Learned super resolution ultrasound for improved breast lesion characterization

An ultrasound scanning system comprising:

• an ultrasound scanner viewing a human patient after intravenous injection of microbubbles, said ultrasound scanner to generate a sequence of ultrasound images, wherein said sequence of ultrasound images is a sequence of B-mode ultrasound images and a sequence of contrast enhanced ultrasound images corresponding to said B-mode ultrasound images;
• a training system to provide synthesized images to an untrained sparse recovery neural network, said training system comprising:
• a high-resolution point source synthesizer to generate synthesized target images with a plurality of randomly selected point sources therein on a high-resolution grid, wherein both a number of said point sources and locations of said point sources on said high-resolution grid are randomly selected to generate each of said synthesized target images;
• and
• a low resolution convolver to convolve each said target image at least with a selected one of a plurality of different point spread functions to generate an associated synthesized low-resolution input image;
• a trained sparse recovery neural network, trained on said synthesized target images and said associated synthesized low-resolution input images, to generate, in response to at least one image of said sequence of contrast-enhanced ultrasound images, a super-resolved microbubbles image of locations of said microbubbles in said at least one image;
• and
• a super resolution image generator to receive said super-resolved microbubbles image and to generate therefrom a super-resolved microvasculature image of a microvasculature through which said microbubbles flowed.