Prior-enhanced attribute embedding for predicting chemoradiotherapy sensitivity in SNSCC patients

Abstract

Preoperative chemoradiotherapy plays a good role in organ preservation in sinonasal tumors, but some patients are not sensitive to chemoradiotherapy. Since the samples of sinonasal squamous cell carcinoma (SNSCC) are insufficient and unbalanced, it is challenging to establish a good model for sensitivity prediction. To solve this problem, this paper proposes an end-to-end Prior Enhancement framework based on Attribute Embedding (PEAE) to predict the sensitivity of SNSCC patients to chemoradiotherapy. The whole prediction task is divided into an image-wise task and a subject-wise task. PEAE is applied to the image-wise task, which can fully mine the prior imaging and non-imaging information in the existing data and can be easily embedded into the mainstream backbone network for end-to-end optimization. Specifically, we propose a multi-level attribute structure to express the prior information of images, which consists of general, spatial, and textual attributes. Furthermore, graph convolutional network is used to establish the relationship between images during prediction, where the adjacency matrix is obtained from the correlation of images calculated according to the multi-level attribute structure. In the subject-wise task, the prediction result of each subject is obtained by averaging the probability values obtained in the image-wise task that each image belongs to each category. The experimental results on SNSCC and an additional public dataset, ADNI-SEG, show that models with PEAE perform better than traditional neural networks. The accuracy, AUC and recall are improved by more than 10% on several mainstream networks.