A Semi-Automated Deep Learning Model for Diagnosing Placenta Accreta Spectrum (COMPAS): Comparison with Radiologists’ Interpretations

Placenta accreta spectrum disorders (PAS) pose significant maternal risks, primarily owing to life-threatening hemorrhage during delivery.However, early and accurate diagnosis of PAS based on patients' clinical information and MRI interpretation requires a high level of clinical expertise from physicians.The present study aimed to evaluate the diagnostic accuracy of the Comprehensive Placenta Accreta Spectrum Diagnosis Deep Learning Model (COMPAS), which integrates T2-weighted MRI images with clinical information in diagnosing PAS.A retrospective analysis of 218 pregnant women (median age, 36.0 years; interquartile range, 32.0-38.0years) who underwent prenatal MRI for suspected PAS was conducted.Among them, 55 patients received a pathological diagnosis of PAS.The cohort was divided into a training/validation group (n = 175) and a test group (n = 43).A specialized deep learning (DL) model, COMPAS, was developed to comprehensively diagnose PAS by integrating clinical information and placental image features derived from semi-automated segmentation on MRI images.The diagnostic efficacy of COMPAS, measured using the area under the curve (AUC), was compared to that of radiologists' assessments using Delong's test.COMPAS demonstrated an AUC of 0.906 (95% confidence interval [CI]: 0.867-0.945)on the test dataset, significantly outperforming the average AUC value of radiologists (0.821 [95% CI: 0.734-0.882],p = 0.023).COMPAS exhibits superior accuracy in diagnosing PAS disorders compared with radiologists, suggesting its potential as a valuable tool for clinical decision support, particularly in settings with limited expertise in these disorders.