Voxel-level machine learning with or without cloud-based support in medical imaging

Patent metadata

Field	Value
Publication number	US-2016110632-A1
Application number	US-201414518138-A
Country	US
Kind code	A1
Filing date	Oct 20, 2014
Priority date	Oct 20, 2014
Publication date	Apr 21, 2016
Grant date	—

Abstract

A single level machine-learnt classifier is used in medical imaging. A gross or large structure is located using any approach, including non-ML approaches such as region growing or level-sets. Smaller portions of the structure are located using ML applied to relatively small patches (small relative to the organ or overall structure of interest). The classification of small patches allows for a simple ML approach specific to a single scale or at a voxel/pixel level. The use of small patches may allow for providing classification as a service (e.g., cloud-based classification) since partial image data is to be transmitted. The use of small patches may allow for feedback on classification and updates to the ML. The use of small patches may allow for the creation of a labeled library of classification partially based on ML. Given a near complete labeled library, a simple matching of patches or a lookup can replace ML classification for faster throughput.

First claim

I (We) claim:

1 . A method for use of machine-learnt classifier in medical imaging, the method comprising:
segmenting, by a processor, an anatomic structure of a patient represented in medical imaging data; locating, by the processor, a region containing the anatomic structure; dividing, by the processor, the region represented in the medical imaging data into a plurality of patches; classifying, with a machine-learnt classifier, each of the patches as including the anatomical structure or not including the anatomical structure, the classifying of each of the patches being independent of classifying the other patches; merging, by the processor, locations for the patches classified as including the anatomical structure to the anatomical structure; and outputting, on a display, a segmented image of the anatomical structure including locations from the locating and the merged locations from the patches.

2 . The method of claim 1 wherein locating comprises region growing.

3 . The method of claim 1 wherein segmenting the anatomical structure comprises region growing and skeletonization.

4 . The method of claim 1 wherein locating comprises segmenting a lung, wherein segmenting comprises locating relatively larger airways in the lung, wherein classifying comprising classifying as including relatively smaller airways or not, wherein merging comprises merging the relatively smaller airways with the relatively larger airways, and wherein outputting comprises outputting an image of the relatively larger and smaller airways.

5 . The method of claim 1 wherein dividing comprises dividing into overlapping patches.

6 . The method of claim 1 wherein classifying comprises classifying with the machine-learnt classifier comprising a neural network.

7 . The method of claim 6 wherein classifying comprises classifying with the machine-learnt classifier comprising a deep learnt, sparse auto-encoder classifier.

8 . The method of claim 1 wherein classifying comprises classifying into as including the anatomical structure or not and at least one other anatomical structure or not.

9 . The method of claim 1 wherein the acts are performed with the machine-learnt classifier for the patches as the only machine-learnt classifier with the patches comprising no more than 50 pixels or voxels along a longest dimension.

10 . The method of claim 1 wherein merging comprises removing the locations for patches classified as including the anatomical structure but not connected to other of the locations directly or through a line or curve fit.

11 . The method of claim 1 wherein segmenting, locating, and dividing are performed by the processor at a local system and wherein classifying is performed with the machine-learnt classifier by a server remote from the local system.

12 . The method of claim 1 further comprising:
receiving from a user input an indication of error in the classifying for at least one of the patches;
transmitting the at least one of the patches and a corrected classification to a remote server; and
updating the machine-learnt classifier using the at least one of the patches and the corrected classification.

13 . The method of claim 12 further comprising:
collecting the patches and additional patches with verified classifications; and
replacing the classifying with the machine-learnt classifier with matching from the collection of the patches and the additional patches.

14 . A method for use of voxel-level machine-learnt classifier in medical imaging, the method comprising:
locating, without using a machine trained operator, a structure from data representing a patient; classifying, by a processor using a machine trained classifier, sub-sets of the data representing the patient within a region around the structure; and expanding the structure with locations corresponding to the sub-sets classified as belonging to the structure.

15 . The method of claim 14 wherein locating comprises locating with region growing and skeletonization, and wherein classifying comprises classifying with a neural network.

16 . The method of claim 14 wherein locating comprises locating a relatively larger part of an airway tree of a lung of the patient, wherein classifying comprises classifying the sub-set as representing a relatively smaller part of the airway tree, and wherein expanding comprises adding the relatively smaller part of the airway tree to the relatively larger part of the airway tree.

17 . The method of claim 14 wherein locating and expanding are performed by another processor and wherein the processor for classifying is remote to the other processor and acts as a server of the other processor.

18 . The method of claim 16 wherein the processor is configured to update the machine trained classifier in response to feedback about the classifying from the other processor.

19 . A method for use of voxel-level machine-learnt classifier in medical imaging, the method comprising:
receiving first labeled patches of scan data of different patients from different computers over time; training a machine-learning classifier with the first labeled patches of the scan data of the patients; serving classifications for other patches with the machine-learnt classifier to the computers; receiving second labeled patches and reclassified other patches; storing the first labeled, second labeled, and reclassified other patches and classifications in a database; once a number of the first, second, and reclassified other patches with a statistically significant variation are stored in the database, ceasing the classifying with the machine-learnt classifier and instead classifying with a match of new patches with the first, second, and reclassified other patches stored in the database; and serving the classifications for the new patches.

20 . The method of claim 19 further comprising:
receiving a correction of one of the classifications of the other patches;
updating the machine-learnt classifier based on the correction.

21 . The method of claim 19 wherein the first patches and the new patches are less than 50 pixels or voxels along a longest dimension.

Assignees

• Siemens Ag

Inventors

• Kiraly Atilla Peter
• Odry Benjamin L

Classifications

• G06V30/194

  References adjustable by an adaptive method, e.g. learning · CPC title

• G06T7/11 Primary

  Region-based segmentation · CPC title

• G06V10/255

  Detecting or recognising potential candidate objects based on visual cues, e.g. shapes · CPC title

• G06V10/267

  by performing operations on regions, e.g. growing, shrinking or watersheds · CPC title

• G06T7/0081

  Physics · mapped topic