Prenatal prediction of neonatal respiratory morbidity: a radiomics method based on imbalanced few-shot fetal lung ultrasound images

Yi Guo, Jing Jiao, Yanran Du, Xiaokang Li, Yi Guo 0002, Yunyun Ren, Yuanyuan WangBmc Medical Imaging 2022 · DOI

TL;DR

This paper introduces a novel radiomics-based approach using fetal lung ultrasound images to predict neonatal respiratory morbidity (NRM). By combining radiomics features with clinical data (gestational age and gestational diabetes mellitus) and employing advanced machine learning techniques to handle imbalanced and limited datasets, the method achieves high diagnostic accuracy, offering a non-invasive alternative to traditional methods.

What Problem Does This Paper Address?

Neonatal respiratory morbidity (NRM), including conditions like respiratory distress syndrome (RDS) and transient tachypnea of the newborn (TTN), is a significant cause of infant mortality. Accurate prenatal prediction of NRM is crucial for timely intervention, such as administering antenatal corticosteroids to promote fetal lung maturation. However, current methods like amniocentesis are invasive and have limitations. Gestational age (GA) is often used as a proxy, but it leads to a high rate of false positives, exposing healthy fetuses to unnecessary treatment. Existing ultrasound-based methods have limitations, particularly in handling imbalanced datasets (where NRM cases are rare) and few-shot scenarios (limited data for NRM cases), which are common in clinical practice and can lead to biased and poorly generalizable models. This paper aims to address these challenges by developing a robust, non-invasive prenatal NRM prediction model using radiomics from fetal lung ultrasound images.

Methods

The study employs a multi-stage approach:

  1. Image Acquisition and Lung Segmentation: Fetal lung ultrasound images were acquired from singleton pregnancies. A radiologist manually delineated the region of interest (ROI) within the fetal lung, which was then confirmed by another physician.
  2. Feature Extraction and Selection:
    • Radiomics Features: A comprehensive set of 380 radiomics features were extracted from the lung ROIs. These included greyscale histogram features, texture features (GLCM, GLRLM, GLSZM, NGTDM), and wavelet features. The goal was to capture subtle textural information independent of ROI size and location.
  3. Clinical Features: Gestational Age (GA) and Gestational Diabetes Mellitus (GDM) were included as clinical predictors, known to be correlated with NRM.
  4. Feature Selection: A random forest-based method was used to rank feature importance, and the top 10 radiomics features were selected for the final model.
  5. Model Building and Evaluation:
    • Addressing Imbalance and Few-Shot Learning: To tackle the imbalanced and few-shot nature of the dataset, several techniques were explored:
  6. Data Augmentation: ADASYN was used to generate synthetic minority class samples.
  7. Cost-Sensitive Learning: Cost-sensitive SVM was employed to penalize misclassifications of the minority class more heavily.
  8. Ensemble Learning: AdaBoost was used to improve generalizability.
  9. Hybrid Methods: RUSBoost (combining undersampling with AdaBoost) and SMOTEBoost (combining oversampling with AdaBoost) were specifically investigated for their ability to handle both imbalance and few-shot learning simultaneously.
  10. Model Comparison: Various models were trained and compared on the original imbalanced dataset, balanced datasets, and using ensemble methods.

Datasets

  • Training and Test Sets: A total of 210 fetal lung ultrasound images from 210 singleton pregnant women were collected from the Obstetrics and Gynecology Hospital Affiliated to Fudan University, Shanghai, China. The dataset was split into training (approximately 80%) and test (approximately 20%) sets, maintaining the same proportion of NRM and normal cases in both.
  • Training set: 167 images (40 NRM, 127 Normal)
    • Test set: 43 images (11 NRM, 32 Normal)
  • New Benchmark: This study utilizes a dataset collected from an Asian population, which differs from previous studies that may have focused on European populations. The dataset is characterized by its imbalanced and few-shot nature, reflecting real-world clinical scenarios.

Key Insights and Ideas

  • Radiomics as a Robust Biomarker: The study demonstrates that radiomics features extracted from fetal lung ultrasound images can effectively capture subtle textural information indicative of NRM, proving robust to variations in ROI shape and location.
  • Clinical Features Enhance Prediction: Incorporating GA and GDM alongside radiomics features significantly improves NRM prediction performance, highlighting the synergistic value of combining imaging and clinical data.
  • Ensemble Learning for Imbalance: RUSBoost emerged as the most effective method for handling the imbalanced and few-shot dataset, outperforming other techniques like cost-sensitive SVM and ADASYN-augmented models. This suggests that ensemble methods that combine undersampling with boosting are well-suited for such challenging datasets.
  • ADASYN's Drawbacks: While data augmentation is a common strategy, ADASYN was found to exacerbate aliasing and introduce noise in this specific context, leading to poorer classification performance. This underscores the importance of carefully selecting data balancing techniques.

Key Results

  • Overall Performance: The best model (RUSBoost with GA, GDM, and radiomics features) achieved a balanced accuracy (bACC) of 0.83 and an Area Under the Curve (AUC) of 0.87 on the test set.
  • Sensitivity and Specificity: The model achieved a sensitivity (SENS) of 0.82 and a specificity (SPEC) of 0.84 on the test set.
  • Feature Set Effectiveness: Combining GA, GDM, and radiomics features yielded the best diagnostic performance, significantly outperforming models using only clinical features.
  • ROI Stability: Using either irregular or square ROIs for feature extraction resulted in similar classification performance, demonstrating the stability of the radiomics features.
  • Comparison to Previous Methods: The developed model's diagnostic efficacy is comparable to invasive amniocentesis tests and shows improved performance over some previous ultrasound-based NRM prediction methods, despite using a smaller training set size.

Limitations and Future Work

  • Retrospective Study: The study is retrospective, which limits definitive causal conclusions.
  • Limited Clinical Information: While GA and GDM were included, other clinical factors that might influence fetal lung development and NRM risk could be explored.
  • Unilateral Lung Analysis: The study focused on one lung. A comparative analysis of right and left lungs is needed to verify generalizability.
  • Multicenter Validation: The proposed method requires robust validation on multicenter datasets from different machines and operators to ensure its stability and applicability in clinical practice.
  • Operator Dependence: While manual delineation was confirmed, future work could explore automated segmentation to reduce operator dependence.
  • Broader Fetal Population: Investigating the model's performance across different gestational week groups and sexes is recommended.

The authors are currently conducting a multicenter study to address these limitations and further validate their findings.

Summary updated

BibTeX
@article{guo2022prenatal,
  author = {Yi Guo and Jing Jiao and Yanran Du and Xiaokang Li and Yi Guo and Yunyun Ren and Yuanyuan Wang},
  title = {{Prenatal prediction of neonatal respiratory morbidity: a radiomics method based on imbalanced few-shot fetal lung ultrasound images}},
  journal = {Bmc Medical Imaging},
  year = {2022},
  doi = {10.1186/s12880-021-00731-z},
}
Figure