An Adaptive Multi-Feature Fusion Network for Predicting Overall Survival of Patients with Head and Neck Cancer

Abstract

Accurate prognostic prediction could allow personalized treatment to achieve optimal clinical outcome. We aimed to develop a highly predictive overall survival model, considering the complementary relationships between clinical information, traditional radiomics and deep image information, to further improve the overall prediction accuracy by constructing a richer feature set and adaptive weighting. A total of 427 patients with Oropharyngeal Cancer (OPC) patients from the TCIA database were included. 341 cases were used for training, 86 cases were used as an independent cohort. Patient characteristics, including TMN, age, gender, HPV status, smoking or drinking status, etc. were considered as potential predictors. Traditional radiomics features of gross tumor volume (GTV) was extracted from planning CT using open-source software. In addition, a two-dimensional convolutional network (2D_CNN) was designed to extract deep image features. An adaptive multi-feature fusion network was developed to predict overall survival of patients based on three types of features. The fusion network integrates an attention mechanism to the channel dimension to obtain proper weighting of each channel in the feature graph through the fully connected network by focusing on effective feature channels and automatic learning according to the loss, thus improving the utilization rate of effective features. The model performance was evaluated using the area-under-ROC-curve (AUC), accuracy, precision, recall, f1-score. The AUCs of predictive models based on clinical features, traditional radiomics features and deep image features were 0.7, 0.61 and 0.72, respectively. Combining patient characteristics, radiomic features and deep imaging features, the AUCs of the prediction models was significantly improved to 0.85 and 0.86 (with attention mechanisms) for the independent test cohort (Table 1). The proposed adaptive multi-channel network assigned effective weights to the potential predictors, selectively enhanced useful features while suppressed irrelevant features, enabling more accurate feature map weights. We demonstrated the improved predictive value, with a multi-channel fusion network integrated with an attention mechanism, for overall survival of OPC patients.