Three-dimensional segmentation from two-dimensional intracardiac echocardiography imaging
For three-dimensional segmentation from two-dimensional intracardiac echocardiography imaging, the three-dimension segmentation is output by a machine-learnt multi-task generator. Rather than the brute force approach of training the generator from 2D ICE images to output a 2D segmentation, the generator is trained from 3D information, such as a sparse ICE volume assembled from the 2D ICE images. Where sufficient ground truth data is not available, computed tomography or magnetic resonance data may be used as the ground truth for the sample sparse ICE volumes. The generator is trained to output both the 3D segmentation and a complete volume (i.e., more voxels represented than in the sparse ICE volume). The 3D segmentation may be further used to project to 2D as an input with an ICE image to another network trained to output a 2D segmentation for the ICE image. Display of the 3D segmentation and/or 2D segmentation may guide ablation of tissue in the patient.
1. A system comprising:
a generator network configured to generate a three-dimensional segmentation when input a sparse intracardiac echocardiography (ICE) volume, the generator network trained adversarially using a first discriminator network and a second discriminator network, the first discriminator network configured to judge three-dimensional segmentations and the second discriminator network configured to judge complete volumes, the adversarial training using a plurality of sparse ICE volumes and ground truth paired three-dimensional segmentations and ground truth complete volumes as training data; and
a display configured to display the three-dimensional segmentation.
2. The system of claim 1 , wherein the generator network is further configured to output a complete volume when input the sparse volume.
3. The system of claim 1 , wherein the ground truth complete volumes are computed from a computed tomography or magnetic resonance scan.
4. The system of claim 1 , wherein the ground truth complete volumes each include more voxels than respective paired sparse ICE volumes.
5. The system of claim 1 , wherein the three-dimensional segmentation comprises boundaries for two or more heart structures of a patient.
6. The system of claim 1 , further comprising:
an image processor configured to generate the sparse ICE volume from a plurality of two-dimensional ICE images.
7. The system of claim 6 , wherein the image processor is further configured to determine a two-dimensional segmentation for one or more two-dimensional ICE images from the three-dimensional segmentation; and the display is configured to display the two-dimensional segmentation.
8. The system of claim 7 , wherein the image processor is configured to determine the two-dimensional segmentation by projecting the three-dimensional segmentation to a plane and generating a two-dimensional segmentation for the two-dimensional ICE images by input of the projection of the three-dimensional segmentation to the plane and input of the ICE image to a machine-learned network.
9. The system of claim 6 , further comprising:
a ICE transducer; and
an ultrasound imager configured to generate the plurality of two-dimensional ICE images of a part of a cardiac system of a patient with the ICE transducer.
10. The system of claim 9 , wherein the ICE transducer is in a catheter and further comprising a position sensor on the catheter, wherein the sparse ICE volume includes position information from the position sensor.
11. A method comprising:
acquiring a plurality of two-dimensional intracardiac echocardiography (ICE) images;
projecting the plurality of two-dimensional ICE images to three-dimensional space to form a sparse ICE volume;
generating a three-dimensional segmentation and a complete volume from the sparse ICE volume with a machine-learned multi-task generative adversarial network;
projecting the three-dimensional segmentation mask onto the plurality of two-dimensional ICE images to generate two-dimensional segmentations; and
displaying the two-dimensional segmentations.
12. The method of claim 11 , further comprising:
generating contours from the two-dimensional segmentations; and
displaying the contours.
13. The method of claim 11 , further comprising performing ablation for atrial fibrillation with an image of at least one of the two-dimensional segmentations as a guide.
14. The method of claim 11 , wherein the machine-learned multi-task generative adversarial network comprises a multi-task architecture with a first task being volume completion and a second task being three-dimensional segmenting.
15. The method of claim 11 , wherein projecting the three-dimensional segmentation mask comprises projecting to a plane and generating a two-dimensional segmentation for a two-dimensional ICE images by input of the projection and a two-dimensional ICE images to a machine-learned network.
16. The method of claim 11 , wherein the complete volume comprises more voxels than the sparse ICE volume.