Automated lesion segmentation from MRI images

Patent metadata

Field	Value
Publication number	US-10049451-B2
Application number	US-201615367349-A
Country	US
Kind code	B2
Filing date	Dec 2, 2016
Priority date	Dec 2, 2015
Publication date	Aug 14, 2018
Grant date	Aug 14, 2018

Abstract

Systems and methods are provided for automated segmentation of lesions within a region of interest of a patient. At least one magnetic resonance imaging (MRI) image of the region of interest is produced. At least one probability map is generated from the at least one MRI image. A given probability map represents, for each of a plurality of pixels, a likelihood that a lesion is present at the location represented by the pixel given the at least one MRI image of the region of interest. The at least one probability map is combined with a plurality of additional probability maps to provide a composite probability map. Lesions are identified from the composite probability map.

First claim

What is claimed is:

1. A method for automated segmentation of lesions within a region of interest of a patient, comprising:
producing at least one magnetic resonance imaging (MRI) image of the region of interest;
generating at least one probability map from the at least one MRI image, a given probability map representing, for each pixel, a likelihood that a lesion is present at the location represented by the pixel given the at least one MRI image of the region of interest, the at least one probability map comprising a first probability map in which the likelihood for each pixel is determined via classification of intensity values of the at least one MRI image of the region of interest at the pixel;
combining the at least one probability map with a plurality of additional probability maps to provide a composite probability map, such that each pixel of the composite probability map is a weighted linear combination of the likelihood values of the corresponding pixels in the at least one probability map and the at least one additional probability map and a set of weights for the weighted linear combination is determined using an expert system trained on prior datasets; and
identifying lesions from the composite probability map.

2. The method of claim 1 , the plurality of additional probability maps comprising a second probability map representing, for each pixel, a likelihood that a lesion is present at the location represented by the pixel given previous data sets from a population of patients, the population of patients including patients other than the patient.

3. The method of claim 2 , wherein the second probability map is a generic MS (Multiple Sclerosis) lesion probability map combining of lesion maps across the population of patients to represent the regions in the brain for which it is most likely that lesions will form.

4. The method of claim 2 , wherein the second probability map is a false positive probability map combining maps of false positive results across the population of patients representing regions having a high likelihood of being false positives across patients.

5. The method of claim 2 , wherein the second probability map is a generic anatomic probability map of gray matter.

6. The method of claim 1 , wherein the at least one probability map includes a cumulative probability map representing longitudinally acquired MRI image data for the patient, such that each of a plurality of pixels comprising the cumulative probability map is determined from a set of probability values at that location across a plurality of probability maps from previous MRI images.

7. The method of claim 1 , wherein the set of weights for the weighted linear combination are different across each of a plurality of regions comprising the image, and the weights for each region are determined via a regression analysis on prior datasets.

8. A system for automated segmentation of lesions within a region of interest of a patient comprising:
a magnetic resonance imaging (MRI) interface that receives at least one MRI image of the region of interest;
a probability map generator that generates at least one probability map, the probability map generator comprising a pattern recognition classifier that determines, for each of a plurality of pixels, a likelihood that a lesion is present at the location represented by the pixel given intensity values of the at least one MRI image of the region of interest to provide a first probability map of the at least one probability map;
a probability reconciliation component that combines the at least one probability map with an additional probability map to provide a composite probability map, the additional probability map representing, for each of a plurality of pixels, a likelihood that a lesion is present at the location represented by the pixel given previous data sets from a population of patients, the population of patients including patients other than the patient; and
a map analysis component that reviews the composite probability map to indicate the location of lesions, the map analysis component selecting connected groups of high-probability pixels as lesions, and excluding groups of high-probability pixels smaller than predetermined number of pixels.

9. The system of claim 8 , wherein the additional probability map is a generic MS (Multiple Sclerosis) lesion probability map combining lesion maps across a large population of patients to represent the regions in the brain for which it is most likely that lesions will form.

10. The system of claim 8 , wherein the additional probability map is a false positive probability map combining maps of false positive results across the population of patients representing regions having a high likelihood of being false positives across patients.

11. The system of claim 8 , wherein the additional probability map is a generic anatomic probability map of gray matter.

12. The system of claim 8 , wherein the additional probability map is a first probability map of a plurality of probability maps.

13. The system of claim 8 , wherein the at least one probability map includes a cumulative probability map representing longitudinally acquired MRI image data for the patient, such that each of a plurality of pixels comprising the cumulative probability map is determined from a set of probability values at that location across a plurality of probability maps from previous MRI images.

14. The system of claim 8 , wherein the at least one MRI image includes at least one of a T2-weighted FLAIR image and a PD-/T2-weighted dual echo image, and the pattern recognition classifier produces a binary map representing a classification of each pixel in the region of interest from intensity values of the at least one of the T2-weighted FLAIR image and the PD-/T2-weighted dual echo image.

15. A system for automated segmentation of lesions within a region of interest of a patient comprising:
a magnetic resonance imaging (MRI) interface that receives at least one MRI image of the region of interest;
a probability map generator that generates at least one probability map, including a cumulative probability map representing longitudinally acquired MRI image data for the patient, such that each of a plurality of pixels comprising the cumulative probability map is determined from a set of probability values at that location across a plurality of probability maps from previous MRI images, and the probability map generator comprising a pattern recognition classifier that determines, for each pixel, a likelihood that a lesion is present at the location represented by the pixel given intensity values of the at least one MRI image of the region of interest to provide a first probability map of the at least one probability map; and
a probability reconciliation component that combines the at least one probability map, the cumulative probability map, and a generic MS (Multiple Sclerosis) lesion probability map combining lesion maps across a large population of patients to represent the regions in the brain for which it is most likely that lesions will form and a false positive probability map combining maps of false positive results across the population of patients representing regions having a high likelihood of being false positives across patients to provide a composite probability map.

16. The system of claim 15 , wherein the probability reconciliation component combines the binary map, the cumulative probability map, the generic MS lesion probability map, the false positive probability map, and a generic anatomic probability map of gray matter to provide the composite ima

Assignees

• Cleveland Clinic Found

Inventors

• Fisher Elizabeth

Classifications

• G06T7/11 Primary

  Region-based segmentation · CPC title

• G06T7/0012 Primary

  Biomedical image inspection · CPC title

• G06F18/256

  of results relating to different input data, e.g. multimodal recognition · CPC title

• G06T2207/20076

  Probabilistic image processing · CPC title

• G06T2207/30016

  Brain · CPC title