Purpose or Case Report: The purpose of this study was to develop a convolutional neural network (CNN)-based deep learning algorithm for the automated detection of developmental dysplasia of the hip (DDH) on conventional radiography and to assess its feasibility and diagnostic performance.
Methods & Materials: From 2,601 hip AP radiographs obtained in three different hospitals January 2011 and June 2018, 5,076 hip images were used to construct the dataset. Two invited radiologists were asked to label hip images as normal or DDH and all 5,076 patched images were divided into training (n = 4,050), validation (n = 513) and test sets (n = 513). Images underwent preprocessing, including cropping and histogram equalization, and were input into a convolutional neural network. To investigate diagnostic performance of the deep learning algorithm, we calculated the receiver operating characteristics (ROC) and precision recall (PRC) plots, accuracy, sensitivity, specificity, positive predictive (PPV) and negative predictive value (NPV) of the deep learning algorithm and they were compared with performances of two human readers with different levels of experience.
Results: The area under the ROC plot of deep learning algorithm and three radiologists were 0.988 and 0.988-0.919, respectively. The AUC of PRC plot of deep learning algorithm and three radiologists were 0.979 and 0.495-0.857, respectively. The accuracy, sensitivity, specificity, PPV and NPV of the proposed deep learning algorithm were 98.4, 94.0, 98.9, 90.4 and 99.4%. In McNemar's test, there was no significant difference between algorithm and experienced radiologist in diagnosis of DDH. On the other hand, the proposed model showed significant difference (P = 1.000) with higher sensitivity, specificity and PPV, compared to inexperienced radiologist. There were three false negative and five false positive cases in 513 test sets by using deep learning algorithm.
Conclusions:
The proposed deep learning algorithm provided an accurate diagnosis of developmental dysplasia of the hip on hip AP conventional radiographs, which was comparable to an experienced radiologist.