In the present study, we evaluated eleven MRI features for the preoperative diagnosis of PAS, including signs that were previously reported and newly proposed criteria; myometrial, and serosal hypervascularity. Accordingly, the significance of serosal hypervascularity together with the other signs previously reported was validated by univariate analysis, and their diagnostic potentials were confirmed. In addition, this study also proposes a decision tree to integrate MRI features, by which PAS can be predicted with high specificity. To the best of our knowledge, this is the first report to propose the use of a decision tree model for preoperative diagnosis of placenta previa with PAS.

In this study, we used newly proposed criteria: myometrial and serosal hypervascularity, as a replacement for the MRI feature “abnormal vascularization of the placental bed”, which was proposed in the SAR-ESUR consensus statement [12]. The criterion “abnormal vascularization of the placental bed” is defined in the SAR-ESUR consensus statement as follows: “Prominent vessels in the placental bed with disruption of the uteroplacental interface. They may extend to the underlying myometrium to a variable degree, reaching up to the uterine serosa; and may be accompanied by extensive neovascularization around the bladder, uterus, and vagina”. According to this description, abnormal vascularization may exist in a variety of tissues, including the muscular or serosal layer of the uterus and adjacent organs. It also may cause confusion as the description includes features used in other criteria, such as myometric disruption and invasion of other organs. As our best efforts to avoid mixed criteria, we divided abnormal vascularization of the placental bed into two categories: “myometrial hypervascularity” and “serosal hypervascularity”. In addition, cases with myometrial disruption by vascularity were categorized in “myometrial thinning/disruption” and those with obvious invasion into other organs were included in “percretism signs”. The univariate analysis revealed a significant correlation between placenta previa with PAS and eight MRI features including serosal hypervascularity. However, myometrial hypervascularity and heterogeneous placenta failed to show a significant correlation with PAS. This finding suggests the importance of distinguishing the layer of the uterus which shows abnormal vascularity.

We developed a decision tree model to integrate the MRI features, which comprised five MRI features: myometrial thinning, placental/uterine bulge, serosal hypervascularity, intraplacental T2 dark bands, and placental ischemic infarction/recess. The sensitivity and the specificity of the decision tree in the validation cohort were 90.0% (95% CI: 53.2–98.9) and 95.5% (95% CI: 89.9–96.8), respectively. Maurea et al. retrospectively evaluated 61 placenta previa patients who underwent pathological diagnosis and reported that the sensitivity and the specificity of MRI to detect PAS were respectively 100 and 53%, when the diagnosis was made in cases with at least one of the MRI features. When the diagnosis was made in cases with at least two of the MRI features observed, the sensitivity and the specificity were respectively 92 and 92%. Although comparison of diagnostic accuracy between reports is difficult because the MRI features included in the reports were different and a limited number of subjects were evaluated in each study, the decision tree that the present study proposed showed favorable accuracy.

In this report, the variable selection for the decision tree excluded three (placental protrusion, loss of T2 hypointense interface, and abnormal intraplacental vascularity) of the MRI features. Variable selection was a necessary step to prevent overfitting of the decision tree to the derivation cohort and improve the predictive accuracy for unknown cases, but it was not for determining the individual diagnostic value of these excluded MRI features. Because the selection of MRI features was determined by the decrease in impurity (Gini), which was calculated using random forest analysis, the relative importance would be influenced by the combinations of MRI features found in each case in the cohort. In other words, if a feature can be substituted by different features, its relative importance in selecting the feature may be reduced. For example, the MRI feature “loss of T2 hypointense interface” showed a significantly high odds ratio in univariate analysis but failed to be selected in the features of the decision tree. We consider it may be because “loss of T2 hypointense interface” frequently observed in combination with myometrial thinning and/or placental/uterine bulge. On the other hand, the MRI feature “placental ischemic infarction/recess” was selected for the decision tree. This feature showed a relatively low odds ratio in the univariate analysis and was also categorized as “uncertain” and not as “recommended” in the consensus statement of SAR-ESUR [12]. However, the consensus statement collected experts’ opinions about each feature and classified its category. Placental ischemic infarction/recess is thought to reflect reduced remodeling of the spiral arteries in the PAS and consequent abnormal blood flow between the uterus and placenta and within the placenta [25,26,27]. Therefore, it is considered an important finding closely related to the pathophysiology of the PAS. Therefore, we consider “placental ischemic infarction/recess” important in combination with other features. Of note, among the 145 cases analyzed in this study, 6 cases (5 cases of heterogenous placenta and 1 case of placental protrusion) were found to have only the three excluded features without the five features included in the decision tree, and the final diagnosis was non-PAS in all cases. This suggests that PAS is unlikely to be present when only these features are present.

Although we conducted the random assignment of the derivation and the validation cohort, this study has limitations from the retrospective nature. Also, the detection of MRI features has a subjective nature and interobserver variability.

In conclusion, the decision tree we propose has the potential to integrate MRI features and predict PAS with high specificity. Further study is needed to evaluate the usefulness of this decision tree for the clinical diagnosis of placenta previa with PAS and comprehensively evaluate the MRI features obtained.