SYNSTITCH: A SELF-SUPERVISED LEARNING NETWORK FOR ULTRASOUND IMAGE STITCHING USING SYNTHETIC TRAINING PAIRS AND...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Ultrasound (US) image stitching expands the field-of-view (FOV) by merging images from different probe positions.
• Registering US images with partial anatomical overlap presents a significant challenge in medical imaging.

Purpose of the Study:

• To introduce SynStitch, a novel self-supervised framework for 2D ultrasound (US) image stitching.
• To address the challenge of US image registration with limited overlapping content.

Main Methods:

• SynStitch employs a synthetic stitching pair generation module (SSPGM) using a patch-conditioned ControlNet.
• SSPGM creates realistic 2D US stitching pairs with known affine matrices from single images.
• An image stitching module (ISM) learns 2D US stitching using the synthetically generated paired data.

Main Results:

• The proposed SynStitch framework demonstrated superior performance in 2D US stitching compared to leading methods.
• Evaluations on a kidney ultrasound dataset showed significant improvements in both qualitative and quantitative analyses.
• The self-supervised approach effectively handles US images with partially overlapping anatomical structures.

Conclusions:

• SynStitch offers an effective self-supervised solution for challenging 2D ultrasound image stitching tasks.
• The framework's ability to generate synthetic data significantly aids in supervised learning for image registration.
• This approach holds promise for improving diagnostic capabilities through expanded FOV ultrasound imaging.