Evaluation of diagnostic accuracy of the system for pulmonary tuberculosis screening based on artificial neural networks

Abstract

The objective of the study: to evaluate the applicability of the automated system for detection of chest diseases during a regular mass screening of the population through assessment of universe parameters of diagnostic accuracy. Subjects and methods. A retrospective diagnostic study was conducted. The index-test (the method being studied) implied distinction and analysis of X-ray films using the software based on convolutional neural networks of U-NET type, which were modified and trained for specific purposes. The reference method used was the double revision of the previously classified X-ray films by two qualified roentgenologists with work experience of 8-10 years. Two depersonalized samplings of digital X-ray films were used: Sample 1 (n = 140), the ratio of the norm and pathology made 50: 50; Sample 2 (n = 150), the ratio of the norm and pathology made 95: 5. Results. The following parameters were set up for Samples 1 and 2 respectively: sensitivity - 87.2 and 75.0%, specificity - 60.0 and 53.5%, the prognostic value of the positive result - 68.6 and 8.3%, the prognostic value of the negative result - 82.4 and 97.5%, the area under characteristic curve - 0.74 and 0.64. Conclusions. The index test can be used only for mass regular screening in the population with low pre-test chances of pathology, which is confirmed by the prognostic value of the negative result (97.5%). This technology was recommended for the semiautomatic formation of pulmonary tuberculosis risk groups for consequent verification of the results by a roentgenologist.