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

Abstract

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.

Noninvasive KRAS mutation estimation in colorectal cancer using a deep learning method based on routine CT imaging

“…Therefore, in the study, data augmentation was used to increase the size of the training dataset and prevent overfitting. Besides, the ResNet with simple architecture and short time consumption was enough to learn the predictive features according to the satisfactory results in our study and the other researchers [26,27].…”

Pattern Recognition in Musculoskeletal Imaging Using Artificial Intelligence

“…Kitamura et al trained a CNN de novo without pretraining on a relatively small data set of ankle radiographs ( 600) and found that using an ensemble of five models, multiple views (three views rather than one view of the ankle) improved the model accuracy of ankle fracture detection from 76% to 81%. 59 The authors attribute their relatively low accuracy to the small sample size but note that it is comparable with the 83% accuracy reported by Olczak et al from a pretrained model using a much larger data set of 256,000 wrist, hand, and ankle radiographs. 60 The vast majority of the AI literature on fracture detection focuses on model performance on radiographs, but some researchers have investigated the potential identification of fractures by CNNs on CT.…”

Comparing performance of deep convolutional neural network with orthopaedic surgeons on identification of total hip prosthesis design from plain radiographs

“…In this study, we implemented DL method to automatically detect the design of THR implants from plain film AP radiographs. Other studies have applied DL methods on plain film radiographs for various orthopedic applications [7][8][9][10][11][12][13][14][15][16] . However, to the best of our knowledge, our pilot study identifying three types of THR femoral implant designs 18 , and this study identifying nine types of THR femoral implant designs are the first applications of DL method to automatically detect THR femoral implant designs.…”

“…We hypothesized that deep learning (DL) based artificial intelligence algorithms could be trained to automatically identify hip implant designs from radiographic images. In recent years, DL methods have been applied to the interpretation of plain film radiographs with high degrees of success for identification and classification of orthopaedic fractures, staging knee osteoarthritis (OA) severity, and detection of aseptic loosening, to name a few [7][8][9][10][11][12][13][14][15][16][17] . In a previous pilot study, we successfully trained a DL method for the first time to classify a given THR radiograph into one of three possible femoral component designs 18 .…”