A computerized diagnostic model for automatically evaluating placenta accrete spectrum disorders based on the combined MR radiomics-clinical signatures

We aimed to establish a computerized diagnostic model to predict placenta accrete spectrum (PAS) disorders based on T2-weighted MR imaging. We recruited pregnant women with clinically suspected PAS disorders between January 2015 and December 2018 in our institution. All preoperative T2-weighted imaging (T2WI) MR images were manually outlined on the picture archive communication system terminal server. A nnU-Net network for automatic segmentation and the corresponding radiomics features extracted from the segmented region were applied to build a radiomics-clinical model for PAS disorders identification. Taking the surgical or pathological findings as the reference standard, we compared this computerized model’s diagnostic performance in detecting PAS disorders. In the training cohort, our model combining both radiomics and clinical characteristics yielded an accuracy of 0.771, a sensitivity of 0.854, and a specificity of 0.750 in identifying PAS disorders. In the testing cohort, this model achieved a segmentation mean Dice coefficient of 0.890 and yielded an accuracy of 0.825, a sensitivity of 0.830 and a specificity of 0.822. In the external validation cohort, this computer-aided diagnostic model yielded an accuracy of 0.690, a sensitivity of 0.929 and a specificity of 0.467 in identifying placenta increta. In the present study, a machine learning model based on preoperative T2WI-based imaging had high accuracy in identifying PAS disorders in respect of surgical and histological findings.


Materials and methods
Patients. Our institutional review board approved this retrospective study. All methods were performed in accordance with the relevant guidelines and regulations. From January 2015 to December 2018, data from 752 patients who underwent MRI, including 551 pregnant women with clinically suspected PAS disorders (either by previous US or external MR reports), were retrospectively retrieved from the PACS at our institution. We excluded 39 of them who had undergone abdominal aorta balloon occlusion surgery before delivery. Finally, 512 pregnant women were included (Fig. 1). We divided the included pregnancies into the following five categories (scores from 0 to 5 points) based on placental position: 0, normal placental position; (1) low-lying placenta; (2) marginal placenta previa; (3) partial placenta previa; (4) central placenta previa; and (5) pernicious placenta previa.
MR examination, imaging reading and lesion segmentation. In our hospital, MR examination was performed using a 1.5-Tesla MR unit (Magnetom Avanto, Siemens). In the external hospital, one MR scan unit was a 3.0-Tesla machine (Ingenia 3.0 T, Philips), and another was a 1.5-Tesla MR unit (Optima MR 360, GE). The detailed scanning parameters for each unit are summarized in Supplementary Table 1. The diagnosis for placental implantation on MRI was established based on the previous well-described criteria 8,15 . First, two observers (each had more than 7 years of PAS diagnosis on MRI) blinded to the US and surgical results analyzed all the MRI datasets of each participant independently on the PACS terminal server. Confidence in identifying the status of placenta accreta spectrum was assessed using a five-point scale as follows: '5' , definitely present; '4' , probably present; '3' , uncertain; '2' , probably absent; and '1' , definitely absent. Second, all conclusions required a consensus agreement between the two observers. All visible placental tissue was examined by an experienced radiologist (H.Z.) on sagittal T2WI using ITk-SNAP software (http:// www. itksn ap. org/ pmwiki/ pmwiki. php?n= Main. HomeP age).
Semantic segmentation and radiomic feature extraction. The pipeline of the imaging process is shown in Fig. 2. A semantic segmentation three dimensional (3D) nnU-Net model was trained by the radiologist, who labeled the placental region to segment the placenta used for radiomics feature extraction 16 . nnU-Net can automatically adapt to arbitrary datasets and take full advantage of the characteristics of the dataset to outperform many U-Net-based models and has been used successfully in 3D biomedical image segmentation. We used a combination of cross entropy loss and dice loss as the loss function. The stochastic gradient descent algorithm was used as the optimizer with an initial learning rate of 0.01, a momentum of 0.9 and a weight decay of 0.005. In brief, a total of 107 features were extracted from each placental region, including the following: (1) 14 shape features; (2) 18 first-order statistical features; and (3) 75 texture features, including the gray level cooccurrence matrix (GLCM), gray-level dependence matrix (GLDM), gray-level run-length matrix (GLRLM), gray-

Results
Baseline characteristics. In the present study, the patients were 21 to 48 years of age and between 18 and 41 weeks of gestation. The average gravida and parity were 3.63 and 1.83, respectively. Among them, 397 (77.54%) had no history of cesarean section, and 115 (22.46%) had undergone a cesarean section 1 to 4 times ( Table 1). Ninety-four (18.36%) women were diagnosed with diabetes mellitus (including pregestational diabetes mellitus and gestational diabetes mellitus), 31 (6.05%) were diagnosed with hypertensive disorders (including gestational hypertension, preeclampsia, and chronic hypertension complicating pregnancy), 19  Clinical data analysis. There were significant differences in some clinical variables between those with PAS disorders and normal placentas (  Fig. 1). The combined model achieved the best performance with AUCs of 0.833 (95% CI: 0.780-0.882) and 0.849 (95% CI: 0.778-0.914) in both the training and testing cohorts, respectively. The statistical analysis for the clinical/radiomics/radiomics-clinical nomogram in the two cohorts is shown in Table 3. The waterfall plot of the testing cohort with an optimal cutoff value of 0.458 for the distribution of the prediction probability is shown in Fig. 6. The DCAs curve indicated that the net benefit of the nomogram was better than the other models when the threshold was in the range between 0.1 and 0.5. The radiomics signature had a higher positive prediction rate when the threshold was greater than 0.5. In external validation, the   Table 4). For our studied cases, either small invasion signs on MRI or large interface area between placenta and myometrium (for example, both anterior and posterior uterine wall involved) mostly attributed to the model's error identification (Fig. 7).

Discussion
With the comprehensive opening of the "third child" policy in China, the number of people who get married at a later age and have children at a later age increases. It can be speculated that the incidence of placental implantation owing to cesarean surgery in China may further increase. In our study, we designed a computerized diagnostic model for evaluating PAS from a large cohort sample from a single institution. Our model yielded an AUC of 0.849 and an ACC of 0.825 in detecting invasive placentation in the testing group. In the external validation group, our model also provided the competitive diagnostic performance in detecting PAS. To our knowledge, this is the first reported study focusing on PAS disorder identification with machine learning techniques using T2WI images.    6,7 . However, in recently published articles, these values varied with an SEN from 66 to 100% and an SPE from 71% to 76.9% 10,12 . Our study's results also did not reach the high diagnostic performance level based on MRI knowledge. Some reasons are explained as follows. First, radiologists' experiences play a vital role in the interpretation of PAS. In a recent study, the authors disclosed that abdominopelvic MRI experience without specific placental MRI experience did not improve diagnostic performance in PAS diagnosis 10 . Knowledge of PAS-related MRI findings definitely varies among various institutions, even in tertiary centers. Second, in our study, the final results included both surgical and histological findings. We could imagine that some objective judgments of PAS diagnosis from individual obstetricians may be incorrect. In our study, we did not use postpartum hemorrhage volume 13 as the reference standard because, on the one hand, the blood loss volume in pregnant women, who underwent previous abdominal aorta balloon occlusion in case massive bleeding occur, may be inaccurate; on the other hand, pregnancies with either accreta or increta placenta diagnosed by obstetricians may not have massive blood loss but may still require some necessary treatments or follow-up after delivery. We believe that our present gold standard is similar to real scenarios where placenta accreta or increta is the greatest concern for obstetricians before invasive procedures.
Herein, we designed a pipeline algorithm aimed at confirming PAS disorders. Our study had several characteristics that were different from those of previous studies 13,14,18 . A trained 3D nnU-Net model was utilized to automatically segment the placental region on T2WI images. Compared with previous radiomics studies 13 , automatic segmentation of targeted regions could reduce the subjectivity, workload of radiologists and negative impact of interreader variance on predictions. This variance is relatively slight due to the good tissue contrast between the placenta and the myometrium on T2WI images. Our present results showed that the radiomics-clinical model achieved a satisfying performance in PAS disorder identification. The radiomics texture features reflect intraplacental heterogeneity, which may indicate physiological changes, such as fibrin deposition. Clinical features,   www.nature.com/scientificreports/ including placenta location, maternal age, and gravida and parity, played an equally important role. Compared to one study using deep learning (DL) method to predict PAS disorders 14 , our study did not apply DL method to extract high-level imaging features. The trained DL network here was only utilized to segment the placental region on MRI. In our opinion, the cascaded modeling process consists of many model building steps including model training, DL feature extraction and combine both the deep learning and radiomics-extracted signatures to get the predicted point and predict placenta invasion. Many combinations of algorithms and hyper-parameters need to be explored to get the best results, which make the process vulnerable to information leaking and overfitting, especially when the studied sample size is small. We compared the model's prediction capability with the radiologists' performance at the two-task level. The model's performance outperformed or was equal to the radiologists' performance in PAS status evaluation. The trained model can process the unsegmented data and get  www.nature.com/scientificreports/ the result within two minutes for each case, which help radiologists, especially less experienced residents, make a correct conclusion with more confidence. In our institution, this computerized model will help our physicians to establish the risk scores for each suspicious PAS pregnancies at the multidisciplinary treatment discussion. Further, this standardized-diagnostic model may also improve the diagnostic confidence and accuracy on PAS identification on MRI through telemedicine, especially for less-trained radiologists in some primary institutions where equip with MR unit. Although, artificial intelligence (represented by machine learning) technology can help physicians to tackle problems effectively, precisely at present; such technology should better incorporate with modern medicine in order to overcome its inherent shortcomings and limitation in the future 19 . Our study has several limitations. First, we train this computerized model using MRI in our single institution, and the external validation samples are limited. The true diagnostic performance still needs to be validated in a larger cohort. Second, in the present study, we only used a single sequence of MRI (sagittal T2WI) to reconstruct the assisted diagnostic model. Theoretically, more selected sequences could help improve the estimated performance to some extent. Third, in this study, 1.5 T MR equipment was applied. Although it has not yet been well evaluated for intrauterine examinations, 3.0 T in utero MR imaging with a high signal-to-noise ratio and fast scanning protocols may improve image resolution 20 .

Conclusion
In the present study, a machine learning model based on preoperative T2WI-based imaging had high accuracy in identifying PAS disorders in respect of surgical and histological findings.