Is Convolutional Neural Network Accurate for Automatic Detection of Zygomatic Fractures on Computed Tomography?

PurposeZygomatic fractures involve complex anatomical structures of the mid-face and the diagnosis can be challenging and labor-consuming. This research aimed to evaluate the performance of an automatic algorithm for the detection of zygomatic fractures based on convolutional neural network (CNN) on spiral computed tomography (CT).Materials and MethodsWe designed a cross-sectional retrospective diagnostic trial study. Clinical records and CT scans of patients with zygomatic fractures were reviewed. The sample consisted of two types of patients with different zygomatic fractures statuses (positive or negative) in Peking University School of Stomatology from 2013 to 2019. All CT samples were randomly divided into three groups at a ratio of 6:2:2 as training set, validation set, and test set, respectively. All CT scans were viewed and annotated by three experienced maxillofacial surgeons, serving as the gold standard. The algorithm consisted of two modules as follows: (1) segmentation of the zygomatic region of CT based on U-Net, a type of CNN model; (2) detection of fractures based on Deep Residual Network 34(ResNet34). The region segmentation model was used first to detect and extract the zygomatic region, then the detection model was used to detect the fracture status. The Dice coefficient was used to evaluate the performance of the segmentation algorithm. The sensitivity and specificity were used to assess the performance of the detection model. The covariates included age, gender, duration of injury, and the etiology of fractures.ResultsA total of 379 patients with an average age of 35.43 ± 12.74 years were included in the study. There were 203 nonfracture patients and 176 fracture patients with 220 sites of zygomatic fractures (44 patients underwent bilateral fractures). The Dice coefficient of zygomatic region detection model and gold standard verified by manual labeling were 0.9337 (coronal plane) and 0.9269 (sagittal plane), respectively. The sensitivity and specificity of the fracture detection model were 100% (p＞.05).ConclusionThe performance of the algorithm based on CNNs was not statistically different from the gold standard (manual diagnosis) for zygomatic fracture detection in order for the algorithm to be applied clinically. Zygomatic fractures involve complex anatomical structures of the mid-face and the diagnosis can be challenging and labor-consuming. This research aimed to evaluate the performance of an automatic algorithm for the detection of zygomatic fractures based on convolutional neural network (CNN) on spiral computed tomography (CT). We designed a cross-sectional retrospective diagnostic trial study. Clinical records and CT scans of patients with zygomatic fractures were reviewed. The sample consisted of two types of patients with different zygomatic fractures statuses (positive or negative) in Peking University School of Stomatology from 2013 to 2019. All CT samples were randomly divided into three groups at a ratio of 6:2:2 as training set, validation set, and test set, respectively. All CT scans were viewed and annotated by three experienced maxillofacial surgeons, serving as the gold standard. The algorithm consisted of two modules as follows: (1) segmentation of the zygomatic region of CT based on U-Net, a type of CNN model; (2) detection of fractures based on Deep Residual Network 34(ResNet34). The region segmentation model was used first to detect and extract the zygomatic region, then the detection model was used to detect the fracture status. The Dice coefficient was used to evaluate the performance of the segmentation algorithm. The sensitivity and specificity were used to assess the performance of the detection model. The covariates included age, gender, duration of injury, and the etiology of fractures. A total of 379 patients with an average age of 35.43 ± 12.74 years were included in the study. There were 203 nonfracture patients and 176 fracture patients with 220 sites of zygomatic fractures (44 patients underwent bilateral fractures). The Dice coefficient of zygomatic region detection model and gold standard verified by manual labeling were 0.9337 (coronal plane) and 0.9269 (sagittal plane), respectively. The sensitivity and specificity of the fracture detection model were 100% (p＞.05). The performance of the algorithm based on CNNs was not statistically different from the gold standard (manual diagnosis) for zygomatic fracture detection in order for the algorithm to be applied clinically.