Deep Learning–Based Segmentation of Fetal Anatomical Structures in the First Trimester

ABSTRACT

Objectives

To develop and evaluate an artificial intelligence (AI) system that automatically identifies and classifies the fetal structures in the first trimester.

Method

Over 20,000 ultrasound images from first‐trimester fetuses were prospectively collected from four university hospitals in the Republic of Korea. Images were annotated according to segmentation‐specific structures by anatomical regions, including the head, neck, thorax, abdomen, extremities, and spine, based on standardized guidelines. The YOLACT model, which enabels real‐time instance segmentation, was used to detect and segment fetal structures. The dataset was randomly divided into a training set (95%) and a test set (5%). Model performance was evaluated using detection accuracy, mean average precision (mAP), and frames per second (FPS).

Results

The YOLACT model achieved an overall anatomical detection accuracy of 98.4%. High segmentation performance (F1‐score > 0.950) was observed for well‐defined structures such as the cranium, heart, and abdominal circumference. Structures like the nasal bone and extremities had relatively lower recall. The model's mAP at IoU 0.5 was 0.622, and real‐time processing was confirmed with a speed of 25.4 FPS.

Conclusions

The YOLACT‐based AI model demonstrated accurate and efficient segmentation of fetal structures in the first trimester, supporting its potential for real‐time clinical application in early anomaly screening.

Key points

• What is already known about this topic?

  • ◦AI technology has been advancing in second‐trimester fetal ultrasound, enabling standard image acquisition, automated measurements, and structural analysis.
  • ◦In the first trimester, AI applications have mainly focused on automatic nuchal translucency (NT) measurement and obtaining the mid‐sagittal standard view.

• What does this study add?

  • ◦This study presents the first comprehensive deep learning model for real‐time segmentation of first‐trimester fetal structures using the YOLACT model.
  • ◦The model enables relatively accurate identification of all major fetal structures in the first trimester using ultrasound imaging.
  • ◦The tool improves understanding of first‐trimester fetal anatomy and supports both clinicians and non‐expert operators in explaining findings to expectant parents.

1. Introduction

The use of artificial intelligence (AI) in big data analysis for disease diagnosis and prognosis prediction has been rapidly advancing. Deep learning technology is an AI tool that analyzes images from ultrasound, computed tomography, and magnetic resonance imaging and enables automatic recognition or interpretation of lesions. In the field of obstetrics, deep learning technology is also employed to extract standard images, measure fetal size, and analyze structures [1, 2, 3]. One of the primary goals of using fetal images with AI technology is to differentiate between normal and abnormal structures for identifying fetal anomalies.

Traditionally, fetal anomalies have been diagnosed using level II ultrasound at 18–24 weeks of gestation. Consequently, most deep learning techniques for fetal measurement and organ evaluation have focused on the second trimester [4, 5, 6, 7]. However, with recent advancements in ultrasound resolution, early detection of fetal anomalies has become possible, enabling an early diagnosis and intervention. While the evaluation of a fetus in the first trimester is important, research using deep learning techniques for first‐trimester fetal ultrasound remains limited [1, 8]. Moreover, a first‐trimester fetus presents more challenges in observation and image interpretation than a second‐trimester fetus owing to its smaller anatomical structures, variable positions, and limited available information.

Meanwhile, instance segmentation, the process of separating and classifying individual objects within an image, plays a vital role in medical image analysis. For this purpose, YOLACT is a model that can perform efficient real‐time instance segmentation and has demonstrated high performance on complex medical image data [9]. Unlike many other instance segmentation models, YOLACT employs a streamlined architecture inspired by Fully Convolutional Networks (FCN) with which it simplifies its design and improves its computational efficiency. This design facilitates faster inference times without compromising segmentation quality. This capability is critical for handling complex first‐trimester ultrasound data, where speed and accuracy are paramount. Using prototype masks and instance coefficients, this model generates high‐quality masks for each detected object, making it particularly effective in tasks requiring detailed structure delineation, such as the analysis of fetal heart anatomy in medical imaging [10, 11].

In this study, we propose a novel deep learning‐based model using YOLACT for automated detection and segmentation of fetal structures in the first trimester. Our approach enables fast and accurate structural delineation, which could enhance early prenatal screening and improve consistency in fetal assessments.

2. Methods

This study was conducted as a prospective multicenter study. Fetal images were obtained from four university hospitals in the Republic of Korea: Seoul St. Mary's Hospital, Eunpyeong St. Mary's Hospital, Uijeongbu St. Mary's Hospital, and Bucheon St. Mary's Hospital. We focused on fetuses aged between 11 + 0 and 13 + 6 weeks of gestation. The acquired images were categorized based on the first‐trimester guidelines of the International Society of Ultrasound in Obstetrics and Gynecology (ISUOG) into the following anatomical categories: head, neck, thorax, abdomen, extremities, and spine. For each category, specific anatomical structures were identified and annotated as required for detailed analysis. The images focused primarily on key anatomical structures, with classifications and definitions of these structures outlined in Table 1. The dataset was then divided into a training set (95%) and a test set (5%). The training set was used for deep learning algorithm development, and the test set was used for algorithm validation. The workflow of the entire study is depicted in Figure 1. This study received approval from the Central Institutional Review Board of The Catholic Medical Center in the Republic of Korea (IRB No. XC22EIDI0054).

TABLE 1.

Classification definitions and segmentation annotations.

Classification	View definition	Essential structure	Variation	Annotation	Test structure
Head	Axial view focused on the cranial bone	Cranium	Thalamus, midline including falx, cerebellum, choroid plexus	Cranium, midline including falx, choroid plexus, extremities	Cranium, midline including falx, choroid plexus
Neck	Sagittal view focused on the neck, from head to chest	One of nasal bone, nuchal translucency, or palate	Nasal bone, nuchal translucency, palate, spine, extremities	Nasal bone, nuchal translucency, palate, spine, extremities	Nasal bone, nuchal translucency, palate
Thorax	Axial view focused on the thorax	Thoracic circumference	Heart, spine, extremities	Thoracic circumference, heart, spine, extremities	Thoracic circumference, heart
Abdomen	Axial view focused on the abdomen	Abdominal circumference	Stomach, cord insertion, spine, extremities	Abdominal circumference, stomach, cord insertion, spine, extremities	Abdominal circumference, stomach
Extremity	Enlarged view of the upper extremities or lower extremities	One of upper arm, forearm, hand, thigh, calf, or foot	Upper arm, forearm, hand, thigh, calf, foot	Upper arm, forearm, hand, thigh, calf, foot	Upper arm, forearm, hand, thigh, calf, foot
Spine	Sagittal or coronal view of the entire fetus or a focused view on the spine	Spine	Stomach	Spine	Spine

FIGURE 1.

The workflow of model development and validation.

The images were collected using two types of ultrasound machines, WS80 A and HERA W10 (Samsung Medison Co., Ltd, Seoul, Korea). Both abdominal and transvaginal transducers were utilized, with scans performed by experienced physicians and sonographers specializing in first‐trimester ultrasound imaging. The acquired images were initially categorized by experts. Segmentation annotations were performed by trained operators and then reviewed for accuracy by experts. To ensure consistency in labeling, images with severe artifacts, unclear anatomical boundaries, or ambiguous anatomical categorization were excluded prior to annotation. These segmentation annotations served as the ground truth for developing and validating the deep learning algorithm. The model's network architecture and training method are described in detail in the Supplementary material (Network architecture and training).

To verify the accuracy of the model's category classification based on essential segmentations, we defined the following evaluation criteria. True positive (TP) is the correct identification of the required segmentation within the correct category. False Negative (FN) is failure to identify the required segmentation. False positive (FP) is incorrect identification of a segmentation in the wrong category or location. True negative (TN) is the absence of detection in irrelevant categories. Examples of TP, FN, FP, and TN are illustrated in Supporting Information S1: Figure S1a–b. When the same segmentation was detected multiple times in a single image, the detection with the highest confidence score was used for evaluation. The model's category classification performance was calculated using accuracy (TP + TN/TP + FP + FN + TN), recall (TP/TP + FN), precision (TP/TP + FP), and F1 score (2/(1/precision + 1/recall)).

The segmentation performance of the model was assessed using the mean average precision (mAP) based on the intersection over union (IoU). IoU measures the overlap between the predicted bounding box or segmented area and the ground truth, while average precision represents the average precision at varying IoU thresholds. mAP is calculated as the mean of the average precision across all classes. We used an IoU threshold of 0.5 and a confidence threshold of 0.3 for the model. These thresholds ensured a balance between segmentation accuracy and practical utility, particularly in clinical settings where precision and speed are equally critical. A confusion matrix was generated and visualized using Python, employing Matplotlib and Seaborn libraries for heatmap representation. For comparative evaluation, we also trained MS‐RCNN and Mask R‐CNN under the same conditions. The results are presented in Supporting Information S1: Table S1 and S2, and discussed in the Discussion section.

To robustly evaluate model performance, we held out a fixed test set of 957 images (5%) for final evaluation. The remaining 19,000 images were repeatedly split into training and validation subsets at a 4:1 ratio using five different random seeds. In each iteration, the model was trained on the training subset with early stopping based on validation performance, and then evaluated on the fixed test set. Structure‐wise box AP and area AP were calculated for each iteration, and the performance was summarized as the mean ± standard deviation (SD) across the five runs. This evaluation protocol is not a conventional k‐fold cross‐validation, but rather a Monte Carlo repeated random split approach designed to assess the model's sensitivity to the train/validation partition while maintaining an unbiased hold‐out test set.

Inference speed, measured in frames per second (FPS), indicates the number of images frames a model can process per second and is a key measure for evaluating a model's real‐time processing capability. The hardware environment included an Intel(R) Xeon(R) CPU E5‐2620 v4 @ 2.10 GHz and a single NVIDIA Tesla P100‐PCIE‐16 GB GPU. Under these conditions, the FPS of the model was calculated.

3. Results

This study included 925 patients, with a total of 20,009 images collected. The average patient age was 34 years, the mean gestational age was 12 weeks 2 days, and the average body mass index was 22.7 kg/m² (Table 2). The number of images in the training and test sets for each classification is summarized in Supporting Information S1: Table S3. The training set included 19,142 images, and the test set comprised 957 images. Figure 2 presents a confusion matrix of the model, and 942 of 957 test images were correctly classified, resulting in an overall accuracy of 98.4%. The classification performance was lowest for the spine and extremity, with the model failing to detect the spine in 6 cases and the extremity in 4 cases.

TABLE 2.

Study population characteristics.

Characteristic	Study population (n = 925)
Age (years)	34.0 ± 3.9
Gestational age at ultrasound exam (weeks)	12.2 ± 0.6
Height (cm)	162.0 ± 5.2
Weight (kg)	59.7 ± 10.0
Body mass index (kg/m2)	22.7 ± 3.5

Note: Data are presented as mean ± SD.

FIGURE 2.

Confusion matrix of the YOLACT model for fetal structure classification. The number in each cell represents the number of images. The vertical axis indicates the actual classification, and the horizontal axis indicates the classification predicted by the model.

Table 3 summarizes the segmentation accuracy for each anatomical structure within its respective classification. The model demonstrated high accuracy in detecting the cranium, midline with falx, palate, thoracic circumference, heart, abdominal circumference, hand, calf, and spine, achieving F1‐scores higher than 0.950 with minimal false positives or false negatives. However, segmentation performance was lower for certain structures, particularly the nasal bone, nuchal translucency, upper arm, and foot, which exhibited higher false negative rates. These structures had recall values of 0.539, 0.607, 0.636, and 0.640, respectively, indicating challenges in their consistent detection.

TABLE 3.

Segmentation performance of the model: accuracy, recall, precision, and F1‐score.

Classification	Structure segmentation	TP	FP	FN	TN	Accuracy	Recall	Precision	F1‐score
Head	Cranium	194	0	1	0	0.995	0.995	1.000	0.997
	Choroid plexus	153	2	3	37	0.974	0.981	0.987	0.984
	Midline with falx	181	0	9	5	0.954	0.953	1.000	0.976
Neck	Nasal bone	89	1	76	33	0.613	0.539	0.989	0.698
	Palate	192	0	1	6	0.995	0.995	1.000	0.997
	Nuchal translucency	119	0	77	3	0.613	0.607	1.000	0.756
Thorax	Thoracic circumference	90	2	1	0	0.968	0.989	0.978	0.984
	Heart	93	0	0	0	1.000	1.000	1.000	1.000
Abdomen	Abdominal circumference	168	0	2	0	0.988	0.988	1.000	0.994
	Stomach	85	5	12	68	0.900	0.876	0.944	0.909
Extremity	Upper arm	28	4	16	105	0.869	0.636	0.875	0.737
	Forearm	50	4	3	96	0.954	0.943	0.926	0.935
	Hand	67	3	3	80	0.961	0.957	0.957	0.957
	Thigh	71	8	8	66	0.895	0.899	0.899	0.899
	Calf	58	2	2	91	0.974	0.967	0.967	0.967
	Foot	32	2	18	101	0.869	0.640	0.941	0.762
Spine	Spine	141	0	6	0	0.959	0.959	1.000	0.979

Note: Data are presented as n. Accuracy is (TP + TN)/(TP + FP + FN + TN). Recall is TP/(TP + FN). Precision is TP/(TP + FP). F1‐score is (2/(1/precision+1/recall)).

Abbreviations: FN, false negative; FP, false positive; TN, true negative; TP, true positive.

Table 4 presents the average precision for area‐ and box‐based segmentation accuracy, summarized as mean ± SD over five independent training runs. Overall, the area‐based average precision was slightly lower than the box‐based precision, with an mAP of 0.622 for area segmentation and 0.662 for box segmentation (± SD shown in Table 4). The YOLACT model achieved high segmentation accuracy in large, well‐defined structures, such as the cranium, with an average precision of 0.965, chest circumference of 0.906, and abdominal circumference of 0.926. However, the model's area average precision was notably lower in complex or fine‐detailed structures. In particular, the brain midline with falx at 0.204 showed a significant decrease compared with its box‐based precision of 0.760, indicating challenges in accurately delineating this structure. Similarly, nasal bone of 0.431, nuchal translucency of 0.472, and thigh of 0.312 displayed lower area‐based precision. Among limb structures, the segmentation accuracy varied. Hand and forearm segmentation performed relatively well, with values above 0.6, whereas upper arm, thigh, and foot segmentation showed moderate accuracy less than 0.6. The‐set performance was stable across five independent random splits, and full trial‐wise results are provided in Supporting Information S1: Table S4 and S5.

TABLE 4.

Structure‐wise segmentation performance over five random‐split training runs.

Classification	Structure	Area average precision	Box average precision
Mean average precision		0.622 ± 0.010	0.662 ± 0.011
Head	Cranium	0.965 ± 0.005	0.993 ± 0.001
	Choroid plexus	0.892 ± 0.049	0.902 ± 0.018
	Midline with falx	0.204 ± 0.024	0.760 ± 0.029
Neck	Nasal bone	0.431 ± 0.011	0.777 ± 0.019
	Nuchal translucency	0.472 ± 0.043	0.529 ± 0.016
	Palate	0.747 ± 0.039	0.496 ± 0.041
Thorax	Thoracic circumference	0.906 ± 0.016	0.958 ± 0.001
	Heart	0.623 ± 0.009	0.574 ± 0.017
Abdomen	Abdominal circumference	0.926 ± 0.011	0.948 ± 0.003
	Stomach	0.785 ± 0.036	0.683 ± 0.058
Extremity	Upper arm	0.447 ± 0.023	0.460 ± 0.019
	Forearm	0.627 ± 0.017	0.611 ± 0.018
	Hand	0.620 ± 0.011	0.623 ± 0.009
	Thigh	0.312 ± 0.014	0.280 ± 0.017
	Calf	0.526 ± 0.016	0.498 ± 0.024
	Foot	0.569 ± 0.017	0.551 ± 0.013
Spine	Spine	0.520 ± 0.015	0.605 ± 0.020

Note: Mean average precision is the mean value of area average precision and box average precision measures across all fetal anatomical structures. Results are calculated on a fixed 957‐image test set and summarized as mean ± standard deviation over five repeated training runs with 4:1 train/validation splits (random seeds). Metrics include box average precision and area average precision. Values range from 0 to 1, with higher values indicating greater precision.

Inference speed is a critical performance metric for real‐time diagnostics. The YOLACT model processed 25.4 FPS, which is sufficient for use in real‐time diagnostic applications. Video 1 demonstrates the model applied to real‐time ultrasound images of a fetus, illustrating effective performance even during rapidly changing frames and consistently maintaining segmentation of fetal structures.

4. Discussion

We developed an algorithm based on YOLACT that accurately identifies first‐trimester fetal structures. Our model achieved an overall classification accuracy of 98.4%, with high segmentation accuracy for well‐defined anatomical structures such as the cranium, heart, and abdominal circumference, all achieving F1‐scores higher than 0.950. However, segmentation performance varied across structures, with lower recall observed for the nasal bone, nuchal translucency, upper arm, and foot. The model attained an area‐based mean average precision of 0.622 and a processing speed of 25.4 FPS. Its high classification accuracy, robust segmentation performance for major fetal structures, and fast processing speed make it well‐suited for real‐time application in ultrasound platforms, aiding in first‐trimester fetal structure identification.

Our model was developed using a prospective collection of more than 20,000 first‐trimester ultrasound images and is the first to segment and classify first‐trimester fetal structures in such fine detail. For each classification, datasets ranging from 1000 to 6000 images provided sufficient data to enable optimal classification and segmentation. Experimental results demonstrated the model's ability to efficiently segment objects even in complex cross‐sectional images, highlighting its robustness.

In general, larger anatomical structures such as the cranium, heart, abdomen, and spine had higher segmentation accuracy. On the other hand, smaller structures such as the nasal bone and nuchal translucency showed notably lower accuracy. Representative examples of successful and ambiguous segmentations are shown in Figure 3, highlighting the model's strength in detecting well‐defined structures and its limitations in cases with suboptimal imaging conditions. A similar trend was on in average precision, with extremities also exhibiting lower performance. This may have been attributed to the limitations of the Feature Pyramid Network in YOLACT, which might not fully exploit multi‐scale features, leading to reduced accuracy for finer anatomical structures [12].

FIGURE 3.

Representative examples of successful and ambiguous segmentation. (A) A correctly segmented case with clear delineation of nasal bone, palate and NT in an optimal imaging view. (B) An ambiguous case in which the nasal bone was annotated by experts but is difficult to visually confirm; the NT view is also suboptimal. These borderline cases illustrate common challenges in early first‐trimester imaging.

Notably, the area‐based precision for the midline falx was significantly lower (0.204) compared to its box‐based precision (0.760). This discrepancy is primarily attributable to the resolution constraints of YOLACT's prototype masks and the inherent difficulty of segmenting thin, elongated structures.

To address these challenges, we propose several architectural enhancements for future implementation. These include incorporating additional high‐resolution feature maps or modifying the FPN, optimizing anchor boxes specifically for fine structures, and employing ensemble models with varied receptive fields [12, 13]. Techniques such as PointRend or refinement modules may also help improve segmentation of complex or narrow anatomical structures. These refinements will be explored in subsequent work to enhance detection accuracy for clinically important fine structures.

We conducted a comparative evaluation of YOLACT against MS‐RCNN and Mask R‐CNN under identical training and evaluation conditions. Nevertheless, YOLACT achieved substantially higher inference speed (25.4 FPS) compared to MS‐RCNN (14.6 FPS) and Mask R‐CNN (11.4 FPS), despite slightly lower segmentation performance (mAP 0.623 vs. 0.701 and 0.722, respectively). Given the practical need for real‐time application in first‐trimester ultrasound, YOLACT was selected as the most suitable model for this task. Detailed comparative results are provided in Supporting Information S1: Table S1 and S2.

Some recent studies have investigated AI models for identifying key fetal structures during the first trimester [14, 15, 16, 17]. Lin et al. developed a model for identifying fetal head structures during the 10–14 14‐weeks gestational period, achieving an AUC of 0.9774 [14]. This study used a hierarchical detection model to classify nine key structures to optimize nuchal translucency screening views. While selecting specific structures or views for model application can improve accuracy, it may be challenging to implement in real‐time applications [15, 16]. Ryou et al. developed an algorithm using FCN to segment and measure the brain, abdomen, and limbs from 3D volumes of fetuses aged 11–14 weeks [17]. This algorithm achieved 99% classification accuracy and 90%‐pixel accuracy for key anatomical structures. Similar to our study, this research targets multiple anatomical regions, including the head, abdomen, and limbs. However, it does not categorize structures in as much detail as in our study. Additionally, while it employs semantic segmentation, instance segmentation is a more effective approach for distinguishing individual structures.

To our knowledge, the present study represents the first attempt to comprehensively identify fetal structures in the first trimester, bridging the gaps between fetal structure segmentation and clinical applications in real‐time ultrasonography. By integrating real‐time segmentation, this technology not only reduces the time required for image acquisition and interpretation but also enhances the ability of clinicians to explain fetal structures to expectant mothers, improving their understanding of ultrasound imaging. This tool is particularly valuable for non‐expert operators with limited experience in first‐trimester fetal anatomy, enabling them to better understand fetal structures and perform examinations with greater confidence. A recent study demonstrated that an AI‐assisted ultrasound system significantly improved the efficiency of obstetric ultrasound training [18]. Similarly, our model, by offering real‐time segmentation, has the potential to serve as an assistive tool for operators. Additionally, its capability to rapidly analyze large datasets opens new opportunities for the broad‐scale study of first‐trimester fetal structures, paving the way for further advancements in prenatal imaging and research.

Our study had several challenges and limitations. Evaluating extremities in first‐trimester fetuses is particularly challenging, as accurately measuring the ossified diaphysis of bones has proven difficult [19]. Therefore, we mainly assessed segmentation accuracy for the boundaries, including soft tissue and skin, rather than strictly evaluating ossified bone structures. However, the accuracy for these boundaries was lower than that for clearly defined structures. Additionally, boundary precision was relatively low owing to the limited resolution of the prototype masks, which hindered the capture of fine boundary details. Increasing the resolution of prototype masks could help address this issue [20].

A few additional technical challenges were also noted. In cases where objects within the same class were densely packed, duplicate detections were observed. This limitation may have emerged from constraints in the non‐maximum suppression (NMS) process. Implementing more efficient NMS algorithms or adopting soft‐NMS could reduce such duplicate detections [21]. Moreover, as a lightweight model optimized for real‐time processing, YOLACT may lack the expressiveness of more complex networks [22]. Future work could explore the balance of real‐time efficiency with enhanced model complexity to improve overall performance. While repeated random splits mitigate sensitivity to a single train/validation partition, the evaluation relies on one fixed test cohort (5%), which may under‐represent distributional shifts. We acknowledge this limitation and recognize that external validation, including a prospective clinical trial, will be necessary to further establish the robustness and real‐world clinical utility of our model.

We also acknowledge that operator‐related factors, maternal characteristics (e.g., BMI), and image quality may affect model performance. However, due to the pseudonymization of clinical data and the multicenter design of our dataset, we were unable to evaluate these variables in the present study. Nonetheless, we performed a subgroup analysis comparing segmentation performance between transvaginal and transabdominal ultrasound approaches and observed comparable results across both modalities. These findings suggest that model performance was not substantially affected by the imaging approach, although transvaginal imaging showed slightly higher performance for certain extremity structures. Detailed results of this sub‐analysis are provided in Supporting Information S1: Table S6.

In conclusion, our developed YOLACT model presents a promising approach for segmenting first‐trimester fetal structures, with significant potential to aid in early detection of fetal anomalies. Its real‐time capabilities and memory efficiency make it well‐suited for application in real‐time sonography. Further refinements addressing the identified limitations could enhance the clinical utility of our model and ensure its broader applicability.

Funding

This study was supported by a grant from the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: RS‐2023‐KH134708) and research funding from Samsung Medison. The funding body played no role in the study design, image acquisition, or model performance testing.

Ethics Statement

This study received approval from the Central Institutional Review Board of The Catholic Medical Center in the Republic of Korea (IRB No. XC22EIDI0054). All participants provided written informed consent prior to inclusion in the study. The research was conducted in accordance with the principles of the Declaration of Helsinki and relevant institutional guidelines for human research ethics.

Consent

Written informed consent was obtained from all adult participants included in the study.

Conflicts of Interest

Jaehong Lee and Kwang Yeon Choi, as research engineers at Samsung Medison, were involved in developing the deep learning model using the provided images. Other authors have no conflicts of interest.

Supporting information

Supporting Information S1

Video S1: Application of the model to real‐time ultrasound imaging of a first‐trimester fetus.

Data Availability Statement

The ultrasound image dataset used in this study contains over 20,000 first‐trimester fetal scans and includes sensitive patient information. Therefore, the data are not publicly available due to ethical and legal restrictions.