Ankle Fracture Detection Utilizing a Convolutional Neural Network Ensemble Implemented with a Small Sample, De Novo Training, and Multiview Incorporation

To determine whether we could train convolutional neural network (CNN) models de novo with a small dataset, a total of 596 normal and abnormal ankle cases were collected and processed. Single- and multiview models were created to determine the effect of multiple views. Data augmentation was performed during training. The Inception V3, Resnet, and Xception convolutional neural networks were constructed utilizing the Python programming language with Tensorflow as the framework. Training was performed using single radiographic views. Measured output metrics were accuracy, positive predictive value (PPV), negative predictive value (NPV), sensitivity, and specificity. Model outputs were evaluated using both one and three radiographic views. Ensembles were created from a combination of CNNs after training. A voting method was implemented to consolidate the output from the three views and model ensemble. For single radiographic views, the ensemble of all 5 models produced the best accuracy at 76%. When all three views for a single case were utilized, the ensemble of all models resulted in the best output metrics with an accuracy of 81%. Despite our small dataset size, by utilizing an ensemble of models and 3 views for each case, we achieved an accuracy of 81%, which was in line with the accuracy of other models using a much higher number of cases with pre-trained models and models which implemented manual feature extraction.