Evaluating convolutional neural network models for automated abdominal aortic calcification scoring in chronic kidney disease patients across multiple centers.

Background: Abdominal aortic calcification (AAC) is a condition particularly common in patients with chronic kidney disease (CKD). This study aims to assess the accuracy and generalizability of convolutional neural networks (CNNs) in automating AAC scoring in CKD patients. Methods: Abdominal aortic X-ray images from five hospitals were employed to evaluate three CNN architectures (DenseNet, RegNet and ResNet) using the internal hold-out and five-fold cross-validation method. Performance metrics included intraclass correlation coefficient (ICC), R-squared value (R2), and root mean square error (RMSE), along with standard binary classification measures, across the independent training, validation, internal test, and external test sets. Results: The present study analyzed 2,853 X-ray images. DenseNet outperformed ResNet and RegNet in the internal test set, with an RMSE of 2.141, a Pearson correlation of 0.926, a Spearman's rank correlation of 0.908, and an ICC of 0.919. On the external test set, the performance of DenseNet was slightly better than that of ResNet. For the AAC severity classification, all models performed comparably in the "none or mild" and "moderate" categories, with an accuracy metric of over 80% and 70%, respectively, and a sensitivity of above 60% and 65%, respectively. In the "severe" category, the sensitivity of DenseNet and ResNet was higher (77.30% and 75.50%, respectively), while the sensitivity of RegNet lagged at 67.30%, although the other metrics exceeded 75% for all models. Conclusions: In conclusion, the DenseNet and ResNet models exhibit strong performance and reliability in AAC scoring for CKD patients, suggesting their viability for clinical use. Their effectiveness in differentiating AAC severity highlights the potential for enhancing diagnostic accuracy in medical settings.