US 12,292,494 B2
Methods and systems for segmenting organs in images using a CNN-based correction network
Deshan Yang, St. Louis, MO (US); and Yabo Fu, St. Louis, MO (US)
Assigned to Washington University, St. Louis, MO (US)
Appl. No. 17/264,620
Filed by Washington University, St. Louis, MO (US)
PCT Filed Jul. 30, 2019, PCT No. PCT/US2019/044118
§ 371(c)(1), (2) Date Jan. 29, 2021,
PCT Pub. No. WO2020/028352, PCT Pub. Date Feb. 6, 2020.
Claims priority of provisional application 62/850,225, filed on May 20, 2019.
Claims priority of provisional application 62/712,619, filed on Jul. 31, 2018.
Prior Publication US 2021/0290096 A1, Sep. 23, 2021
Int. Cl. G01R 33/56 (2006.01); A61B 5/00 (2006.01); A61B 5/20 (2006.01); G01R 33/48 (2006.01); G06F 18/21 (2023.01); G06F 18/214 (2023.01); G06N 3/0455 (2023.01); G06N 3/0895 (2023.01); G06N 20/20 (2019.01); G06T 7/12 (2017.01); G06V 10/82 (2022.01)
CPC G01R 33/5608 (2013.01) [A61B 5/201 (2013.01); A61B 5/4238 (2013.01); A61B 5/4244 (2013.01); A61B 5/4255 (2013.01); A61B 5/4836 (2013.01); A61B 5/7267 (2013.01); G01R 33/4808 (2013.01); G06F 18/2148 (2023.01); G06F 18/2185 (2023.01); G06F 18/2193 (2023.01); G06N 3/0455 (2023.01); G06N 3/0895 (2023.01); G06N 20/20 (2019.01); G06T 7/12 (2017.01); G06V 10/82 (2022.01); G06T 2207/10088 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30056 (2013.01); G06T 2207/30084 (2013.01); G06T 2207/30092 (2013.01); G06V 2201/031 (2022.01)] 9 Claims
OG exemplary drawing
 
1. A method of segmenting contours of one or more abdominal cavity organs comprising:
(i) providing an image dataset;
(ii) applying a first convolutional neural network (sub-CNN1) to the image dataset, resulting in a label probability map of sub-CNN1 (P1) (segmentation results);
(iii) applying a correction network comprising a second convolutional neural network (sub-CNN2) and a third convolutional neural network (sub-CNN3) to P1 comprising:
(a) applying the second convolutional neural network (sub-CNN2) to the label probability map of sub-CNN1 (P1), resulting in a label probability map of sub-CNN2 (P2);
(b) concatenating the image datasets, P1, and P2; and
(c) applying the third convolutional neural network (sub-CNN3) to the image datasets, P1 and P2;
wherein applying the first convolutional neural network and the correction network results in accurately segmented organs in images; and
(iv) training sub-CNN1, sub-CNN2, and sub-CNN3, wherein training comprises:
training each of sub-CNN1, sub-CNN2, and sub-CNN3 parameters using whole 3D images to incorporate anatomical contextual information;
randomizing sub-CNN1, sub-CNN2, and sub-CNN3 parameters using a Gaussian distribution;
calculating a cross entropy loss function using a softmax classifier for sub-CNN1, wherein sub-CNN2 and sub-CNN3 are constant;
calculating a cross entropy loss function using a softmax classifier for sub-CNN2, wherein sub-CNN1 and sub-CNN3 are constant; and
calculating a cross entropy loss function using a softmax classifier for sub-CNN3, wherein sub-CNN1 and sub-CNN2 are constant.