Multi-Level fusion graph neural network: Application to PET and CT imaging for risk stratification of head and neck cancer

Abstract

Background and objectiveRisk stratification of head and neck cancer (HNC) patients is important to settle individualized treatment strategies. This study proposed a multi-level fusion graph neural network (MLF-GNN) that fuses PET/CT image features and clinical data for improved prognosis prediction of HNC patients. MethodsThis study exploited 642 HNC patients from 7 different centers collected from The Cancer Imaging Archive (TCIA), which were arbitrarily separated into 507 patients from 4 centers for training and internal validation, and 135 patients from 3 centers for external testing. We constructed a population graph with edges representing the similarity between patients and the vertices representing radiomics features. Single-modality GNN models and MLF-GNN models by using feature-fusion and network-fusion strategies were constructed for progression-free survival (PFS) prediction, respectively. Model performance was evaluated by the concordance index (C-index), area under the receiver operator characteristic curve (ROC-AUC), and Kaplan–Meier curves. ResultsCompared with the single-modality GNN models and feature-fusion based MLF-GNN model, the network-fusion based MLF-GNN model achieved the highest C-index of 0.788 (95 % CI: 0.727–0.848) and AUC of 0.807 (95 % CI: 0.731–0.870) for PFS prediction in the external testing set. Besides, it also showed good performance on the secondary endpoints of overall survival (OS), recurrence-free survival (RFS), and metastasis-free survival (MFS) in the external testing set, with C-index of 0.800 (95 % CI: 0.729–0.871), 0.823 (95 % CI: 0.763–0.884) and 0.758 (95 % CI: 0.673–0.844), respectively. Subgroup analysis stratified with pathogenic site and treatment type showed that the proposed method has good prediction performance on subgroup risk stratification. ConclusionsThe proposed MLF-GNN model could capture the topological relationships among patients by taking full advantage of multi-modality imaging data and clinical data, which achieved improved prognostic performance and was beneficial to guide individual treatment.