Feedback Attention to Enhance Unsupervised Deep Learning Image Registration in 3D Echocardiography

Area of Science:

• Medical imaging
• Artificial intelligence
• Cardiovascular imaging

Background:

• Cardiac motion estimation is crucial for assessing heart health, with 3D echocardiography offering advantages over 2D.
• Deep learning image registration (DLIR) provides speed and precision for motion estimation but faces challenges in 3D echocardiography.
• Existing unsupervised 2D DLIR methods often fail in 3D, and few 3D implementations exist.

Purpose of the Study:

• To introduce a novel spatial feedback attention (FBA) module to enhance unsupervised 3D DLIR for cardiac motion estimation.
• To demonstrate the effectiveness and flexibility of the FBA module in improving 3D echo DLIR performance.
• To identify an optimal 3D DLIR configuration incorporating FBA for superior results.

Main Methods:

• Development of a spatial feedback attention (FBA) module that uses registration results to generate spatial co-attention maps, guiding DLIR to minimize errors.
• Integration of the FBA module with various DLIR network architectures, including those with transformer enhancements.
• Evaluation of the FBA module's performance on both fetal and adult 3D echocardiography datasets.

Main Results:

• The FBA module significantly improves the performance of multiple 3D DLIR designs, including transformer-enhanced networks.
• The FBA module demonstrates broad applicability across fetal and adult 3D echocardiography.
• An optimal configuration combining FBA, a spatial transformer, and an attention-modified DLIR backbone outperformed existing 3D DLIR methods.

Conclusions:

• The FBA module effectively enables and enhances unsupervised 3D DLIR for cardiac motion estimation.
• Spatial attention mechanisms are key to scaling DLIR from 2D to 3D.
• Focusing on post-registration image quality is a promising strategy for improving DLIR performance in 3D echocardiography.