Placenta Accreta Spectrum and Hysterectomy Prediction Using MRI Radiomic Features

Abstract

In women with placenta accreta spectrum (PAS), patient management may involve cesarean hysterectomy at delivery. Magnetic resonance imaging (MRI) has been used for further evaluation of PAS and surgical planning. This work tackles two prediction problems: predicting presence of PAS and predicting hysterectomy using MR images of pregnant patients. First, we extracted approximately 2,500 radiomic features from MR images with two regions of interest: the placenta and the uterus. In addition to analyzing two regions of interest, we dilated the placenta and uterus masks by 5, 10, 15, and 20 mm to gain insights from the myometrium, where the uterus and placenta overlap in the case of PAS. This study cohort includes 241 pregnant women. Of these women, 89 underwent hysterectomy while 152 did not; 141 with suspected PAS, and 100 without suspected PAS. We obtained an accuracy of 0.88 for predicting hysterectomy and an accuracy of 0.92 for classifying suspected PAS. The radiomic analysis tool is further validated, it can be useful for aiding clinicians in decision making on the care of pregnant women.

1. INTRODUCTION

During pregnancy, the placenta and uterus must work in harmony to ensure optimal living conditions for the mother and her developing fetus. Placenta accreta spectrum (PAS) is a pathological condition in which the placenta invades into the uterine wall and thus fails to separate after delivery. The occurrence of PAS is steadily rising¹, due to increasing number of cesarean deliveries and can have significant maternal morbidity² including emergency hysterectomy, blood transfusion, and intensive care admission. Timely diagnosis of PAS allows planning and the necessary support to reduce the chance of maternal morbidity³.

Radiomics is an emerging machine learning approach in the field of radiology aiming to extract a high-dimensional set of features from clinical data for quantitative analysis of the data⁴. In previous work, nearly 40% of radiomic features were found to be consistent between expert segmented and deep-learning segmented placenta and uterus volumes^5–7. Hysterectomy was predicted with an area under the curve (AUC) of 0.69 for manually segmented data, and 0.78 for deep-learning segmented data using radiomic features, magnetic resonance imaging (MRI) data, and placenta and uterus segmentation masks. Another group achieved 0.94 accuracy when predicting placental invasion on T2-Weighted MRI using a mixture of radiomics and deep learning⁸. Color doppler ultrasound is the first line of clinical PAS diagnosis. Using ultrasound, significant differences (P < 0.05) are found between pregnant women that require hysterectomy and those who do not⁹. This work investigates PAS and hysterectomy prediction using MRI radiomics.

2. METHODS

2.1. MRI Acquisition

We performed an IRB approved review of 241 pregnancies. Images were acquired on a 1.5T MR scanner (Avanto, Siemens Healthcare, Erlangen, Germany). Half Fourier single shot turbo spin echo (HASTE) T2-weighted axial imaging sequence covering the entire gravid uterus was evaluated. Each image volume had 28 to 52 two-dimensional (2D) transverse slices. Each MRI slice was configured to be 256 × 256 pixels. The axial image slice spacing is 7.0 mm and the pixels are isotropic across slices. In plan resolution ranges from 1.05 × 1.05 mm² to 1.72 × 1.72 mm².

2.2. Database

Of the 241 pregnant women, 100 were normal and 141 were PAS-suspected; of these, 89 required hysterectomy and 52 did not. For each patient, an expert radiologist manually segmented the uterine cavity and placenta. As a result, each patient had two corresponding masks, or regions of interest. In addition, we were provided data regarding the suspected presence of PAS and corresponding hysterectomy treatment outcome status.

2.3. Image Processing

For each patient, the MR images were dilated using 5, 10, 15, and 20 mm in MATLAB (MathWorks, Inc., Natick, MA). Figure 1 illustrates the effects of dilation on the placenta mask (A) and uterus mask (B). The placenta mask (A) is dilated to mimics PAS by simulating the invasion of the placenta into the myometrium. Typically, the myometrium is approximately 7 mm thick during pregnancy¹⁰. Therefore, dilation from 0–10 mm resembles normal thickness, 10–20 represents the various stages of PAS.

Figure 1.

Example changes in area from one patient slice at 10 mm dilation, where the uterus mask is red and the placenta mask is yellow. (A) The pink area denotes the dilation of the placenta mask by 10 mm. (B) The blue are denotes the dilation of the uterus mask by 10 mm.

2.4. Radiomic Feature Extraction and Selection

Both non-dilated and dilated MRI were loaded into PyRadiomics where 2,463 unique features were extracted. These features are the result of several image filters, feature class, and feature setting parameters being manipulated. The parameters that were tested are summarized in feature extraction settings section. All radiomic feature definitions were established by the Imaging Biomarker Standardization Initiative (IBSI)¹¹.

After feature extraction, radiomic feature selection was performed. The purpose of feature selection is to avoid issues presented by high dimensionality, algorithm performance, noisy or ambiguous data, and replicability of the study¹². The goal of feature selection is to identify the relationship between the radiomic feature(s) and the presence of PAS or hysterectomy¹³. Three feature selection methods were implemented: analysis of variance (ANOVA), LASSO, and Ridge Regression. These feature selection methods can be categorized into filter and embedded tools. ANOVA is a filter method. Filter methods are useful due to their quick computation time and intolerance to overfitting. LASSO and Ridge Regression are examples of embedded feature selection tools. These methods work by combining filter and wrapper methods into a learning algorithm^14–17.

2.5. Machine Learning

Seventeen machine learning algorithms were tested for each of the three feature extraction methods and five dilation groups (0, 5, 10, 15, 20 mm), for a total of 255 combinations. All machine learning algorithms were implemented using the python scikit-learn package with their default parameters¹². Classifiers were trained using an independent training set. The predictive performance was evaluated based on an independent testing set using accuracy analysis. Both the independent training and testing sets were randomly selected on a patient basis. Each dilation group was tested within the group. For example, a 5 mm dilated training set was used to train the algorithm for the 5 mm dilated testing set. In the study, 157 patients were used for training, 24 patients for validation, and 60 patients for testing. No patient overlaps occurred between the training and testing groups.

3. RESULTS

PAS was classified with 92% accuracy using both uterus and placenta masks (Table 1). Uterus mask resulted in 92% accuracy without dilation, and placenta mask required 20 mm dilation. For the placenta mask, the feature selection algorithm was ANOVA and the algorithm was Extra Gradient Boost Classifier. For the uterus mask, the feature selection and the algorithms were the same.

Table 1.

Accuracy as a relationship between mask type and dilation for predicting PAS on the testing set.

	Dilation (mm)
	None	5	10	15	20
Placenta	0.88	0.88	0.87	0.90	0.92
Uterus	0.92	0.86	0.82	0.87	0.85

Hysterectomy clinical outcome was classified with 88% accuracy using the placenta mask without dilation on the whole test set including the normal cases (Table 2). To obtain 88% with the placenta mask, LASSO feature selection tool was used in conjunction with the Gradient Boost Classifier.

Table 2.

Accuracy as a relationship between mask type and dilation for predicting hysterectomy on the testing set (including the normal cases).

	Dilation (mm)
	None	5	10	15	20
Placenta	0.88	0.85	0.80	0.85	0.83
Uterus	0.83	0.80	0.82	0.82	0.85

The most important features for predicting PAS using the placenta mask are voxel-based calculated volume (hereafter called ‘voxel-based volume’), maximum 3D diameter, and mesh volume (i.e., the volume that is represented by a surface mesh). The most important features for predicting PAS using the uterus mask are voxel-based volume, mesh volume, and minor axis length. The most important features for predicting hysterectomy using the placenta mask are voxel-based volume, maximum 3D diameter, and mesh volume. Table 3 summarizes these features where voxel-based volume is the most frequently chosen feature among predicting PAS and hysterectomy treatment response. Voxel-based volume is calculated by multiplying the number of voxels in the region of interest by the volume of a single voxel.

Table 3.

Top three radiomic features chosen for each prediction (PAS or hysterectomy).

PAS (Placenta Mask)	PAS (Uterus Mask)	Hysterectomy
Voxel-based Volume	Voxel-based Volume	Voxel-based Volume
Maximum 3D Diameter	Mesh Volume	Maximum 3D Diameter
Mesh Volume	Minor Axis Length	Mesh Volume

4. DISCUSSION & CONCLUSION

We developed an MRI-based radiomic method for PAS and hysterectomy prediction. PAS was independently classified with an accuracy of 0.92 without image augmentation of the uterus mask. Hysterectomy clinical outcome was classified with an accuracy of 0.88 when using the placenta mask without image augmentation. The voxel-based volume feature was found to be important for the prediction.

The major limitation of this study is the small data set. The number of patients (N=241) may not be sufficient to generalize the applicability of this study in a clinical environment. Although we implemented an independent testing set, we cannot conclude that this data is representative of all cases that may be presented. If we had access to a larger dataset, it is likely that the model presented here would be more universal in its applicability. For hysterectomy prediction, we tested the algorithm on a test set with both PAS and normal cases. In future work, we are going to follow a decision tree and predict the hysterectomy using the cases that were classified as PAS during the first step. This approach is more compatible with the current clinical workflow and could be helpful to have a better evaluation over the whole workflow. We also anticipate a higher hysterectomy prediction rate using this approach.

In addition to the voxel-based volume feature, the data suggest that maximum 3D diameter, mesh volume, and minor axis length are connected to determining PAS and hysterectomy. Voxel-based volume combined with maximum 3D diameter can be related to the sphericity of the region of interest. For a specific 3D diameter value, the closer the volume is to a sphere, the greater the volume. Mesh volume and minor axis length follow similar patterns. When taking a closer look at the placenta radiomic features, we found that hysterectomy and PAS positive individuals were more likely to have features corresponding to spherical placentas. With this understanding, we may be able to conclude that placenta shapes that are more spherical tend to develop PAS and require hysterectomy. More research must be done in order to make this assertion. However, this study has demonstrated promise for using radiomics and MRI to aid in early detection of PAS and prepare patients for potential hysterectomy planning.

APPENDIX: FEATURE EXTRACTION SETTINGS

Extraction Settings	weightingNorm = Manhattan
	distances = 1:1:10
	interpolator 1:1:10
	padDistance = 3
	preCrop = True, False
	resegmentMode = relative
	resegmentShape = True, False
	correctMask = True False
	kernelRadius = 0:1:9
	maskedKernel = True, False
Image Filters	LoG
	Wavelet
	Square
	SquareRoot
	Logarithm
	Exponential
	Gradient
	LBP3D
Image Filter Settings	LoG settings: sigma = 0.25:0.25:2.5
	Gradient settings: gradientUseSpacing = True
	Wavelet settings: start level = 0:1:9
	Wavelet settings: level = 5, 10, 1, 20, 30, 40, 50, 60, 70, 80
	Wavelet settings: wavelet = all possible wavelets
Feature Settings	voxelArrayShift = 0:2000
	symmetricalGLCM = True, False
	gldm_a = 0:.25:0.25:2.5