Purpose: Development and validation of a deep learning network to automatically detect non-diagnostic digital radiograph (DR) lateral airway/soft tissue neck examinations, at the point of care, that should be repeated before the technologist completes the examination.
Methods: Two radiologists independently reviewed 1121 pediatric lateral airway/soft tissue neck radiographs and classified them into cases that required a repeat examination (bad scans, due to poor visualization of the airway and/or relevant soft tissues) and that were diagnostic (good scans). When the two radiologists did not come to a consensus, a third radiologist reviewed the image and the majority consensus was used for the final classification. The review process resulted in 682 no-repeat cases and 439 repeat cases. A deep learning model based on the GoogleNet convolutional neural network architecture with inception modules was used to train the classifier. The training set consisted of 366 no-repeat cases and 249 repeat cases. Data augmentation was used both at the training and validation phases to increase the training images and reduce over-fitting. The Keras framework with Tensorflow-backend was used for implementation. Diagnostic performance of the final model was evaluated using sensitivity and specificity of the model to detect non-diagnostic scans.
Results: Discriminating between airway examinations requiring a repeat scan achieved sensitivity, specificity and accuracy values of 0.85, 0.83 and 0.83, respectively when tested on a dataset of 514 images (421 no-repeat and 92 repeat cases) with a median age of 4 Â± 9 years and male to female ratio of 0.56.
Conclusion: A deep learning model that automatically detects non-diagnostic pediatric DR airway X-ray examinations has been developed and validated. The initial results show potential for a point-of-care model that can flag bad radiographic images that require repeating immediately without radiologist intervention.