3D deep-learning radiomics from MR-T2WI for predicting placenta accreta spectrum disorders: A multicenter study

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Placenta accreta spectrum (PAS) disorders pose significant risks during pregnancy.
• Accurate prediction of PAS is crucial for effective management and improved maternal outcomes.
• Current diagnostic methods have limitations in sensitivity and specificity.

Purpose of the Study:

• To develop and validate a 3D deep-learning radiomics model for predicting PAS risk.
• To compare the performance of the deep learning model against traditional radiomics and clinical models.
• To evaluate the diagnostic accuracy of the model compared to expert radiologists.

Main Methods:

• Retrospective multicenter study involving 601 suspected PAS cases.
• Development of a 3D deep learning model using T2-weighted MRI scans (T2WI).
• Extraction of radiomics and deep features, followed by feature selection and model training.
• External validation on an independent cohort and comparison with radiologist diagnoses.

Main Results:

• The 3D deep learning (DL3D) model achieved high AUCs (train: 0.912, validation: 0.864, test: 0.817).
• The DL3D model significantly outperformed traditional radiomics and clinical models.
• A combined multimodal model integrating DL3D, radiomics, and clinical features yielded the highest AUCs (train: 0.927, validation: 0.867, test: 0.847).
• Radiologist diagnostic accuracy (AUC range: 0.586-0.623) was substantially lower than all developed models.

Conclusions:

• The standalone DL3D model demonstrated superior performance compared to expert radiologists.
• The combined multimodal model provided the most significant performance advantage.
• 3D deep-learning radiomics offers a promising tool for enhancing PAS risk prediction.