Intelligent multimodal classification of infantile vascular diseases based on ultrasound images

Infantile hemangiomas (IH) and venous malformations (VMs) often exhibit overlapping echogenic features in grayscale ultrasound images, necessitating color Doppler ultrasound for accurate diagnosis. Furthermore, some venous malformation patients develop painful phleboliths (vein stones), whose early detection is critical for treatment. However, in real-world primary care and emergency scenarios, color Doppler imaging is frequently unavailable. Current single-modal or single-task deep learning models struggle to extract effective features under these compounded clinical challenges. A classification dataset was constructed comprising 1,379 grayscale-color Doppler ultrasound image pairs (507 pairs in the IH group, 507 pairs in the VM group, and 365 pairs in the normal skin group), along with a detection dataset of 95 ultrasound images of phleboliths. We propose IVDNet, a task-driven integrated framework for the specific clinical scenario of pediatric vascular anomalies. It integrates an EfficientNet-B0 backbone to extract structural and blood flow features, employing squeeze-and-excitation attention, a pyramid pooling module, and multi-scale decoding. Noted that a dynamic score suppression mechanism is utilized to adapt to scenarios with missing color Doppler data. IVDNet achieved an accuracy of 95.6% (95% CI: 91.2%–98.1%) and a macro-F1 score of 0.953, alongside a mean mAP50 of 0.934 (95% CI: 0.802–0.961) for phlebolith detection. Under strictly fair multimodal baseline comparisons, IVDNet maintained a stable advantage. In missing modality stress tests, the model maintained a robust 94.5% accuracy even when Doppler inputs were masked. The proposed IVDNet demonstrates preliminary feasibility in simultaneously classifying infantile vascular anomalies and detecting phleboliths on an internal retrospective dataset. By mathematically adapting to color Doppler absence, it provides a highly resilient auxiliary diagnostic reference for resource-variable clinical workflows. Diagnosing certain vascular conditions in infants, like infantile hemangiomas and venous malformations, can be challenging because they often look similar on standard black-and-white ultrasound scans. To get a clearer picture, doctors frequently use an additional ultrasound method called color Doppler, which visualizes blood flow. A further complication is that some infants with venous malformations develop painful vein stones, also known as phleboliths, which are important to identify for proper treatment. To address these diagnostic challenges, we created a new artificial intelligence tool called IVDNet. This tool analyzes both black-and-white and color Doppler ultrasound images to automatically distinguish between the two conditions and to detect the presence of vein stones. By combining information from both types of images, IVDNet makes more accurate diagnoses. A significant advantage of our tool is its flexibility; it can still function effectively using only black-and-white images if color Doppler scans are unavailable. In our tests, IVDNet demonstrated a 95.6% accuracy rate in classifying these conditions, outperforming other AI models, and was also highly effective at finding vein stones. This tool has the potential to help doctors make faster, more reliable diagnoses, reducing the chance of errors and helping guide better treatment decisions for infants.