This study develops an end-to-end Vision Transformer (ViT)-based artificial intelligence system for ultrasound-based diagnosis of placenta accreta spectrum (PAS), aiming to improve the accuracy and efficiency of prenatal screening using standardized ultrasound video inputs.
Placenta accreta spectrum (PAS) is a life-threatening obstetric disorder involving abnormal placental invasion into the uterine wall, which is associated with severe maternal and neonatal complications. Despite advances in imaging, prenatal diagnosis remains challenging due to variability in ultrasound interpretation and reliance on operator expertise. This study will establish a standardized ultrasound video acquisition protocol and develop a deep learning-based model using Vision Transformer (ViT) architecture to process dynamic ultrasound sequences. The model will be trained using clinically confirmed postpartum outcomes as reference labels. The diagnostic performance of the system will be systematically evaluated, with the goal of improving consistency in interpretation and supporting more efficient clinical decision-making in prenatal PAS screening.
Study Type
OBSERVATIONAL
Enrollment
561
An end-to-end ultrasound AI model based on the Vision Transformer (ViT) architecture was developed for the diagnosis of placenta accreta spectrum (PAS) using standardized ultrasound video inputs. Ultrasound Video Acquisition Protocol: With the patient in the supine position, the operator scanned the lower abdomen using a conventional grayscale probe. Video recording was performed in gray-scale mode for approximately 20-30 seconds, ensuring that the entire scanning region from the lower uterine segment to the uterine fundus was comprehensively captured.
The Third Affiliated Hospital, Guangzhou Medical University
Guangzhou, Guangdong, China
RECRUITINGDiagnostic performance of the end-to-end Vision Transformer (ViT)-based ultrasound AI model for placenta accreta spectrum (PAS)
Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC) of the model.
Time frame: At delivery (following confirmation of PAS status by surgical and/or pathological findings)
Clinical Feasibility of the Standardized Ultrasound Video Recording Method
Completion Rate (Proportion of patients who successfully complete the standardized recording) and time consumption (Mean recording time)
Time frame: At enrollment during the ultrasound examination
Consistency and Efficiency Between the AI Model and Physician Diagnosis
Diagnostic Consistency: Kappa coefficient used to evaluate the consistency between the AI model's diagnoses and those of experienced ultrasound physicians (Kappa \> 0.75 indicates good agreement).
Time frame: At enrollment during the ultrasound examination
Safety of the AI Model
False Negative Rate: Proportion of missed PAS-positive patients and the impact on patient outcomes .
Time frame: At delivery, when PAS status and maternal outcomes are assessed
This platform is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional.