Method and system for approximating deep neural networks for anatomical object detection

Patent metadata

Field	Value
Publication number	US-9633306-B2
Application number	US-201514706108-A
Country	US
Kind code	B2
Filing date	May 7, 2015
Priority date	May 7, 2015
Publication date	Apr 25, 2017
Grant date	Apr 25, 2017

Abstract

A method and system for approximating a deep neural network for anatomical object detection is discloses. A deep neural network is trained to detect an anatomical object in medical images. An approximation of the trained deep neural network is calculated that reduces the computational complexity of the trained deep neural network. The anatomical object is detected in an input medical image of a patient using the approximation of the trained deep neural network.

First claim

The invention claimed is:

1. A method for anatomical object detection in a medical image comprising:
training a deep neural network to detect the anatomical object in medical images;
calculating an approximation of the trained deep neural network that reduces the computational complexity of the trained deep neural network; and
detecting the anatomical object in a received medical image of a patient using the approximation of the trained deep neural network, wherein calculating an approximation of the trained deep neural network that reduces the computational complexity of the trained deep neural network comprises:
for each of a plurality of nodes in each of a plurality of layers of the trained deep neural network, reconstructing a trained weight matrix for the node using 1-D Haar wavelet bases and wavelet coefficients.

2. The method of claim 1 , wherein training a deep neural network to detect the anatomical object in medical images comprises training a respective filter for each of the plurality of nodes in each of the plurality of layers of the deep neural network, wherein each respective filter is a weight matrix comprising a plurality of weights that weight node outputs of the nodes of a previous one of the plurality of layers, and calculating an approximation of the trained deep neural network that reduces the computational complexity of the trained deep neural network further comprises:
sparsifying the weights of the filters for each of the plurality of layers of the trained deep neural network.

3. The method of claim 2 , wherein sparsifying the weights of the filters for each of the plurality of layers of the trained deep neural network comprises:
reducing a number of non-zero weights in each filter for each of the plurality of layers in the trained deep neural network by setting a predetermined percentage of non-zero weights with lowest magnitudes in each filter equal to zero; and
refining the remaining non-zero weights in each filter for each of the plurality of layers to alleviate an effect of reducing the number of non-zero weights in each filter.

4. The method of claim 3 , wherein refining the remaining non-zero weights in each filter for each of the plurality of layers to alleviate an effect of reducing the number of non-zero weights in each filter comprises:
performing one or more iterations of back-propagation on the approximation of the trained deep neural network resulting from reducing the number of non-zero weights in each filter to refine the remaining non-zero weights in each filter to reduce a cost function that measures an error between predicted anatomical object locations using the approximation of the trained deep neural network and ground truth anatomical object locations in a set of training data.

5. The method of claim 2 , wherein sparsifying the weights of the filters for each of the plurality of layers of the trained deep neural network comprises:
performing re-weighted L1-norm regularization on the weights of the filters for each of the plurality layers of the trained deep neural network, wherein the re-weighted L1-norm regularization drives a plurality of non-zero weights of the filters to zero; and
refining the remaining non-zero weights in the filters for each of the plurality of layers to alleviate an effect of driving the plurality of non-zero weights to zero.

6. The method of claim 5 , wherein performing re-weighted L1-norm regularization on the weights of the filters for each of the plurality layers of the trained deep neural network, wherein the re-weighted L1-norm regularization drives a plurality of non-zero weights of the filters to zero comprises:
adding a term that re-weights the L1-norm to a cost function that measures an error between predicted anatomical object locations and ground truth anatomical object locations in a set of training data; and
performing back-propagation on the trained deep neural network to refine the weights in the filters for each of the plurality of layers of the trained deep neural network to reduce the cost function with the added term that re-weights the L1-norm.

7. The method of claim 6 , wherein refining the remaining non-zero weights in the filters for each of the plurality of layers to alleviate an effect of driving the plurality of non-zero weights to zero comprises:
performing one or more iterations of back-propagation on the approximation of the trained deep neural network resulting from driving the plurality of non-zero weights to zero to refine the remaining non-zero weights in the filters to reduce the cost function that measures an error between predicted anatomical object locations and ground truth anatomical object locations in the set of training data, without the added term that re-weights the L1-norm.

8. The method of claim 1 , wherein calculating an approximation of the trained deep neural network that reduces the computational complexity of the trained deep neural network further comprises:
determining a subset of nodes of the plurality nodes in a current layer of the trained deep neural network that linearly approximate the plurality of nodes in the current layer of the trained deep neural network and removing the plurality of nodes in the current layer that are not in subset of nodes from the trained deep neural network; and
updating weights for a next layer of the trained deep neural network based on the subset of nodes remaining in the current layer of the trained deep neural network.

9. The method of claim 8 , wherein determining a subset of nodes of a plurality nodes in a current layer of the trained deep neural network that linearly approximate the plurality of nodes in the current layer of the trained deep neural network and removing the plurality of nodes in the current layer that are not in subset of nodes from the trained deep neural network comprises:
determining the subset of nodes in the current layer and a mixing matrix that best minimizes an error between each of the plurality of nodes in the current layer and a respective approximation for each of the plurality of nodes in the currently layer calculated by linearly combining the subset of nodes using the mixing matrix, subject to a constraint on a size of the subset of nodes.

10. The method of claim 9 , wherein updating weights for a next layer of the trained deep neural network based on the subset of nodes remaining in the current layer of the trained deep neural network comprises:
removing filters for the next layer of the trained deep neural network whose indices are not in the subset of nodes in the current layer; and
updating the remaining filters for the next layer of the trained deep neural network with weights generated by linearly combining weights of the subset of nodes in the current layer using the mixing matrix to approximate weighted inputs to the next layer from the removed ones of the plurality of nodes in the current layer.

11. The method of claim 8 , wherein calculating an approximation of the trained deep neural network that reduces the computational complexity of the trained deep neural network further comprises:
repeating the steps of determining a subset of nodes of a plurality nodes in a current layer of the trained deep neural network that linearly approximate the plurality of nodes in the current layer of the trained deep neural network and removing the plurality of nodes in the current layer that are not in subset of nodes from the trained deep neural network and updating weights for a next layer of the trained deep neural network based on the subset of nodes remaining in the current layer of the trained deep neural network, for each of a plurality of layers in the trained deep neural network, result

Assignees

• Siemens Healthcare Gmbh

Inventors

• Liu David
• Lay Nathan
• Zhou Shaohua Kevin
• Kretschmer Jan
• Nguyen Hien
• Singh Vivek Kumar
• Zheng Yefeng
• Georgescu Bogdan
• Comaniciu Dorin

Classifications

• G06V10/446

  using Haar-like filters, e.g. using integral image techniques · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06F18/2433

  Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title