AI supported fetal echocardiography with quality assessment

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Fetal echocardiography quality assessment is crucial for accurate diagnosis.
• Deep learning models show promise but require validation on diverse datasets.
• Real-world data variability can impact AI model performance.

Purpose of the Study:

• To develop and prospectively validate a deep learning model for assessing fetal echocardiography quality.
• To evaluate the impact of image quality on model performance and clinical agreement.
• To emphasize the importance of using real-world data for AI model development in medical imaging.

Main Methods:

• Trained a deep learning model on 5363 fetal echocardiography images from 2551 pregnancies (2008-2018).
• Performed prospective validation on 192 images from 100 patients.
• Assessed model accuracy using segmentation and a quality score (QS).
• Compared clinician and expert agreement on image quality and segmentation.

Main Results:

• The model achieved an average accuracy of 0.91 (SD 0.09), with higher accuracy (0.97, SD 0.03) for images with above-average QS.
• Clinicians rated 44.8% of prospective validation images as equal to manual capture, favoring auto-capture in 18.69% and manual in 36.51%.
• Higher QS correlated with better agreement on segmentation and QS among experts (p < 0.001).
• Low QS negatively affected model performance and clinician agreement.

Conclusions:

• Deep learning models for fetal echocardiography quality assessment are feasible.
• Image quality significantly influences AI model performance and clinical utility.
• Developing and validating AI models on real-world, 'noisy' data is essential for clinical translation.