This study investigates the application of deep learning techniques for the classification of ultrasound images in prenatal and maternal care. The objective was to develop a robust convolutional neural network (CNN) model capable of accurately distinguishing between various anatomical structures including the fetal abdomen, brain, femur, thorax, and maternal cervix. A dataset comprising a diverse range of ultrasound images was used for training and evaluation purposes. The CNN model demonstrated exceptional performance in classification tasks, achieving average precision, recall, and F1 score metrics exceeding 98% across all classes. This indicates the model's capability to efectively identify and diferentiate critical features within ultrasound images relevant to fetal and maternal health. The results highlight the potential of deep learning in enhancing diagnostic accuracy and eficiency in prenatal and maternal health monitoring.
Ultrasound imaging is a critical component of prenatal care, providing invaluable insights into the
health and development of the fetus. It is widely used due to its non-invasive nature, safety, and ability
to ofer real-time visualization of the fetus [
Given these challenges, there is a growing interest in applying artificial intelligence (AI) to the ifeld of medical imaging. AI, particularly deep learning, has shown remarkable success in various (Co-located with the 17th International Conference on Verification and Evaluation of Computer and Communication Systems (H. Bellali)
CEUR
ceur-ws.org
image analysis tasks, from object detection to classi-fication[
Zhang et al. [
In this research, we harness the capabilities of the InceptionResNetV2 model, a sophisticated convolutional neural network, to tackle the challenges of fetal ultra-sound image classification. This study is distinguished by the integration of a meticulously curated dataset and advanced preprocessing techniques to maximize the performance of the AI model proposed. The dataset, collected in 2020, includes 12,400 2D ultrasound images from 896 pregnant women. These images are categorized into six anatomical regions: maternal cervix, thorax, femur, abdomen, brain, and other.
Our innovative approach incorporates several key steps. We rigorously cleaned the dataset to remove poor quality images and irrelevant information, ensuring a high standard of input data. We precisely cropped the images to focus on the relevant anatomical regions, enhancing the model’s ability to learn important features. We also applied the SMOTE techniques to balance the dataset, to ensure an even representation of all categories during training. The core of our methodology is the InceptionResNetV2 model, known for its hybrid architecture combining Inception modules and Residual connections. This design allows the model to eficiently handle high-dimensional data and capture intricate patterns in ultrasound images.
The remainder of this paper is structured as follows.Section II details the proposed model and technique, including model training and parameters. Section III presents the findings of the proposed model. Finally, Section IV concludes with a discussion of our results and potential future work.
The methodology proposed for fetal ultrasound image classification involved several key steps to ensure robust model training and accurate results. We began by compiling a dataset consisting of 12,400 2D ultrasound images from 896 pregnant women, sourced from hospitals in Barcelona, Spain, in 2020. Each image was meticulously annotated by specialist fetal doctors to label anatomical regions such as the maternal cervix, thorax, femur, abdomen, brain, and other structures.Data preprocessing was crucial to enhance the dataset quality. We conducted rigorous cleaning to remove images with artifacts or poor quality that could interfere with model train-ing. Additionally, we employed cropping techniques to focus the model’s attention on relevant anatomical features while removing extraneous elements from the images.We adopted the InceptionResNetV2 architecture for model development, leveraging its pretrained weights from the ImageNet dataset. This allowed the model to extract complex features from ultrasound images eficiently.Addressing class imbalance was essential, achieved through SMOTE to ensure each anatomical category had suficient representation during training. Evaluation of model performance included standard metrics such as accuracy and F1-score, providing comprehensive insights into its efectiveness.By integrating these methodologies, our study aims to contribute to advancements in prenatal care by developing a reliable automated system for fetal health assessment through ultrasound imaging. This approach enhances diagnostic capabilities and supports more efective patient care in clinical settings. the proposed methodology is presented in 1.
The dataset used, collected in 2020 from various hospitals in Barcelona, Spain, is comprehensive and
diverse, encompassing 12,400 2D ultrasound images from 896 pregnant women[
To prepare the dataset for efective model training, we implemented several key data preprocessing steps.
In this study on fetal ultrasound image classification, data cleaning played a pivotal role in ensuring the quality and reliability of our dataset. The primary objectives of our data cleaning process were twofold, ifrst, to enhance the clarity and relevance of the ultrasound images. Second, to standardize the dataset for consistent model training. Here’s how we approached data cleaning: • Artifact Removal: Ultrasound images often contain artifacts such as shadows, speckles, and machine specific annotations. We applied image processing techniques, such as noise reduction iflters and artifact removal algorithms, to eliminate these distractions. By doing so, we ensured that the model focused only on the anatomical structures relevant to fetal health assessment. • Normalization and Standardization: To facilitate efective model training, we normalized the intensity values of the images and standardized their dimensions. This preprocessing step ensured uniformity across the dataset, enabling the model to learn consistent features irrespective of variations in image acquisition parameters.
Preprocessing stepsensure that the model focuses on the most pertinent anatomical regions, thereby improving the accuracy and eficiency of the classification process. The advantages of using preprocessing in our research are significant. It enhances the model’s ability to concentrate on the essential features of each anatomical region, such as the maternal cervix, thorax, femur, abdomen, and brain parts. This focused learning improves the model’s capability to diferentiate between similar anatomical structures and identify subtle anomalies. This consistency is crucial for achieving reliable performance across diferent ultrasound machines and varying image acquisition conditions.
In this research, we employed Synthetic Minority Over-sampling Technique (SMOTE) [
The use of SMOTE in our dataset, which consists of 12,400 2D ultrasound images categorized into five anatomical regions, is particularly advantageous. Fetal ultrasound images often exhibit significant class imbalances, with some anatomical regions being underrepresented compared to others. This imbalance can lead to biased model training, where the model becomes more accurate in predicting the majority classes while underperforming on the minority ones. By applying SMOTE, we efectively mitigated this issue, ensuring that each class is adequately represented in the training process.
In this research, the choice of model played anessential role in achieving accurate and reliable results.
We opted for the InceptionResNetV2 architecture[
In our implementation, we leveraged the pretrained weights of InceptionResNetV2 trained on the ImageNet dataset. Transfer learning from ImageNet provides a significant advantage by initializing the model with weights that have already learned general features from a vast and diverse set of natural images. Fine-tuning the model on our specific fetal ultrasound dataset further tailored its parameters to better recognize anatomical structures such as the maternal cervix, thorax, femur, abdomen, and brainregions to fetal health assessment. After importing the pretrained InceptionResNetV2 without its fully connected layers, we introduced dropout layers strategically placed after the convolutional base and dense layers. These dropout layers mitigate overfitting by randomly deactivating neurons during training, promoting better generalization of the model. Following the dropout layers, we incorporated a flatten layer to reshape the output into a 1D vector, preparing it for input into subsequent dense layers. These dense layers, equipped with ReLU activation functions, facilitate the learning of complex, nonlinear relationships within the data, crucial for distinguishing between diferent structures in fetal ultrasound images. To stabilize and accelerate training, batch normalization layers were added after each dense layer, normalizing the activations and improving the model’s convergence speed and overall performance. This tailored approach not only leverages the powerful feature extraction capabilities of InceptionResNetV2 but also optimizes it for the nuanced requirements of medical image classification, ultimately advancing automated diagnostic tools for prenatal care.
In our study, we optimized various hyperparameters to enhance the performance of our model for fetal ultrasound image classification. The Adam optimizer was selected due to its adaptive learning rate capabilities, which help in eficiently handling the sparse gradients encountered in our dataset. We set the learning rate at 0.001, striking a balance between training speed and convergence stability. The categorical cross entropy loss function was employed to handle the multiclass classification task efectively. Our dataset was divided into training, validation, and test sets in a ratio of 64:16:20, ensuring a robust evaluation of our model’s generalizability. A batch size of 64 was chosen to make the training process manageable while maintaining computational eficiency. Finally, the model was trained for 50 epochs, allowing suficient iterations to learn complex patterns in the data without overfitting. These carefully selected hyperparameters contributed significantly to the superior performance metrics achieved in our study.
The proposed approach yielded promising results, demonstrating significant improvements in the accuracy and robustness of fetal ultrasound image classification compared to existing methods. The InceptionResNetV2 model achieved high classification accuracy across all five categories, with notable performance in distinguishing between anatomically similar regions such as the thorax and abdomen.
The training process spanned 50 epochs, during which both training and validation datasets were utilized to optimize the model. Initially, the model achieved a validation loss of 0.8232 and an accuracy of 95.23% in the first epoch. This initial validation performance provided a benchmark for subsequent epochs.
As training progressed, there was a consistent improvement in both training and validation metrics, indicating that the model efectively learned to generalize to unseen data. The validation accuracy regularly increased, reaching a peak of 99.60% by the 31st epoch. Notably, the validation loss continued to decrease, suggesting that the model’s predictions became more precise and aligned with ground truth labels. Despite the model’s impressive performance, there were instances where the training accuracy exceeded the validation accuracy slightly, suggesting some degree of over-fitting. However, the consistent decrease in validation loss and increase in validation accuracy throughout most epochs indicate that the model learned to generalize well to new data, mitigating overfitting to a considerable extent.
The model achieved a peak validation accuracy of 99.60%, indicating that it correctly classified the majority of samples in the validation set. The validation loss decreased from 0.8232 in the first epoch to as low as 0.0142 by the last epoch, highlighting the model’s improved predictive accuracy and consistency.
Figure 4 presents the evaluation of performance metrics throughout the training and validation phases, focusing on loss and accuracy.
Throughout the training, both precision and recall metrics consistently improved. Precision measures the model’s ability to correctly identify positive instances among all predicted positive instances, while recall measures the model’s ability to correctly identify positive instances among all actual positive instances. By the end of training, precision and recall values were consistently above 99%, indicating high confidence in the model’s predictions.
Overall, the model exhibits excellent performance across all evaluated metrics. It shows high precision and recall for each class, typically above 98%, indicating accurate predictions and efective identification of relevant instances. The F1 score also reflect a balanced trade of between precision and recall, consistently above 98%. With an overall accuracy of 99.09%, the model correctly predicts class labels for the majority of instances, demonstrating its efectiveness in distinguishing between diferent classes. The support (number of instances) for each class is well-distributed, which helps ensure the model learns efectively without bias towards any specific class.In summary, the model’s strong performance across accuracy, precision, recall, and F1 scores indicates its robustness and reliability in classifying instances into their respective classes. It is well suited for practical applications where accurate classification is essential. For future work, several avenues could be explored to potentially enhance the model’s performance and address broader considerations, experimenting with diferent neural network architectures or exploring more advanced models (e.g., deeper networks, attention mechanisms) could potentially capture more intricate patterns in the data and further boost performance. Thoroughly validating the model’s performance in real-world settings and monitoring its performance post-deployment to ensure consistency and reliability in diverse conditions.
The authors have not employed any Generative AI tools.