Habitat Radiomics Outperforms Whole-Placenta Analysis for Predicting Placenta Accreta Spectrum Disorders on T2WI: A Multicenter Study

Background Current MRI diagnosis of Placenta Accreta Spectrum is limited by subjective interpretation and whole-placenta radiomics that miss key local features. This multicenter study tests whether radiomics targeting the specific placental invasion site ("placental habitat") on T2WI outperforms both whole-placenta analysis and expert assessment. Materials and Methods This multicenter retrospective study enrolled three cohorts for model training, internal validation, and independent external testing. A total of 597 pregnant women with suspected placenta accreta spectrum disorders were included, with Center A contributing 481 cases, Center B providing 68 cases, and Center C contributing 48 cases. Among these, 10 cases from Center A and all cases from Center C were combined to form an independent external test set. The remaining 539 cases from Centers A and B constituted the development cohort, which was randomly divided in a stratified manner into a training set of 377 cases and a validation set of 162 cases. Radiomics features were extracted from two regions: (1) whole-placenta regions and (2) habitat-defined subregions (identified via K-means clustering on T2WI). Four predictive models were developed: a clinical model (incorporating clinical risk factors), a Radiomics model (using whole-placenta features), a habitat model (utilizing habitat-specific features), and a combined model (integrating clinical and habitat features). Performance was benchmarked against assessments by junior and senior radiologists. Model evaluation was conducted using receiver operating characteristic (ROC) analysis with the area under the curve (AUC) as the primary metric. Results In the training cohort, the AUC values were as follows: junior radiologists (0.601), senior radiologists (0.620), the clinical model (0.841), the Radiomics model (0.777), the habitat model (0.885), and the combined model (0.896). In the validation cohort, the corresponding AUC values were: junior radiologists (0.532), senior radiologists (0.551), the clinical model (0.819), the Radiomics model (0.717), the habitat model (0.862), and the combined model (0.876). In the external test cohort, the AUC values were: junior radiologists (0.570), senior radiologists (0.625), the clinical model (0.814), the Radiomics model (0.751), the habitat model (0.812), and the combined model (0.823). Conclusion Habitat Radiomics consistently outperformed both whole-placenta analysis and radiologist assessment across all cohorts. The habitat model achieved the highest standalone performance (AUC: 0.862–0.885), demonstrating superior robustness and generalizability for PAS prediction. Integration of clinical features further enhanced performance (combined model AUC: 0.823–0.896).