Methods and systems for 3D object detection using learning

1. A method of detecting tumors in 3D medical images comprising:

using a learning procedure for feature selection from a feature set based on an annotated medical image-volume database,

generating a set of selected features;

building a classifier using a classification scheme to distinguish between tumor location and a non-tumor location and using the set of selected features; and

applying the classifier at a candidate volume to determine whether the candidate volume contains a tumor.

2. The method of claim 1 , wherein the feature set comprises at least one of 3D features, 2D features, 1D features, or complex features.

3. The method of claim 1 , wherein at least one of the steps of using a learning procedure for feature selection or building a classifier is performed using a feature selection method.

4. The method of claim 3 , wherein the feature selection method is a boosting method.

5. The method of claim 4 , wherein the boosting method incorporates feature cost to simultaneously optimize detection performance and cost.

6. The method of claim 5 , wherein detection performance comprises classification accuracy, and wherein cost comprises at least one of feature acquisition cost or feature computational cost.

7. The method of claim 4 , wherein the boosting method comprises the steps of:

(a) inputting a set of training samples having a set of features, feature computational costs for the set of features, and a predetermined number L that denotes the maximal number of strong classifiers;

(b) generating a set of weak learners based on the set of features;

(c) initializing a set of final strong classifiers (FSCs) and a set of candidate strong classifiers (CSCs), the set of FSCs being initialized to null, the set of CSCs being initialized to contain the set of weak learners;

(d) applying a selection criterion, selecting the best CSCs from the set of CSCs, generating a next set of CSCs;

(e) updating the set of FSCs;

(f) if the magnitude of the set of FSCs is equal to L, performing an output module; and

(g) using a children function, generating at least one new CSC, wherein the at least one new CSC is added to the next set of CSCs.

8. The method of claim 7 , further comprising: (h) repeating steps (d)

through (g) if the magnitude of the set of FSCs is less than L.

9. The method of claim 7 , wherein the updating step comprises:

determining if a termination criterion is satisfied; and

if the termination criterion is not satisfied, for each CSC in the next set of CSCs, adding a CSC to the set of FSCs and removing the CSC from the next set of CSCs.

10. The method of claim 9 , wherein the termination criterion is based on at least one of a validation set or a negligible improvement on a class separation margin using a training set.

11. The method of claim 7 , wherein the children function is used to select at least one weak learner from the next set of CSCs.

12. The method of claim 7 , wherein the step of performing the output module comprises selecting the best strong classifier from the set of FSCs based a balancing method.

13. The method of claim 12 , wherein the balancing method comprises balancing classification accuracy and feature cost.

14. The method of claim 1 , wherein the candidate volume is obtained by scanning.

15. The method of claim 14 wherein scanning is performed at a plurality of locations, varying at least one of scale, rotation or aspect ratio.

16. A boosting method for training classifiers that distinguish tumor locations from non-tumor location in 3D medical images, wherein said method incorporates feature cost to simultaneously optimize detection performance and cost, said method comprising the steps of:

(a) inputting a set of training samples having a set of features, feature computational costs for the set of features, and a pre-determined number L that denotes the maximal number of strong classifiers;

(b) generating a set of weak learners based on the set of features;

(c) initializing a set of final strong classifiers (FSCs) and a set of candidate strong classifiers (CSCs), the set of FSCs being initialized to null, the set of CSCs being initialized to contain the set of weak learners;

(d) applying a selection criterion, selecting the best CSCs from the set of CSCs, generating a next set of CSCs;

(e) updating the set of FSCs;

(f) if the magnitude of the set of FSCs is equal to L, performing an output module; and

(g) using a children function, generating at least one new CSC, adding the at least one new CSC to the next set of CSCs.

17. The method of claim 16 , further comprising: (h) repeating steps (d)

through (g) if the magnitude of the set of FSCs is less than L.

18. The method of claim 16 , wherein updating the set of FSCs comprises:

determining if a termination criterion is satisfied; and

if the termination criterion is not satisfied, for each CSC in the next set of CSCs, adding a CSC to the set of FSCs and removing the CSC from the next set of CSCs.

19. The method of claim 18 , wherein the termination criterion is based on at least one of a validation set or a negligible improvement on a class separation margin using a training set.

20. The method of claim 16 , wherein the children function is used to select at least one weak learner from the next set of CSCs.

21. The method of claim 16 wherein the step of performing the output module comprises selecting the best strong classifier from the set of FSCs based a balancing method.

22. The method of claim 21 , wherein the balancing method comprises balancing the classification accuracy and feature cost.

23. A program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform the method steps for detecting tumors in 3D medical images, said method comprising the steps of:

using a learning module for feature selection from a feature set based on an annotated medical image-volume database,

generating a set of selected features;

building a classifier using a classification scheme to distinguish between tumor location and a non-tumor location and using the set of selected features; and

applying the classifier at a candidate volume to determine whether the candidate volume contains a tumor.

24. The computer readable program storage device of claim 23 , wherein at least one of using a learning module for feature selection or building a classifier includes performing a feature selection method.

25. The computer readable program storage device of claim 23 , wherein the feature selection method is a boosting method.

26. The computer readable program storage device of claim 25 , wherein the boosting method incorporates feature cost to simultaneously optimize detection performance and cost.

27. The computer readable program storage device of claim 26 , wherein detection performance comprises classification accuracy, and wherein cost comprises at least one of feature acquisition cost or feature computational cost.

28. The computer readable program storage device of claim 25 , wherein the boosting method comprises the steps of:

(a) inputting a set of training samples having a set of features, feature computational costs for the set of features, and a pre-determined number L that denotes the maximal number of strong classifiers;

(b) generating a set of weak learners based on the set of features;

(c) initializing a set of final strong classifiers (FSCs) and a set of candidate strong classifiers (CSCs), the set of FSCs being initialized to null, the set of CSCs being initialized to contain the set of weak learners;

(d) applying a selection criterion, selecting the best CSCs from the set of CSCs, generating a next set of CSCs;

(e) updating the set of FSCs;

(f) if the magnitude of the set of FSCs is equal to L, performing an output module; and

(g) using a children function, generating at least one new CSC, wherein the at least one new CSC is added to the next set of CSCs.

29. The computer readable program storage device of claim 28 , further comprising:

(h) repeating steps (d) through (g) if the magnitude of the set of FSCs is less than L.

30. The computer readable program storage device of claim 28 , wherein the updating step comprises:

determining if a termination criterion is satisfied; and

if the termination criterion is not satisfied, for each CSC in the next set of CSCs, adding a CSC to the set of FSCs and removing the CSC from the next set of CSCs.

31. The computer readable program storage device of claim 30 , wherein the termination criterion is based on at least one of a validation set or a negligible improvement on a class separation margin using a training set.

32. The computer readable program storage device of claim 28 , wherein the children function is used to select at least one weak learner from the next set of CSCs.

33. The computer readable program storage device of claim 28 , wherein the step of performing the output module comprises selecting the best strong classifier from the set of FSCs based a balancing method.

34. The computer readable program storage device of claim 33 , wherein the balancing method comprises balancing classification accuracy and feature cost.

35. A boosting method for training classifiers that distinguish tumor locations from non-tumor location in 3D medical images, wherein said method incorporates feature cost to simultaneously optimize detection performance and cost, said method comprising the steps of:

(a) inputting a set of training samples having a set of features and feature computational costs for the set of features;

(b) generating a set of weak learners based on the set of features; and

(c) building a boosting classifier based on a subset of the weak learners to minimize a cost function of the detection performance and the feature computational costs.

36. A program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform the method steps for training classifiers that distinguish tumor locations from non-tumor location in 3D medical images, wherein said method incorporates feature cost to simultaneously optimize detection performance and cost, said method comprising the steps of:

(a) inputting a set of training samples having a set of features, feature computational costs for the set of features, and a pre-determined number L that denotes the maximal number of strong classifiers;

(b) generating a set of weak learners based on the set of features;

(c) initializing a set of final strong classifiers (FSCs) and a set of candidate strong classifiers (CSCs), the set of FSCs being initialized to null, the set of CSCs being initialized to contain the set of weak learners;

(d) applying a selection criterion, selecting the best CSCs from the set of CSCs, generating a next set of CSCs;

(e) updating the set of FSCs;

(f) if the magnitude of the set of FSCs is equal to L, performing an output module; and

(g) using a children function, generating at least one new CSC, adding the at least one new CSC to the next set of CSCs.

37. The computer readable program storage device of claim 36 , the method further comprising: (h) repeating steps (d) through (g) if the magnitude of the set of FSCs is less than L.

38. The computer readable program storage device of claim 36 , wherein updating the set of FSCs comprises:

determining if a termination criterion is satisfied; and

if the termination criterion is not satisfied, for each CSC in the next set of CSCs, adding a CSC to the set of FSCs and removing the CSC from the next set of CSCs.

39. The computer readable program storage device of claim 38 , wherein the termination criterion is based on at least one of a validation set or a negligible improvement on a class separation margin using a training set.