A deep learning-based automated segmentation model successfully discriminated higher aortic arch calcification burden on chest radiographs with an AUC of 0.873, and identified AoAC as an independent risk factor for hypertension (OR 26.48).
Observational (n=347)
No
Does a deep learning-based automated segmentation model accurately detect and quantify aortic arch calcification on chest radiographs compared to manual scoring?
A deep learning model can automatically segment and quantify aortic arch calcification on routine chest radiographs, which is independently associated with hypertension.
Effect estimate: AUC 0.873 (95% CI 0.717-1.029)
Aortic arch calcification (AoAC) is an established independent predictor of coronary heart disease and broader cardiovascular outcomes. We have developed a deep convolutional neural network that enables automated detection and quantification of AoAC from routine chest radiographs (X-ray). In this retrospective study, three radiologists annotated AoAC on chest radiographs, including image-level AoAC status, manual AoAC severity score, and pixel-wise segmentation masks. We trained a deep-learning segmentation model to acquire the contour of the aortic arch calcification. The model was further evaluated for its ability to discriminate radiographs with higher AoAC burden (manual AoAC score > 25%) from the remaining cases using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. In addition, radiomics features extracted from the predicted AoAC regions were analyzed as post-segmentation quantitative phenotypes, and their associations with the manual AoAC severity score were assessed using Spearman’s ρ. Multivariable logistic regression was further used to evaluate the association between AoAC and clinical variables. A total of 347 patients with a median age of 77 years (interquartile range 77 ± 6), 224 women, were included and partitioned in a training set of 242 cases, and tseting dataset included 69 cases. Radiomics parameters characteristics of the aorta arch, which area, perimeter, area to perimeter ratio, sphericity, sphere imbalance, maximum 2D diameter, minimum 2D diameter, long axis length, short axis length, entropy, and energy showed correlation with the AoAC. Moreover, multivariable logistic regression analysis demonstrated that AoAC was independently associated with hypertension (OR = 26.48; 95% confidence interval CI: 3.29–213.10; P = 0.002). We built an automated segmentation and quantification framework to assess the AoAC at chest radiographs, which could be used for the community checkup populations. It can quickly identify the degree of aortic arch calcification and assess cardiovascular risk.
Yin et al. (Mon,) conducted a observational in Aortic arch calcification (n=347). Deep learning-based automated segmentation model (U-net) vs. Manual radiologist annotation was evaluated on Discrimination of higher aortic arch calcification burden (manual AoAC score >25%) (AUC 0.873, 95% CI 0.717-1.029). A deep learning-based automated segmentation model successfully discriminated higher aortic arch calcification burden on chest radiographs with an AUC of 0.873, and identified AoAC as an independent risk factor for hypertension (OR 26.48).