AI-Based Chest X-Ray Tool to Detect Lung Nodule in a Tuberculosis Endemic Population: A Screening Clinical Series Study

Introduction: In countries with high tuberculosis prevalence, early lung cancer (LC) detection can be confounded by nodules from latent TB (LTB), delaying diagnosis. Limited evidence exists on TB and LC in integrated screening, and AI performance in TB-endemic settings remains unknown. This study evaluates AI’s effectiveness in screening LC in such settings. Methods: This is a screening clinical series where qXR Lung Nodule, Lung Nodule Malignancy Score (LNMS), and TB algorithm were used. Patients who were flagged for nodule (high or low-risk by LNMS) and gave consent were enrolled. Overall, AI processed CXR of 25166 patients. 1683 were flagged for nodules; 318 gave consent. 36 patients who had other cancer were excluded. Agreement between AI and radiologist is reported. For patients who underwent CT, the positive and negative predictive value of AI is reported with Lung-RADS as reference standard (>=4A as high-risk). Results: 282 (180 high-risk and 102 low-risk) were included. Mean age was 54.4 and 145 (51.4%) were females. 155 (55%) were also flagged as TB presumptive by AI. The proportion agreement between radiologist and AI for nodule was 53.5% (47.7-59.3). 94 (54 high-risk and 40 low-risk) underwent CT scan. PPV of high-LNMS was 44.4% (32.0-57.6). NPV of low-LNMS was 77.5% (62.5-87.7). PPV/NPV in those with concurrent AI’s TB flag 46/73 and for those with TB negative it was 42/100. 36 patients underwent biopsy and 14 (9 high-LNMS, 5 low-LNMS) were confirmed as cancer. Six of the 14 cancer cases were confirmed TB negatives and in four of these patients AI detected high-risk nodule. Conclusions: There is considerable overlap in AI prediction of TB and nodules. PPV was comparable when analyzed by AI’s TB results. However, NPV was higher when there is no concurrent TB prediction. LC investigations may benefit high-risk LNMS patients once TB is ruled out.