[Clinical application of convolutional neural network in pathological diagnosis of metastatic lymph nodes of gastric cancer].

Abstract

To examine the value and clinical application of convolutional neural network in pathological diagnosis of metastatic lymph nodes of gastric cancer. Totally 124 patients with advanced gastric cancer who underwent radical gastrectomy plus D2 lymphadenectomy at Affiliated Hospital of Qingdao University from July 2016 to December 2018 were selected in the study. According to the chronological order, the first 80 cases were served as learning group. The remaining 44 cases were served as verification group. There were 45 males and 35 females in the study group, with average age of 57.6 years. There were 29 males and 15 females in the validation group, with average age of 9.2 years. The pre-training convolutional neural network architecture Resnet50 was trained and fine-tuned by 21 352 patches with cancer areas and 14 997 patches without cancer areas in the training group. A total of 78 whole-slide image served as a test dataset including positive (38) and negative (40) lymph nodes. The convolutional neural network computer-aided detection (CNN-CAD) system was used to analyze the ability of convolutional neural network system to screen metastatic lymph nodes at the level of slice by setting threshold, and evaluate the system's classification accuracy by calculating its sensitivity, specificity, positive predictive value, negative predictive value and area under the receiver operating characteristic curve (AUC). The classification accuracy of CNN-CAD system at slice level was 100%.The AUC for the CNN-CAD system was 0.89. The sensitivity was 0.778, specificity was 0.995, overall accuracy was 0.989. Positive and negative predictive values were 0.822 and 0.994, respectively. The CNN-CAD system achieved the same classification results as pathologists. The CNN-CAD system has been constructed to distinguished benign and malignant lymph node slides with high accuracy and specificity. It could achieve the similar classification results as pathologists.