RESUMO
Objective: To explore the application value of convolution neural network in quality control (QC) of chest digital radiology (DR) images. Methods: We classified and labeled 1 618 chest DR images taken by different machines, 1 294 of which were used as training set for convolution neural network and 324 as test set for detection effect. The sensitivity, specificity, positive predicted value (PPV), negative predicted value (NPV) and overall accuracy of the test results were calculated using the confusion matrix of two and four classifications. Results: The sensitivity, specificity, PPV and NPV of the two-classification results were 73.53%, 97.93%, 80.65%, 96.93%, and 95.37%, respectively; the total accuracy of the four-classification results was 75.93%. The overall accuracy of the two-classification results was significantly higher than that of the four-classification results (P<0.05). Conclusion: Convolutional neural network can satisfy the requirement of image QC to meet the minimum standard. However, in order to carry out high-level image quality scoring and assessment management, larger data sets and more detailed feature markers are needed.