Method and system for accurate boundary delineation of tubular structures in medical images using infinitely recurrent neural networks

Patent metadata

Field	Value
Publication number	US-10258304-B1
Application number	US-201715825304-A
Country	US
Kind code	B1
Filing date	Nov 29, 2017
Priority date	Nov 29, 2017
Publication date	Apr 16, 2019
Grant date	Apr 16, 2019

Abstract

A method and apparatus for automated boundary delineation of a tubular structure in a 3D medical image of a patient using an infinitely recurrent neural network (IRNN) is disclosed. An unraveled cross-section image corresponding to a portion of a tubular structure is extracted from 3D medical image. The unraveled cross-section image is divided into a plurality of image chunks. A boundary of the portion of the tubular structure is detected based on the plurality of image chunks using a trained IRNN. The trained IRNN repeatedly inputs a sequential data stream, including the plurality of image chunks of the unraveled cross-section image, for a plurality of iterations while preserving a memory state between iterations, and detects, for each image chunk of the unraveled cross-section image input to the trained IRNN in the sequential data stream, a corresponding section of the boundary of the portion of the tubular structure.

First claim

The invention claimed is:

1. A method for automated boundary delineation of a tubular structure in a 3D medical image of a patient, comprising:
extracting, from a 3D medical image of a patient, an unraveled cross-section image corresponding to a portion of a tubular structure in the 3D medical image;
dividing the unraveled cross-section image into a plurality of image chunks;
detecting a boundary of the portion of the tubular structure based on the plurality of image chunks using a trained infinitely recurrent neural network, wherein the trained infinitely recurrent neural network repeatedly inputs a sequential data stream including the plurality of image chunks of the unraveled cross-section image for a plurality of iterations, while preserving a memory state between iterations, and detects, for each image chunk of the unraveled cross-section image input to the trained infinitely recurrent neural network in the sequential data stream, a corresponding section of the boundary of the portion of the tubular structure.

2. The method of claim 1 , wherein extracting, from a 3D medical image of a patient, an unraveled cross-section image corresponding to a portion of a tubular structure in the 3D medical image comprises:
extracting a 2D cross-section image at a point on a centerline of the tubular structure in the 3D medical image; and
generating a 2D unraveled cross-section image by unraveling a circular region of the 2D cross-section image having a predetermined radius about the point on the centerline of the tubular structure.

3. The method of claim 2 , wherein dividing the unraveled cross-section image into a plurality of image chunks comprises:
dividing the 2D unraveled cross-section image into a plurality of 2D image patches, each of which corresponds to an arc length of the tubular structure in the 2D cross-section image.

4. The method of claim 3 , wherein the trained infinitely recurrent neural network repeatedly inputs a sequential data stream, including the plurality of 2D image patches of the 2D unraveled cross-section image, for a plurality of iterations while preserving a memory state between iterations, and detects, for each 2D image patch of the 2D unraveled cross-section image input to the trained infinitely recurrent neural network in the sequential data stream, a boundary of the corresponding arc length of the tubular structure in the 2D cross-section image.

5. The method of claim 1 , wherein extracting, from a 3D medical image of a patient, an unraveled cross-section image corresponding to a portion of a tubular structure in the 3D medical image comprises:
extracting a 3D tubular region including a cross-section of the tubular structure over a certain length of a centerline of the tubular structure in the 3D medical image; and
generating a 3D unraveled cross-section volume by unraveling the 3D tubular region about the centerline of the tubular structure over the certain length.

6. The method of claim 5 , wherein dividing the unraveled cross-section image into a plurality of image chunks comprises:
dividing the 3D unraveled cross-section volume into a plurality of 3D sub volumes, each of which corresponds to an arc length of the tubular structure over the certain length of the centerline.

7. The method of claim 6 , wherein the trained infinitely recurrent neural network repeatedly inputs a sequential data stream, including the plurality of 3D sub volumes of the 3D unraveled cross-section volume, for a plurality of iterations while preserving a memory state between iterations, and detects, for each 3D sub volume of the 3D unraveled cross-section volume input to the trained infinitely recurrent neural network in the sequential data stream, a boundary of the corresponding arc length of the tubular structure over the certain length of the centerline.

8. The method of claim 1 , wherein detecting a boundary of the portion of the tubular structure based on the plurality of image chunks using a trained infinitely recurrent neural network comprises:
detecting an inner boundary and an outer boundary for the portion of the tubular structure based on the plurality of image chunks using the trained infinitely recurrent neural network.

9. The method of claim 8 , wherein the trained infinitely recurrent neural network detects, for each image chunk of the unraveled cross-section image input to the trained infinitely recurrent neural network in the sequential data stream, a surface between a corresponding section of the inner boundary of the portion of the tubular structure and a corresponding section of the outer boundary of the portion of the tubular structure.

10. The method of claim 1 , wherein the tubular structure is an airway.

11. The method of claim 1 , wherein the tubular structure is a vascular structure.

12. An apparatus for automated boundary delineation of a tubular structure in a 3D medical image of a patient, comprising:
means for extracting, from a 3D medical image of a patient, an unraveled cross-section image corresponding to a portion of a tubular structure in the 3D medical image;
means for dividing the unraveled cross-section image into a plurality of image chunks; and
means for detecting a boundary of the portion of the tubular structure based on the plurality of image chunks using a trained infinitely recurrent neural network, wherein the trained infinitely recurrent neural network repeatedly inputs a sequential data stream, including the plurality of image chunks of the unraveled cross-section image, for a plurality of iterations while preserving a memory state between iterations, and detects, for each image chunk of the unraveled cross-section image input to the trained infinitely recurrent neural network in the sequential data stream, a corresponding section of the boundary of the portion of the tubular structure.

13. The apparatus of claim 12 , wherein the means for extracting, from a 3D medical image of a patient, an unraveled cross-section image corresponding to a portion of a tubular structure in the 3D medical image comprises:
means for extracting a 2D cross-section image at a point on a centerline of the tubular structure in the 3D medical image; and
means for generating a 2D unraveled cross-section image by unraveling a circular region of the 2D cross-section image having a predetermined radius about the point on the centerline of the tubular structure.

14. The apparatus of claim 13 , wherein the means for dividing the unraveled cross-section image into a plurality of image chunks comprises:
means for dividing the 2D unraveled cross-section image into a plurality of 2D image patches, each of which corresponds to an arc length of the tubular structure in the 2D cross-section image.

15. The apparatus of claim 14 , wherein the trained infinitely recurrent neural network repeatedly inputs a sequential data stream, including the plurality of 2D image patches of the 2D unraveled cross-section image, for a plurality of iterations while preserving a memory state between iterations, and detects, for each 2D image patch of the 2D unraveled cross-section image input to the trained infinitely recurrent neural network in the sequential data stream, a boundary of the corresponding arc length of the tubular structure in the 2D cross-section image.

16. The apparatus of claim 12 , wherein the means for extracting, from a 3D medical image of a patient, an unraveled cross-section image corresponding to a portion of a tubular structure in the 3D medical image comprises:
means for extracting a 3D tubular region including a cross-section of the tubular structure over a certain len

Assignees

• Siemens Healthcare Gmbh

Inventors

• Kiraly Atilla Peter
• Novak Carol L
• Odry Benjamin L

Classifications

• G16H50/30

  for calculating health indices; for individual health risk assessment · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title