US 12,333,677 B2
Methods and systems for super-resolution with progressive sub-voxel up-sampling
Rohan Keshav Patil, Karnataka (IN); and Sudhanya Chatterjee, Karnataka (IN)
Assigned to GE PRECISION HEALTHCARE LLC, Wauwatosa, WI (US)
Filed by GE Precision Healthcare LLC, Wauwatosa, WI (US)
Filed on Jun. 30, 2022, as Appl. No. 17/810,271.
Prior Publication US 2024/0005451 A1, Jan. 4, 2024
Int. Cl. G06T 3/4076 (2024.01); G06T 3/4046 (2024.01); G16H 30/40 (2018.01)
CPC G06T 3/4076 (2013.01) [G06T 3/4046 (2013.01); G16H 30/40 (2018.01)] 20 Claims
OG exemplary drawing
 
1. A method, comprising:
training a deep neural network, wherein training the deep neural network comprises:
generating, using a ground truth image having a first resolution, a second image having a second resolution, wherein the first resolution is higher than the second resolution;
predicting, using the second image, a third image at a third resolution higher than the second resolution; and
computing loss, wherein computing loss comprises comparing the third image to the ground truth image or an image generated from the ground truth image; and
progressively up-sampling an input image to generate a super-resolution output image by:
generating N intermediate images based on the input image, where N is equal to at least one, including a first intermediate image by providing the input image to the trained deep neural network, where a resolution of the first intermediate image is a multiple of a resolution of the input image, higher than the resolution of the input image, and can be any positive real value and not necessarily an integer value;
generating the super-resolution output image based on the N intermediate images, the super-resolution output image having a resolution higher than a respective resolution of each intermediate image of the N intermediate images and the resolution of the input image; and
displaying the super-resolution output image via a display device and/or storing the super-resolution output image to a computer memory.