Distal radius fracture classification on dual-view radiography using ensemble deep learning framework.

Abstract

Purpose Distal radius fractures (DRFs) are one of the most common fractures treated surgically. To examine the injured wrists, standard radiographs, including posteroanterior (PA) and lateral views, are often taken in the emergency department. DRFs can be classified into intraand extra-articular fractures. In extra-articular fractures, the fracture line does not extend to the joint, while the intra-articular fractures involve the articular surface, which may require further evaluation and more complex treatments. Identifying DRFs as intra- or extraarticular can be useful for guiding further treatment. However, radiographic classification of DRFs is challenging due to the extreme variability of fracture patterns, complex anatomy of the wrist and variability in imaging quality of radiography. The aim of this study is to propose a deep learning (DL) framework incorporating both PA and lateral view X-rays for automatic DRF classification and evaluate the framework on clinically acquired wrist X-ray dataset. Methods The proposed framework consists of a distal radius region of interest (ROI) detection stage and a DRF classification stage as shown in Fig. 1. The distal radius ROI detection stage used an ensemble model of 10 YOLOv5 [1] base networks which is a recent release of the YOLO object detection network. This step allows the framework to zoom in on the relevant regions on PA and lateral view X-rays for fracture pattern analysis. Following the ROI extraction, an ensemble model of 10 dual-branch EfficientNet (DB-EffiNet) was applied to classify the DRFs into intra- or extra-articular fracture. The DBEffiNet is a novel adaptation of the EfficientNet [2] constructed in this study, which consists of two EfficientNet-b0 branches taking PA and lateral view X-rays as input respectively. The two branches were fused at the last linear layer via summation, followed by an additional linear layer to generate the final classification output. The dataset used for evaluating the DL framework contains 302 cases of clinically retrieved wrist X-rays. The dataset was randomly split into a training set of 251 cases with 257 fractures and a testing set of 51 cases with 52 fractures. There are 193 and 38 intra-articular DRFs in training and testing set respectively. The training set was randomly partitioned into 10 folds for cross-validation. For distal radius ROI detection, the YOLOv5s variant was trained on the PA and lateral view X-rays separately for 100 epochs within tenfold cross-validation, generating 10 YOLO base models for each view. The batch size was set as 8 and image size as 1280 x 1280, with stochastic gradient descent (SGD) used as the optimizer. Translation, scaling, horizontal flip and mosaic augmentations were adopted during training. For DRF classification, the DB-EffiNet was trained on the PA-lateral ROI pairs for 50 epochs within tenfold cross-validation. Each EfficientNet-b0 branch was pre-loaded with ImageNet pretrained weights. The ROI images were resized to 256 x 256 and normalized to the mean and standard deviation of ImageNet. The Adam optimizer was used with a learning rate of 0.0001 and the batch size was set as 16. Horizontal flip, rotation and brightness adjustment based on Power-Law transformations were used for augmentation during training. The model with the best area under the receiver operating characteristic curve (AUROC) on the validation set was saved as the base model in each cross-validation iteration. Given an unseen testing instance, the distal radius ROI was detected on each view through merging the 10 YOLO base models by enabling the model ensemble feature of YOLOv5. The ROIs on the PA and lateral views were then passed into the 10 DB-EffiNet base models. The ensemble probability was computed by averaging the probabilities across all base models. Results When evaluated on the testing data, the YOLO ensemble model successfully detected all distal radius ROIs on PA and lateral view X-rays with no false positives. As for differentiating intra- from extraarticular DRFs, the DB-EffiNet ensemble model achieved an AUROC of 0.90, an accuracy of 0.87, a sensitivity of 0.87 and a specificity of 0.86. Conclusion This study proposed a dual-view DL framework for automatic DRF classification on wrist radiography. Evaluated on a clinically acquired wrist X-ray dataset, this framework attained promising performance in identifying intra- and extra-articular DRFs using both PA and lateral view X-rays, which demonstrated its potential to assist clinicians in fracture pattern analysis and further treatment planning.