Feedback Attention for Unsupervised Cardiac Motion Estimation in 3D Echocardiography

Md. Kamrul Hasan, Guang Yang, Choon Hwai Yap

Abstract: Echocardiography motion estimation is essential for clinical cardiac health assessments, such as myocardial strain and ejection fraction. Attaining reliable performance with deep learning image registration (DLIR) is traditionally challenging due to intrinsic noise and fuzzy anatomic boundaries in echocardiography. It is further advantageous to achieve DLIR in 3D, as the cardiac anatomy has complex 3D structures and motions that are difficult to understand in 2D, especially with disease and malformations. However, successful regularization strategies for 2D DLIR are often not as effective in 3D. To date, there have been few 3D DLIR implementations for adult echocardiography and none in fetal echocardiography, where small fetal cardiac structures pose additional challenges. Here, we propose a self-supervision module for an unsupervised DLIR network for 3D+time echocardiography and demonstrate its good performance in both adult and fetal echocardiography images. We introduce a novel feedback spatial transformer module where the registration outputs are used to generate a co-attention map that describes the remaining registration errors to guide the network’s spatial emphasis during DLIR training. This effectively facilitates self-supervision. This feedback attention approach could be added to existing transformer-based approaches, including the co-attention spatial transformer with and without the spatial and channel attention in the DLIR backbone, bringing about non-overlapping benefits with existing approaches. Our results with 3D images here suggest that a focus on the resulting image after registration warping is key to good DLIR performance, and this is consistent with our earlier 2D DLIR investigations. Codes are available at https://github.com/kamruleee51/Feedback_DLIR.