Improved lymph node metastasis prediction from preoperative esophageal squamous cell cancer CT by graph attention convolutional neural network (GACNN)

e16093 Background: Lymph node (LN) metastasis is the most important factor for decision making in esophageal squamous cell carcinoma (ESCC). A more accurate prediction model for LN metastatic status in ESCC patients is needed. Methods: In this retrospective study, 397 ESCC patients who took Contrast-Enhanced CT (CECT) within 15 days before surgery between October 2013 and November 2018 were collected. There are 924 (798 negative and 126 positive) LNs with pathologically confirmed status after surgery. All LNs were randomly divided into a training set (n = 663) and validation set (n = 185). Data augmentation including shifting and rotation was performed in the training set, resulting in 1326 negative and 1140 positive LN samples. The GACNN model was trained over CT volumetric patches centred at manually segmented LN samples. GACNN was composed of a 3D UNet encoder to extract deep features, and a graph attention layer to integrate morphological features extracted from segmented LN. The model was validated using the validation set (135 negative and 50 positive) and measured by area under ROC curve (auc), sensitivity (sen), and specificity (spe). Results: GACNN achieved better auc, sen, and spe of 0.802, 0.765, and 0.826, when compared to 3 other models including CT radiomics model (auc 0.733, sen 0.689, spe 0.765), 3D UNet encoder (auc 0.778, sen 0.722, spe 0.767), and our model without morphological features (auc 0.796, sen 0.754, spe 0.803). The improvement was statistically significant (p < 0.001). Conclusions: Our prediction model improved the prediction of LN metastasis, which has the potential to assist LN metastasis risk evaluation and personalized treatment planning in ESCC patients for surgery or radiotherapy.