FetalDenseNet: multi-scale deep learning for enhanced early detection of fetal anatomical planes in prenatal ultrasound

Introduction

Prenatal ultrasound [1], [2], [3] is an important, simple imaging method that clinicians use widely to diagnose prenatal defects by assessing the fetus’s health, development, and position. Despite its widespread use, the traditional interpretation of ultrasound images remains complex, time-consuming, and dependent on professional expertise. Structural defects impact about one out of every 40 infants, and early detection is typically critical to effective intervention and care [4]. However, the complexity of embryonic anatomical components [5], the variety in picture quality, and the subjectivity of human interpretation all pose significant challenges to reliable and precise diagnosis.

In recent years, deep learning (DL) [6] has demonstrated itself as a significant method in healthcare, particularly in medical image analysis, by offering automated solutions with noteworthy performance in tasks such as disease detection, segmentation, and classification. Among the DL models, Convolutional Neural Networks (CNNs) [7] have shown outstanding success due to their ability to extract hierarchical features from image data, making them ideal candidates for addressing challenges in fetal ultrasound imaging. Existing studies have applied CNNs to classify fetal planes and detect anatomical structures [8], but several continuous difficulties still limit clinical deployment. Here, Rauf et al. [9] have introduced two novel convolutional neural network architectures, namely the 3-residual and the 4-residual-block models, each designed to enhance efficiency compared to ResNet 18 and ResNet 50. But they face challenges for training a deep learning model in an imbalanced dataset and can lead to the extraction of irrelevant information from deeper layers. Sarno et al. [10] evaluates the potential of AI in obstetrics for improving diagnosis but faces challenges such as the need for larger datasets, clinician training, and evidence-based guidelines for clinical integration. Xie et al. [11] verified the feasibility of classifying fetal brains using DL algorithms for the binary classification of normal or abnormal standard fetal ultrasound brain images in axial planes. However, their used dataset was not multicenter, which limited data diversity. Moreover, they focused only on binary classification (normal vs. abnormal) of fetal brain, lacking generalizability to other anatomical planes. These include limited documented datasets, imbalanced class distributions, difficulty in distinguishing visually similar anatomical regions, and the necessity for model interpretability in clinical decision-making.

This study aims to address these challenges by implementing and evaluating five different CNN architectures [12], such as VGG16, ResNet50, InceptionV3, DenseNet169, and MobileNetV2, on a large dataset (FETAL_PLANES_DB) [13] consisting of over 12,400 ultrasound images from 1,792 patients. In contrast to many existing studies, this research focuses on classification accuracy, the clinical scalability, and explainability of the models, providing the findings more applicable to real-world medical use. The study also demonstrates that high-performance DL in fetal ultrasound analysis can be achieved without the need for expensive computing infrastructure, making it more accessible to low-resource healthcare environments. The main objectives of this study are:

1. To present a reproducible and efficient DL method for classifying fetal ultrasound images across key anatomical planes using a clinically validated dataset.

2. To evaluate and compare the performance of five innovative CNN architectures like VGG16, ResNet50, InceptionV3, DenseNet169, and MobileNetV2.

3. To evaluate the clinical applicability and scalability of the proposed models, providing their real-world diagnostic utility, computational feasibility, and possible incorporation into prenatal care processes.

Several studies have used DL models to enhance prenatal ultrasound image interpretation, focusing on fetal plane categorization, anomaly diagnosis, and structure segmentation. This section reviews some existing methods, findings, and challenges. Rathika et al. [14] present a Radial Basis Function Neural Network (RBFNN) method and achieves a classification accuracy of 98.07 %. The study needs high computational resources and doesn’t explore deployment for real-time clinical applications. Yang et al. [15] collected 1779 normal and abnormal fetal US cardiac pictures in five standard views of the heart. They employed five You Only Look Once version 5 (YOLOv5) networks as their primary model to categorize photos as “normal” or “abnormal”. According to the study, their model achieved an overall accuracy of 90.67 %. Gong et al. [16] developed a novel generative adversarial network (GAN) model by integrating deep anomaly detection (DANomaly) and generative adversarial CNN (GACNN) architectures to detect fetal congenital heart disease (FCH) from echocardiography images. Using a modified WGAN-GP, they created the DGACNN model, achieving 85 % accuracy in detecting FCH.

Zhou et al. [17] introduced the Category Attention-Instance Segmentation Network for segmenting fetal cardiac four-chamber ultrasound images. By adding a Category Attention Module, the model reduces misclassification errors. However, they faced challenges in distinguishing individual instances and needed better context modeling on a more diverse dataset. Nurmaini et al. [18] analyzed four common heart images and three congenital heart abnormalities while utilizing a Mask-RCNN architecture to predict 24 features in fetal heart sectors. The model achieved a DICE score of 89.70 % and an IoU of 79.97 %. However, it was tested on a small dataset of 1,149 fetal heart images. Moradi et al. [19] proposed Multi-Feature Pyramid U-net (MFP-Unet) for automated left ventricle segmentation in 2D echocardiography. Trained on an augmented dataset of 1,370 images, MFP-Unet achieved a Dice score of 0.945 and Hausdorff Distance of 1.62. However, the study was limited by dataset size, image resolution, and computational resources. Qu et al. [20] introduced a differential CNN method for identifying six fetal brain planes by computing the element-wise difference between input images. Although they achieved 92.93 % accuracy using data augmentation, the approach is sensitive to equipment and human operational variations, which can affect model robustness and diagnostic precision. Xie et al. [21] developed a model using DCNN, U-Nets, and VGG-Net to identify embryonic brain abnormalities in fetal images, achieving an overall accuracy of 91.5 % and high F1-scores. Class activation mapping (CAM) helped visualize effectively, but the low IoU values indicate that better object detection methods are needed for more accurate localization.

Wang et al. [22] developed FB-ZWUNet to improve the cerebellum segmentation in fetal brain ultrasounds, achieving a Dice coefficient of 0.8743 and IoU of 0.7813. However, its performance on different ultrasound devices and poor-quality images remains untested. Ciobanu et al. [23] developed an automated method for classifying fetal abdominal planes using MobileNetV3Large and EfficientNetV2S, with accuracies of 79.89 % and 79.19 %. Table 1 presents a comparison of the existing methods and their respective performance metrics.

Methods

The study addresses the challenging task of classifying ultrasound images into five distinct anatomical categories. Figure 1 illustrates the entire process, from data collection and preprocessing to model training and performance evaluation. A detailed step-by-step process for fetal ultrasound image processing – covering data collection, model training, evaluation, and prediction – is presented in Algorithm 1 (Table 2).

Figure 1:

Architectural diagram of fetal ultrasound classification.

Data collection

This study utilizes the (FETAL_PLANES_DB) [13] consists of 1,792 patient records with 12,000+ ultrasound images in a real clinical environment. The six classes represent clinically significant fetal structures: (i) fetal abdomen, (ii) fetal brain, (iii) fetal femur, (iv) fetal thorax, (v) maternal cervix, and (vi) Other. The distribution of each type of fetal ultrasound image is as follows: fetal abdomen: 5.7 %, fetal brain: 24.9 %, maternal cervix: 13.1 %, fetal femur: 8.4 %, fetal thorax: 13.9 %, Other: 33.9 %. Figure 2 shows the different maternal–fetal anatomical planes [24] and data representation before and after preprocessing. Each image was labelled by an Expert clinician to ensure high-quality annotations. The images in the dataset vary in size, which can lead to inconsistent visual representation and negatively impact the quality of analysis and model performance. Therefore, resizing them to a uniform dimension is essential for maintaining clarity and consistency.

Figure 2:

Visual representation of six different fetal ultrasound images before and after data preprocessing. (A) Different maternal–fetal anatomical planes and data representation before preprocessing, (B) different maternal–fetal anatomical planes after preprocessing.

Model training

This study optimized the models to separate fetal ultrasound images into five anatomical groups. Each model has been configured with ImageNet [27] trained weights, enabling transfer learning [27] by utilizing pretrained visual features. The final layers of each model are customized to the current classification task by replacing the original fully connected layers with new dense layers, followed by a softmax activation function for multi-class prediction. The Adam optimizer [28] is chosen here due to its adjustable learning rate and computing efficiency, which make it most effective in medical imaging tasks with a large number of parameters. Each model is trained over 25 to 50 epochs with a batch size of 16 or 32, based on model size, early stopping criteria, and memory usage. Figure 3 shows the structured model training process of this study.

Figure 3:

Comprehensive architecture of model training process with parameter specifications.

VGG16

The VGG16 [12] architecture is initialized using weights obtained from training on the ImageNet dataset and configured without its top classification layer. Equation (1) represents the fundamental convolutional operation used in the VGG16 architecture. Here, Z^(l) represent the output of the lth layer in the network. The terms W^(l) and b^(l) denote the weights and biases associated with that layer, respectively. The symbol ∗ refers to the convolution operation applied between the input and the kernel. The activation function used is the Rectified Linear Unit (ReLU), mathematically defined as ReLU(x) = max(0, x), which introduces non-linearity into the model while maintaining computational efficiency. It uses small 3 × 3 kernels to capture fine-grained spatial characteristics and recognize features such as the head, abdomen, or femur to distinguish fetal planes.

(1) Z ( l ) = ReLU ( W ( l ) ∗ Z ( l − 1 ) + b ( l ) )

ResNet50

ResNet50 [12] is initialized with trained weights of ImageNet and modified by appending a custom classification head. Equation (2) represents the fundamental operation used in the ResNet50 architecture. Here, Z^(l) is the input to the residual block, and Z^(l+1) is the output. The function F(Z^(l), W^(l)) represents the residual mapping learned by the block, typically consisting of a series of convolutional layers with ReLU activations and batch normalization. The original input Z^(l) is added directly to the output of F, enabling the network to learn modifications to the identity function. It is used to preserve important details in deep layers, effectively detect small differences in fetal images, and accurately distinguish between similar fetal planes.

(2) Z ( l + 1 ) = F ( Z ( l ) , W ( l ) ) + Z ( l )

InceptionV3

InceptionV3 [12] model is initialized with trained ImageNet weights, and its top layers are excluded to allow the addition of a task-specific classification head. Equation (3) represents the fundamental operation used in the DenseNet169 architecture. Here, X represents the input, Conv_k×k denote convolution with a k × k kernel, Pool_3×3 represents 3 × 3 max or average pooling, and Concat indicates depth-wise concatenation of feature maps. The model improves performance by using a learning rate scheduler, early stopping, model checkpointing, Adam optimization, and categorical cross-entropy.

(3) Z = Concat Conv 1 × 1 ( X ) , Conv 3 × 3 ( X ) ,   Conv 5 × 5 ( X ) , Pool 3 × 3 ( X )

InceptionV3 incorporates multiple responses in parallel, making it highly effective for capturing features at different levels. This is particularly relevant in fetal ultrasound where organs and anatomical regions vary significantly in size and orientation. Its efficiency and robustness help in learning discriminative features across varying fetal positions and scanning angles.

DenseNet169

The DenseNet169 [12] model is initialized with ImageNet pre-trained weights and configured without its top classification layers. Eq. (4) represents the fundamental operation used in the DenseNet169 architecture. Here, H^(l) is a composite function (BatchNorm → ReLU → Convolutionlayer) and [Z⁽⁰⁾, Z⁽¹⁾, …, Z^(l−1)] represents concatenation of all previous layer outputs. DenseNet169 improves feature reuse and gradient to detect subtle anatomical differences in fetal ultrasound images.

(4) Z ( l ) = H ( l ) [ Z ( 0 ) , Z ( 1 ) , … , Z ( l − 1 ) ]

MobileNetV2

The MobileNetV2 [12] model is trained using a transfer learning approach, where the base convolutional layers are initialized with pre-trained weights from the ImageNet dataset and all layers are halted to allow fine-tuning. Equation (5) represents the fundamental operation used in the MobileNetV2 architecture. Here, Conv_Depthwise(X) applies spatial filtering per channel, Conv_Pointwise performs 1 × 1 convolution for cross-channel fusion, and Conv_DS(X) represents the output.

(5) Conv DS ( X ) = Conv Pointwise Conv Depthwise ( X )

Results

All experiments are conducted on a computational system equipped with an Intel(R) Core(TM) i5-10400 CPU @ 2.90 GHz, 16 GB DDR4 RAM, and a 64 bit Windows 11 operating system (×64-based architecture), ensuring reproducibility and alignment with standard hardware configurations for model evaluation. The model’s performance [29] is evaluated using accuracy, precision, recall, F1-score, and confusion matrices. Each metric is computed based on the number of correctly and incorrectly predicted samples, categorized as True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN). This ensured a thorough and reliable assessment of its classification effectiveness. Equation (6)–(9) shows the calculating procedure of the evaluation metrics.

Performance metrics

The VGG16 model, fine-tuned with pre-trained ImageNet weights, achieved a test accuracy of 90.22 %. The ResNet50 model, with a test accuracy of 88 %, is close to VGG16, while InceptionV3 also performs similarly with an accuracy close to 90 %. MobileNetV2, optimized for speed and efficiency, achieved a 71 % accuracy. Among all models, DenseNet169 outperforms the others with the highest test accuracy of 92 % presented in Table 3.

Table 3:

Comparison of classification reports for different deep learning models.

VGG16 classification report
Class	Precision	Recall	F1-score	Support
Fetal abdomen	0.77	0.72	0.74	154
Fetal brain	0.97	0.97	0.97	625
Fetal femur	0.85	0.84	0.84	199
Fetal thorax	0.83	0.85	0.84	319
Maternal cervix	0.99	0.98	0.98	339
Other	0.88	0.89	0.88	828
ResNet50 classification report
Class	Precision	Recall	F1-score	Support
Fetal abdomen	0.66	0.60	0.63	154
Fetal brain	0.96	0.96	0.96	625
Fetal femur	0.83	0.80	0.82	199
Fetal thorax	0.77	0.87	0.82	319
Maternal cervix	0.98	0.99	0.98	339
Other	0.88	0.86	0.87	828
InceptionV3 classification report
Class	Precision	Recall	F1-score	Support
Fetal abdomen	0.76	0.77	0.76	154
Fetal brain	0.98	0.97	0.98	625
Fetal femur	0.82	0.80	0.81	199
Fetal thorax	0.85	0.83	0.84	319
Maternal cervix	1.00	0.99	0.99	339
Other	0.88	0.89	0.89	828
MobileNetV2 classification report
Class	Precision	Recall	F1-score	Support
Fetal abdomen	1.00	0.41	0.58	142
Fetal brain	0.71	1.00	0.83	615
Fetal femur	0.62	0.20	0.31	207
Fetal thorax	1.00	0.45	0.62	274
Maternal cervix	0.43	1.00	0.60	322
Other	1.00	0.64	0.78	839
DenseNet169 classification report
Class	Precision	Recall	F1-score	Support
Fetal abdomen	0.79	0.84	0.81	154
Fetal brain	0.97	0.99	0.98	625
Fetal femur	0.88	0.77	0.82	199
Fetal thorax	0.89	0.89	0.89	319
Maternal cervix	0.99	1.00	0.99	339
Other	0.89	0.89	0.89	828

Figure 4 shows the confusion matrix of all five CNN models. Figure 5 shows high classification accuracy for all classes, with AUC values between 0.87 and 1.00.

Figure 4:

Confusion matrix of various CNN models’. (A) VGG16 model performance. (B) ResNet50 model performance. (C) InceptionV3 model performance. (D) MobileNetV2 model performance. (E) DenseNet169 model performance.

Figure 5:

ROC curves of various CNN architectures used for fetal anatomical plane classification. (A) VGG16 model ROC curve. (B) ResNet50 model ROC curve. (C) InceptionV3 model ROC curve. (D) MobileNetV2 model ROC curve. (E) DenseNet169 model ROC curve.

Discussions

This study conducts a comprehensive evaluation of five deep learning models, VGG16, ResNet50, InceptionV3, DenseNet169, and MobileNetV2 for the classification of fetal ultrasound images across six anatomical planes. Among these models, DenseNet169 demonstrated superior performance, achieving a classification accuracy of 92 %, thereby highlighting its exceptional capability in recognizing intricate fetal structures, particularly the fetal brain and maternal cervix. InceptionV3 and VGG16 also exhibited competitive performance. In contrast, MobileNetV2, despite its lightweight and computationally efficient design, significantly underperformed with an accuracy of 65.83 %, especially in complex anatomical regions such as the fetal abdomen and femur. This study identified deeper CNN architectures, particularly DenseNet169, outperforming lighter models such as MobileNetV2 in classifying complex fetal ultrasound images. This indicates deeper networks are more effective at preserving subtle spatial features and fine-grained anatomical distinctions, which are important for accurate classification in medical imaging tasks. These findings are important because they demonstrate the clinical feasibility of incorporating advanced deep learning models into diagnostic workflows. DenseNet169, with high F1-scores across most classes, may be utilized as an option for sonographers, lowering diagnostic errors, particularly in resource-constrained environments. MobileNetV2’s low performance, despite being fast and efficient, shows the limitations of using lightweight models in high-precision tasks without customized optimizations. The models also show that almost all models performed better when recognizing the fetal brain and maternal cervix, suggesting that these anatomical planes are visually separate and simpler to segment. On the other hand, the fetal abdomen consistently appeared as the weakest class across all models, with significantly lower precision and recall. The inability of models to learn robust features is probably caused by both class imbalance and increased variation between different classes. The DenseNet169 model’s strong performance is significant not just for accuracy but also for usage in real-world medical contexts. Using this type of DL model in prenatal care can help in various ways, including reducing diagnostic time by automatically detecting important anatomical views, improving consistency by reducing differences between different doctors’ interpretations, and assisting doctors in making correct decisions. These advantages are particularly useful in clinics with limited resources, where professional doctors may not always be available.

Furthermore, while some studies have examined prenatal defects, authors frequently focused on specific structures; this research has utilized a complete multi-class categorization method. Compared to previous studies that primarily addressed limited anatomy or binary classification tasks, this study takes a more comprehensive approach. Our methodology covers multi-class categorization across six essential fetal anatomical planes, hence increasing its therapeutic value in real-world circumstances. Furthermore, previous research frequently did not include robust testing across varied anatomical complexity and failed to incorporate multi-plane detection. Our findings highlight the relevance of employing deeper CNN designs such as DenseNet169, which remains successful across almost all classes. This contributes to broader diagnostic applicability and sets a benchmark for automated fetal ultrasound classification systems.

Although the dataset is clinically labeled and large, it may not cover every aspect of fetal morphological changes among populations. This limitation could affect the model’s generalization when applied to new clinical settings. The training and evaluation were conducted on resource-constrained hardware. While this demonstrates the feasibility of Deep Learning on modest systems, it may have limited the exploration of larger architectures or more complex hyperparameter tuning. A slight imbalance in the number of samples across classes, particularly for the fetal abdomen and fetal femur, may have influenced the model’s performance. Models may have been biased toward classes with more examples, resulting in reduced recall and precision in underrepresented categories. To extend the scope and applicability of this study, future research will focus on several key areas. Firstly, expanding the dataset with fetal ultrasound images from multiple hospitals, different ultrasound machines, and varied patient populations will enhance the model’s generalization and robustness across diverse clinical settings. Secondly, exploring model compression and optimization strategies such as pruning, quantization, or the use of inherently lightweight architectures like MobileNetV2 can facilitate real-time deployment in environments with limited computational resources. Third, improving the explainability of model decisions through visualization techniques like Grad-CAM or SHAP can increase transparency and help clinicians understand and trust the model’s outputs. Lastly, integrating these models into real-time clinical workflows should be investigated. Real-time deployment during ultrasound scans has the potential to support practitioners with immediate feedback, thus improving both the speed and consistency of prenatal diagnostics.

Conclusions

This study proposes a deep learning based method for the classification of fetal ultrasound images using five customized CNN architectures. These models are trained and evaluated on the FETAL_PLANES_DB dataset, which contains six clinically significant anatomical planes. The objective is to evaluate the effectiveness of current deep learning approaches in accurately identifying fetal structures from ultrasound images. Among the models, DenseNet169 achieves the highest test accuracy of 92 %. The results show the need for deep learning models to support automated prenatal diagnostics, enabling clinicians to identify fetal anatomical features more efficiently and with reduced variability. This method performs effectively even on basic resource equipment, demonstrating its suitability for deployment in low-resource clinical environments. To provide higher reliability in real-world applications, future research should focus on increasing dataset diversity, optimizing model performance, and improving interpretability. Improving these fields is essential for incorporating deep models into reliable and efficient medical procedures.