US 12,254,538 B2
Devices and process for synthesizing images from a source nature to a target nature
Kumar Shreshtha, Paris (FR); Aurelien Lombard, Paris (FR); and Nikos Paragios, Paris (FR)
Assigned to THERAPANACEA, Paris (FR)
Filed by THERAPANACEA, Paris (FR)
Filed on Jan. 14, 2022, as Appl. No. 17/576,041.
Claims priority of application No. 21305043 (EP), filed on Jan. 14, 2021.
Prior Publication US 2022/0222873 A1, Jul. 14, 2022
Int. Cl. G06T 11/00 (2006.01); A61B 34/10 (2016.01); G06N 20/20 (2019.01); G06T 1/20 (2006.01); G06T 7/00 (2017.01); G06T 7/30 (2017.01); G06T 7/64 (2017.01); G06V 10/774 (2022.01); G16H 30/40 (2018.01)
CPC G06T 11/008 (2013.01) [A61B 34/10 (2016.02); G06N 20/20 (2019.01); G06T 1/20 (2013.01); G06T 7/0012 (2013.01); G06T 7/30 (2017.01); G06T 7/64 (2017.01); G06V 10/7747 (2022.01); G16H 30/40 (2018.01); G06T 2207/10081 (2013.01); G06T 2207/10088 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30004 (2013.01)] 15 Claims
OG exemplary drawing
 
1. A device for synthesizing images from a source imaging modality to a target imaging modality through unsupervised machine learning, on the basis of an original training set of unaligned original source images compliant with the source imaging modality and original target images compliant with the target imaging modality, said device comprising:
at least one input adapted to receive the original training set,
at least one processor configured for training a first machine learning architecture through an unsupervised first learning pipeline applied to the original training set, so as to generate a trained model of the first machine learning architecture, adapted to receive images compliant with the source imaging modality and to yield respectively associated images compliant with the target imaging modality, and representations of a plurality of said original source images compliant with the target imaging modality, called induced target images,
wherein said at least one processor is configured for training a second machine learning architecture through an at least partly supervised second learning pipeline applied at least to an induced training set of aligned image pairs, each of said aligned image pairs comprising a first item corresponding to one of said original source images, called a kept source image, and a second item corresponding to the induced target image associated with said kept source image, so as to generate a trained model of the second machine learning architecture, adapted to receive images compliant with the source imaging modality and to yield respectively associated images compliant with the target imaging modality,
said device further comprising:
at least one output adapted to produce at least part of said trained model of the second machine learning architecture, so as to carry out image syntheses from the source imaging modality to the target imaging modality.