Artificial intelligence in radiology is no longer a “future” technology. Today, AI helps radiologists detect pathological findings faster, prioritize studies by urgency, reduce routine workload, and make reporting more standardized. It is important to emphasize that AI is not a replacement for the physician. It is a clinical decision-support tool — a “second opinion” that analyzes medical images and delivers its results to the specialist within the familiar diagnostic workflow.
Radiology is one of the most mature areas for the use of medical AI. The reason is straightforward: X-ray, fluorography, mammography, and computed tomography produce digital images by design, which makes them suitable for computer-vision analysis. According to a systematic review of FDA-authorized AI/ML medical devices, by June 2024 the United States had authorized 950 medical AI solutions, 723 of which were related to radiology — approximately 76% [1]. This demonstrates that medical imaging has become the main area of practical AI implementation in healthcare.
Market data also confirms the shift from pilots to industrial-scale use. According to MarketsandMarkets, the global radiology AI market was valued at USD 0.61 billion in 2024 and is expected to reach USD 2.27 billion by 2030, with a compound annual growth rate of 24.5% [2]. The key growth drivers include rising imaging volumes, shortage of specialists, the need for early disease detection, and automation of patient routing.
The first practical scenario is study triage. An algorithm can flag examinations with probable critical findings in advance: signs of stroke on head CT, significant lung abnormalities, or suspicious areas on mammograms. This helps the physician move faster to the most urgent cases, which is especially important in emergency care, cancer screening, and high-volume preventive examinations.
The second scenario is screening and early disease detection. The strongest evidence base has been accumulated in mammography. In the MASAI randomized trial, AI-supported breast cancer screening reduced the reading workload for physicians by approximately 44% without compromising the safety of the approach [3]. In practical terms, the algorithm can take on part of the preliminary analysis, while the radiologist focuses on clinically significant or ambiguous cases.
The third scenario is quantitative assessment. In chest CT, AI can not only highlight an area of interest, but also calculate the extent of involvement, determine the location of findings, and help compare studies over time. This is particularly valuable when departments handle large volumes of data and physicians need to quickly assess the severity of changes and decide on the patient pathway.
The fourth scenario is quality standardization. Algorithms follow consistent rules and do not experience fatigue, which helps reduce variability in the initial analysis. For a medical organization, this means a more manageable flow of studies, the ability to perform retrospective analysis of imaging archives, and support for departments where radiologists’ workload is growing faster than staffing resources.
Celsus develops medical AI services for key areas of radiology: mammography, fluorography and chest X-ray, chest CT, and head CT. A major strength of Celsus is the combination of clinically oriented algorithms with integration into real-world diagnostic workflows. The company’s solutions are used in 44 regions of the Russian Federation, are integrated with Moscow’s Unified Radiology Information Service (ERIS), and are deployed in more than 200 radiology departments [4].
A core principle is that AI should not create an additional “screen” or a new routine for the physician. Celsus can be integrated with PACS, RIS, and MIS via API, supports HL7/FHIR, and sends analysis results back to the physician’s usual workstation [5]. This approach increases the likelihood that the technology will actually be used: the radiologist receives decision support exactly where they already work with the study.
Celsus.Mammography detects malignant and benign lesions, calcifications, lymph nodes, and signs of fibrocystic mastopathy; assesses breast tissue density according to ACR; and interprets results using BI-RADS [4]. This helps standardize reporting and supports physicians in screening workflows. In 2026, results of a retrospective study conducted by the L7 Mammology Center were published: the analysis included 24,150 women who underwent digital mammography from January 2020 through June 2025, and the false-negative rate for the Celsus system was 0.28% [6].
For fluorography, Celsus analyzes images for pathological changes in accordance with codes 1–23 of the “Digital Codes for Fluorography Reporting”; the company’s website indicates 93% accuracy and industrial-operation status [4]. In mass screening, the key advantage of such a solution is rapid separation of normal studies from those that require a physician’s attention. This is especially important for regional healthcare systems, mobile units, telemedicine workflows, and institutions with high volumes of preventive examinations.
Celsus.Chest CT analyzes lung CT studies, detects signs of pathology, and calculates both the volume and localization of lesions. The solution page states that one study is processed in approximately 60 seconds, pathology-detection accuracy exceeds 95%, and use cases include triage, diagnosis, screening, retrospective analysis, and remote diagnostics [5]. For a clinic, this means not just “highlighting” findings, but supporting the patient pathway from the moment the study is received to the delivery of results to the physician.
In head CT, Celsus helps analyze studies for signs of hemorrhage and ischemic stroke, as well as perform key morphometric measurements [4]. This area is especially important for emergency medicine: when stroke is suspected, the time required to make a decision is critical, and AI can help identify studies that require priority review.
The effectiveness of AI depends not only on model accuracy. Equally important are the quality of source data, validation in a real patient population, post-implementation monitoring, a clear user interface, and feedback from physicians. AI must be embedded into a clinical protocol: who sees the result, how it is considered, how discrepancies are reviewed, and how its impact on reporting time, detection rates, and specialist workload is evaluated.
Artificial intelligence already works in radiology wherever there is a clearly defined clinical task: finding a suspicious area, prioritizing studies, measuring the extent of involvement, accelerating screening, and reducing the risk of missed findings. This is the approach implemented by Celsus: AI becomes not an experimental technology, but a practical tool for the physician, the medical organization, and the patient.