Automated, standardized, quantitative analysis of cardiovascular borders on chest X-rays using deep learning for assessing cardiovascular disease
Bibliographic record
Abstract
ABSTRACTS OBJECTIVE The analysis of cardiovascular borders (CVBs) on chest X-rays (CXRs) has traditionally relied on subjective assessment, and the cardiothoracic (CT) ratio, its sole quantitative marker, does not reflect great vessel changes and lacks established normal ranges. This study aimed to develop a deep learning-based method for quantifying CVBs on CXRs and to explore its clinical utility. DESIGN Diagnostic/prognostic study SETTING Pre-validated deep learning for quantification and z-score standardization of CVBs: the superior vena cava/ascending aorta (SVC/AO), right atrium (RA), aortic arch, pulmonary artery, left atrial appendage (LAA), left ventricle (LV), descending aorta, and carinal angle. PARTICIPANTS A total of 96,129 normal CXRs from 4 sites were used to establish age- and sex-specific normal ranges of CVBs. The clinical utility of the z-score analysis was tested using 44,567 diseased CXRs from 3 sites. MAIN OUTCOMES MEASURES The area under the curve (AUC) for detecting disease, differences in z-scores for classifying subtypes, and hazard ratio (HR) for predicting 5-year risk of death or myocardial infarction. RESULTS A total of 44,567 patients with disease (9964 valve disease; 32,900 coronary artery disease; 1299 congenital heart disease; 294 aortic aneurysm; 110 mediastinal mass) were analyzed. For distinguishing valve disease from normal controls, the AUC for the CT ratio was 0.79 (95% CI, 0.78-0.80), while the combination of RA and LV had an AUC of 0.82 (95% CI, 0.82-0.83). Between mitral and aortic stenosis, z-scores of CVBs were significantly different in LAA (1.54 vs. 0.33, p<0.001), carinal angle (1.10 vs. 0.67, p<0.001), and SVC/AO (0.63 vs. 1.02, p<0.001), reflecting distinct disease pathophysiology (dilatation of LA vs. AO). CT ratio was independently associated with a 5-year risk of death or myocardial infarction in the coronary artery disease group (z-score ≥2, adjusted HR 3.73 [95% CI, 2.09-6.64], reference z-score <-1). CONCLUSIONS Fully automated, deep learning-derived z-score analysis of CXR showed potential in detecting, classifying, and stratifying the risk of cardiovascular abnormalities. Further research is needed to determine the most beneficial clinical scenarios for this method. What is already known on this topic? Previous deep learning research in the diagnosis of cardiovascular disease using chest X-rays has focused on predicting specific disease categories, forecasting cardiovascular outcomes, and automatically measuring the cardiothoracic (CT) ratio. The end-to-end learning methods that predict disease categories or outcomes are typically limited to specific conditions and often lack explainability. While the CT ratio is traditionally used in chest X-ray analysis, it often lacks well-defined normal ranges and may not effectively detect conditions such as aortic dilatation or pulmonary trunk enlargement. What this study adds To the best of our knowledge, this is the first study to propose age- and sex-specific normal values for all cardiovascular borders (CVBs) as well as the CT ratio. Utilizing 96,129 normal chest X-rays from multiple centers, we have established normal ranges for CVBs and standardized these values into z-score mapping. This approach simplifies and enhances the practicality of clinical application. The z-score mapping of CVBs has demonstrated clinical utility in diagnosing and categorizing diseases, as well as in predicting prognosis. The AI software that automatically analyzes CVBs from CXR is available for external validation and free trial use through our dedicated research website ( www.adcstudy.com ). This study has transformed the interpretation of cardiovascular configuration on chest X-ray from subjective expert assessments to objective, quantifiable, and standardized measurements expressed as z-scores.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.001 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.003 | 0.010 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".