major Tumor Margin Extraction (The "Ring Map")

Instead of analyzing the entire slide blindly, the model first identifies the boundary of the primary tumor. It then extracts a "ring" of tissue extending both inward and outward from this margin. This "Ring Map" restricts the model’s attention to the specific region where STAS is pathologically defined to occur, drastically reducing noise from the tumor core and surrounding healthy tissue.

Stage 2: Curvature-Based Graph Network

Within the Ring Map, patches are treated as nodes in a graph. To address the "over-squashing" effect, the model utilizes Ollivier-Ricci Curvature (ORC). This geometric measure identifies bottlenecks in the graph (areas of negative curvature). The model adaptively adjusts edge weights based on ORC to ensure that information flows correctly between distant patches, preserving the signal of tiny STAS clusters. The Pool-Refined Curvature-based Graph (PRCG) layer then captures long-distance correlations between these clusters to make a final prediction.

Key Results and Interpretability

ORCGT was evaluated on a dataset of 284 pathologically confirmed lung adenocarcinoma WSIs. The model significantly outperformed several state-of-the-art (SOTA) methods, including ABMIL, TransMIL, and PatchGCN:

  • AUROC: 0.753 (vs. 0.712 for the next best, ABMIL).
  • AUPRC: 0.556 (a 4.8% improvement over SOTA, critical for imbalanced datasets).
  • F1 Score: 0.502 (vs. 0.422 for PatchGCN).

Ablation studies confirmed that the "Ring Map" is the single most important factor for this task; removing it caused the AUPRC to drop by 8.2%. Beyond performance metrics, ORCGT provides superior interpretability.

Comparative attention maps show that ORCGT (d) accurately highlights sparse STAS clusters (red areas) near the tumor margin, aligning with clinical definitions, whereas baseline models (b, c) fail to capture these subtle signals.

As shown in the attention maps, ORCGT specifically highlights the sparse clusters in the air spaces, aligning perfectly with the clinical definition of STAS. This is a marked improvement over previous models that often highlighted the dense center of the tumor.

Conclusion

ORCGT represents a significant advancement in computational pathology for lung cancer. By combining domain-specific spatial constraints with advanced graph geometry, the model overcomes the limitations of traditional deep learning in detecting sparse, microscopic features. While the current results are promising, future work will focus on validating the model on larger, multi-center cohorts to ensure generalization across different laboratory preparations and scanners.

Summary updated

Abstract

Tumor Spread Through Air Spaces (STAS), identified as a mechanism of invasion, has been substantiated by multiple studies to be associated with lower survival rates, underscoring its significant prognostic implications. In clinical practice, pathological diagnosis is regarded as the gold standard for STAS examination. Nonetheless, manual STAS diagnosis is characterized by labor-intensive and time-consuming processes, which are susceptible to misdiagnosis. In this paper, we attempt for the first time to identify the underlying features from histopathological images for the automatic prediction of STAS. Existing deep learning-based methods usually produce undesirable predictive performance with poor interpretability for this task, as they fail to identify small tumor cells spread around the main tumor and their complex correlations. To address these issues, we propose a novel Ollivier-Ricci Curvature-based Graph model for STAS prediction (ORCGT), which utilizes the information from the major tumor margin to improve both the accuracy and interpretability. The model first extracts the major tumor margin by a tumor density map with minimal and coarse annotations, which enhances the visibility of small tumor regions to the model. Then, we develop a Pool-Refined Ollivier-Ricci Curvature-based module to enable complex interactions between patches regardless of long distances and reduce the negative impact of the over-squashing phenomenon among patches linked by negative curvature edges. Extensive experiments conducted on our collected dataset demonstrate the effectiveness and interpretability of the proposed approach for predicting lung STAS. Our code is available at https://github.com/zhengwang9/ORCGT.
BibTeX
@inproceedings{cen2024orcgt,
  author = {Min Cen and Zheng Wang and Zhenfeng Zhuang and Hong Zhang and Dan Su and Zhen Bao and Weiwei Wei and Baptiste Magnier and Lequan Yu and Liansheng Wang},
  title = {{ORCGT: Ollivier-Ricci Curvature-Based Graph Model for Lung STAS Prediction}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2024},
  doi = {10.1007/978-3-031-72086-4\_52},
}
Figure