A multi-stage deep learning network for prenatal diagnosis of coarctation of the aorta

Background: Coarctation of the aorta (CoA) is a common congenital cardiovascular disorder. Severe cases may cause neonatal shock or heart failure. Accurate prenatal diagnosis is critical to reducing mortality and enabling timely medical intervention. Fetal echocardiography is widely used for prenatal CoA detection, but its diagnostic performance is limited by relatively high false-positive and false-negative rates. Recent advances have proposed deep learning to assist diagnosis. However, current approaches struggle to extract meaningful features from the complex distributions of fetal echocardiographic images. These limitations emphasize the urgent need for more advanced methodologies to improve diagnostic accuracy in this field. Purpose: To develop a deep learning network that effectively integrates local and global features for extracting discriminative characteristics from fetal echocardiograms, thereby enabling precise prenatal diagnosis of fetal CoA. Methods: This study introduces CoA-Net, a multi-stage deep learning network for prenatal CoA diagnosis through local-global feature integration. CoA-Net includes four sequential stages, each equipped with a local feature extractor and a global feature extractor to ensure comprehensive feature extraction. The local extractor adopts lightweight convolutional layers and a local attention mechanism—this mechanism assigns input-adaptive importance weights, allowing accurate lesion localization and detailed local feature extraction. The global extractor incorporates a sparse global attention mechanism to capture the most discriminative global features, supporting reliable contextual modeling between the aortic arch and adjacent structures. To validate CoA-Net, a dataset of 488 samples (including CoA cases and healthy controls) was constructed, with five-fold cross-validation for train-test partitioning. Extensive experiments were conducted on this dataset to evaluate model performance, using multiple state-of-the-art deep learning classification methods. Results were also compared with clinician diagnoses. Cross-validation t-tests were used to determine significant differences between CoA-Net and other approaches, and the Benjamini–Hochberg False Discovery Rate (BH-FDR) method was applied to correct for multiple comparisons when analyzing significant results. Results: CoA-Net outperforms other state-of-the-art networks, with statistically significant improvements in balanced accuracy, F1 score, and Matthews correlation coefficient (MCC) (p < 0.05 after BH-FDR correction) and a large effect size (Effect Size > 0.8). This indicates CoA-Net not only has statistical superiority but also delivers substantial, perceptible improvements in CoA diagnostic accuracy. Specifically, CoA-Net achieves 82.63% balanced accuracy, 68.81% F1 score, 69.73% MCC, and 0.867 area under the curve (AUC). To further evaluate the clinical utility of CoA-Net, we compared its performance against diagnoses made by experienced clinicians. The clinicians achieved a balanced accuracy of 77.11%, an F1 score of 53.76%, an MCC of 44.54%, and an AUC of 0.863. Notably, CoA-Net demonstrated comparable or even superior performance across these metrics, underscoring its potential as an effective decision-support tool in real-world clinical settings. Finally, ablation studies conclusively validated the indispensable role of the proposed local and global feature extractors, confirming that their synergistic contribution is critical to the model's overall high performance. Conclusions: This study proposes CoA-Net for fetal CoA diagnosis. The model exhibits excellent classification performance and holds potential to support clinicians in achieving more precise and reliable fetal CoA diagnoses.