Deep learning super-resolution enhances MRI radiomics for predicting placenta accreta spectrum and invasiveness.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, The First Affiliated Hospital of Guangxi Medical University, Nanning, Guangxi Zhuang Autonomous Region, 530021, People's Republic of China; Department of Radiology, Guangxi Hospital Division of the First Affiliated Hospital, Sun-Yat-Sen university, Nanning, Guangxi Zhuang Autonomous Region, 530022, People's Republic of China.
- Department of Radiology, The First Affiliated Hospital of Guangxi Medical University, Nanning, Guangxi Zhuang Autonomous Region, 530021, People's Republic of China.
- Department of Radiology, Maternal and Child Health Hospital of Guangxi Zhuang Autonomous Region, Nanning, Guangxi Zhuang Autonomous Region, 530023, People's Republic of China.
- Department of Radiology, Maternal and Child Health Hospital of Guangxi Zhuang Autonomous Region, Nanning, Guangxi Zhuang Autonomous Region, 530023, People's Republic of China. Electronic address: [email protected].
- Department of Radiology, The First Affiliated Hospital of Guangxi Medical University, Nanning, Guangxi Zhuang Autonomous Region, 530021, People's Republic of China; Key Laboratory of Early Prevention and Treatment for Regional High Frequency Tumor (Guangxi Medical University), Ministry of Education, Nanning, Guangxi Zhuang Autonomous Region, 530021, People's Republic of China. Electronic address: [email protected].
Abstract
Placenta Accreta Spectrum (PAS) is a severe obstetric complication. This study evaluated whether retrospective, post-processing deep learning-based super-resolution reconstruction (SR) was associated with improved MRI radiomics performance for predicting PAS and invasive PAS. This retrospective study involved 245 patients who had placental MRI scans at three different institutions. Of these, 158 cases from Institution I for training set and internal validation set, while 34 patients from Institution II and 53 patients from Institution III served as external validation sets. SR using a generative adversarial network (GAN) was applied to generate SR T2WI (SRT2WI). Predictive models for PAS and invasive PAS were developed based on T2WI, SRT2WI, and the SRT2WI-Clinical model. Two radiologists independently diagnosed PAS using T2WI to serve as a reference for evaluating model performance. Model performance was assessed using the area under the curve (AUC), calibration curves, decision curve analysis (DCA), Delong test, and net reclassification improvement (NRI). SHAP values interpreted the models and quantified feature contributions. Across all four cohorts, SRT2WI yielded numerically higher AUCs than T2WI, and SRT2WI-Clinical achieved the highest AUCs. For PAS, SRT2WI-Clinical achieved AUCs of 0.953, 0.887, 0.793, and 0.788 in the training, internal validation, and external validation cohorts 1 and 2, respectively (T2WI: 0.738, 0.697, 0.654, and 0.652; SRT2WI: 0.879, 0.820, 0.696, and 0.768). Corresponding AUCs for invasive PAS were 0.895, 0.875, 0.843, and 0.815 (T2WI: 0.774, 0.699, 0.575, and 0.745; SRT2WI: 0.861, 0.794, 0.729, and 0.808). For PAS, SRT2WI-Clinical significantly outperformed T2WI (P < 0.001) and SRT2WI (P = 0.004) only in training, with no significant validation differences. For invasive PAS, it significantly outperformed T2WI in training, internal validation, and external validation cohort 1 (P = 0.003, 0.021, and 0.019, respectively), but not external validation cohort 2 (P = 0.398); comparisons with SRT2WI were nonsignificant. Calibration showed good agreement, DCA favored SRT2WI-Clinical, and SHAP quantified feature contributions. The combination of SR and radiomics shows promise for improving the prediction of PAS and invasive PAS. Integrating clinical features may further enhance model performance, supporting its potential clinical utility.