US 12,299,849 B2
Noise preserving models and methods for resolution recovery of x-ray computed tomography images
Roman Melnyk, New Berlin, WI (US); Madhuri Mahendra Nagare, Karmala (IN); Jie Tang, Merion Station, PA (US); Obaidullah Rahman, South Bend, IN (US); Brian E Nett, Wauwatosa, WI (US); Ken Sauer, South Bend, IN (US); and Charles Addison Bouman, Jr., West Lafayette, IN (US)
Assigned to GE Precision Healthcare LLC, Waukesha, WI (US); Purdue Research Foundation, West Lafayette, IN (US); and University of Notre Dame du Lac, South Bend, IN (US)
Filed by GE Precision Healthcare LLC, Wauwatosa, WI (US); Purdue Research Foundation, West Lafayette, IN (US); and University of Notre Dame du Lac, South Bend, IN (US)
Filed on Jun. 20, 2022, as Appl. No. 17/807,779.
Prior Publication US 2023/0410259 A1, Dec. 21, 2023
Int. Cl. G06T 5/70 (2024.01); A61B 6/03 (2006.01); G06T 7/00 (2017.01)
CPC G06T 5/70 (2024.01) [A61B 6/032 (2013.01); G06T 7/0012 (2013.01); G06T 2207/10081 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A system, comprising:
a memory that stores computer executable components; and
a processor that executes at least one of the computer executable components that:
generates a first set of pairs of training images from a second set of pairs of images, wherein each pair of images of the second set comprises an input image and a noise free ground truth image, wherein the input image is a blurred version of the ground truth image, and wherein the generating comprises, for each pair of images of the second set:
generating a pair of training images of the first set, comprising:
generating a training input image by adding a first amount of noise from a noise sample to the input image based on an input image scaling factor, and
generating a training ground truth image by adding a second amount of noise from the noise sample to the input image based on a ground truth scaling factor, wherein the second amount of noise is scaled to be substantially similar in intensity to the first amount of noise according to a defined threshold;
trains a machine learning based sharpening algorithm to generate sharpened images, using the first set of pairs of training images, by approximately minimizing a loss function associated with reducing image blur that determines respective errors between the training input images and the training ground truth images of the pairs of training images, and maintains a substantially similar intensity of noise between the training input image and the training ground truth image according to the defined threshold; and
sharpens, using the sharpening algorithm, a new input image to generate a sharpened image that has reduced image blur as compared to the new input image, wherein the new input image and the sharpened image comprises a substantially similar intensity of noise according to the defined threshold.