Automatic liver segmentation using adversarial image-to-image network

Patent metadata

Field	Value
Publication number	US-10600185-B2
Application number	US-201815877805-A
Country	US
Kind code	B2
Filing date	Jan 23, 2018
Priority date	Mar 8, 2017
Publication date	Mar 24, 2020
Grant date	Mar 24, 2020

Abstract

A method and apparatus for automated liver segmentation in a 3D medical image of a patient is disclosed. A 3D medical image, such as a 3D computed tomography (CT) volume, of a patient is received. The 3D medical image of the patient is input to a trained deep image-to-image network. The trained deep image-to-image network is trained in an adversarial network together with a discriminative network that distinguishes between predicted liver segmentation masks generated by the deep image-to-image network from input training volumes and ground truth liver segmentation masks. A liver segmentation mask defining a segmented liver region in the 3D medical image of the patient is generated using the trained deep image-to-image network.

First claim

The invention claimed is:

1. A method for automated liver segmentation in a 3D medical image of a patient, comprising:
receiving a 3D medical image of a patient;
inputting the 3D medical image of the patient to a trained deep image-to-image network, wherein the trained deep image-to-image network is trained in an adversarial network together with a discriminative network based on a segmentation loss calculated as a voxel-wise cross entropy between predicted liver segmentation masks generated by the deep image-to-image network from input training volumes and ground truth liver segmentation masks, wherein the discriminative network distinguishes between the predicted liver segmentation masks and the ground truth liver segmentation masks; and
generating, using the trained deep image-to-image network, a liver segmentation mask defining a segmented liver region in the 3D medical image of the patient.

2. The method of claim 1 , wherein the adversarial network including the deep image-to-image network and the discriminator network is trained by iteratively alternating discriminator network training based on a first loss function and deep image-to-image network training based on a second loss function for a plurality of iterations.

3. The method of claim 2 , wherein the discriminator network training updates weights of the discriminator network to minimize the first loss function having a first component based on classification by the discriminator network of the ground truth liver segmentation masks and a second component based on classification by the discriminator network of the predicted liver segmentation masks generated by the deep image-to-image network from the input training volumes, and wherein minimizing the first loss function increases a probability of the discriminator network correctly classifying the ground truth liver segmentation masks as positive and increases a probability of the discriminator network correctly classifying the predicted liver segmentation masks as negative.

4. The method of claim 3 , wherein the deep image-to-image network training updates weights of the deep image-to-image network to minimize the second loss function having a first component that calculates the segmentation loss between the predicted liver segmentation masks generated by the deep image-to-image network from the input training volumes and the ground truth liver segmentation masks and a second component based on classification by the discriminator network of the predicted liver segmentation masks generated by the deep image-to-image network from the input training volumes, and wherein minimizing the second loss function decreases the segmentation loss calculated between the predicted liver segmentation masks and the ground truth liver segmentation mask and increases the probability of the discriminator network incorrectly classifying the predicted liver segmentation masks as positive.

5. The method of claim 4 , wherein the deep image-to-image network is a deep encoder-decoder network having a plurality of encoder layers, a plurality of decoder layers, and a plurality of branches connected to corresponding decoder layers with each of the plurality of branches terminating in a respective output layer that matches a size of the input training volumes, and the second loss function calculates the segmentation loss as a weighted combination of respective binary voxel-wise cross entropy loss terms for all of the output layers of the plurality of branches and a final output layer of the deep image-to-image network.

6. The method of claim 2 , wherein the deep image-to-image network is pre-trained based on the training volumes and corresponding ground truth liver segmentation maps prior to the adversarial network including the deep image-to-image network and the discriminator network being trained.

7. The method of claim 1 , wherein the 3D medical image of the patient is a 3D computed tomography (CT) volume of the patient.

8. An apparatus for automated liver segmentation in a 3D medical image of a patient, comprising:
means for receiving a 3D medical image of a patient;
means for inputting the 3D medical image of the patient to a trained deep image-to-image network, wherein the trained deep image-to-image network is trained in an adversarial network together with a discriminative network based on a segmentation loss calculated as a voxel-wise cross entropy between predicted liver segmentation masks generated by the deep image-to-image network from input training volumes and ground truth liver segmentation masks, wherein the discriminative network distinguishes between the predicted liver segmentation masks and the ground truth liver segmentation masks; and
means for generating, using the trained deep image-to-image network, a liver segmentation mask defining a segmented liver region in the 3D medical image of the patient.

9. The apparatus of claim 8 , wherein the adversarial network including the deep image-to-image network and the discriminator network is trained by iteratively alternating discriminator network training based on a first loss function and deep image-to-image network training based on a second loss function for a plurality of iterations.

10. The apparatus of claim 9 , wherein the discriminator network training updates weights of the discriminator network to minimize the first loss function having a first component based on classification by the discriminator network of the ground truth liver segmentation masks and a second component based on classification by the discriminator network of the predicted liver segmentation masks generated by the deep image-to-image network from the input training volumes, and wherein minimizing the first loss function increases a probability of the discriminator network correctly classifying the ground truth liver segmentation masks as positive and increases a probability of the discriminator network correctly classifying the predicted liver segmentation masks as negative.

11. The apparatus of claim 10 , wherein the deep image-to-image network training updates weights of the deep image-to-image network to minimize the second loss function having a first component that calculates the segmentation loss between the predicted liver segmentation masks generated by the deep image-to-image network from the input training volumes and the ground truth liver segmentation masks and a second component based on classification by the discriminator network of the predicted liver segmentation masks generated by the deep image-to-image network from the input training volumes, and wherein minimizing the second loss function decreases the segmentation loss calculated between the predicted liver segmentation masks and the ground truth liver segmentation mask and increases the probability of the discriminator network incorrectly classifying the predicted liver segmentation masks as positive.

12. The apparatus of claim 11 , wherein the deep image-to-image network is a deep encoder-decoder network having a plurality of encoder layers, a plurality of decoder layers, and a plurality of branches connected to corresponding decoder layers with each of the plurality of branches terminating in a respective output layer that matches a size of the input training volumes, and the second loss function calculates the segmentation loss as a weighted combination of respective binary voxel-wise cross entropy loss terms for all of the output layers of the plurality of branches and a final output layer of the deep image-to-image network.

13. The apparatus of claim 9 , wherein the deep image-to-image network is pre-trained based on the training volumes and corresponding ground truth l

Assignees

• Siemens Healthcare Gmbh

Inventors

• Yang Dong
• Xu Daguang
• Zhou Shaohua Kevin
• Georgescu Bogdan
• Chen Mingqing
• Comaniciu Dorin

Classifications

• G06T7/11 Primary

  Region-based segmentation · CPC title

• G06T7/143

  involving probabilistic approaches, e.g. Markov random field [MRF] modelling · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T7/0012

  Biomedical image inspection · CPC title