Method and system for automatic landmark detection using discriminative joint context
View Patent ↗A method and system for detecting anatomic landmarks in medical images is disclosed. In order to detect multiple related anatomic landmarks, a plurality of landmark candidates are first detected individually using trained landmark detectors. A joint context is then generated for each combination of the landmark candidates. The best combination of landmarks in then determined based on the joint context using a trained joint context detector.
1. A method for detecting anatomic landmarks of a left ventricle (LV) in a magnetic resonance (MR) long axis image slice, comprising:
detecting a plurality of apex candidates in the MR long axis image slice using a trained apex detector;
detecting a plurality of base plane candidates in the MR long axis image slice using a trained base plane detector;
generating a joint context for each apex-base plane candidate pair; and
determining a best apex-base plane candidate pair based on the generated joint context using a trained joint context detector.
2. The method of claim 1 , wherein said step of detecting a plurality of base plane candidates in the MR long axis image slice using a trained base plane detector comprises:
detecting a plurality of basal annulus point candidates in the MR long axis image slice using a trained annulus point detector;
generating base plane hypotheses by generating a joint context for combinations of detected basal annulus point candidates; and
detecting the plurality of base plane candidates from the base plane hypotheses using the trained base plane detector.
3. The method of claim 1 , wherein the apex candidates and the base plane candidates are detected using marginal space learning (MSL).
4. The method of claim 1 , wherein the joint context of the apex-base plane candidate pairs is detected using probabilistic boosting tree (PBT).
5. The method of claim 1 , wherein said step of generating a joint context for each apex-base plane candidate pair comprises, for each apex-base plane candidate pair:
mapping parameters of the apex candidate and parameters of the base plane candidate using a mapping function to determine parameters of the joint context.
6. The method of claim 5 , wherein said mapping function defines a geometric relationship between the parameters of the apex candidate and the parameters of the base plane candidate.
7. The method of claim 1 , wherein said step of determining a best apex-base plane candidate pair based on the generated joint context using a trained joint context detector comprises:
determining the best apex-base plane candidate based on fusion of a probability determined by the joint context detector, a probability determined by the apex detector, and a probability determined by the base plane detector.
8. The method of claim 7 , wherein said step of determining the best apex-base plane candidate based on fusion of a probability determined by the joint context detector, a probability determined by the apex detector, and a probability determined by the base plane detector comprises:
selecting an apex-base plane candidate with a best probability score:
p=p j *( p a +p b )/2,
where p j denotes the probability determined by the joint context detector, p a denotes the probability determined by the apex detector, and p b denotes the probability determined by the base plane detector.
9. The method of claim 1 , wherein detected positions of the apex and basal annulus points of the LV are defined by the best apex-base plane pair, further comprising:
refining the detected positions of the apex and basal annulus points using local searches constrained based on the detected positions of the apex and basal annulus points.
10. An apparatus for detecting anatomic landmarks of a left ventricle (LV) in a magnetic resonance (MR) long axis image slice, comprising:
means for detecting a plurality of apex candidates in the MR long axis image slice using a trained apex detector;
means for detecting a plurality of base plane candidates in the MR long axis image slice using a trained base plane detector;
means for generating a joint context for each apex-base plane candidate pair; and
means for determining a best apex-base plane candidate pair based on the generated joint context using a trained joint context detector.
11. The apparatus of claim 10 , wherein said means for detecting a plurality of base plane candidates in the MR long axis image slice using a trained base plane detector comprises:
means for detecting a plurality of basal annulus point candidates in the MR long axis image slice using a trained annulus point detector;
means for generating base plane hypotheses by generating a joint context for combinations of detected basal annulus point candidates; and
means for detecting the plurality of base plane candidates from the base plane hypotheses using the trained base plane detector.
12. The apparatus of claim 10 , wherein the apex candidates and the base plane candidates are detected using marginal space learning (MSL).
13. The apparatus of claim 10 , wherein the joint context of the apex-base plane candidate pairs is detected using probabilistic boosting tree (PBT).
14. The apparatus of claim 10 , wherein said means for generating a joint context for each apex-base plane candidate pair comprises:
means for mapping parameters of an apex candidate and parameters of a base plane candidate using a mapping function to determine parameters of the joint context for the apex candidate and the base plane candidate.
15. The apparatus of claim 14 , wherein said mapping function defines a geometric relationship between the parameters of the apex candidate and the parameters of the base plane candidate.
16. The apparatus of claim 10 , wherein said means for determining a best apex-base plane candidate pair based on the generated joint context using a trained joint context detector comprises:
means for determining the best apex-base plane candidate based on fusion of a probability determined by the joint context detector, a probability determined by the apex detector, and a probability determined by the base plane detector.
17. The apparatus of claim 16 , wherein said means for determining the best apex-base plane candidate based on fusion of a probability determined by the joint context detector, a probability determined by the apex detector, and a probability determined by the base plane detector comprises:
means for selecting an apex-base plane candidate with a best probability score:
p=p j *( p a +p b )/2,
where p j denotes the probability determined by the joint context detector, p a denotes the probability determined by the apex detector, and p b denotes the probability determined by the base plane detector.
18. The apparatus of claim 10 , wherein detected positions of the apex and basal annulus points of the LV are defined by the best apex-base plane pair, further comprising:
means for refining the detected positions of the apex and basal annulus points using local searches constrained based on the detected positions of the apex and basal annulus points.
19. A non-transitory computer readable medium encoded with computer executable instructions for detecting anatomic landmarks of a left ventricle (LV) in a magnetic resonance (MR) long axis image slice, the computer executable instructions when executed on a processor configures the processor to implement steps comprising:
detecting a plurality of apex candidates in the MR long axis image slice using a trained apex detector;
detecting a plurality of base plane candidates in the MR long axis image slice using a trained base plane detector;
generating a joint context for each apex-base plane candidate pair; and
determining a best apex-base plane candidate pair based on the generated joint context using a trained joint context detector.
20. The computer readable medium of claim 19 , wherein the computer executable instructions defining the step of detecting a plurality of base plane candidates in the MR long axis image slice using a trained base plane detector comprise computer executable instructions defining the steps of:
detecting a plurality of basal annulus point candidates in the MR long axis image slice using a trained annulus point detector;
generating base plane hypotheses by generating a joint context for combinations of detected basal annulus point candidates; and
detecting the plurality of base plane candidates from the base plane hypotheses using the trained base plane detector.
21. The computer readable medium of claim 19 , wherein the apex candidates and the base plane candidates are detected using marginal space learning (MSL).
22. The computer readable medium of claim 19 , wherein the joint context of the apex-base plane candidate pairs is detected using probabilistic boosting tree (PBT).
23. The computer readable medium of claim 19 , wherein the computer executable instructions defining the step of generating a joint context for each apex-base plane candidate pair comprise computer executable instructions defining the step of, for each apex-base plane candidate pair:
mapping parameters of the apex candidate and parameters of the base plane candidate using a mapping function to determine parameters of the joint context.
24. The computer readable medium of claim 23 , wherein said mapping function defines a geometric relationship between the parameters of the apex candidate and the parameters of the base plane candidate.
25. The computer readable medium of claim 19 , wherein the computer executable instructions defining the step of determining a best apex-base plane candidate pair based on the generated joint context using a trained joint context detector comprise computer executable instructions defining the step of:
determining the best apex-base plane candidate based on fusion of a probability determined by the joint context detector, a probability determined by the apex detector, and a probability determined by the base plane detector.
26. The computer readable medium of claim 25 , wherein the computer executable instructions defining the step of determining the best apex-base plane candidate based on fusion of a probability determined by the joint context detector, a probability determined by the apex detector, and a probability determined by the base plane detector comprise computer executable instructions defining the step of:
selecting an apex-base plane candidate with a best probability score:
p=p j *( p a +p b )/2,
where p j denotes the probability determined by the joint context detector, p a denotes the probability determined by the apex detector, and p b denotes the probability determined by the base plane detector.
27. The computer readable medium of claim 19 , wherein detected positions of the apex and basal annulus points of the LV are defined by the best apex-base plane pair, further comprising computer executable instructions defining the step of:
refining the detected positions of the apex and basal annulus points using local searches constrained based on the detected positions of the apex and basal annulus points.
28. A method of detecting a plurality of anatomic landmarks in a medical image comprising:
individually detecting candidates for each of the plurality of landmarks in the medical image using trained detectors corresponding to each of the plurality of landmarks;
generating a joint context for each combination of the candidates detected for the plurality of landmarks; and
detecting positions of the plurality of landmarks in the medical image by determining a best combination of landmarks based on the generated joint context using a trained joint context detector.
29. The method of claim 28 , wherein the plurality of anatomic landmarks are semantically related to each other.
30. The method of claim 29 , wherein the plurality of anatomic landmarks are parts of the same anatomical structure.
31. The method of claim 29 , wherein said step of generating a joint context for each combination of the candidates detected for the plurality of landmarks comprises:
mapping parameters of the plurality of landmarks using a mapping function that defines the semantic relationship between the landmarks to generate parameters of the joint context.