Generative augmentations for improved cardiac ultrasound segmentation using diffusion models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Segmentation in cardiac ultrasound faces challenges due to limited, varied labeled datasets and inconsistent annotation conventions.
• This hinders the development of robust segmentation models that generalize effectively to new datasets.

Purpose of the Study:

• To improve the generalizability of cardiac ultrasound segmentation models using generative data augmentation.
• To enhance segmentation robustness and accuracy of ejection fraction estimation without requiring additional annotated data.

Main Methods:

• Utilized diffusion models to create generative augmentations for cardiac ultrasound datasets.
• Applied generative augmentations alongside traditional augmentation techniques.
• Conducted a visual Turing test with experts to assess the realism of generated images.

Main Results:

• Generative augmentations significantly improved dataset diversity.
• Experts could not reliably distinguish between real and generated cardiac ultrasound images.
• Segmentation robustness improved by over 20 mm in Hausdorff distance on external datasets.
• Limits of agreement for ejection fraction estimation improved by up to 20% for out-of-distribution cases.

Conclusions:

• Generative augmentations from diffusion models enhance segmentation model performance and generalizability.
• This approach addresses data scarcity and annotation variability in cardiac ultrasound imaging.
• The method offers a viable strategy for improving AI model reliability in medical diagnostics.