C Jin1, Y Jiang1, H Yu1, W Wang2, B Li1, C Chen3, Q Yuan3, Y Hu4,5, Y Xu3, Z Zhou2, G Li4,5, R Li1. 1. Department of Radiation Oncology, Stanford University School of Medicine, Stanford California, USA. 2. Department of Gastric Surgery, Sun Yat-sen University Cancer Centre, Guangzhou, China. 3. Departments of Medical Imaging, Nanfang Hospital, Southern Medical University, Guangzhou, China. 4. General Surgery, Nanfang Hospital, Southern Medical University, Guangzhou, China. 5. Guangdong Provincial Key Laboratory on Precision and Minimally Invasive Medicine for Gastrointestinal Cancers, Guangzhou, China.
Abstract
BACKGROUND: Lymph node metastasis (LNM) in gastric cancer is a prognostic factor and has implications for the extent of lymph node dissection. The lymphatic drainage of the stomach involves multiple nodal stations with different risks of metastases. The aim of this study was to develop a deep learning system for predicting LNMs in multiple nodal stations based on preoperative CT images in patients with gastric cancer. METHODS: Preoperative CT images from patients who underwent gastrectomy with lymph node dissection at two medical centres were analysed retrospectively. Using a discovery patient cohort, a system of deep convolutional neural networks was developed to predict pathologically confirmed LNMs at 11 regional nodal stations. To gain understanding about the networks' prediction ability, gradient-weighted class activation mapping for visualization was assessed. The performance was tested in an external cohort of patients by analysis of area under the receiver operating characteristic (ROC) curves (AUC), sensitivity and specificity. RESULTS: The discovery and external cohorts included 1172 and 527 patients respectively. The deep learning system demonstrated excellent prediction accuracy in the external validation cohort, with a median AUC of 0·876 (range 0·856-0·893), sensitivity of 0·743 (0·551-0·859) and specificity of 0·936 (0·672-0·966) for 11 nodal stations. The imaging models substantially outperformed clinicopathological variables for predicting LNMs (median AUC 0·652, range 0·571-0·763). By visualizing nearly 19 000 subnetworks, imaging features related to intratumoral heterogeneity and the invasive front were found to be most useful for predicting LNMs. CONCLUSION: A deep learning system for the prediction of LNMs was developed based on preoperative CT images of gastric cancer. The models require further validation but may be used to inform prognosis and guide individualized surgical treatment.
BACKGROUND: Lymph node metastasis (LNM) in gastric cancer is a prognostic factor and has implications for the extent of lymph node dissection. The lymphatic drainage of the stomach involves multiple nodal stations with different risks of metastases. The aim of this study was to develop a deep learning system for predicting LNMs in multiple nodal stations based on preoperative CT images in patients with gastric cancer. METHODS: Preoperative CT images from patients who underwent gastrectomy with lymph node dissection at two medical centres were analysed retrospectively. Using a discovery patient cohort, a system of deep convolutional neural networks was developed to predict pathologically confirmed LNMs at 11 regional nodal stations. To gain understanding about the networks' prediction ability, gradient-weighted class activation mapping for visualization was assessed. The performance was tested in an external cohort of patients by analysis of area under the receiver operating characteristic (ROC) curves (AUC), sensitivity and specificity. RESULTS: The discovery and external cohorts included 1172 and 527 patients respectively. The deep learning system demonstrated excellent prediction accuracy in the external validation cohort, with a median AUC of 0·876 (range 0·856-0·893), sensitivity of 0·743 (0·551-0·859) and specificity of 0·936 (0·672-0·966) for 11 nodal stations. The imaging models substantially outperformed clinicopathological variables for predicting LNMs (median AUC 0·652, range 0·571-0·763). By visualizing nearly 19 000 subnetworks, imaging features related to intratumoral heterogeneity and the invasive front were found to be most useful for predicting LNMs. CONCLUSION: A deep learning system for the prediction of LNMs was developed based on preoperative CT images of gastric cancer. The models require further validation but may be used to inform prognosis and guide individualized surgical treatment.