Prevalence of COVID-19-associated pneumonia signs on chest computed tomography in cancer patients: the ARILUS study

Abbreviations:

• ACOD – Arkhangelsk Clinical Oncology Dispensary
• AI – artificial intelligence
• COVID-19 – CoronaVirus Disease 2019
• CI – Confidence Interval
• CT – Computed Tomography
• ICD-10 – International Classification of Diseases, 10th Revision
• INIPA – insurance number of individual personal account
• GIT – Gastrointestinal Tract
• MN – Malignant Neoplasm
• NAO – Nenets Autonomous Okrug
• OR – Odds Ratio
• PR – Prevalence Ratio
• PBCR – Population-Based Cancer Registry

The 2019 coronavirus disease (COVID-19) pandemic significantly affected the diagnosis and treatment of malignant neoplasms (MN). During the pandemic, a substantial decline in global MN incidence rates was observed [1][2], including in the Russian Federation [3], primarily due to quarantine measures. Breast and cervical cancer screening programs, along with other cancer screenings, were suspended during the pandemic and later gradually resumed with scheduled visit intervals to reduce staff density and enhance infection control protocols [4]. In the Arkhangelsk region, the decrease in MN incidence during the COVID-19 pandemic was largely attributed to reduced detection of early-stage cervical, lung, and colorectal cancers [5].

Compared to other visualization methods, chest computed tomography (CT) has one of the highest sensitivity rates in detecting lung abnormalities associated with COVID-19 pneumonia. CT can identify characteristic lung changes in COVID-19 patients even before positive laboratory test results are obtained [6]. Meanwhile, artificial intelligence (AI) algorithms enable highly accurate detection of minimal lung changes on CT scans in asymptomatic and mild COVID-19 cases that do not require hospitalization. An independent evaluation of one such algorithm, developed by IRA Labs LLC (Moscow, Russia), demonstrated high diagnostic performance in detecting COVID-19 pneumonia signs: Receiver Operating Characteristic Area Under the Curve (ROC AUC) – 0.98, sensitivity – 0.95, specificity – 0.94, and accuracy – 0.94 [7].

During the COVID-19 pandemic, population-based cancer registries (PBCRs), with their large sample sizes and broad population coverage, were well-suited for monitoring shifts in cancer stage distribution at initial diagnosis and survival analysis. However, challenges arose in determining the exact stage at diagnosis due to delays in surgical interventions and pathological assessments. Given that chest CT scans were frequently performed for MN patients during outpatient visits and hospitalizations in the COVID-19 era, signs of pneumonia—including clinically silent cases – could also be tracked.

The Arkhangelsk Regional and Nenets Autonomous Okrug (NAO) Cancer Registry was established in 1998 and has maintained satisfactory completeness in recording and tracking MN patients from initial diagnosis to outcome since 2000. Data on deceased MN patients are updated monthly by cross-referencing mortality records from the Arkhangelsk Regional Medical Information and Analytical Center with the registry database. The registry’s completeness, accuracy, and timeliness have undergone multiple international audits, including within the “Cancer on Five Continents”, CONCORD, and VENUSCANCER programs [8–10]. The registry also contains codes for the immediate causes of death in cancer patients, enabling the estimation of cancer-specific survival and non-cancer mortality rates.

Study objective: to assess the prevalence of COVID-19-associated pneumonia signs in a population-based cohort of MN patients using AI-assisted chest CT analysis.

MATERIALS AND METHODS

A cross-sectional study was conducted as part of the Arkhangelsk Research on the Impact of Multitarget Artificial Intelligence for Computed Tomography on Reducing Non-Cancer Lethal Outcomes in Patients with Malignant Neoplasms (ARILUS) project [11].

Data Collection

To achieve the study’s objective, all chest CT scan series from the central medical imaging archive of the Arkhangelsk Clinical Oncology Dispensary (AСOD) were extracted for the period from April 1, 2020, to December 31, 2021, corresponding to the COVID-19 pandemic.

Chest CT scans during this period were performed for two main indications: routine diagnostic workup and staging of MN; exclusion of viral pneumonia signs in hospitalized patients who developed COVID-19 symptoms (at admission, all patients were required to have no respiratory symptoms and a negative nasopharyngeal/oropharyngeal swab test for COVID-19). A total of 11,173 CT scans were performed during this period, of which 3533 were conducted for healthcare workers or private-pay patients without MN.

Selection of valid images for processing was performed considering AI algorithm limitations, specifically excluding cases with lung atelectasis based on radiologists’ interpretation reports, contrast-enhanced CT series based on DICOM tag analysis (ProtocolName Tag 0018,1030).

After deidentification, all valid CT series were sent for analysis by IRA Labs LLC (Moscow, Russia) AI system. At this stage, cases with severe motion artifacts, slice thickness >1.5 mm and incomplete lung scanning area were excluded. The total number of patients with valid CT studies was 1542.

After processing all images by the AI algorithm, they were sent via secure channel to ACOD using the key – insurance number of individual personal account (INIPA), and merged with the database of PBCR of Arkhangelsk Region and NAO, extracted on 15.04.2024. The total PBCR database contained information on 137,773 patients registered with MN diagnosis at the data extraction date, with INIPA data available for 62,988 patients. INIPA data in PBCR of Arkhangelsk Region and NAO have been recorded since January 1, 2021 for all newly registered patients, and during 2021 INIPA numbers were added for follow-up category patients. For the entire registration period (01.01.2000–15.04.2024), completeness of INIPA data was 45.7%. Data extraction from PBCR was necessary to establish causes of death in patients with available INIPA data. Among patients with valid CT studies, 394 had cancer but their INIPA data were unavailable, making outcome assessment impossible (the ARILIS study part on outcome assessment is considered separately and will be presented independently). Patients without INIPA data were excluded from the study.

The final analysis included 1148 patients (Fig. 1).

The combined database for analysis included the following variables: patient identification code, INIPA, age at the time of CT scan, sex, type of residential locality (urban/rural), MN diagnosis code under the International Classification of Diseases, 10th Revision (ICD-10). The data for all the variables for the study period were 100 per cent complete.

AI Algorithm for COVID-19-Associated Pneumonia Diagnosis

To detect qualitative and quantitative (percentage of lung involvement) infiltrative changes characteristic of COVID-19 viral pneumonia (classified as U07 under ICD-10) [12], we used a medical AI-based software developed by IRA Labs LLC (Moscow, Russia): “Software for CT Scan Analysis Using AI Technology ‘Intelligent Radiology Assistants’”, Technical Specifications (TU): 58.29.32-001-44270315-2021, Registration Certificate (Roszdravnadzor): No. RZN 2024/22895¹ (Fig. 2).

Depending on the presence or absence of signs of pneumonia associated with COVID-19, all patients included in the database (n = 1148) were divided into two groups: without signs of pneumonia (n = 592, 51.6%) and with signs of pneumonia (n = 556, 48.4%). Notably, in most patients in the second group – 530 (95.3%) – the lesion volume was less than 25% (CT-1), signs of moderate (CT-2) and moderate (CT-3) pneumonia were found in 22 (4%) and 4 (0.7%) patients.

The AI-based software utilized in this study features a multitarget algorithm, enabling concurrent assessment of pulmonary emphysema, aortic and pulmonary trunk diameters, coronary artery calcification, bone mineral density of thoracic vertebrae (Fig. 2).

Statistical Methods

For ease of analysis and interpretation, all variables were categorized as ordinal (age, coronary artery calcification, bone mineral density), nominal (ICD-10 diagnosis codes/MN groups), binomial (sex, pulmonary emphysema, aortic aneurysm/dilation, pulmonary trunk dilation). Frequency distribution comparisons were performed using Pearson’s chi-square test.

Association strength was quantified through unadjusted (univariate) and adjusted (multivariate) prevalence ratios (PR) with 95% confidence intervals (CI). PR was calculated using Poisson regression (both univariate and multivariate forms) with the standard formula. This approach was selected over logistic regression to avoid overestimation of effect measures given the high prevalence of the outcome variable. Only predictor variables showing an association with CT-detectable pneumonia signs at significance level (p) less than 0.15 in preliminary analyses were retained in the final adjusted model.

The reference categories included age below 40 years, urban residence, absence of signs of pulmonary emphysema, aortic dilation and osteoporosis, MNs of the female genital organs (ICD-10 codes C51–58), as well as absence of signs of coronary artery calcification. For the rank variables, the significance level for the linear trend was calculated by including the categories listed in the table as continuous variables.

In all statistical procedures, the critical significance level (p) was set at 0.05.

Statistical analysis was performed using Stata v.18 (Stata Corp., TX, USA).

RESULTS

The distribution of characteristics by study groups is presented in Table 1.

In the analyzed cohort, sex distribution was approximately equal, while CT signs of pneumonia were detected significantly more frequently in men than women. The distribution of age categories was skewed towards older age groups in patients with COVID-19 pneumonia (Table 1).

In the study population, the most common tumor sites were the lower gastrointestinal tract (GIT) (20.7%), lung cancer (16.1%), breast cancer (13.2%), and upper GIT cancer (12.2%). The frequency of pneumonia detection varied significantly depending on cancer location: the highest rate was observed in lung cancer patients (70.3%), followed by head and neck malignancies (59.6%) and upper GIT cancers (57.1%). In patients with skin and soft tissue malignancies, the frequency of COVID-19 pneumonia was 50%; breast cancer (47.7%), urinary tract (48.7%), prostate cancer (42%), lower GIT (37.8%), female genital tract (34.4%), and other sites (29.2%) (Table 1).

Indicators of cardiovascular and pulmonary pathology, as well as signs of osteoporosis detected by AI-based multitarget CT analysis, were unevenly distributed. Signs of pulmonary emphysema, aortic dilation, presence and severity of coronary artery calcification, and osteoporosis were more prevalent in the group with COVID-19 pneumonia. However, no significant differences were found between the two groups in the distribution of aortic aneurysm/dilation and pulmonary trunk dilation.

The results of univariate and multivariate Poisson regression analyses are presented in Table 2.

In the unadjusted model, pneumonia sign prevalence was significantly higher in men (by 40%) than women, but after adjustment for other factors, gender differences became insignificant. Prevalence of COVID-19 signs on CT progressively increased with age (p for trend <0.001), but intergroup differences reached statistical significance only for age group ≥80 years, where COVID-19 signs were recorded 1.8 times more frequently than in reference group (<40 years). However, after including other variables in the multivariate model, all age differences in COVID-19 sign prevalence disappeared.

Rural residents had 13% higher prevalence of COVID-19 signs on CT in univariate analysis, but differences did not reach statistical significance (p = 0.058). The multivariate model showed no differences between urban and rural cancer patients in COVID-19 sign prevalence.

In the multivariate model, prevalence of COVID-19–associated pneumonia signs was 51–87% higher in lung cancer, head and neck tumors, and upper GIT tumors compared to female genital tumors. It was also significantly higher in breast cancer. For lower GIT tumors and prostate cancer, pneumonia sign prevalence did not differ significantly from the reference group.

Pulmonary emphysema signs were significantly associated with COVID-19 in both univariate and multivariate analyses. In patients with emphysema, after adjustment for all available factors, prevalence of COVID-19-associated pneumonia signs was 25% higher than in those without emphysema.

Radiological signs of aortic aneurysm or dilation were significantly associated with COVID-19 in univariate analysis, but no independent associations between this sign and pneumonia signs on CT were found. Another potential predictor of increased cardiovascular mortality – coronary artery calcification level measured by Agatston score – was associated with higher probability of COVID-19 pneumonia in both univariate and multivariate models. Adjusted PRs for pneumonia signs increased from 1.24 (95% CI 1.05–1.47) to 1.61 (95% CI 1.36–1.90) for scores 1–99 to ≥300 compared to the no–calcification reference group. Aortic aneurysm and pulmonary trunk dilation were associated with pneumonia signs on CT at significance level >0.15 in univariate modeling and were not included in the multivariate model.

Prevalence of COVID-19 signs progressively increased with osteoporosis severity (p = 0.015 for trend) in univariate analysis. After including other factors in the model, PRs decreased to statistically insignificant levels.

DISCUSSION

Our study found that half of MN patients admitted to AСOD in 2020-2021 for specialized treatment had signs of COVID-19-associated pneumonia, with most cases showing ≤25% lung involvement. We identified independent factors associated with COVID-19 pneumonia detection: cancer topography in lungs, head/neck, upper GIT, and breast, as well as presence of emphysema and coronary calcification on CT.

The COVID-19 pandemic significantly impacted cancer diagnosis and treatment organization: mortality reached 23.4% in hospitalized patients [13], mainly from pneumonia against 3-15-fold increased thrombosis rates [14]. Approximately 15–30% of hospitalized patients developed acute respiratory distress syndrome, increasing mortality risk [15]. Lung involvement volume on CT predicts COVID-19-associated pneumonia mortality [16]. COVID-19-associated pneumonia can be asymptomatic in 50% of patients [17][18], which is consistent with our findings.

Risk factors for severe disease include older age, male sex and comorbidities. In our study, age and male sex were not independent pneumonia risk factors. However, older patients and men in general population have higher risk of coronary vessel atherosclerosis and emphysema – factors that showed independent effects on pneumonia risk in our study. This highlights the need for comprehensive risk assessment.

Our analysis showed substantially higher pneumonia prevalence in lung, upper GIT, and head/neck cancer patients versus cohort average. Smoking is the most important modifiable risk factor for these MNs. While some early studies suggested protective association between smoking and COVID-19 severity (“smoker’s paradox” [19]), most authors link tobacco smoking with increased risk of symptomatic SARS-CoV-2 infection and disease progression [20][21]. A recent US analysis found current tobacco smoking significantly associated with increased hospitalization (Odds Ratio (OR) 1.72; 95% CI: 1.62–1.82; p < 0.001), ICU admission (OR 1.22; 95% CI: 1.10–1.34; p < 0.001) and all-cause mortality (OR 1.37; 95% CI: 1.20–1.57; p < 0.001) after adjustment [22]. Our assumption about smoking and CT opacity probability is further supported by more frequent pneumonia signs in men, who smoke more.

Among healthcare AI services, radiology has the most products. Russia’s largest AI radiology project is the Experiment on Using Innovative Computer Vision Technologies for Medical Image Analysis in Moscow Healthcare System, which processed >12 million radiological studies. In this project, IRA Labs LLC leads the maturity matrix for comprehensive chest CT AI by performance quality (ROC AUC) [7]. Using this algorithm detects pneumonia signs more frequently than unaided interpretation.

Study strengths include population-based design, using all available CTs for analysis. PBCR data completeness was previously validated at high levels. For example, the mortality/incidence ratio in Arkhangelsk Region and NAO PBCR for 2008–2017 is 0.58, comparable to Eastern European registries. Death certificate only (DCO) rate is 4.5%, higher than in European countries, but this can be explained by high autopsy rates (>60% in MN patients) and, subsequently, a high rate of accident postmortem cancers. The difference between cases registered in PBCR during 2008–2017 and annual reports versus 5-year updated data is <3% [10].

Reliable PBCR patient data enabled assessment of lung infection prevalence in a representative MN patient cohort and provided high statistical power. Another key advantage is using objective AI-derived radiological criteria, revealing higher viral pneumonia probability with chronic lung (emphysema) and coronary vessel diseases.

Study Limitations

Chest CT was used both for routine MN staging/diagnosis and COVID-19 symptom evaluation. Separating these streams was impossible in this design. However, hospitalization rules during pandemic required baseline disease exclusion, making hospital-acquired infection risk random and dependent only on analyzed factors.

Correlation of CT pneumonia signs with SARS-CoV-2 PCR results was impossible. However, during the pandemic, other viral infections rarely caused pneumonia [23]. Moreover, many cases with clear clinical manifestations had initial negative nasopharyngeal PCR, with COVID-19 later confirmed by repeat PCR or seroconversion [24].

Incomplete INIPA data in PBCR. However, we assume most patients undergoing chest CT during 04/2020–12/2021 were in follow-up category by 2021. Among 1542 patients with validated CTs and AI analysis, 74.4% matched PBCR records.

The cancer registry doesn’t collect smoking data, preventing independent assessment of this factor’s impact on CT pneumonia sign risk.

Future Research Directions

Clinical significance of incidentally detected CT pneumonia signs in MN patients will be assessed in survival and cause-of-death analyses. Since age was an independent mortality risk factor in some COVID-19 pneumonia studies [25][26], we plan to analyze the overall cohort survival by age.

CONCLUSION

In this analysis, half of the patients admitted to Arkhangelsk oncology center during COVID-19 pandemic showed CT signs of pneumonia. Independent factors associated with COVID-19 pneumonia detection include tumor location (lung, head/neck, upper GIT, breast) and CT signs of emphysema and coronary calcification.

AUTHORS CONTRIBUTION

Andrey A. Dyachenko – development of the research concept, development of the methodology, analysis of the work, drafting the manuscript, critical revision of the text, interpretation of the research results, final approval of the manuscript. Andrej M. Grjibovski – development of the methodology, statistical analysis and its interpretation, critical revision of the text, interpretation of the research results. Maxim A. Bogdanov, Dmitriy V. Bogdanov, Valeria Yu. Chernina – analysis of the work, critical revision of the text, interpretation of the research results. Ekaterina A. Nazarova – analysis of the work, critical revision of the text, interpretation of the research results. Anna A. Meldo, Mikhail G. Belyaev – critical revision of the text, interpretation of the research results. Victor A. Gombolevsky, Mikhail Yu. Valkov – scientific supervision, development of the research concept, development of the methodology, critical revision of the text, interpretation of the research results. All authors approved the final version of the article.

ВКЛАД АВТОРОВ

А.А. Дяченко – разработка концепции исследования, разработка методологии, анализ работы, составление черновика рукописи, критический пересмотр текста, интерпретация результатов исследования, окончательное утверждение рукописи. А.М. Гржибовский – разработка методологии, статистический анализ и его интерпретация, критический пересмотр текста, интерпретация результатов исследования. М.А. Богданов, Д.В. Богданов, В.Ю. Чернина – анализ работы, критический пересмотр текста, интерпретация результатов исследования. Е.А. Назарова – анализ работы, критический пересмотр текста, интерпретация результатов исследования. А.А. Мелдо, М.Г. Беляев – критический пересмотр текста, интерпретация результатов исследования. В.А. Гомболевский, М.Ю. Вальков – научное руководство, разработка концепции исследования, развитие методологии, критический пересмотр текста, интерпретация результатов исследования. Все авторы утвердили окончательную версию статьи.