Prognostic Analysis Of Head And Neck Cancer Based On Graph Convolutional Network And Multimodal Data

Head and neck cancer(HNC)ranks as the seventh most common malignancy worldwide,which incidence is on the rise in China.Moreover,HNC lacks typical symptoms in the early stage and is easy to metastasize in the late stage,which brings great challenges to clinical diagnosis and treatment.Accurate risk stratification is of great significance for personalized treatment of HNC and improvement of survival rate.At present,PET/CT imaging is an important imaging tool for early diagnosis,grading and prognosis prediction of HNC.Traditional clinical indicators such as SUV and TNM staging are difficult to characterize tumor heterogeneity,and the prognostic performance is also unsatisfactory.Radiomics can directly extract high-throughput quantitative features from images and characterize tumor heterogeneity.It has been widely used in the classification,staging,efficacy evaluation and prognosis prediction of HNC,with good performances.However,the radiomics method extracts features directly from the grid Euclidian space,the radiomics method can’t well format the relationship/difference among different sub-regions for each HNC patient due to the extensive lesion distribution and complex pathological types of HNC.In view of this,aiming at the above problems,focusing on the graph convolution neural network which can directly learn features from the graph structure data and the multi-modal data fusion method,the following two studies are carried out in this paper:(1)Predictive analysis of HNC prognosis based on a functional-structural sub-region graph convolutional network(FSGCN).In order to overcome the inherent defect of traditional machine learning methods,which extract features directly based on grid Euclidean space,this article proposes a functional-structural sub-region graph convolutional network(FSGCN)for accurate risk stratification of HNC.By taking advantage of the high internal heterogeneity of HNC,PET and CT voxel information was used to cluster tumors into multiple sub-regions.The radiomics features were extracted from each sub-region to characterize its functional and structural information,and graphs were constructed in non-European space to represent the relationship/difference between different sub-regions of each patient.Graph outcome data representing patients were then input into the graph convolutional neural network model was then constructed to generate prognostic scores to predict progression-free survival(PFS)of patients with head and neck cancer(HNC).The results show that the proposed method is superior to the traditional radiomics analysis method in performance,and the experiment of multi-center data also shows that the proposed method has better generalization ability in predicting the prognosis of HNC compared with the radiomics method.(2)Predictive analysis of HNC prognosis based on population graph using multimodality fusion graph convolution neural network.In order to make full use of multi-modality imaging data and clinical data to capture the relationship between HNC patients,this article proposes a multi-modality fusion graph neural network(MF-GNN)based on node classification to capture the topological relationship among patients through the analysis of multi-modal medical data.In this experiment,PET/CT image information was fused by similarity matrix algorithm to construct patient population graph,and population graph was input into the constructed MF-GNN model for prognostic prediction of HNC.Two multimodality data fusion strategies were proposed:feature level fusion and network level fusion.The results show that the network level fusion strategy achieved higher prognostic performance and showed great potential in predicting the prognosis of patients with HNC.