[Paper Review] Lymph Node Gross Tumor Volume Detection in Oncology Imaging via Relationship Learning Using Graph Neural Network

Chun-Hung Chao, Zhuotun Zhu+11 more|arXiv (Cornell University)|Aug 29, 2020

Radiomics and Machine Learning in Medical Imaging28 references4 citations

TL;DR

This paper proposes a 3D CNN-GNN framework that jointly models lymph node (LN) appearance and inter-LN relationships to improve detection of lymph node gross tumor volume (OSLN) in oncology imaging. By leveraging graph neural networks to encode spatial priors and metastatic spread pathways, the method reduces false positives while boosting sensitivity, achieving a 13.1% improvement in average sensitivity at 6 false positives per patient over SOTA methods on a 142-patient esophageal radiotherapy dataset.

ABSTRACT

Determining the spread of GTV$_{LN}$ is essential in defining the respective resection or irradiating regions for the downstream workflows of surgical resection and radiotherapy for many cancers. Different from the more common enlarged lymph node (LN), GTV$_{LN}$ also includes smaller ones if associated with high positron emission tomography signals and/or any metastasis signs in CT. This is a daunting task. In this work, we propose a unified LN appearance and inter-LN relationship learning framework to detect the true GTV$_{LN}$. This is motivated by the prior clinical knowledge that LNs form a connected lymphatic system, and the spread of cancer cells among LNs often follows certain pathways. Specifically, we first utilize a 3D convolutional neural network with ROI-pooling to extract the GTV$_{LN}$'s instance-wise appearance features. Next, we introduce a graph neural network to further model the inter-LN relationships where the global LN-tumor spatial priors are included in the learning process. This leads to an end-to-end trainable network to detect by classifying GTV$_{LN}$. We operate our model on a set of GTV$_{LN}$ candidates generated by a preliminary 1st-stage method, which has a sensitivity of $>85\%$ at the cost of high false positive (FP) ($>15$ FPs per patient). We validate our approach on a radiotherapy dataset with 142 paired PET/RTCT scans containing the chest and upper abdominal body parts. The proposed method significantly improves over the state-of-the-art (SOTA) LN classification method by $5.5\%$ and $13.1\%$ in F1 score and the averaged sensitivity value at $2, 3, 4, 6$ FPs per patient, respectively.

Motivation & Objective

• To address the challenge of detecting small, metastasis-suspicious lymph nodes (OSLN) that are often missed in radiotherapy CT scans due to low contrast and high anatomical variability.
• To improve the precision of OSLN detection by reducing false positives in high-sensitivity candidate sets generated by a first-stage detection system.
• To model the biological prior that cancer spreads along connected lymphatic pathways by learning inter-LN relationships using graph neural networks.
• To develop an end-to-end trainable framework that integrates 3D CNN-based appearance features with GNN-based relational reasoning for improved OSLN classification.

Proposed method

• A 3D convolutional neural network (CNN) is used to extract instance-wise appearance features from CT and PET images of candidate lymph nodes.
• Inter-LN relationships are modeled using a graph neural network (GNN), where nodes represent candidate LNs and edges encode 3D spatial distances and angles relative to the primary tumor.
• Global spatial priors—such as distance and angular relationships—are incorporated into the GNN to reflect the anatomical connectivity of the lymphatic system.
• An element-level feature fusion strategy is employed in the GNN to aggregate multi-scale relational features, improving discrimination between true OSLN and false positives.
• The entire pipeline is end-to-end trainable, combining 3D CNN feature extraction with GNN-based relational reasoning for joint classification of OSLN.
• The method is evaluated on 142 paired PET/CT scans from esophageal cancer patients, using a high-sensitivity candidate set with >85% sensitivity but >15 FPs per patient.

Experimental results

Research questions

• RQ1 Can modeling inter-lymph node relationships improve the detection of small, metastasis-suspicious lymph nodes (OSLN) in non-contrast-enhanced radiotherapy CT?
• RQ2 Does incorporating global spatial priors—such as distance and angular relationships relative to the primary tumor—enhance OSLN classification beyond appearance-based features alone?
• RQ3 How does a graph neural network compare to conventional instance-wise classification in reducing false positives while maintaining high sensitivity for OSLN detection?
• RQ4 To what extent does fusing 3D CNN features with relational GNN representations improve performance over SOTA methods in OSLN detection?

Key findings

• The proposed CNN+GNN p model achieved a 5.5% absolute improvement in F1 score over the SOTA 2.5D CNN method on the same dataset.
• At 6 false positives per patient, the method improved average sensitivity by 13.1% compared to the SOTA baseline, demonstrating strong robustness to high FP rates.
• The model achieved a 6.9% improvement in mFROC (0.595 to 0.664) under CT+PET input with spatial priors, outperforming both the 3D CNN and node-level GNN fusion baselines.
• The element-level feature fusion in the GNN led to a 3.4% and 6.9% mFROC gain over the 3D CNN under CT and CT+PET settings, respectively, confirming the effectiveness of relational reasoning.
• Qualitative results showed that the GNN correctly reclassified false positives near low-probability neighbors and correctly identified small true OSLN when multiple strong nearby candidates were present.
• The addition of PET imaging consistently improved mFROC across all models, indicating its critical role in enhancing OSLN detection performance.