Investigation of the capabilities of algorithms for automated quality assurance of DICOM metadata of chest X-ray examinations

Abstract

Purpose. To evaluate the quality of filling DICOM tags responsible for the orientation, scanning area and photometric interpretation of the image, as well as to develop and test algorithms for automatically determining the true values of these tags for chest x-rays and fluorograms.Materials and methods. To assess the quality of filling DICOM tags, were used 1885 studies obtained from the Unified Radiological Information Service of the Unified Medical Information and Analysis System (ERIS EMIAS). For training and validation of algorithms for automatic determination of the true values of tags, were used datasets of radiographs in standard frontal and lateral projections, from open databases and from ERIS EMIAS (12,920 studies in total). The deep neural network architecture VGG 19 was chosen as the basis for creating algorithms.Results. We found that the frequency of missing values in DICOM tags can range from 6 to 75%, depending on the tag. At the same time, up to 70% of filled tag values have errors. We obtained next models: a model for determining the anatomical area of x-ray examination, a model for determining the projection on the chest x-ray, a model for determining the photometric interpretation of the image. All of the obtained algorithms have high classification quality indicators. The AUC for each of the obtained models was more than 0.99.Conclusions. Our study shows that a large number of studies in diagnostic practice contain incorrect values of DICOM tags, which can critically affect the implementation of software based on artificial intelligence technology in clinical practice. Our obtained algorithms can be integrated into the development process of such software and used in the preprocessing of images before their analysis.