Artificial Intelligence-Based Estimation of Fetal Head Circumference With Biparietal and Occipitofrontal Diameters Using Two-Dimensional Ultrasound Images

Ultrasound imaging is widely used in the medical field due to its ease of use and simplicity. Furthermore, it has no harmful or radioactive effect on the human body, unlike computed tomography, which exposes the body to high levels of ionizing radiation, causing side effects ranging from rash or irritation to an increased risk of cancer [1]. Ultrasound is widely used in various medical scans such as abdominal, cardiac, maternity, gynecological, urological, and cerebrovascular assessments [2,3]. Out of these, ultrasound imaging is extensively used to monitor the health of pregnant women and to track the development of their fetuses. It is popular among radiologists and gynecologists since it is an affordable and rapid imaging approach. At the time of sonography, various fetal biometric parameters [4,5] are assessed in which most commonly are biparietal diameter (BPD), head circumference (HC), abdominal circumference (AC), and femur diaphysis length, which is further utilized to estimate fetal weight, identify the gestational age (GA), and track the fetus's evolution [6,7]. Out of these parameters, HC and BPD are two of the most commonly used measurements in detecting various prenatal disorders as well as monitoring fetal growth. The circumference of a child's head is measured around its biggest portion. The measurement of HC is crucial in detecting anomalies and syndromes like microcephaly [8,9]. Microcephaly is a disorder in which the fetal head is much smaller than expected. It usually happens because the fetal brain did not develop adequately during pregnancy or stopped developing after delivery, resulting in a smaller head size [10]. Microcephaly can cause fetal concerns or disabilities that range from minor to extreme and are typically lifelong, such as learning disability, speech delay, impulsivity, irritability, hearing loss, visual disorder, poor coordination, and aberrant muscular functions. There is no reliable treatment for microcephaly, according to current studies [11]. As a result, it is essential to recognize microcephaly before birth or throughout pregnancy so that doctors and parents can give supportive treatment once the baby is born. During pregnancy, an ultrasound screening can detect microcephaly [12]. The ultrasound test should be performed late in the second or third trimester to detect microcephaly during pregnancy [13–15].

In addition, HC is an important biomaker in the identification of fetal heart disorders. Larger abdomen relative to HC in congenital heart disease fetuses are related to better neurodevelopment [16]. Manual HC measurement is time-consuming, which may delay the treatment of fetal anomalies. Therefore, it is crucial to develop procedures that can efficiently and automatically carry out different measurements without additional interaction and offer the details of numerous key measures in real-time to doctors, radiologists, and gynecologists. Medical intervention is possible with early detection of congenital defects and prevents the mother or fetus from getting compromised. An automated approach could aid radiologists in collecting precise measurements and early anomalies diagnosis. In this paper, we focus on measuring the HC since it can be used to establish the gestational age and track the fetus' progress. This study proposes an intelligent, real-time, and lightweight technique based on image processing and segmentation to address the discussed shortcomings of manual measurements and existing ultrasonography techniques. We deployed our proposed image processing algorithm which utilizes biometry and estimate HC, head orientation (angle) along with BPD and occipitofrontal (OFD) diameter by using segmented image which is generated by any segmentation model. For segmented image or to get binary mask of image, we used SegNET (semantic segmentation) [17], GCN (global convolutional neural network) [18], HRNet (High-resolution network) [19], yet it is not limited to use particular any model or technique to get segmented image. The proposed algorithm can provide fetal parameters based on segmented image or binary mask of image, which is generated by any technique or model. Our proposed algorithm is capable with all segmentation models and techniques. The contributions of this study are summarized as follows:

1. To the best of our knowledge, this is the first work that adopts a segmented image from any model or technique and provides HC, head orientation (angle) along with BPD and OFD diameter.

2. The suggested technique is lightweight, real-time, and easily integrate with any segmentation model or technique.

3. The proposed techniques provide head orientation (angle), and HC with BFD and OFD diameter in a single shot.

This paper is subdivided into the six sections listed below: The Sec. 1 represents the significance of various parameters in prenatal, as well as key concepts and the significance of this study. The essential literature regarding previous comparable studies and techniques is included in Sec. 2. Section 3 includes our suggested approach in depth. Section 4 illustrates the dataset details and evaluation of the proposed technique. Section 5 exhibits the experimental results based on our proposed technique. Section 6 summarizes our results.

Artificial intelligence (AI) shows various benefits in medical ultrasonography [20] from pregnancy surveillance to fetal heart monitoring [21], weight estimation [22]. AI shows diverse applications, such as recognition of fetal facial expressions [23], which help assess brain function and development in the latter half of pregnancy. Integration of AI with ultrasonography for obstetrics and gynecology opens the door to new scientific research, such as “AI science of fetal brain.” In ultrasonography, the HC boundary plays a significant role in various use cases such as fetal weight estimation [22] or prediction of the date of the delivery [24]. Campbell and Newman [25] presented a research study based on an experiment in which the author measured and examined fetal BPD by ultrasonography during normal pregnancy. The observation includes BPD measurements ranging from 13 weeks until the term for each week of pregnancy, along with longitudinal data used to describe the variation in growth rate in accordance with fetal maturity, head size, and weight, which indicates the relation between fetal development and BPD during pregnancy. In many cases, HC is essential in determining the mode of delivery as either vaginal delivery or C-section. Vaginal delivery is painful, exhausting, and physically grueling for patients. In contrast, a C-section is pain-free, less time-consuming, and safer for newborn babies. Additionally, medical professionals, patients, and health systems are concerned about the consistently increasing cesarean delivery rates. Prelabor assessment can be improved by finding characteristics that predict a patient's risk of cesarean delivery. Such elements could influence resource allocation, health system planning, patient satisfaction, and patient safety. Lipschuetz et al. [26] performed a study analysis involving a multicenter electronic medical record of birth result of primiparous women with the term (37–42 weeks) singleton fetuses presenting for an ultrasound with fetal biometry within 1 week of delivery. Based on this study, the author demonstrates the relationship between larger fetal HC and the risk of cesarean delivery. Another paper [27] shows techniques to predict gestational age using ultrasound images. Additionally, AI overcomes the limitations of ultrasonography during acquisition or scanning and provides better-scanned images to radiologists and doctors at the time of diagnosis. Chang et al. [28] developed a two-stage deep convolutional neural network (CNN) that incorporates information about the noise to manage distinct noise distributions and levels, thereby dismissing the appropriate noise (Table 1).

wherein,

I: Original input image, X: Pixel to be defined, W: Weight matrix

To perform annotation, we used the VGG image annotator tool [33]. We annotated the boundaries of the head region from an image and generated a binary mask of the input image, which can be used as ground truth. The ground truth is stored in a JSON file, which contains the name of the image and coordinates of the head region. Using segmentation, we were able to generate a binary mask of a given input head image. A binary mask is a region of interest (ROI) of an image. ROI is generally used to focus on a particular region through boundary, shape, and size. For segmentation (binary mask), we have used three different models as SegNET [17], GCN [18] and HRNet [19]. These all models are trained for 300 epochs with a learning rate as 0.01.

To calculate tilt ellipse circumference (Fig. 2), we need to derive a formula for it, as follows.

The predicted mask for the fetal head is in the form of a tilted ellipse (ellipse with some angle), We have proposed geometrical approach as in Fig. 3 which gives four endpoints. The major axis and minor axis length are calculated by using these endpoints coordinates, using the distance formula.

The image dataset is publicly available by Radboud University Medical Center's Department of Obstetrics for Grand Challenge (HC'18) [32]. The images are in grayscale format and Each two-dimensional ultrasound image is 800 by 540 pixels in dimension, with pixels ranging from 0.052 to 0.326 mm in size. 1335 fetal head ultrasound images are sorted for training, validation, and testing in a ratio of 60:20:20, which is 800:200:335 images, respectively.

The model pipeline architecture (Fig. 4) consists of various processing domains. The ultrasound-scanned image is used as input to the segmentation model, which generates a binary mask and boundary mask of the fetal head. The combination of binary mask and boundary mask will add an extra layer to the original/input image. On that image using image processing and biometry the ellipse fitting techniques are performed. It will give the major axis length, which is the BPD. Also, it is providing the minor axis length which is the OFD and orientation of the fetal head. Using the major axis length (BPD), minor axis length (OFD), and orientation of the fetal head (angle), the fetal HC is calculated (Fig. 5).

The dice score represents how similar the input image is with respect to the predicted image. Its value ranges between 0 and 1. It is measured as the overlap of the two segmentation divided by the total size of the two objects. Table 2 shows the evaluation results using the average dice score as a metric for all three segmentation models.

Out of 1335 total images, we used 200 images for validation and 335 images for testing. We generated a binary mask for each image and tried to measure the accuracy of the binary mask image with respect to the original image using a dice score. To perform testing on given semantic segmentation models, we generated ground truth by manually annotating the head ROI and calculated the dice score between ground truth and predicated images (Table 2).

Based on visual indicators, it was observed that GCN [18] performs poorly on the task of segmentation (Fig. 6), while SegNet [17] shows comparatively better results (Fig. 6) & HRNet [19] gives the best results (Fig. 7).

Using the HRNet [19] model to predict the segmentation masks on the test dataset, the average dice score value was found to be approximately 0.96 for all images (Table 2). HRNet [19] Semantic segmentation model performs the best among the considered models on the task of generating binary segmentation masks for the fetal head. Due to that, out of three considered models, we have chosen HRNet [19] model segmentation results for further process due to high dice score. As per the biometry of the fetal head, the HC, diameter, and orientation are very important for doctors and gynecologists to predict and diagnose various parameters of pregnant women as well as for their baby in the womb. It helps to estimate fetal weight and medical analysis of the fetus. For given fetal head biometry, applying image processing techniques on segmentation binary masks, we are able to calculate all the required parameters (Fig. 5).

The BPD is the longest diameter or axis from the ellipse, whereas the OFD is perpendicular to BPD. So, after several mathematical and geometrical operations as well as image processing, we can be able to get the circumference, major axis-BPD, minor axis-OFD, and angle of orientation (Fig. 5). All values obtained are pixel-wise distances. To get the exact length at the time of image acquisition need to take the extra parameter, which shows pixel-wise distance to the actual distance in the ultrasound machine or from DICOM properties. So by multiplying the output values of BPD or Circumference with the pixel-wise actual distance unit, we are able to get the exact actual length.

The proposed technique demonstrates remarkable efficiency even in the presence of variations in fetal positioning, as observed across various test cases. Moreover, our observations indicate that variations in fetal size and position have negligible impact on the outcome of our proposed technique. This inherent resilience to variations in fetal anatomy and positioning enhances the reliability, robustness, and flexibility of our approach, rendering it suitable for deployment in real-world clinical settings [34–40]. By maintaining consistent performance regardless of fetal characteristics, our technique exhibits a high degree of adaptability and suitability for diverse clinical scenarios, contributing to its potential as a valuable tool in obstetric care [41–50].

This research suggested a method for measuring fetal HC from two-dimensional ultrasound images using novel image processing algorithm which is based on biometry and segmentation. The segmentation model was used to predict the segmentation mask for the fetal head. Image processing and computer vision techniques were used to measure the HC from predicted segmentation masks. The performances of several models such as SegNet [17], GCN [18], and HRNet [19] were compared on the task of semantic segmentation using dice score as a performance validation metric. It was observed that HRNet [19] gave the best results on the test dataset with an average dice score of 0.96, which is used for further process. Our proposed algorithm adopts segmented image from any model or technique and provides fetal parameters. Using the proposed image processing algorithm, it is possible to predict HC, BPD, OFD, and head orientation/angle of fetus using two-dimensional ultrasound images.

The data and information that support the findings of this article are freely available online.²