To evaluate the value of the MobileNetV4 model in identifying aortic arch and its branch anomalies in prenatal ultrasound images.

A total of 12 284 prenatal ultrasound images of vascular rings diagnosed at Shenzhen Maternity and Child Healthcare Hospital from March 2019 to March 2024 were retrospectively collected, including 8913 images with vascular ring anomalies and 3371 normal control images. The images were divided into training, testing, and validation sets at a ratio of 8∶1∶1 according to disease type. A deep learning model based on MobileNetV4 was constructed for the automatic diagnosis of aortic arch branch anomalies, and its performance was compared with five mainstream models applied in different scenarios, including MSPAnet-50, MedMamba, RepVit, ResNet-50, and ViT. Evaluation metrics included accuracy, sensitivity, specificity, precision, area under the receiver operating characteristic (ROC) curve (AUC), and F1 score. The model's comprehensive performance in diagnosing vascular ring anomalies was assessed from multiple perspectives by plotting ROC curves, confusion matrices, Grad-CAM heatmaps, and t-SNE feature visualizations.

In the diagnostic tasks involving double aortic arch (DAA), right aortic arch (RAA), aberrant right subclavian artery (ARSA), and normal images, the MobileNetV4 model demonstrated superior performance, with accuracy, specificity, sensitivity, precision, AUC, and F1 score reaching 94.10%, 98.03%, 94.09%, 94.21%, 99.12%, and 94.04%, respectively. In the heatmaps, only the MobileNetV4 model showed highly concentrated attention on key anatomical structures such as the aortic arch and its branches, exhibiting clearly defined attention regions and sharp boundaries. The confusion matrix revealed that the MobileNetV4 model's predictions were strongly clustered along the diagonal for all categories, indicating high overall diagnostic accuracy, with 100% accuracy achieved on normal control samples. The ROC curves showed that the AUC values for DAA, RAA, ARSA, and normal images were 99.49%, 98.47%, 100%, and 98.51%, respectively, outperforming other comparison models. In the t-SNE visualization, the feature distribution of MobileNetV4 presented more distinct clustering boundaries.

The MobileNetV4 model demonstrates excellent performance in the intelligent recognition of aortic arch branch anomalies in ultrasound images, providing effective technical support for prenatal ultrasound examinations and promoting the intelligent screening and standardized workflow of vascular ring anomalies.