Beyond Qualitative Interpretation to Quantitative Biomarkers: Clinical Potential and Challenges of AI-Based Cardiovascular-Border Analysis on Chest Radiographs

See the article, “Assessment of Pulmonary Thromboembolism on Chest Radiographs Using Quantitative AI-Based Cardiovascular-Border Analysis”, in volume 87 on page 895-908 (https://doi.org/10.3348/jksr.2025.0131).

Chest radiography (CXR) is one of the most accessible first-line imaging modalities. However, in acute cardiovascular conditions such as pulmonary thromboembolism (PTE), its clinical role has traditionally been limited to excluding alternative diagnoses rather than directly assessing hemodynamic compromise (1). Although acute pulmonary vascular occlusion can alter the cardiovascular silhouette, qualitative interpretation of CXR findings, or reliance on a simple cardiothoracic ratio (CTR), is insufficient to precisely characterize these changes (1).

In this issue, the study by Ok et al. (2) is methodologically notable for applying a deep-learning model to quantify specific cardiovascular borders on CXR, including the pulmonary trunk, superior vena cava/ascending aorta, left atrial appendage, and left ventricle, expressed as age- and sex-adjusted Z-scores in patients with PTE. The Z-scores, together with the CTR, were significantly elevated in the PTE cohort compared with controls and they decreased significantly after treatment in patients with central PTE.

This study had several limitations that warrant careful interpretation. Since the control group consisted of health-screening participants, common cardiopulmonary comorbidities and clinical confounders in hospitalized patients, such as acute fluid administration, may not have been adequately accounted for. The follow-up interval also varied substantially (12–57 days), and without objective stratification by treatment response, the reported changes in cardiovascular contours cannot be conclusively attributed to therapeutic effect. In addition, the diagnostic performance for distinguishing PTE from controls was modest (area under the curve, 0.506–0.658), with a sensitivity of only 44.6% at the optimal cutoff. At this stage, this approach appears to be better suited to the support of treatment monitoring than for primary diagnosis.

Despite these limitations, the study represents a meaningful methodological advance by demonstrating automated segmentation and quantification of cardiovascular borders on CXR using AI. Transforming linear distances from the spinal midline into standardized Z-scores, which effectively extends CXR into a quantitative biomarker, illustrates the growing feasibility of AI-based integration in thoracic imaging (3,4). Further validation of the robustness to patient positioning, respiratory phase, and variability across acquisition techniques, is essential. Even so, this study provides compelling evidence of the clinical potential of AI-based CXR analysis for the characterization of complex cardiovascular states.