Рossibilities of automatic systems in interpretation of lung X-rays in patients with suspicion for round formations

To analyze the information content of automatic recognition of formations in lungs at digital radiography on the example of one of the commonly available diagnostic algorithms. Materials and methods. This study is based on the results of recognition and analysis of digital radiographs from test bases by software product, based on convolutional neural networks on the example of one of the generally available diagnostic algorithms. The database consisted of anonymized digital radiographs of 240 patients whose health condition was confirmed by histological data and computed tomography (CT) results, the ratio between normal and pathology was 50:50%. Results. At application of automatic recognition of formations in the lungs on digital radiographs, good possibilities were revealed for detecting rounded formations in the lung tissue (93% of pathological changes). Recognition results with a solid structure of formations were higher than with a subsolid — 94% and 88%, respectively. A significant dependence of the possibility of detecting formations in the lungs on their size was revealed. For the focal lesions smaller than 10 mm in none of the cases the algorithm determined them with a probability of more than 50%, and only in 33% of cases the probability was 10–50%. Indicators are significantly higher at the size of formations from 10 to 305 mm (61% of solid formations and 25% of subsolid formations with a probability indicator more than 90%). When the size of solid formations is more than 30 mm, all the 100% of cases are revealed, while in 73% the probability is indicated more than 90%. There was also identified 100% of the subsolid formations. Сonclusion. The application of automated systems as a method of interpretation of radiographs is a promising direction, requiring further improvement and more detailed study of the results obtained. The diagnostic effectiveness of these algorithms is close to effectiveness of radiologists. Currently, the results obtained by the algorithm as a software product for identifying of rounded formations in the lungs cannot be used as a reliable diagnostic method, but it might be considered as an auxiliary «second reading» for the radiologist.

X-ray of the lungs, lung formations, screening, artificial intelligence, convolutional neural networks

1. Tyurin I.E. Radiological diagnostics in the Russian Federation in 2016. Herald of radiology and radiology, 2017, Vol. 98, No. 4, рр. 219–226 (In Russ.)

2. Luchkevich V.S., Hasanova E.A. The trends of the epidemiological situation of tuberculosis in Russia at the present stage (review). Medical Alliance, 2016, No. 3, рр. 20–23 (In Russ.)

3. Phthisiatry. National clinical guidelines / ed. P. K. Yablonsky. Moscow: GEOTAR-Media, 2015, 240 р. (In Russ.)

4. Yablonskij P.K., Vasil’eva I.A., Ergeshov A.E. Clinical recommendations for the diagnosis and treatment of respiratory tuberculosis in adults. Moscow, 2013 (In Russ.)

5. Koreckaya N.M., Narkevich A.N. Newly diagnosed pulmonary tuberculosis in individuals who comply with and violate the prescribed timeframe for fluorosurveys in an epidemically unfavorable region. Tuberculosis and lung diseases, 2013, Vol. 90, No. 9, рр. 21–24 (In Russ.)

6. Gavrilov P., Ushkov A, Smol’nikova U. Detection of lumps in the lungs with digital X-ray: the role of the work experience of the radiologist. Medical Alliance, 2019, No. 2, рр. 51–56 (In Russ.)

7. Yerushalmy J. et al. The role of dual reading in mass radiography // American Review of Tuberculosis. 1950. Vol. 61. Р. 443–464.

8. Nakamura K. et al. Studies on the diagnostic value of 70 mm radiophotograms by mirror camera and the reading ability of physicians // Kekkaku. 1970. Vol. 45. Р. 121–128.

9. Morozov S.P., Vladzimirskij A.V., Ledihova N.V., Sokolina I.A., Kul’berg N.S., Gombolevskij V.A. Evaluation of the diagnostic accuracy of the screening system of pulmonary tuberculosis based on artificial intelligence. Tuberculosis and lung diseases, 2018, Vol. 95, No. 8, рр. 42–49 (In Russ.)

10. Lakhani Р. et al. Deep Learning at Chest Radiography: Automated Classification of Pulmonary Tuberculosis by Using Convolutional Neural Networks // Radiology. 2017. Vol. 284, Iss. 2. Р. 574–582.

11. Jaeger S. et al. Automatic Tuberculosis Screening Using Chest Radiographs // IEEE transactions on medical imaging. 2014. Vol. 33, Iss. 2. P. 233–245.

12. Gavrilov P.V., Baulin I.A., Lukina O.V. Standardized interpretation and control of identified single lung formations using the Lung Imaging Reporting and Data System (Lung-RADS™). Medical Alliance, 2017, No. 3, рр. 17–27 (In Russ)

13. Garland L.H. On the reliability of roentgen survey procedures // American Journal of Rentgenology and Radium Therapeutic Nuclear Medicine. 1950. Vol. 64. Р. 32–41.

14. Garland L.H. Studies on the accuracy of diagnostic procedures // American Journal of Rentgenology and Radium Therapeutic Nuclear Medicine. 1959. Vol. 82. Р. 25–38.

15. Newell R.R. Chamberlain WE. Rigler L. Descriptive classification of pulmonary shadows: a revelation of unreliability in the roentenographic diagnosis of tuberculosis // American Review of Tuberculosis. 1954. Vol. 69. Р. 566–584.

16. Sone S. et al. Characteristics of small lung cancers invisible on conventional chest radiography and detected by population based screening using spiral CT // The British journal of radiology. 2000. Vol. 73. No. 866. Р. 137–145.

17. Dobbins J.T. et al. Digital tomosynthesis of the chest for lung nodule detection: Interim sensitivity results from an ongoing NIH‐sponsored trial // Medical physics. 2008. Vol. 35. No. 6. Part 1. Р. 2554–2557.

18. Meldo A.A., Utkin L.V., Trofimova T.N., Ryabinin M.A., Moiseenko V.M., Shelekhova K.V. Novel approaches to development of artificial intelligence algorithms in the lung cancer diagnostics // Diagnostic radiology and radiotherapy 2019. Vol. 10. No. 1. pp. 8–18 (In Russ.) DOI: 10.22328/2079-5343-2019-10-1-8-18.