Construction of a convolutional neural network classifier developed by computed tomography images for pancreatic cancer diagnosis

BACKGROUND Efforts should be made to develop a deep-learning diagnosis system to distinguish pancreatic cancer from benign tissue due to the high morbidity of pancreatic cancer. AIM To identify pancreatic cancer in computed tomography (CT) images automatically by constructing a convolutional neural network (CNN) classifier. METHODS A CNN model was constructed using a dataset of 3494 CT images obtained from 222 patients with pathologically confirmed pancreatic cancer and 3751 CT images from 190 patients with normal pancreas from June 2017 to June 2018. We established three datasets from these images according to the image phases, evaluated the approach in terms of binary classification (i.e., cancer or not) and ternary classification (i.e., no cancer, cancer at tail/body, cancer at head/neck of the pancreas) using 10-fold cross validation, and measured the effectiveness of the model with regard to the accuracy, sensitivity, and specificity. RESULTS The overall diagnostic accuracy of the trained binary classifier was 95.47%, 95.76%, 95.15% on the plain scan, arterial phase, and venous phase, respectively. The sensitivity was 91.58%, 94.08%, 92.28% on three phases, with no significant differences (chi(2)= 0.914,P= 0.633). Considering that the plain phase had same sensitivity, easier access, and lower radiation compared with arterial phase and venous phase , it is more sufficient for the binary classifier. Its accuracy on plain scans was 95.47%, sensitivity was 91.58%, and specificity was 98.27%. The CNN and board-certified gastroenterologists achieved higher accuracies than trainees on plain scan diagnosis (chi(2)= 21.534,P< 0.001;chi(2)= 9.524,P< 0.05; respectively). However, the difference between CNN and gastroenterologists was not significant (chi(2)= 0.759,P= 0.384). In the trained ternary classifier, the overall diagnostic accuracy of the ternary classifier CNN was 82.06%, 79.06%, and 78.80% on plain phase, arterial phase, and venous phase, respectively. The sensitivity scores for detecting cancers in the tail were 52.51%, 41.10% and, 36.03%, while sensitivity for cancers in the head was 46.21%, 85.24% and 72.87% on three phases, respectively. Difference in sensitivity for cancers in the head among the three phases was significant (chi(2)= 16.651,P< 0.001), with arterial phase having the highest sensitivity. CONCLUSION We proposed a deep learning-based pancreatic cancer classifier trained on medium-sized datasets of CT images. It was suitable for screen