Deformable detection transformers for domain adaptable ultrasound localization microscopy with robustness to point...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Super-resolution imaging in diagnostic ultrasound is advancing rapidly.
• Ultrasound Localization Microscopy (ULM) uses microbubbles (MBs) for sub-wavelength microvasculature imaging.
• Challenges include dynamic point spread functions (PSFs) and domain generalization issues in deep learning models.

Purpose of the Study:

• To develop a novel approach for precise microbubble localization in Ultrasound Localization Microscopy (ULM).
• To address challenges related to dynamic PSFs and simulation-to-in vivo data discrepancies.
• To improve the accuracy and reliability of super-resolution ultrasound imaging.

Main Methods:

• Proposed a DEformable DEtection TRansformer (DE-DETR) object detection network.
• Utilized multi-scale feature maps and a deformable attention module to handle object deformations.
• Employed a KDTree algorithm for efficient microbubble tracking across frames.

Main Results:

• Demonstrated improved precision and recall compared to existing methods.
• Successfully evaluated the approach using both simulated and in vivo ultrasound data.
• Showcased enhanced microbubble localization accuracy.

Conclusions:

• The DE-DETR approach offers a significant improvement for microbubble localization in ULM.
• This method shows potential for enhancing the performance of super-resolution ultrasound in clinical settings.
• Overcomes key limitations in current ULM techniques, paving the way for more accurate diagnostics.