Scanner Agnostic Ultrasound Image Interpretation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Ultrasound scanners imprint unique textures on images, acting as anatomical markers.
• Deep Learning (DL) models learn these scanner-specific textures, limiting their portability across different scanner types.
• This texture bias leads to suboptimal performance in tasks like image segmentation when models are applied to data from different scanners.

Purpose of the Study:

• To improve the portability of Deep Learning (DL) models for ultrasound image analysis across different scanner types.
• To adapt existing DL models to new data distributions without the need for retraining.
• To reduce the complexity of algorithmic pipelines for style transfer in medical imaging.

Main Methods:

• Utilized neural style transfer to transform the texture of unseen ultrasound data.
• Employed feature maps from a pre-trained DL model (for image interpretation tasks like segmentation) to guide the style transfer process.
• Avoided retraining the DL model by adapting the input data's texture to match the model's training distribution.

Main Results:

• Demonstrated successful texture transformation of ultrasound images to match a target scanner's distribution.
• Showcased significant improvement in segmentation outcomes after applying style transfer to images from a different scanner type.
• Validated the effectiveness of using the interpretation model's own feature maps for style transfer, simplifying the pipeline.

Conclusions:

• Texture style transfer is an effective method for enhancing the portability of DL models in ultrasound imaging.
• Leveraging the interpretation model's features for style transfer offers a computationally efficient and effective approach.
• This technique enables the use of pre-trained DL models on diverse ultrasound datasets without costly retraining, improving clinical applicability.