Multi-view texture transfer for super-resolution reconstruction in murine echocardiography

Image-based models of cardiac contractility have enhanced the understanding of cardiac function, adding new depth and dimensionality to traditional functional diagnosis. Regional strain analysis through multi-view imaging has provided structural indices sensitive to subclinical changes in cardiac motion. In particular, deep-learning-based super-resolution reconstruction (SRR) has become increasingly prevalent for complete spatiotemporal analysis of cardiac motion. In this study, we propose a non-local attention network to generate super-resolution cardiac images from multi-view echocardiography in mice. The network addresses limitations in capturing (i) high-frequency spatial features in echocardiography and (ii) high rates of cardiac motion in mice through the normalized cross-correlation (NCC). The network takes as input a pair of 4DUS images or volumes. Each image pair is a concatenated representation of the low-resolution (LR) and high-resolution (HR) images. Prior to concatenation, the LR images were resampled from 128 × 128 × 7 to the target size of 128 × 128 × 32 using linear interpolation. The decoder stage of the CNN progressively refined the deformation field used to map the LR image to the HR image, maximizing spatial alignment between the two. SRR performance was evaluated through structural similarity and mean-squared analysis. Key structural details of the heart were produced with minimal blurring. The structural similarity index (SSIM) confirmed high preservation of local textures, while the peak RMSE indicated strong overall similarity to the HR reference (SSIM: SR vs. HR = 0.918 ± 0.056; RMSE: 3.006 ± 1.990). Overall, the approach closely approximated the reference HR images, presenting a pathway towards high-fidelity, complete spatiotemporal strain imaging.