Method and system for anatomical object detection using marginal space deep neural networks
A method and system for anatomical object detection using marginal space deep neural networks is disclosed. The pose parameter space for an anatomical object is divided into a series of marginal search spaces with increasing dimensionality. A respective deep neural network is trained for each of the marginal search spaces, resulting in a series of trained deep neural networks. Each of the trained deep neural networks can evaluate hypotheses in a current parameter space using discriminative classification or a regression function. An anatomical object is detected in a medical image by sequentially applying the series of trained deep neural networks to the medical image.
1. A method for anatomical object detection in a medical image of a patient, comprising:
receiving a medical image of a patient including a target anatomical object; and
detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces, wherein each respective trained deep neural network inputs image patches of the medical image, samples pixels or voxels in each input image patch in a sparse sampling pattern adaptively learned during training of the respective trained deep neural network, and operates directly on the pixels or voxels sampled from each input image patch to calculate one or more candidates for the target anatomical object in a respective one of the marginal parameter spaces and wherein, for each of the marginal parameter spaces, a respective cascade of trained shallow sparse neural networks filters image patches of the medical image to remove negative image patches prior to the remaining image patches being input to the respective trained deep neural network for that marginal parameter space.
2. The method of claim 1 , wherein the respective trained deep neural network for at least one of the marginal parameter spaces comprises a discriminative deep neural network that inputs hypothesis image patches in the respective marginal parameter space, and calculates, for each of the hypothesis image patch, a probability that the hypothesis image patch is an image patch of the target anatomical object in the respective marginal parameter space.
3. The method of claim 1 , wherein the respective trained deep neural network for at least one of the marginal parameter spaces comprises a deep neural network regressor that inputs hypothesis image patches in the respective marginal parameter space, and calculates, for each hypothesis image patch, a displacement vector that maps the hypothesis image patch to a predicted image patch of the target anatomical object in the respective marginal parameter space.
4. The method of claim 3 , wherein the deep neural network regressor further calculates, for each hypothesis image patch, a probability for the corresponding predicted image patch of the target anatomical object in the respective marginal parameter space.
5. The method of claim 3 , wherein the deep neural network regressor iteratively refines the predicted image patches of the target anatomical object in the respective marginal parameter space by inputting each predicted image patch back into the deep neural network regressor and calculating, for each predicted image patch, a displacement vector that maps the predicted image patch to a refined predicted image patch in the respective marginal parameter space.
6. The method of claim 1 , wherein the series of marginal parameter spaces comprises a position parameter space, a position-orientation parameter space, and a position-orientation-scale parameter space.
7. The method of claim 6 , wherein detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces comprises:
detecting position candidates in the medical image using a first trained deep neural network that inputs image patches centered at each of a plurality of voxels in the medical image;
generating position-orientation hypotheses based on the detected position candidates;
detecting position-orientation candidates in the medical image using a second trained deep neural network that inputs image patches corresponding to the position-orientation hypotheses;
generating position-orientation-scale hypotheses based on the detected position-orientation candidates; and
detecting the pose of the target anatomical object in the medical image using a third trained deep neural network that inputs image patches corresponding to the position-orientation-scale hypotheses.
8. The method of claim 7 , wherein generating position-orientation hypotheses based on the detected position candidates comprises:
rotating a respective image patch centered at each detected position candidate to a plurality of orientations sampled from a predetermined range of orientations for the target anatomical object.
9. The method of claim 7 , wherein generating position-orientation-scale hypotheses based on the detected position-orientation candidates comprises:
scaling a respective image patch corresponding to each detected position-orientation candidate to a plurality of scales sampled from a predetermined range of scales for the target anatomical object.
10. The method of claim 7 , wherein the second trained deep neural network is a trained deep neural network regressor, and generating position-orientation hypotheses based on the detected position candidates comprises:
generating a respective position-orientation hypothesis for each detected position candidate by using an image patch centered at each detected position candidate as the respective position-orientation hypothesis.
11. The method of claim 7 , wherein the second trained deep neural network is a trained deep neural network regressor, and generating position-orientation-scale hypotheses based on the detected position-orientation candidates comprises:
generating a respective position-orientation-scale hypothesis for each detected position-orientation candidate by using an image patch corresponding to each detected position-orientation candidate as the respective position-orientation-scale hypothesis.
12. The method of claim 1 , wherein the respective deep neural network for each of the marginal parameter spaces is a multi-layer deep neural network trained using a stacked denoising auto-encoder (DAE).
13. The method of claim 1 , wherein the respective deep neural network for each of the marginal parameter spaces is a multi-layer deep neural network trained using one of a convolutional neural network (CNN), a stacked restricted Boltzmann machine (RBM), or a stacked auto-encoder (AE).
14. The method of claim 1 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained by sequentially training each of a plurality of hidden layers to reconstruct respective inputs to each hidden layer in an unsupervised pre-training stage, and training the deep neural network to detect the target anatomical object in the respective marginal parameter space based on ground truth training image patches in the respective marginal parameter space in a supervised training stage.
15. The method of claim 14 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained in the supervised training stage by adding an output layer for a target output for the respective marginal parameter space and training the plurality of hidden layers and the output layer using back-propagation to minimize an error between detected image patches for the anatomical object and the ground truth training image patches in the respective marginal parameter space.
16. The method of claim 14 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained in the supervised training stage by treating an output of the plurality of hidden layers as high-level image features and training a discriminative classifier based on the high-level image features output by the plurality of hidden layers.
17. The method of claim 1 , wherein the trained deep neural network for a first one of the marginal parameter spaces is trained using training image patches sampled from a set of training images, and the trained deep neural network for each subsequent one of the marginal parameter spaces is trained using training image patches generated by augmenting training image patches detected by the trained deep neural network in a preceding marginal parameter space with additional parameters of a current marginal parameter space sampled from a range of the current marginal parameter space in the training images.
18. The method of claim 1 , wherein detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces comprises:
for at least one of the marginal parameter spaces:
calculating an approximation of the respective trained deep neural network for that marginal parameter space; and
detecting one or more candidates for the target anatomical object in that marginal parameter space using the approximation of the respective trained deep neural network for that marginal parameter space.
19. The method of claim 18 , wherein calculating an approximation of the respective trained deep neural network for that marginal parameter space comprises:
calculating a reduced rank approximation of trained weights of the respective trained deep neural network for that marginal parameter space.
20. The method of claim 19 , wherein calculating a reduced rank approximation of trained weights of the respective trained deep neural network for that marginal parameter space comprises:
calculating a low rank tensor decomposition to approximate the trained weights of the respective trained deep neural network for that marginal parameter space.
21. The method of claim 19 , wherein calculating a reduced rank approximation of trained weights of the respective trained deep neural network for that marginal parameter space comprises:
approximating the trained weights of the respective trained deep neural network for that marginal parameter space using a sparse approximation matrix.
22. An apparatus for anatomical object detection in a medical image of a patient, comprising:
a processor; and
a memory storing computer program instructions which when executed by the processor cause the processor to perform operations comprising:
receiving a medical image of a patient including a target anatomical object; and
detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces, wherein each respective trained deep neural network inputs image patches of the medical image, samples pixels or voxels in each input image patch in a sparse sampling pattern adaptively learned during training of the respective trained deep neural network, and operates directly on the pixels or voxels sampled from each input image patch to calculate one or more candidates for the target anatomical object in a respective one of the marginal parameter spaces, and wherein, for each of the marginal parameter spaces, a respective cascade of trained shallow sparse neural networks filters image patches of the medical image to remove negative image patches prior to the remaining image patches being input to the respective trained deep neural network for that marginal parameter space.
23. The apparatus of claim 22 , wherein the respective trained deep neural network for at least one of the marginal parameter spaces comprises a discriminative deep neural network that inputs hypothesis image patches in the respective marginal parameter space, and calculates, for each of the hypothesis image patch, a probability that the hypothesis image patch is an image patch of the target anatomical object in the respective marginal parameter space.
24. The apparatus of claim 22 , wherein the respective trained deep neural network for at least one of the marginal parameter spaces comprises a deep neural network regressor that inputs hypothesis image patches in the respective marginal parameter space, and calculates, for each hypothesis image patch, a displacement vector that maps the hypothesis image patch to a predicted image patch of the target anatomical object in the respective marginal parameter space.
25. The apparatus of claim 24 , wherein the deep neural network regressor further calculates, for each hypothesis image patch, a probability for the corresponding predicted image patch of the target anatomical object in the respective marginal parameter space.
26. The apparatus of claim 24 , wherein the deep neural network regressor iteratively refines the predicted image patches of the target anatomical object in the respective marginal parameter space by inputting each predicted image patch back into the deep neural network regressor and calculating, for each predicted image patch, a displacement vector that maps the predicted image patch to a refined predicted image patch in the respective marginal parameter space.
27. The apparatus of claim 22 , wherein the series of marginal parameter spaces comprises a position parameter space, a position-orientation parameter space, and a position-orientation-scale parameter space.
28. The apparatus of claim 27 , wherein detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces comprises:
detecting position candidates in the medical image using a first trained deep neural network that inputs image patches centered at each of a plurality of voxels in the medical image;
generating position-orientation hypotheses based on the detected position candidates;
detecting position-orientation candidates in the medical image using a second trained deep neural network that inputs image patches corresponding to the position-orientation hypotheses;
generating position-orientation-scale hypotheses based on the detected position-orientation candidates; and
detecting the pose of the target anatomical object in the medical image using a third trained deep neural network that inputs image patches corresponding to the position-orientation-scale hypotheses.
29. The apparatus of claim 22 , wherein the respective deep neural network for each of the marginal parameter spaces is a multi-layer deep neural network trained using a stacked denoising auto-encoder (DAE).
30. The apparatus of claim 22 , wherein the respective deep neural network for each of the marginal parameter spaces is a multi-layer deep neural network trained using one of a convolutional neural network (CNN), a stacked restricted Boltzmann machine (RBM), or a stacked auto-encoder (AE).
31. The apparatus of claim 22 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained by sequentially training each of a plurality of hidden layers to reconstruct respective inputs to each hidden layer in an unsupervised pre-training stage, and training the deep neural network to detect the target anatomical object in the respective marginal parameter space based on ground truth training image patches in the respective marginal parameter space in a supervised training stage.
32. The apparatus of claim 31 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained in the supervised training stage by adding an output layer for a target output for the respective marginal parameter space and training the plurality of hidden layers and the output layer using back-propagation to minimize an error between detected image patches for the anatomical object and the ground truth training image patches in the respective marginal parameter space.
33. The apparatus of claim 31 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained in the supervised training stage by treating an output of the plurality of hidden layers as high-level image features and training a discriminative classifier based on the high-level image features output by the plurality of hidden layers.
34. The apparatus of claim 22 , wherein the trained deep neural network for a first one of the marginal parameter spaces is trained using training image patches sampled from a set of training images, and the trained deep neural network for each subsequent one of the marginal parameter spaces is trained using training image patches generated by augmenting training image patches detected by the trained deep neural network in a preceding marginal parameter space with additional parameters of a current marginal parameter space sampled from a range of the current marginal parameter space in the training images.
35. The apparatus of claim 22 , wherein detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces comprises:
calculating an approximation of the respective trained deep neural network for at least one of the marginal parameter spaces; and
detecting one or more candidates for the target anatomical object in the at least one of the marginal parameter spaces using the approximation of the respective trained deep neural network for the at least one of the marginal parameter spaces.
36. The apparatus of claim 35 , wherein calculating an approximation of the respective trained deep neural network for at least one of the marginal parameter spaces comprises:
calculating a reduced rank approximation of trained weights of the respective trained deep neural network for the at least one of the marginal parameter spaces.
37. A non-transitory computer readable medium storing computer program instructions for anatomical object detection in a medical image of a patient, the computer program instructions when executed by a processor cause the processor to perform operations comprising:
receiving a medical image of a patient including a target anatomical object; and
detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces, wherein each respective trained deep neural network inputs image patches of the medical image, samples pixels or voxels in each input image patch in a sparse sampling pattern adaptively learned during training of the respective trained deep neural network, and operates directly on the pixels or voxels sampled from each input image patch to calculate one or more candidates for the target anatomical object in a respective one of the marginal parameter spaces, and wherein, for each of the marginal parameter spaces, a respective cascade of trained shallow sparse neural networks filters image patches of the medical image to remove negative image patches prior to the remaining image patches being input to the respective trained deep neural network for that marginal parameter space.
38. The non-transitory computer readable medium of claim 37 , wherein the respective trained deep neural network for at least one of the marginal parameter spaces comprises a discriminative deep neural network that inputs hypothesis image patches in the respective marginal parameter space, and calculates, for each of the hypothesis image patch, a probability that the hypothesis image patch is an image patch of the target anatomical object in the respective marginal parameter space.
39. The non-transitory computer readable medium of claim 37 , wherein the respective trained deep neural network for at least one of the marginal parameter spaces comprises a deep neural network regressor that inputs hypothesis image patches in the respective marginal parameter space, and calculates, for each hypothesis image patch, a displacement vector that maps the hypothesis image patch to a predicted image patch of the target anatomical object in the respective marginal parameter space.
40. The non-transitory computer readable medium of claim 39 , wherein the deep neural network regressor further calculates, for each hypothesis image patch, a probability for the corresponding predicted image patch of the target anatomical object in the respective marginal parameter space.
41. The non-transitory computer readable medium of claim 39 , wherein the deep neural network regressor iteratively refines the predicted image patches of the target anatomical object in the respective marginal parameter space by inputting each predicted image patch back into the deep neural network regressor and calculating, for each predicted image patch, a displacement vector that maps the predicted image patch to a refined predicted image patch in the respective marginal parameter space.
42. The non-transitory computer readable medium of claim 37 , wherein the series of marginal parameter spaces comprises a position parameter space, a position-orientation parameter space, and a position-orientation-scale parameter space.
43. The non-transitory computer readable medium of claim 42 , wherein detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces comprises:
detecting position candidates in the medical image using a first trained deep neural network that inputs image patches centered at each of a plurality of voxels in the medical image;
generating position-orientation hypotheses based on the detected position candidates;
detecting position-orientation candidates in the medical image using a second trained deep neural network that inputs image patches corresponding to the position-orientation hypotheses;
generating position-orientation-scale hypotheses based on the detected position-orientation candidates; and
detecting the pose of the target anatomical object in the medical image using a third trained deep neural network that inputs image patches corresponding to the position-orientation-scale hypotheses.
44. The non-transitory computer readable medium of claim 43 , wherein generating position-orientation hypotheses based on the detected position candidates comprises:
rotating a respective image patch centered at each detected position candidate to a plurality of orientations sampled from a predetermined range of orientations for the target anatomical object.
45. The non-transitory computer readable medium of claim 43 , wherein generating position-orientation-scale hypotheses based on the detected position-orientation candidates comprises:
scaling a respective image patch corresponding to each detected position-orientation candidate to a plurality of scales sampled from a predetermined range of scales for the target anatomical object.
46. The non-transitory computer readable medium of claim 43 , wherein the second trained deep neural network is a trained deep neural network regressor, and generating position-orientation hypotheses based on the detected position candidates comprises:
generating a respective position-orientation hypothesis for each detected position candidate by using an image patch centered at each detected position candidate as the respective position-orientation hypothesis.
47. The method non-transitory computer readable medium claim 43 , wherein the second trained deep neural network is a trained deep neural network regressor, and generating position-orientation-scale hypotheses based on the detected position-orientation candidates comprises:
generating a respective position-orientation-scale hypothesis for each detected position-orientation candidate by using an image patch corresponding to each detected position-orientation candidate as the respective position-orientation-scale hypothesis.
48. The non-transitory computer readable medium of claim 37 , wherein the respective deep neural network for each of the marginal parameter spaces is a multi-layer deep neural network trained using a stacked denoising auto-encoder (DAE).
49. The non-transitory computer readable medium of claim 37 , wherein the respective deep neural network for each of the marginal parameter spaces is a multi-layer deep neural network trained using one of a convolutional neural network (CNN), a stacked restricted Boltzmann machine (RBM), or a stacked auto-encoder (AE).
50. The non-transitory computer readable medium of claim 37 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained by sequentially training each of a plurality of hidden layers to reconstruct respective inputs to each hidden layer in an unsupervised pre-training stage, and training the deep neural network to detect the target anatomical object in the respective marginal parameter space based on ground truth training image patches in the respective marginal parameter space in a supervised training stage.
51. The non-transitory computer readable medium of claim 50 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained in the supervised training stage by adding an output layer for a target output for the respective marginal parameter space and training the plurality of hidden layers and the output layer using back-propagation to minimize an error between detected image patches for the anatomical object and the ground truth training image patches in the respective marginal parameter space.
52. The non-transitory computer readable medium of claim 50 , wherein the respective trained deep neural network for each of the marginal parameter spaces is trained in the supervised training stage by treating an output of the plurality of hidden layers as high-level image features and training a discriminative classifier based on the high-level image features output by the plurality of hidden layers.
53. The non-transitory computer readable medium of claim 37 , wherein the trained deep neural network for a first one of the marginal parameter spaces is trained using training image patches sampled from a set of training images, and the trained deep neural network for each subsequent one of the marginal parameter spaces is trained using training image patches generated by augmenting training image patches detected by the trained deep neural network in a preceding marginal parameter space with additional parameters of a current marginal parameter space sampled from a range of the current marginal parameter space in the training images.
54. The non-transitory computer readable medium of claim 37 , wherein detecting a pose of the target anatomical object in the medical image in a series of marginal parameter spaces of increasing dimensionality using a respective trained deep neural network for each of the marginal parameter spaces comprises:
for at least one of the marginal parameter spaces:
calculating an approximation of the respective trained deep neural network for that marginal parameter space; and
detecting one or more candidates for the target anatomical object in that marginal parameter space using the approximation of the respective trained deep neural network for that marginal parameter space.
55. The non-transitory computer readable medium of claim 54 , wherein calculating an approximation of the respective trained deep neural network for that marginal parameter space comprises:
calculating a reduced rank approximation of trained weights of the respective trained deep neural network for that marginal parameter space.
56. The non-transitory computer readable medium of claim 55 , wherein calculating a reduced rank approximation of trained weights of the respective trained deep neural network for that marginal parameter space comprises:
calculating a low rank tensor decomposition to approximate the trained weights of the respective trained deep neural network for that marginal parameter space.
57. The non-transitory computer readable medium of claim 55 , wherein calculating a reduced rank approximation of trained weights of the respective trained deep neural network for that marginal parameter space comprises:
approximating the trained weights of the respective trained deep neural network for that marginal parameter space using a sparse approximation matrix.