TL;DR
This paper introduces an AI model based on nnU-NetV2 for automatically segmenting critical anatomical structures in fetal four-chamber view ultrasound images. The model achieves high accuracy in segmentation and can automatically measure the cardiac axis (CAx) and cardiothoracic ratio (CTR), demonstrating strong agreement with experienced sonographers and offering potential to improve prenatal cardiac screening efficiency and reliability.
What Problem Does This Paper Address?
Fetal echocardiography is crucial for detecting congenital heart defects (CHDs), with the four-chamber view being particularly important. Key metrics like the cardiac axis (CAx) and cardiothoracic ratio (CTR) are used to assess cardiac position and function. However, manual measurement of these parameters is time-consuming, prone to significant inter-observer variability, and can lead to increased patient anxiety and potentially misguided clinical decisions. While AI has shown promise in medical imaging, its application to automated fetal cardiac parameter measurement, especially with robust clinical validation against expert sonographers, has been limited. This study aims to bridge this gap by developing and validating an AI model for accurate and automated CAx and CTR measurements.
Methods
The core of this research is an AI-based model built upon the nnU-NetV2 framework. This state-of-the-art architecture is specifically designed for medical image segmentation and features a U-shaped design that integrates high-level semantic features with low-level detailed features. The training process employed a combined Dice loss and cross-entropy loss to balance pixel-wise accuracy and region-based similarity.
For measurement extraction, the study utilized advanced digital image processing techniques on the segmentation masks generated by nnU-NetV2. Specifically:
- Cardiac Axis (CAx): A skeleton line algorithm was used to accurately determine the long axis of the interventricular septum. This was combined with the anterior-posterior axis of the thoracic cavity, derived from the centers of mass of thoracic and spine masks, to calculate the CAx angle.
- Cardiothoracic Ratio (CTR): The model calculated the ratio of the heart area to the thoracic cavity area. This involved fitting ellipses to the segmented heart and chest masks using OpenCV's
fitEllipsemethod to obtain precise area measurements.
The model was trained from scratch, without pre-trained weights, to tailor it specifically to the fetal ultrasound dataset.
Datasets
The study utilized a dataset of 1,083 high-quality fetal four-chamber view ultrasound images. These images were acquired from various manufacturers (Samsung, GE, and Philips) at a single hospital. The inclusion criteria were pregnant women with a gestational age between 18 and 32 weeks and a singleton pregnancy, excluding cases with suspected or known fetal congenital heart disease. The dataset was divided into training/validation (867 images) and test (216 images) sets at an 8:2 ratio. For clinical validation, 100 images from the test set were randomly selected.
Key Insights and Ideas
- nnU-NetV2's Superiority: The study found that nnU-NetV2 outperformed other state-of-the-art semantic segmentation models (U-Net, U-Net++, DeepLabV3+, and SAN) in segmenting most anatomical structures, indicating its effectiveness for complex medical image segmentation.
- Automated Measurement Potential: The research demonstrates that AI can not only segment anatomical structures but also accurately perform complex measurements like CAx and CTR, which traditionally require expert manual input.
- Bridging the Expertise Gap: The AI model showed strong agreement with senior sonographers, suggesting it can serve as a valuable tool to assist less experienced sonographers and standardize measurements across different skill levels.
- Efficiency Gains: The automated measurement process is significantly faster than manual methods, potentially freeing up sonographer time and improving workflow efficiency.
- Clinical Relevance: The focus on CAx and CTR, which are critical parameters for prenatal screening, highlights the direct clinical applicability of this AI approach.
Key Results
- Segmentation Performance: The nnU-NetV2 model achieved a mean Dice coefficient (mDice) of 87.11% and a mean Intersection over Union (mIoU) of 77.68% for the segmentation of critical anatomical structures.
- CAx and CTR Agreement:
- Senior Sonographers: The AI model showed strong agreement with senior sonographers, with an Intraclass Correlation Coefficient (ICC) of 0.83 for CAx and 0.81 for CTR.
- Intermediate Sonographers: ICCs were 0.73 for CAx and 0.81 for CTR.
- Junior Sonographers: ICCs were 0.68 for CAx and 0.75 for CTR.
- Bland-Altman Analysis: Confirmed high agreement, with the senior sonographer exhibiting the smallest mean bias for CAx (1.15°) and CTR (0.0012). While junior sonographers showed larger deviations, the overall consistency with experienced sonographers was notable.
- Comparison with Other Models: nnU-NetV2 consistently outperformed other tested models across most evaluated classes in terms of Dice scores and IoU.
Limitations and Future Work
- Dataset Limitations: The current dataset, while substantial, may not fully represent the diversity of anatomical variations and pathologies encountered in clinical practice. Future work should involve expanding the dataset with larger, multicenter samples to improve generalizability.
- Interpretability: The decision-making process of the AI model may not be easily interpretable by clinicians, which could hinder adoption. Incorporating explainable AI techniques is suggested.
- Real-time Performance: The current implementation is for offline analysis. Adapting the model for real-time performance during live ultrasound examinations requires further optimization and integration with ultrasound systems.
- Scope Extension: The model is currently focused on the four-chamber view. Future research should extend its capabilities to other cardiac views and additional measurements for a more comprehensive fetal cardiac evaluation.
- Validation: Further validation with constrained sample sizes and multicenter assessments is needed to confirm generalizability and clinical utility.
Summary updated