Establish a Normal Fetal Lung Gestational Age Grading Model and Explore the Potential Value of Deep Learning Algorithms in Fetal Lung Maturity Evaluation
Prenatal evaluation of fetal lung maturity (FLM) remains a significant challenge in obstetrics. Lung immaturity, leading to surfactant deficiency, is a major cause of neonatal respiratory morbidity (NRM) and mortality, particularly in preterm and early-term infants. Despite advancements in neonatal care, NRM continues to be a leading complication in these populations. Additionally, gestational conditions such as gestational diabetes mellitus (GDM), pre-eclampsia (PE), oligohydramnios, and fetal intrauterine growth restriction can further impair fetal lung development. Accurate assessment of FLM is crucial, especially in the third trimester, to guide clinical decisions regarding antenatal corticosteroid (ACS) therapy and the timing of elective deliveries. Current methods for evaluating FLM rely on invasive procedures such as amniocentesis, which carry risks of complications. Non-invasive methods have been explored but often lack the diagnostic accuracy required for clinical use. This study aims to establish a normal fetal lung gestational age (GA) grading model using deep learning (DL) algorithms and evaluate its potential in assessing FLM.
The study utilized a dataset of 7013 ultrasound images obtained from 1023 normal pregnancies between 20 and 41+6 weeks of gestation. All pregnancies were free of complications that could affect fetal lung development, and all infants were born without neonatal respiratory diseases. The images were categorized into three classes based on GA: class I (20 to 29+6 weeks), class II (30 to 36+6 weeks), and class III (37 to 41+6 weeks). The dataset included 3323, 2142, and 1548 images in each class, respectively. The images were collected using eight different ultrasound machines from six manufacturers, ensuring a diverse and representative dataset.
The study employed a DL-based approach to classify fetal lung ultrasound images according to GA. The process involved three main steps: image pre-processing, building a classification network, and validating the model. In the pre-processing step, irrelevant information such as machine parameters was removed from the images. This was achieved using a threshold-based segmentation method to isolate the region of interest (ROI). The classification network was designed based on the DenseNet architecture, which includes convolutional layers, pooling layers, dense blocks, transition layers, and fully connected layers. The network was trained using cross-entropy loss and optimized using stochastic gradient descent with momentum. Data augmentation techniques, including random rotation, cropping, and flipping, were applied to prevent overfitting. The model’s performance was validated using ten-fold cross-validation, and metrics such as accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUC) were used to evaluate its effectiveness.
The results demonstrated that the DL-based model achieved high accuracy in classifying fetal lung ultrasound images according to GA. The sensitivities for classes I, II, and III in the independent test set were 91.7%, 69.8%, and 86.4%, respectively. The specificities for these classes were 76.8%, 90.0%, and 83.1%, respectively. The total accuracy of the model was 83.8%. The AUC values for classes I, II, and III were 0.982, 0.907, and 0.960, respectively. The micro-average AUC was 0.957, and the macro-average AUC was 0.949. These results indicate that the model can effectively distinguish between different GA categories based on fetal lung ultrasound images.
The study also compared the DL-based model with traditional machine learning algorithms, including random forests (RF), support vector machines (SVM), and naïve Bayes (NB). The DL model outperformed these traditional methods in terms of accuracy, sensitivity, specificity, and AUC. For example, the macroF1 and microF1 scores for the DL model were 81.8% and 83.8%, respectively, compared to 51.4%/52.9% for NB, 65.8%/69.3% for RF, and 71.1%/73.2% for SVM. The DL model also showed higher AUC values for all three classes when compared to NB, RF, and SVM. These findings highlight the superior performance of DL algorithms in classifying fetal lung ultrasound images.
The established normal fetal lung GA grading model has several clinical applications. It can help identify abnormal lung development caused by gestational diseases such as GDM and PE. Additionally, the model can assess lung maturity following ACS therapy, providing valuable information for clinical decision-making. The model’s ability to automatically extract and classify subtle image features related to GA makes it a promising non-invasive tool for FLM assessment. This approach reduces the reliance on invasive procedures and minimizes the risks associated with amniocentesis.
The study also addressed the limitations of traditional quantitative imaging methods, which often lack robustness due to variations in acquisition conditions. The DL-based model, however, is robust to small changes in image acquisition settings, such as depth and gain adjustments. This robustness enhances the model’s reliability and potential for widespread clinical use. Furthermore, the model’s end-to-end implementation allows for automatic feature extraction and classification, eliminating the need for manual feature design and improving classification accuracy.
Despite its strengths, the study has some limitations. The retrospective nature of the study and the relatively small dataset, particularly for classes II and III, may affect the generalizability of the results. Future research should aim to include larger and more diverse datasets to further validate the model’s performance. Additionally, the study focused on normal pregnancies, and further investigation is needed to evaluate the model’s effectiveness in pregnancies with complications affecting fetal lung development.
In conclusion, this study successfully established a normal fetal lung GA grading model using DL algorithms. The model demonstrated high accuracy in classifying fetal lung ultrasound images according to GA and has the potential to serve as a non-invasive method for assessing FLM. The model’s ability to identify abnormal lung development and evaluate lung maturity following ACS therapy makes it a valuable tool for clinical practice. The study also highlighted the superior performance of DL algorithms compared to traditional machine learning methods in this context. Future research should focus on expanding the dataset and exploring the model’s applicability in high-risk pregnancies. The findings of this study lay the foundation for the use of DL algorithms in prenatal FLM assessment and support further research in this area.
doi.org/10.1097/CM9.0000000000001547
Was this helpful?
0 / 0