Role of Artificial Intelligence Enabled Chest X-ray Interpretation to Detect Lung Nodules in a Low-income Country

Abstract Introduction: Lung cancer is one of the most widely diagnosed cancers at the global and national level. It is the leading cause of cancer-related deaths worldwide. Low-dose computed tomography (LDCT) is recommended as an early screening method for high-risk groups. The use of AI for chest X-ray (CXR) analysis is increasing globally. This study aimed to determine the role of AI-based CXR analysis for early lung nodule detection. Methods and Material: This prospective single-arm cohort study was conducted at Tikur Anbessa Specialized Hospital, Addis Ababa, Ethiopia from January to September 2024. Ethical approval was obtained from the Institutional Review Board of the College of Health Sciences, Addis Ababa University. Consecutive CXRs considered technically acceptable from adults 18 years and above who consented based on AI-based CXR reading were included in the study. Confirmed TB cases and patients not willing to come were excluded. Data on socio-demographics, indications for CXR, risk factors, and AI-based CXR findings were collected. The AI-based CXR software (Qure.ai India) processed all chest X-rays that met the eligibility requirements throughout the study period. Once AI-based CXR software flagged an image with a pulmonary nodule, the image was reevaluated by senor radiologist. If the radiologist agreed, LDCT was performed with a diagnostic biopsy done if lung cancer was suspected. Result: A total of 6978 patients were included, with a median age of 48 years, including 52% women. 6803 CXRs pushed to the AI-based CXR software were valid and processed. Among these images, 807 (11.6%) scans were flagged for nodules (high risk in 158 and low risk in 649). Senior radiologists evaluated a total of 392 (48.6%) scans, of which 49 (12.5%) agreed in interpretation and were eligible for LDCT scan. From those aligned with the radiologist's interpretation, only 9 (18.4%) were high risk. The remaining 40 (81.6%) were low-risk. LDCT was performed for 36 (73.5%). From 36 suspected nodules with LDCT performed, 28 patients were classified by lung-RADS: 11 patients were classified as 0 lung-RADS, 3 patients as 1 lung-RADS, 7 patients classified as 2 lung-RADS, 2 patients classified as 3 lung-RADS, 2 patients classified as 4A lung-RADS, 1 patient classified as 4B lung-RADS, 2 patients classified as 4X lung-RADS, and no patient classified as Lung-RADS classification S. For two patients, lung biopsy was done and lung cancer confirmed. Conclusion: This ongoing study clearly shows that AI-based CXR software offers improved detection of pulmonary nodules on CXR.