ROI-Free Assessment of In-vivo Image Quality with Feature Extraction and the Earth Mover's Distance

Conventional image quality metrics require manual selection of regions of interest and can lead to a subjective, myopic, and inefficient image assessment. We propose an automated, image-based metric that compares a test image to a reference distribution composed of Field II simulations free of phase aberration, reverberation, and off-axis scattering. By training an autoencoder on channel data of in-vivo and simulated images, we extracted a 64-dimensional feature vector for each pixel. We aggregated all pixels of an image into a 64-dimensional distribution and compressed it with K-means clustering. Lastly, we computed the similarity between test and reference distributions with the earth mover's distance (EMD). Simulation experiments demonstrate that EMD decreases monotonically as signal to clutter ratio increases. In-vivo experiments on cardiac cineloops suggest that EMD is relative consistent across frames, and appears to correlate with human perception of image quality.