MRI-based Deep Radiomics Ensemble Learning for Predicting Postpartum Hemorrhage in Patients with Suspected Placenta Accreta Spectrum

Motivation: Accurately predicting postpartum hemorrhage (PPH) risk in patients with suspected placenta accreta spectrum (PAS) disorders remains a clinical challenge. Goal(s): To develop an ensemble model integrating deep learning (DL) and radiomics to predict PPH risk in suspected PAS patients. Approach: 538 patients from multiple medical centers were recruited retrospectively. Using T2WI, a multitask DL model was trained for placenta segmentation, PPH classification and estimated blood loss (EBL) regression. Radiomics features were extracted and integrated with the DL to enhance predictive power. Results: The ensemble model demonstrated satisfying performance, significantly enhancing PPH risk assessment and EBL prediction in both internal and external testing sets. Impact: This study demonstrates an automated pipeline integrating deep learning and radiomics to predict postpartum hemorrhage in high-risk pregnancies using MRI. It enhances prenatal risk assessment, potentially improving clinical outcomes by enabling timely intervention and better resource allocation in obstetric care.