US 11,790,496 B2
Image normalization increasing robustness of machine learning applications for medical images
Christian Huemmer, Lichtenfels (DE); Ramyar Biniazan, Nuremberg (DE); Andreas Fieselmann, Erlangen (DE); and Steffen Kappler, Effeltrich (DE)
Assigned to Siemens Healthcare GmbH, Erlangen (DE)
Filed by Siemens Healthcare GmbH, Erlangen (DE)
Filed on Mar. 17, 2021, as Appl. No. 17/204,030.
Claims priority of application No. 20166951 (EP), filed on Mar. 31, 2020.
Prior Publication US 2021/0304361 A1, Sep. 30, 2021
Int. Cl. G06T 5/00 (2006.01); G06N 20/00 (2019.01); G06N 3/04 (2023.01); G06V 10/50 (2022.01); G06V 30/24 (2022.01); G06T 3/40 (2006.01); G06T 5/40 (2006.01); G06T 7/00 (2017.01)
CPC G06T 5/007 (2013.01) [G06N 3/04 (2013.01); G06N 20/00 (2019.01); G06T 3/4084 (2013.01); G06T 5/009 (2013.01); G06T 5/40 (2013.01); G06T 7/0012 (2013.01); G06V 10/50 (2022.01); G06V 30/2504 (2022.01)] 18 Claims
OG exemplary drawing
 
1. A computer-implemented method for normalizing images from a type of image acquisition device, the computer-implemented method comprising:
receiving, at a decomposition unit, a set of image data with images, the image data having been generated by being converted from detector signals acquired at a detector of an image acquisition device, wherein the detector signals were converted by applying different settings of image acquisition device-specific processing algorithms;
decomposing, at the decomposition unit, each of the images of the set of image data into components by incorporating at least information from the different settings of the image acquisition device-specific processing algorithms;
normalizing, at a normalizing unit, each of the components via a first machine learning unit to produce normalized components, by processing at least information from the different settings of the image acquisition device-specific processing algorithms to provide a set of normalized images with a decreased variability score; and
receiving the normalized components at an input interface of a second machine learning unit and executing a machine learning algorithm with the normalized components for a defined task, wherein
the normalized components are entered as feature maps of a convolutional layer of a neural network, and
during training of the first machine learning unit and the second machine learning unit, at least a part of a plurality of parameters of the first machine learning unit are jointly trained with parameters used and processed in the second machine learning unit using a same cost function.