Neonatal spinal ultrasound is a critical tool for screening Tethered Cord Syndrome (TCS), but its clinical application is constrained by the limited field-of-view of standard probes. This necessitates a fragmented, multi-image examination process that can lead to diagnostic inconsistencies and prolonged assessment times. To address this limitation, we present a fully automated framework for the seamless stitching of lumbar and sacral ultrasound images and subsequent vertebral identification. Our approach first employs a segmentation model, based on a U-Net architecture with an EfficientNetB7 encoder and a hybrid Dice-Focal loss function, to accurately delineate vertebral bodies and the spinal cord. We then introduce a novel alignment and stitching pipeline that uses these segmented anatomical structures as robust anchors. For unannotated images, a similarity-based vertebral level prediction algorithm was developed, achieving an accuracy of 94% in our tests. Preliminary clinical validation demonstrates that the proposed system achieves high fidelity in image stitching and robust accuracy in detecting spinal cord abnormalities, offering significant technical support for the early and reliable diagnosis of neonatal spinal pathologies.
Chen et al. (Thu,) studied this question.