DefinePK

Abstract

Introduction: This project presents a successful implementation of a CNN and RNN-based deep learning model for detecting congenital heart defects in fetal echocardiography. The model achieved high accuracy and demonstrated good interpretability, making it a potential decision-support tool for clinicians. It can help reduce diagnostic errors, especially in resource-limited settings where expert interpretation may not always be available. Future work may involve expanding the model to detect different types of CHDs, real-time analysis, and integration into clinical workflows to support routine prenatal care.
Aim and Objective: The main aim of this project is to develop a deep learning-based diagnostic tool for detecting congenital heart defects in fetal echocardiography using CNN and RNN models.
Material and Methods: The study employed a dataset of fetal echocardiographic images labeled as normal or CHD-affected. Preprocessing steps included image resizing, normalization, and data augmentation to improve model generalization. A CNN architecture (such as VGG-19 or ResNet) was used to extract spatial features from the images. These features were then fed into an RNN model (LSTM or GRU) to capture temporal dynamics within image sequences. The hybrid CNN-RNN model was trained using the Adam optimizer and binary cross-entropy as the loss function. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC-ROC. Grad-CAM was used to visualize the regions of interest that contributed most to the model’s decision-making process.
Results: The deep learning model demonstrated high performance in distinguishing between normal and CHD-affected fetal hearts. It achieved an accuracy of 94.3%, precision of 92.1%, recall of 95.5%, F1-score of 93.8%, and an AUC-ROC value of 0.97. Grad-CAM visualizations showed that the model focused on clinically relevant heart regions, confirming the interpretability and clinical reliability of the approach. These results suggest that the CNN-RNN-based method can effectively support the diagnosis of congenital heart defects in fetal echocardiography.
Conclusion: This project presents a successful implementation of a CNN and RNN-based deep learning model for detecting congenital heart defects in fetal echocardiography. The model achieved high accuracy and demonstrated good interpretability, making it a potential decision-support tool for clinicians. It can help reduce diagnostic errors, especially in resource-limited settings where expert interpretation may not always be available. Future work may involve expanding the model to detect different types of CHDs, real-time analysis, and integration into clinical workflows to support routine prenatal care