MRI of the Placenta Accreta Spectrum (PAS) Disorder: Radiomics Analysis Correlates With Surgical and Pathological Outcome

Abstract

Background: Placenta accreta spectrum (PAS) in women with previous cesarean delivery has become increasingly prevalent. Depending on the severity, patient management may involve cesarean hysterectomy. Purpose: To investigate textural analyses as the radiomics in MRI of the placenta in predicting the PAS requiring cesarean hysterectomy in a high‐risk population. Study Type: Retrospective. Population: Sixty‐two women with prior cesarean delivery referred for MRI because of sonographic suspicion for PAS. Field Strength/Sequence: 1.5T with T 1 W, T 2 W, and diffusion‐weighted imaging (DWI). Assessment: Two reviewers independently evaluated MR images based on five established PAS variables. Placental regions of interest (ROIs) were generated on T 2 W, DWI, and an apparent diffusion coefficient (ADC) map, based on definitions of areas of placenta in proximity to and remote from previous surgical incision sites. Statistical Tests: Reader agreement was assessed by simple kappa and prevalence adjusted bias adjusted kappa (PABAK). T‐tests and chi‐square analyses between the primary outcome (hysterectomy vs. no hysterectomy) were performed. Thirteen Haralick texture features calculated from gray‐level co‐occurrence matrixes were extracted from manually drawn placental ROIs within each of three MR acquisitions. Univariate and multivariable logistic regression were used to assess the association with cesarean hysterectomy. Results: Of 62 pregnancies at risk for PAS, 40 required cesarean hysterectomy (65%), with excellent correlation between need for hysterectomy and pathology confirmation of PAS in the hysterectomy specimen [κ = 0.82 (0.62, 1)]. Reader agreement was fair to moderate. Of the 13 Haralick variables within each of three acquisition groups, significant differences (P < 0.05) were seen in 22 of 39 parameters comparing placental ROIs in proximity to incision scar(s) to those ROIs remote from scar. A stepwise selection algorithm indicated that the combination of T 2 W Fcm.sum.var, ADC Fcm.diff.entr, and DWI Fcm.energy gave the highest leave‐one‐out‐AUC of 0.80 (0.68, 0.91). Data Conclusion: Assessment of PAS severity is subjective and dependent on radiologist expertise. We identified textural features on placental MR images in the region of the prior uterine scar that differentiated pregnancies requiring cesarean hysterectomy based on clinical suspicion of PAS from those that did not, suggesting predictive capabilities of these objective radiomics features. Level of Evidence: 3 Technical Efficacy Stage: 1 J. Magn. Reson. Imaging 2020;51:936–946.