ANALYSIS OF THE EFFECTIVENESS OF APPLICATION OF THE ARTIFICIAL INTELLIGENCE PROGRAM IN THE ASSESSMENT OF FLUOROGRAMS
- Authors: Abasov A.R.1, Bezborodov F.A.2, Klokova E.I.3
-
Affiliations:
- Воронежский государственный медицинский университет им. Н.Н.Бурденко
- ВГМУ им Н.Н.Бурденко
- ВГМУ им. Н.Н.Бурденко
- Issue: Vol 10 (2021): Материалы XVII Международной Бурденковской научной конференции 22-24 апреля 2021 года
- Pages: 111-113
- Section: Внутренние болезни
- URL: https://new.vestnik-surgery.com/index.php/2415-7805/article/view/6384
Cite item
Full Text
Abstract
Annotation. Computerization and technical progress have not spared medicine either. More and more technologies are coming to help doctors who are loaded with a constant stream of patients. The use of one of them - artificial intelligence (AI), is currently the most discussed topic in radiation diagnostics. This article discusses the prospect of using AI in medical imaging. The authors, relying on the experience of last year's study at the Voronezh Regional Clinical Consultative and Diagnostic Center (VOKKDC), studied the capabilities of AI in structured analysis of chest fluorograms. The study was conducted during the first three weeks of January 2021. The study used an artificial intelligence program using deep learning algorithms based on a multi-level convolutional neural network. Artificial intelligence and qualified radiographers analyzed 50 fluorographic images provided to the authors of the VOKKDTS, with the condition that the images were divided into groups of "normal", "pathology" and "alarming". As a result of the study, the following data were obtained: 34% of fluorograms were marked as “normal”, 36% of fluorograms were marked as “pathology” and 30% of images were classified as “alarming”, which exactly coincides with the opinion of radiologists. Such a result should be regarded as successful. The study shows that today artificial intelligence, with deep learning algorithms, is a promising point in the development of modern radiation diagnostics.
Full Text
Relevance. Artificial intelligence (AI) is the most talked about topic in medical imaging research today. The main reason for the emergence of AI in medical imaging was the desire for greater efficiency and effectiveness in radiation diagnostics. X-ray imaging data continues to grow at a disproportionate rate compared to the number of trained physicians, and the decline in imaging costs has forced healthcare providers to compensate for this with increased productivity. These factors have contributed to a dramatic increase in the burden on radiologists. Since radiation diagnostics includes visual perception, as well as decision making under conditions of uncertainty, mistakes become inevitable [1].
It becomes clear that the most interesting is the possibility of using AI in screening methods of radiation diagnostics, in particular in fluorography.
Fluorography is a method of X-ray examination, which consists in obtaining a full-size shadow image from an X-ray screen or from an electronic transducer screen onto a small format film. At present, it is difficult to overestimate the importance of this screening study in the diagnosis of respiratory diseases, in particular tuberculosis and lung cancer. Every person is obliged to undergo fluorography annually, so radiologists have to work with massive amounts of fluorograms in a very short period. Undoubtedly, when a doctor works with such a volume of information, diagnostic errors can also occur. Therefore, now more than ever, it is necessary to use artificial intelligence to improve the quality of diagnostics of lung diseases.
Traditional AI methods rely heavily on predefined algorithms for engineering functions with explicit parameters based on expert knowledge. Recent advances in artificial intelligence research have led to the emergence of new, non-deterministic deep learning algorithms that build on the capabilities of convolutional neural networks (CNN). Convolutional neural networks are capable of generating a highly efficient representation of input data, well suited for tasks focused on evaluating the properties of an image provided to a program. CNN has several levels of convolutions, consisting of activation layers that simplify input data by dividing it into simple constituent parts; union layers, completely connecting different levels of neural networks, which calculate the final data of the analysis. Such a multi-level system allows at the output to obtain a detailed analysis of the entered graphic data and their distribution according to predetermined criteria [2].
Purpose of the study: analysis and evaluation of the data obtained using the artificial intelligence program in the diagnosis of diseases of the chest organs through fluorography with a further assessment of the effectiveness and prospects of its use in screening methods of radiation diagnostics.
Materials and methods. The study was conducted in January 2021. The Voronezh Regional Clinical Consultative and Diagnostic Center (VOKKDC) provided the authors with 50 random chest cavity fluorograms in frontal projection. The AI that performed the processing and analysis was represented by the software RU.96876180.62.01.29-01.
The program is designed to analyze digital fluorographic images in direct anterior projection and identify the presence of possible pathologies. Artificial intelligence is capable of automatically detecting type 1 pathologies (dangerous) and type 2 pathologies (non-dangerous). Type 1 pathologies include: infiltration (more than 1.5 cm), cavity, pneumothorax, hydrothorax, focus, pathological changes in the roots of the lungs, fluid level and foci; type 2 pathology will include interstitial changes in the lung parenchyma, liver cirrhosis, fibrothorax, pleural changes, calcifications, diaphragmatic hernia, changes in bones, metal VIII-chain, foreign bodies, an area of increased transparency, atelectasis and changes in the mediastinal organs. The analysis of chest fluorograms was carried out with the installation of the division of images into "normal", corresponding to images of normal unchanged organs of the chest cavity, and "pathology" - images with clear signs of pathology [3].
For an objective assessment of the results of the study, we used the clinical recommendations of the Scientific and Practical Clinical Center for Diagnostics and Telemedicine Technologies of the Moscow Department of Healthcare (NPKTs DT DZM), which also include a basic set of metrics that determine diagnostic efficiency: sensitivity, specificity, general validity, likelihood ratio of a positive result, likelihood ratio of a negative result, predictive value of a positive result, predictive value of a negative result, frequency of false positives [4].
In addition, a consultation was held with qualified radiologists of the VOKKDTS, who, in parallel with the AI, analyzed the provided fluorograms with the further formulation of an expert opinion on each image, which the authors relied on when describing the results of the effectiveness of this software.
Statistical processing of the obtained data was carried out using the methods of parametric and nonparametric analysis. Statistical analysis was performed using the STATISTICA 13.3 software. The accumulation, adjustment, systematization of the initial data and the visualization of the results obtained were carried out in Microsoft Office Excel 2016 spreadsheets.
Results. When working with AI, the authors obtained the following data: 17 (34%) fluorograms were marked as "normal" and 18 (36%) fluorograms were marked as "pathology", which exactly coincides with the opinion of radiologists of the VOKKDC, who divided, according to According to the recommendations of the NPKTs DT DZM, the results obtained by the program were true positive (36%) - the existing pathology was revealed and true negative (34%) - the absent pathology was not identified. 15 (30%) fluorograms were marked by AI as “alarming” requiring additional in-depth analysis.
All metric parameters, except for the likelihood ratio of a negative result, were 100%.
Discussion. The results obtained confirm the diagnostic accuracy of artificial intelligence, with the further prospect of introducing artificial intelligence using a multilevel convolutional neural network into the work of a radiologist when conducting screening methods for studying radiation diagnostics.
Conclusion. The research carried out by the authors shows that the program RU.96876180.62.01.29-01 has shown its effectiveness as an additional way to control the quality of the description of fluorograms. Today, artificial intelligence using deep learning algorithms with a multilevel CNN system is a promising point in the development of modern radiation diagnostics. Using the example of an accurate assessment of the provided fluorographic images of the chest organs, the consistency of the application of the program in the field of mass prophylactic X-ray examinations is shown. This practice allows one to think about the possibility of using AI for other research methods, in particular, computed tomography of the lungs in order to detect early signs of damage to the respiratory parenchyma. This is extremely important at the time of the COVID-19 pandemic, when the question of early diagnosis of pathology in the maximum number of patients in a minimum period of time arises. However, there are also a number of difficulties. Modern CNN models require large amounts of data for their training, which have the properties of completeness and high quality, which are far from always available even with the modern level of informatization. Before being introduced into practical healthcare, these technologies must undergo thorough clinical trials and prove their effectiveness and sustainability [5].
About the authors
Asadula Raufovich Abasov
Воронежский государственный медицинский университет им. Н.Н.Бурденко
Author for correspondence.
Email: asadula_abasov@mail.ru
ORCID iD: 0000-0003-2135-6906
SPIN-code: 5422-1412
Russian Federation
Fedor Alexandrovich Bezborodov
ВГМУ им Н.Н.Бурденко
Email: tauser2012@mail.ru
ORCID iD: 0000-0002-9642-5809
SPIN-code: 9284-6010
Ekaterina Igorevna Klokova
ВГМУ им. Н.Н.Бурденко
Email: Eklokova12@gmail.com
ORCID iD: 0000-0003-4221-4712
SPIN-code: 5957-2014
References
- Tang X. The role of artificial intelligence in medical imaging research | BJR|Open | Vol 2, No 1. Birpublications.org. https://www.birpublications.org/doi/full/10.1259/bjro.20190031. Published 2019. Accessed November 13, 2019.
- Lundervold A., Lundervold A. An overview of deep learning in medical imaging focusing on MRI. Zeitschrift für Medizinische Physik. V.29. 2019 May
- Баранов И.А., Титова Л.А., Толстых Е.М., Суслова Е.Ю. Перспективы внедрения программы искусственного интеллекта в процесс диагностики заболеваний легких // Системный анализ и управление в биомедицинских системах – 2020. – Т.19 №3 – С. 165-170. [Baranov I.A., Titova L.A., Tolstykh L.A., Suslova E.Yu. Prospects for implementing an artificial intelligence program in the diagnosis process of lung diseases // Sistemniy analiz I upravlenie v biomedicinskih sistemax – 2020;(19): 165-170 (in Russ).]
- Морозов С.П., Владзимирский А.В., Кляшторный В.Г., и др. Клинические испытания программного обеспечения на основе интеллектуальных технологий (лучевая диагностика)/Серия «Лучшие практики лучевой и инструментальной диагностики». – Вып. 57 – Москва, 2019. – 51 с.
- Борисов Д.Н., Кушнирчук И.О., Севрюков В.В., Коваленко Е.И. Использование искусственного интеллекта в клинической практике // Клиническая патофизиология – 2019. –Т.25 №2 –С.26-31. [Borisov D., Kushnirchuck I., Sevryukov V., Kovalenko E. Using artificial intelligence in clinical practice // Klinicheskayz patofiziolgiya – 2019;(25): 26-31 (in Russ)]