US 11,704,799 B2
Systems and methods for medical image style transfer using deep neural networks
Zhijin Li, Paris (FR); Serge Muller, Guyancourt (FR); Giovanni Palma, Chaville (FR); and Razvan Iordache, Paris (FR)
Assigned to GE Precision Healthcare LLC, Wauwatosa, WI (US)
Filed by GE Precision Healthcare LLC, Wauwatosa, WI (US)
Filed on Apr. 20, 2022, as Appl. No. 17/725,066.
Application 17/725,066 is a division of application No. 16/752,461, filed on Jan. 24, 2020, granted, now 11,348,243.
Prior Publication US 2022/0245814 A1, Aug. 4, 2022
Int. Cl. G06K 9/00 (2022.01); G06T 7/00 (2017.01); G16H 30/40 (2018.01)
CPC G06T 7/0014 (2013.01) [G16H 30/40 (2018.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30096 (2013.01); G06T 2207/30168 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A method comprising:
selecting a set of medical images, wherein the set of medical images are in a first style;
selecting a set of target images, wherein the set of target images are in a second style, and wherein the set of medical images and the set of target images are unpaired;
selecting a clinical quality estimator;
selecting a style similarity estimator;
selecting an image content regularizer;
training a style transfer network using the set of medical images, the set of target images, the clinical quality estimator, the style similarity estimator, and the image content regularizer, to produce a trained style transfer network;
receiving a medical image, wherein the medical image is in the first style;
mapping the medical image to a style transferred medical image using the trained style transfer network, wherein the style transferred medical image is in the second style;
displaying the style transferred medical image via a display device; and
wherein training the style transfer network using the set of medical images, the set of target images, the clinical quality estimator, the style similarity estimator, and the image content regularizer, to produce the trained style transfer network comprises training the style similarity estimator using the set of target images, and
wherein the style similarity estimator comprises a neural network, and wherein training the style similarity estimator using the set of target images comprises:
selecting at random an image from either the set of target images or a set of non-target images;
mapping the image to a probability score using the style similarity estimator, wherein the probability score indicates a probability of the image belonging to the set of target images;
calculating a loss based on the probability score; and
updating parameters of the style similarity estimator to reduce the loss using a gradient descent algorithm.