Long-short diffeomorphism memory network for weakly-supervised ultrasound landmark tracking

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Ultrasound imaging offers low-cost, real-time acquisition valuable for clinical applications.
• Accurate anatomical landmark tracking is crucial for procedures like minimally invasive surgery and radiation therapy.
• Challenges in ultrasound landmark tracking include deformation, visual ambiguity, and partial observation.

Purpose of the Study:

• To develop an advanced deep learning framework for accurate anatomical landmark tracking in ultrasound videos.
• To address the challenges of landmark deformation and ambiguity in real-time ultrasound imaging.

Main Methods:

• Proposed a novel long-short diffeomorphism memory network (LSDM) with a learnable deformation prior.
• Introduced a diffeomorphic representation storing long and short temporal information in separate memory banks.
• Developed an expectation maximization memory alignment (EMMA) algorithm for iterative memory optimization.
• Enabled weakly-supervised training with minimal landmark annotations.

Main Results:

• The LSDM model demonstrated superior or competitive performance in landmark tracking accuracy.
• The method showed strong generalization capabilities across diverse ultrasound scanner types and modalities.
• Experimental results validated the effectiveness on both public and private ultrasound datasets.

Conclusions:

• The proposed LSDM framework effectively improves anatomical landmark tracking in challenging ultrasound environments.
• The novel memory-based approach and EMMA algorithm mitigate cumulative errors and local ambiguities.
• LSDM offers a robust and generalizable solution for critical clinical applications relying on precise ultrasound guidance.