Published
RSNA Annual Meeting (2024)
Authors
H. Shin1, D. Shin1, J. Ra2
Affiliations
1AI Engineering Division, Radisen Co. Ltd., Seoul, KOREA, REPUBLIC OF
2San Francisco, CA
Purpose
To present an AI model measuring cardiothoracic ratio (CTR) on chest x-ray radiographs (CXRs) and examine the correlation between CTR and echocardiographic diagnoses of severe left ventricular hypertrophy (SLVH) and dilated left ventricle (DLV).
Methods and Materials
71,589 CXRs of 24,689 patients were used in our study (CheXchoNet). The data was approved by the IRB and is publicly available. Patients underwent echocardiograms within one year of their X-ray scans, which diagnosed them with either SLVH or DLV. We constructed a composite binary label based on the presence of either condition (9,861/61,728 composite positive/negative). We used commercially available AI software to measure CTRs on individual CXRs. We first examined the histograms of CTR values according to the composite labels to check whether there was a significant difference between the two groups using a t-test. Then, we calculated sensitivity, specificity, area under the curve (AUC), and the Youden index with ground truth to the composite labels by changing the threshold of CTR for binary classification. Finally, we developed a classification AI model (multilayer perceptron with two fully connected layers) that takes a CTR value, age, and sex of a patient as inputs and predicts a binary composite label. We used 80% of CXRs for training and the remaining data for evaluation. We repeated the AUC and Youden index calculations.
Results
The average CTR value was significantly higher in composite SLVH/DLV positive patients compared to negative patients (mean ± std: 0.56 ± 0.07 for positive; 0.51 ± 0.07 for negatives; p-value < 0.001). The binary classification of the composite label showed an AUC of 0.69, and the Youden index of 0.30 was maximized when CTR > 0.53, reporting a sensitivity of 0.70 and specificity of 0.60. When sex, age, and CTR were used as inputs in the classification AI model, an AUC was 0.71, and the Youden index was 0.32, with a sensitivity of 0.74 and specificity of 0.58. When CTR values were not included as inputs, the AUC dropped significantly to 0.54, implying the importance of CTR measurements for accurate prediction of composite SLVH/DLV labels.
Conclusions
We propose an AI model that provides automated measurements of CTR on chest radiographs and shows its potential to predict the echocardiographic diagnoses of left ventricular structural abnormalities.
Clinical Relevance/Application
Automated and precise measurement of CTR using AI models can assist radiologists in the identification of cardiomegaly on chest radiographs. This may facilitate clinical decision-making to pursue confirmatory imaging with echocardiography for earlier recognition of left ventricular hypertrophy and dilation.