Abstract
Objective: to evaluate the diagnostic accuracy of machine learning systems and the analysis of digital x-ray images on the example of three commonly available software products.
Methods: The study is based on the recognition and analysis of digital radiographs using 3 software products based on convolutional neural networks using a sample of 300 digital radiographs of the lungs, the ratio is normal:pathology-50:50%-150 digital radiographs with nodule/mass in the lungs and 150 digital radiographs without significant radiological pathology in the lungs.
Results: On average, 80% of the images were correctly interpreted. Specificity-71-99%, sensitivity-55-87%. The indicator of overdiagnosis is 14%. The discrepancy rate of the results of the interpretation of radiographs was 51%. The hypodiagnostic indicator averaged 28%, while 50% were radiographs of patients with verified lung cancer. Pathology was missed by all software products in 21% of radiographs. Most of the erroneous interpretations were among images with nodule of a solid structure (69%), nodule 10-30mm in size (69%), and localization of pathology in 2.4 segments of the right lung and 2.6 segments of the left lung, which corresponded to the summation of the shadow of the nodule and shadow of ribs, clavicles and roots of the lungs on the radiographs.
Conclusions: The results obtained indicate the feasibility of training and testing systems on different datasets in order to increase their diagnostic efficiency, as well as the possibility of using similar systems as a “second reading” in screening studies with a pre-known prevalence of radiographs without significant radiological pathology.
Footnotes
Cite this article as: European Respiratory Journal 2020; 56: Suppl. 64, 857.
This abstract was presented at the 2020 ERS International Congress, in session “Respiratory viruses in the "pre COVID-19" era”.
This is an ERS International Congress abstract. No full-text version is available. Further material to accompany this abstract may be available at www.ers-education.org (ERS member access only).
- Copyright ©the authors 2020