Deep Learning-Assisted Diagnosis of Placenta Accreta Spectrum Using the DenseNet-121 Model: A Multicenter, Retrospective Study.

Abstract

To explore the diagnostic value of deep learning (DL) imaging based on MRI in predicting placenta accreta spectrum (PAS) in high-risk pregnant women. A total of 263 patients with suspected placenta accreta from Institution I and Institution II were retrospectively analyzed and divided into training (n = 170) and external verification sets (n = 93). Through imaging acquisition, feature extraction, and radiomic data processing, 15 radiomic features were used to train support vector machine (SVM), K-nearest neighbor (KNN), random forest (RF), light gradient boosting machine (LGBM), and DL models to predict PAS. The diagnostic performances of the models were evaluated in the training set using the area under the curve (AUC) and accuracy and further validated in the external verification set. Univariate and multivariate logistic regression analysis revealed that a history of cesarean section, placental thickness, and placenta previa were independent clinical risk factors for predicting PAS. Among machine learning (ML) models, SVM demonstrated the highest diagnostic power (AUC = 0.944), with an accuracy of 0.876. The diagnostic efficiency of the DL model was significantly better than that of other models, with an AUC of 0.956 (95% CI 0.931-0.981) in the training set and 0.863 (95% CI 0.816-0.910) in the external verification set. In terms of specificity, the DL model outperformed the ML models. The DL model based on MRI may demonstrate better performance in the diagnosis of PAS compared to traditional clinical models or ML radiomics models, as further confirmed by the external verification set.