Refining Convolutional Neural Network Detection of Small-Bowel Obstruction in Conventional Radiography.

OBJECTIVE: The purpose of this study was to evaluate improvement of convolutional neural network detection of high-grade small-bowel obstruction on conventional radiographs with increased training set size.

MATERIALS AND METHODS: A set of 2210 abdominal radiographs from one institution (image set 1) had been previously classified into obstructive and nonobstructive categories by consensus judgments of three abdominal radiologists. The images were used to fine-tune an initial convolutional neural network classifier (stage 1). An additional set of 13,935 clinical images from the same institution was reduced to 5558 radiographs (image set 2) primarily by retaining only images classified positive for bowel obstruction by the initial classifier. These images were classified into obstructive and nonobstructive categories by an abdominal radiologist. The combined 7768 radiographs were used to train additional classifiers (stage 2 training). The best classifiers from stage 1 and stage 2 training were evaluated on a held-out test set of 1453 abdominal radiographs from image set 1.

RESULTS: The ROC AUC for the neural network trained on image set 1 was 0.803; after stage 2, the ROC AUC of the best model was 0.971. By use of an operating point based on maximizing the validation set Youden J index, the stage 2-trained model had a test set sensitivity of 91.4% and specificity of 91.9%. Classification performance increased with training set size, reaching a plateau with over 200 positive training examples.

CONCLUSION: Accuracy of detection of high-grade small-bowel obstruction with a convolutional neural network improves significantly with the number of positive training radiographs.