ABSTRACT
Objective: To study whether a deep residual neural network can detect small bowel obstruction patterns on upright abdominal radiographs. Methods:The data of training set and test set used in this study were obtained from The First Affiliated Hospital of Xi'an Jiaotong University and No.215 Hospital of Shaanxi Nuclear Industry; the data of validation set came from No.215 Hospital of Shaanxi Nuclear Industry. Totally 3 298 clinical upright abdominal radiographs obtained from two hospitals were classified into obstructive and non-obstructive categories independently by two radiologists on the basis of the four signs on upright abdominal radiographs, who discussed and reached consensus when disagreements arose. Among them, 569(17.3%) images were found to be consistent with small bowel obstruction, and 2 729 (82.7%) images had no small bowel obstruction. A total of 2 305 training sets and 993 test sets (training set: test set = 2.3:1) were composed of data from the two groups, including 405 cases (17.6%) of small bowel obstruction, 1 900 cases (82.4%) of non-small bowel obstruction, 164 cases (16.5%) of small bowel obstruction, and 829 cases (83.5%) of non-small bowel obstruction. The diagnosis of small bowel obstruction in training and testing sets was based on experienced radiologists' evaluation. Totally 861 abdominal upright abdominal radiographs constituted the validation set (99 with small bowel obstruction and 762 with no small bowel obstruction); the surgical results and clinical diagnosis were set as the gold standard. In this study, the image 2012 large-scale visual recognition challenge data set (ILSVRC2012) was used for pre-training the deep residual neural network (ResNet38). The retraining of deep residual network (ResNet38) with training set data was used to establish the diagnostic model. The test set was mainly used in the learning algorithm process to adjust the algorithm parameters to modify the network, so as to make the network model more efficient. Results: After training, the deep residual neural network achieved an AUC of 0.83 on the test set (95% CI 0.82-0.92). The sensitivity of the system for small bowel obstruction was 84.1%, with a specificity of 65.0%. And on validation set it achieved an AUC of 0.87 (95% CI 0.82-0.92), the sensitivity of the system for small bowel obstruction was 89.9%, with a specificity of 68.0%. Conclusion: Transfer learning with deep residual neural network may be used to train a detector for small bowel obstruction on upright abdominal radiographs even with limited training data.