The Efficacy of a Super-Resolution Reconstruction Radiomics Model Based on T2WI for Predicting Placenta Accreta...

Placenta accreta spectrum (PAS) disorders threaten maternal and fetal health. This multicenter study aimed to evaluate whether super-resolution reconstruction (SRR) of T2-weighted magnetic resonance images improves the diagnostic performance of a radiomics model for predicting PAS compared with conventional images.

We retrospectively analyzed 603 suspected PAS cases from three centers. Center A (n=480, 224 PAS vs. 256 non-PAS) served as the training dataset; Centers B (n=66) and C (n=57) were external validation sets. Deep learning-based SRR generated 2× and 4× super-resolution T2WI. An automated nnUNet model segmented the placenta. From each resolution, 107 radiomics features were extracted and reduced by LASSO regression. Three classifiers (KNN, AdaBoost, and Gradient Boosting) were developed. Model performance was assessed using AUC and DeLong's test.

The nnUNet segmentation achieved Dice coefficients of 0.863 and 0.883 on the two external sets. In the training set, the Gradient Boosting model on 4× images yielded the highest AUC of 0.874 (95% CI: 0.8441-0.9047). However, DeLong tests showed no statistically significant differences among the 1×, 2×, and 4× models for any classifier (all P > 0.05). External validation AUCs ranged from 0.542 to 0.724, indicating only moderate generalizability.

Although SRR significantly enhanced image resolution, it did not provide incremental diagnostic value for PAS prediction within the radiomics framework. The diagnostic information relevant to PAS may be adequately captured at original resolution, or the radiomics features used are insensitive to SRR-enhanced textural changes.

Radiomics models based on 2× or 4× super-resolution T2WI were not superior to those based on pre-super-resolution images for predicting PAS. Routine SRR is not recommended for this clinical application.