Integrating Preoperative CT and Clinical Factors for Lymph Node Metastasis Prediction in Esophageal Cancer by Feature-Wise Attentional Graph Neural Network (FAGNN)

• DOI: 10.1016/j.ijrobp.2021.07.545
• Publication Date: 2021-10-25T14:44:39Z
• Authors: M. Ding; H. Cui; B. Li; B. Zou; Y. Xu; B. Fan; W. Li; L. Ma; J. Yu; L. Wang
• Keywords: 03 medical and health sciences; 0302 clinical medicine; 3. Good health

Purpose/Objective(s) This study aims to propose an adaptive integration model to predict lymph node metastasis (LNM) in esophageal squamous cell carcinoma (ESCC) patients by learning from preoperative CT imaging and non-imaging clinical parameters. Materials/Methods Contrast enhanced CT (CECT) scans taken two weeks before surgery and 20 clinical factors including general, pathological, hematological and diagnostic information were collected from 397 ESCC patients between October 2013 and November 2018. Of these, there are 924 lymph nodes (LN) (798 negative: 126 positive) with pathologically confirmed status after surgery. All LNs were randomly divided into a training set (n = 663) and validation set (n = 185). Data augmentations generated 1326 negative and 1140 positive LN samples used for training. FAGNN was composed of: a) deep image feature extraction by 3D CNN encoder and high-level non-imaging factors representation by a feed-forward neural network; b) a feature-wise attention module for feature embedding with learnable adaptive weights; and c) a graph convolutional layer to integrate the embedded features for final LN level metastasis prediction. Results FAGNN achieved area under ROC curve (AUC) of 0.816, sensitivity (sen) of 0.794, specificity (spe) of 0.861 in validation, which outperformed our model without adaptive embedding (AUC 0.810, sen 0.785, spe 0.857), our model using CT alone (AUC 0.802, sen 0.765, and spe 0.826), radiomics integration model (AUC 0.762, sen 0.739, spe 0.818), CT radiomics model (AUC 0.733, sen 0.689, spe 0.765). The improvement was statistically significant (P Conclusion Our adaptive integration model improved the LNM prediction performance based on CT and clinical parameters. Our model has the potential to foster effective fusion of multi-sourced parameters and support early prognosis and personalized surgery or radiotherapy planning in ESCC patients.