System and method for automated detection and segmentation of tumor boundaries within medical imaging data

• US 8,265,355 B2
• Filed: 11/18/2005
• Issued: 09/11/2012
• Est. Priority Date: 11/19/2004
• Status: Active Grant

• Current Assignee: Koninklijke Philips Electronics N.V. (Koninklijke Philips N.V.)
• Original Assignee: Koninklijke Philips Electronics N.V. (Koninklijke Philips N.V.)
• Inventors: Zhao, Luyin, Lee, Kwok Pun
• Primary Examiner(s): Akhavannik, Hadi
• Application Number:
  US11/719,214
• Publication Number:
  US 20090148007A1
• Time in Patent Office: 2,489 Days
• Field of Search: 382/128, 382/228
• US Class Current: 382/128
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06T 2207/10081 Computed x-ray tomography [CT]
• CPC Class Codes: G06T 2207/10132 Ultrasound image
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 2207/30068 Mammography; Breast
• CPC Class Codes: G06T 2207/30096 Tumor; Lesion
• CPC Class Codes: G06T 7/0012 Biomedical image inspection
• CPC Class Codes: G06T 7/12 Edge-based segmentation
• CPC Class Codes: G06V 10/772 Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries

Abstract

A method for segmenting regions within a medical image includes evaluating a set of candidate segmentations generated from an initial segmentation. Based on distance calculations for each candidate using derivative segmentations, the best candidate is recommended to clinician if it is better than the initial segmentation. This recommender realizes a most stable segmentation that will benefit follow-up computer aided diagnosis (i.e. classifying lesion to benign/malignant).

Claims

1. A method for classifying regions of interest within a medical image, comprising the steps of:

training a classifier on a set of medical image training data, which training data includes segmented regions where a clinical ground truth classifying the segmented regions is known;

acquiring non-training medical image data for investigation;

generating an initial segmentation for a region of interest of the medical image;

generating a plurality of candidate segmentations based on the initial segmentation;

comparing the initial and the plurality of candidate segmentations with each other;

selecting a best segmentation from a set of segmentations including the initial and the plurality of candidate segmentations based on the comparisons, where the selected best segmentation is used to train the classifier;

processing the segmented regions to extract a full feature set for each of the segmented regions; and

classifying the regions of interest using the full feature set;

wherein, the step of training includes using a recommender to realize a stable segmentation.

2. The method as set forth in claim 1, where the plurality of candidate segmentations are generated by generating perturbations in the parameters defining the initial segmentation, to realize a plurality of candidate segmentations, where one of the candidate segmentations is recommended if the determination is that a candidate segmentation is better than the initial segmentation and the other candidate segmentations, and where the initial segmentation is recommended if the determination is that the initial segmentation is better than the candidate segmentations.

3. The method set forth in claim 2, comprising:

generating a derivative segmentation for the initial and candidate segmentations.

7. The method set forth in claim 2, wherein generating perturbations in the parameters defining the initial segmentation includes at least one of increasing or decreasing a parameter by a different pre-determined amount for each of the perturbations.

8. The method set forth in claim 7, wherein the different predetermined amounts are different pre-determined percentages.

9. The method set forth in claim 7, wherein the parameters are values representing a diameter of tissue of interest.

10. The method set forth in claim 2, wherein a candidate segmentation causing less of a change in a shape of the initial segmentation is identified as better than a candidate segmentation causing more of a change in the shape of the initial segmentation.

11. The method set forth in claim 1, comprising:

determining distance values between the initial segmentation and each of the candidate segmentations;

sorting the distance values based on corresponding absolute values; and

recommending a segmentation with the lowest distance value.

12. The method set forth in claim 11, wherein the segmentation with the lowest distance value is provided to a clinician.

4. A system, comprising:

a segmenter configured to;

segment a region of interest within a medical image to produce an initial segmentation, and

generate a plurality of co-existing candidate segmentations for the region of interest by varying or perturbing boundaries of the initial segmentation,

wherein if it is determined that one of the plurality of co-existing segmentations is better suited for post segmentation processing than the initial segmentation, then the segmenter recommends changing the initial segmentation, and

wherein if it is determined that the initial segmentation is better suited for post segmentation processing than the plurality of co-existing candidate segmentations, then a recommendation to change the initial segmentation is not made.

5. The system of claim 4, wherein the segmenter perturbs each candidate segmentation variation to determine if it is a more desirable segmentation.

6. The system of claim 4, comprising:

a CAD system; and

a false positive reduction system, in communication with the CAD system and segmenter, which includes;

a feature extractor;

a genetic algorithm in communication with the feature extractor which generates an optimal feature subset extracted by, and provided from the feature extractor; and

a support vector machine (SVM) in communication with the feature extractor for classifying each segmented region in accord with the subset of features, with a minimum of false positives.

13. A method, comprising:

obtaining an initial segmentation;

generating a candidate segmentation set for segmentation of a region of interest in a medical image, where the candidate segmentation set comprises at least a first candidate segmentation and a second candidate segmentation; and

evaluating the initial segmentation and individual segmentations of the candidate segmentation set against one another in a non-iterative action to produce an evaluation result; and

making a determination of a better segmentation among the initial segmentation and individual candidate segmentations of the candidate segmentation set based, at least in part, on the evaluation result.

14. The method of claim 13, where if an individual candidate segmentation is determined as the better segmentation, then the individual candidate segmentation is recommended.

15. The method of claim 13, where the initial segmentation is initially recommended and where if the initial segmentation is determined as the better segmentation, then no subsequent recommendation is made.

16. The method of claim 13, further comprising:

training a classifier based, at least in part, on the determination.

17. The method of claim 16, further comprising:

using the classifier to segment a region of interest within a second medical image after training the classifier based, at least in part, on the determination.

18. The method of claim 13, further comprising:

recommending the better segmentation to a clinician when a candidate segmentation of the candidate segmentation set is the better segmentation, where the clinician can choose to replace the initial segmentation with the better segmentation or choose to keep the initial segmentation without replacing or changing the initial segmentation with the better segmentation.

19. The method of claim 13, further comprising:

determining distance values between the initial segmentation and each of the individual candidate segmentations; and

sorting the distance values based on corresponding absolute values, where the better segmentation is determined as a segmentation with the lowest distance value.

20. The method of claim 13, where individual segmentations of the candidate segmentation set are derived from the initial segmentation.