Photoacoustic tomography: a machine-learning-based image comparator corrects motion blur, even on real palms

Rather than assessing the absolute quality of an image, the network learns to determine which of two reconstructions is the best, thereby guiding an iterative search for the optimal parameters.

In 3D photoacoustic tomography, prolonged mechanical scanning of sparse sensor arrays causes motion artifacts that degrade the images. Because each subreconstruction is of poor quality, directly tracking the patient's movement is difficult, and conventional regularizations do not reflect the actual quality of a photoacoustic image.

Compare rather than rate

The authors propose "deep preferential evolution" (DPE), a gradient-free optimization in which a learned comparator directs evolutionary research on reconstruction parameters. Trained to judge the difference in quality between two images of the same target rather than to assign an absolute score, this comparator is better able to switch from simulation to in vivo images, after a simple unsupervised calibration. The method corrects synthetic movements on various anatomies and backgrounds and attenuates artifacts due to natural movements during palm examinations. This is a promising avenue for restoring medical images when annotated training data is missing.