Method and system for anatomic landmark detection using constrained marginal space learning and geometric inference
View Patent ↗A method and apparatus for detecting multiple anatomical landmarks in a 3D volume. A first anatomical landmark is detected in a 3D volume using marginal space learning (MSL). Locations of remaining anatomical landmarks are estimated in the 3D volume based on the detected first anatomical landmark using a learned geometric model relating the anatomical landmarks. Each of the remaining anatomical landmarks is then detected using MSL in a portion of the 3D volume constrained based on the estimated location of each remaining landmark. This method can be used to detect the anatomical landmarks of the crista galli (CG), tip of the occipital bone (OB), anterior of the corpus callosum (ACC), and posterior of the corpus callosum (PCC) in a brain magnetic resonance imaging (MRI) volume.
1. A method for detecting a plurality of anatomical landmarks in an image comprising:
detecting a first landmark of the plurality of anatomical landmarks in the image using marginal space learning (MSL);
estimating locations of remaining landmarks of the plurality of anatomical landmarks in the image based on the detected first landmark using a learned geometric model relating the plurality of anatomical landmarks; and
detecting each of said remaining landmarks using MSL in a portion of the image constrained based on the estimated location of each of said remaining landmarks.
2. The method of claim 1 , wherein said step of detecting a first landmark of the plurality of anatomical landmarks in the image using MSL comprises:
detecting position candidates in the image for the first landmark using a trained position classifier;
generating position-orientation hypotheses from said position candidates;
detecting position-orientation candidates from said position-orientation hypotheses using a trained position-orientation classifier;
generating similarity transformation hypotheses from said position-orientation candidates; and
detecting at least one similarity transformation candidate from said similarity transformation hypotheses using a trained similarity transformation classifier.
3. The method of claim 2 , wherein the position classifier, the position-orientation classifier, and the similarity transformation classifier are each trained based on training data using a probabilistic boosting tree (PBT).
4. The method of claim 1 , wherein the learned geometric model is a mean shape relating the plurality of anatomical landmarks trained based on training data.
5. The method of claim 1 , wherein said step of estimating locations of remaining landmarks of the plurality of anatomical landmarks in the image on the detected first landmark using a learned geometric model relating the plurality of anatomical landmarks comprises:
transforming a mean shape of the learned geometric model based on a similarity transformation of the detected first landmark.
6. The method of claim 1 , wherein said step of detecting each of said remaining landmarks using MSL in a portion of the image constrained based on the estimated location of each of said remaining landmarks comprises, for each of said remaining landmarks:
detecting position candidates in a portion of the image within a certain distance of the estimated location of the landmark using a trained position classifier;
generating position-orientation hypotheses from said position candidates;
detecting position-orientation candidates from said position-orientation hypotheses using a trained position-orientation classifier;
generating similarity transformation hypotheses from said position-orientation candidates; and
detecting at least one similarity transformation candidate from said similarity transformation hypotheses using a trained similarity transformation classifier.
7. The method of claim 1 , wherein the image is a 3D volume.
8. The method of claim 1 , wherein the image comprises a brain MRI volume and the plurality of anatomical landmarks comprises a crista galli (CG), a tip of the occipital bone (OB), an anterior of the corpus callosum (ACC), and a posterior of the corpus callosum (PCC).
9. The method of claim 8 , further comprising:
constraining detected locations for each of CG, OB, ACC, and PCC to be on a plane.
10. An apparatus for detecting a plurality of anatomical landmarks in an image comprising:
means for detecting a first landmark of the plurality of anatomical landmarks in the image using marginal space learning (MSL);
means for estimating locations of remaining landmarks of the plurality of anatomical landmarks in the image based on the detected first landmark using a learned geometric model relating the plurality of anatomical landmarks; and
means for detecting each of said remaining landmarks using MSL in a portion of the image constrained based on the estimated location of each of said remaining landmarks.
11. The apparatus of claim 10 , wherein said means for detecting a first landmark of the plurality of anatomical landmarks in the image using MSL comprises:
means for detecting position candidates in the image for the first landmark using a trained position classifier;
means for generating position-orientation hypotheses from said position candidates;
means for detecting position-orientation candidates from said position-orientation hypotheses using a trained position-orientation classifier;
means for generating similarity transformation hypotheses from said position-orientation candidates; and
means for detecting at least one similarity transformation candidate from said similarity transformation hypotheses using a trained similarity transformation classifier.
12. The apparatus of claim 10 , wherein the learned geometric model is a mean shape relating the plurality of anatomical landmarks trained based on training data.
13. The apparatus of claim 10 , wherein said means for estimating locations of remaining landmarks of the plurality of anatomical landmarks in the image based on the detected first landmark using a learned geometric model relating the plurality of anatomical landmarks comprises:
means for transforming a mean shape of the learned geometric model based on a similarity transformation of the detected first landmark.
14. The apparatus of claim 10 , wherein said means for detecting each of said remaining landmarks using MSL in a portion of the image constrained based on the estimated location of each of said remaining landmarks comprises:
means for detecting position candidates in a portion of the image within a certain distance of an estimated location of a landmark using a trained position classifier;
means for generating position-orientation hypotheses from said position candidates;
means for detecting position-orientation candidates from said position-orientation hypotheses using a trained position-orientation classifier;
means for generating similarity transformation hypotheses from said position-orientation candidates; and
means for detecting at least one similarity transformation candidate from said similarity transformation hypotheses using a trained similarity transformation classifier.
15. The apparatus of claim 10 , wherein the image is a 3D volume.
16. The apparatus of claim 10 , wherein the image comprises a brain MRI volume and the plurality of anatomical landmarks comprises a crista galli (CG), a tip of the occipital bone (OB), an anterior of the corpus callosum (ACC), and a posterior of the corpus callosum (PCC).
17. A computer readable medium encoded with computer executable instructions for detecting a plurality of anatomical landmarks in an image, the computer executable instructions defining steps comprising:
detecting a first landmark of the plurality of anatomical landmarks in the image using marginal space learning (MSL);
estimating locations of remaining landmarks of the plurality of anatomical landmarks in the image based on the detected first landmark using a learned geometric model relating the plurality of anatomical landmarks; and
detecting each of said remaining landmarks using MSL in a portion of the image constrained based on the estimated location of each of said remaining landmarks.
18. The computer readable medium of claim 17 , wherein the computer executable instructions defining the step of detecting a first landmark of the plurality of anatomical landmarks in the image using MSL comprise computer executable instructions defining the steps of:
detecting position candidates in the image for the first landmark using a trained position classifier;
generating position-orientation hypotheses from said position candidates;
detecting position-orientation candidates from said position-orientation hypotheses using a trained position-orientation classifier;
generating similarity transformation hypotheses from said position-orientation candidates; and
detecting at least one similarity transformation candidate from said similarity transformation hypotheses using a trained similarity transformation classifier.
19. The computer readable medium of claim 17 , wherein the learned geometric model is a mean shape relating the plurality of anatomical landmarks trained based on training data.
20. The computer readable medium of claim 17 , wherein the computer executable instructions defining the step of estimating locations of remaining landmarks of the plurality of anatomical landmarks in the image based on the detected first landmark using a learned geometric model relating the plurality of anatomical landmarks comprise computer executable instructions defining the step of:
transforming a mean shape of the learned geometric model based on a similarity transformation of the detected first landmark.
21. The computer readable medium of claim 17 , wherein the computer executable instructions defining the step of detecting each of said remaining landmarks using MSL in a portion of the image constrained based on the estimated location of each of said remaining landmarks comprise computer executable instructions defining the steps of, for each of said remaining landmarks:
detecting position candidates in a portion of the image within a certain distance of the estimated location of the landmark using a trained position classifier;
generating position-orientation hypotheses from said position candidates;
detecting position-orientation candidates from said position-orientation hypotheses using a trained position-orientation classifier;
generating similarity transformation hypotheses from said position-orientation candidates; and
detecting at least one similarity transformation candidate from said similarity transformation hypotheses using a trained similarity transformation classifier.
22. The computer readable medium of claim 17 , wherein the image is a 3D volume.
23. The computer readable medium of claim 17 , wherein the image comprises a brain MRI volume and the plurality of anatomical landmarks comprises a crista galli (CG), a tip of the occipital bone (OB), an anterior of the corpus callosum (ACC), and a posterior of the corpus callosum (PCC).