Automatic condition diagnosis using an attention-guided framework

Patent metadata

Field	Value
Publication number	US-11875898-B2
Application number	US-202117331251-A
Country	US
Kind code	B2
Filing date	May 26, 2021
Priority date	May 26, 2021
Publication date	Jan 16, 2024
Grant date	Jan 16, 2024

Abstract

Methods and systems for training computer-aided condition detection systems. One method includes receiving a plurality of images for a plurality of patients, some of the images including an annotation associated with a condition; iteratively applying a first deep learning network to each of the images to produce an attention map, a feature map, and an image-level probability of the condition for each of the images; iteratively applying a second deep learning network to each feature map produced by the first network to produce a plurality of outputs; training the first network based on the attention map produced for each image; and training the second network based on the output produced for each of the patients. The second network includes a plurality of convolution layers and a plurality of convolutional long short-term memory (LSTM) layers. Each of the outputs includes a patient-level probability of the condition for one of the patients.

First claim

What is claimed is:

1. A method of training a computer-aided condition detection system, the method comprising:
receiving a plurality of medical images for a plurality of patients, a portion of the plurality of medical images including an annotation associated with a condition;
iteratively applying a first deep learning network to each of the plurality of medical images to produce an attention map, a feature map, and an image-level probability of the condition for each of the plurality of medical images;
iteratively applying a second deep learning network to the feature map produced by the first deep learning network for each of the plurality of medical images to produce a plurality of patient outputs;
training the first deep learning network based on the attention map produced by the first deep learning network for each image included in the portion of the plurality of medical images; and
training the second deep learning network based on the patient output produced by the second deep learning network for each of the plurality of patients,
wherein the second deep learning network includes a plurality of convolution layers and a plurality of convolutional long short-term memory (LSTM) layers, and
wherein each of the plurality of patient outputs includes a patient-level probability of the condition for one of the plurality of patients.

2. The method of claim 1 , wherein the plurality of medical images include a plurality of two-dimensional (2-D) images, each of a plurality of subsets of the plurality of 2-D images constituting a series of images corresponding to a three-dimensional (3-D) image of one of the plurality of patients, wherein each 3-D image is associated with a label indicating whether one of the plurality of patients has the condition.

3. The method of claim 2 , wherein training the first deep learning network includes training the first deep learning network using back propagation to reduce a loss function.

4. The method of claim 3 , wherein training the first deep learning network includes:
for each 2-D image included in the portion of the plurality of 2-D medical images including an annotation of the condition:
comparing the annotation of the condition to the attention map produced by first deep learning network for the 2-D image to determine an attention loss, and
comparing the image-level probability produced by the first deep learning network for the 2-D image with a label of the 2-D image to produce a classification loss; and
updating a first set of parameters of the first deep learning network using one or both of the classification loss and the attention loss.

5. The method of claim 4 , wherein training the second deep learning network includes:
freezing the first set of parameters;
updating a second set of parameters of the second deep learning network based on comparing the patient-level probability of the condition of each of the plurality of patient outputs to the label of one of the plurality of 3-D images.

6. The method of claim 4 , wherein training the second deep learning network includes:
updating the first set of parameters of the first deep learning network and a second set of parameters of the second deep learning network based on comparing the patient-level probability of the condition of each of the plurality of patient outputs to the label of one of the plurality of 3-D images.

7. The method of claim 4 , wherein determining the attention loss includes determining the attention loss based on the attention map and a down-sampled version of the corresponding 2-D image with an annotation associated with the condition.

8. The method of claim 4 , wherein determining the attention loss includes determining the attention loss based on an up-sampled version of the attention map and the corresponding 2-D image with an annotation associated with the condition.

9. The method of claim 1 , wherein the condition is pulmonary embolism.

10. The method of claim 1 , wherein the first deep learning network is one selected from a group consisting of ResNet, DenseNet, and SqueezeNet.

11. The method of claim 1 , wherein iteratively applying the first deep learning network to produce the attention map includes producing the attention map based on Gradient-weighted Class Activation Mapping (GradCAM).

12. The method of claim 1 , wherein the first deep learning includes one or more convolutional layers followed by a global average pooling layer and one or more dense layers, wherein the image-level probability is an output of the one or more dense layers and the feature map is an output of a last of the one or more convolutional layers before the global average pooling layer.

13. The method of claim 1 , wherein iteratively applying the second deep learning network to the feature map produced for each of the plurality of medical images to produce the plurality of patient outputs includes pooling, for each of the plurality of patients, outputs produced by the second deep learning network for a subset of the plurality of medical images.

14. The method of claim 13 , wherein pooling the outputs for a subset of the plurality of medical images includes pooling the outputs using an aggregation function, wherein the aggregation function is one selected from a group consisting of a mean function, a max function, a mode function, and a self-attention function.

15. The method of claim 14 , wherein the aggregation function is the self-attention function, and the self-attention function is a third deep learning network including a plurality of fully connected dense layers and a plurality of non-linear activation function layers.

16. The method of claim 1 , wherein each of the plurality of convolutional long short-term memory (LSTM) layers is one selected from a group consisting of a unidirectional LSTM layer and a bidirectional LSTM layer.

17. The method of claim 1 , wherein training the second deep learning network includes training the second deep learning network to minimize a classification loss using at least one selected from a group consisting of a binary cross-entropy loss and a focal loss as an objective function.

18. The method of claim 1 , wherein the method further comprises, after training the first deep learning network and after training the second deep learning network:
receiving a three-dimensional (3-D) medical image of a patient, the 3-D image including a plurality of two-dimensional (2-D) medical images of the patient;
iteratively applying the first deep learning network to each of the plurality of 2-D medical images of the patient to produce a plurality of feature maps associated with the patient, wherein each of the plurality of feature maps associated with the patient is associated with one of the plurality of 2-D medical images of the patient;
iteratively applying the second deep learning network to each of the plurality of feature maps associated with the patient to produce a plurality of classification outputs associated with the patient;
pooling the plurality of classification outputs associated with the patient to generate a condition probability of the patient, the condition probability of the patient including a probability of the patient having the condition; and
outputting the condition probability of the patient.

19. A system for training a computer-aided condition detection program, the system comprising:
a computing device including an electronic processor configured to:
receive a plurality of medical images for a plurality of patients, a portion of the plurality of medical images including

Assignees

• Merative Us L P

Inventors

• Shi Luyao
• Beymer David James
• Marvast Ehsan Dehghan
• Rajan Deepta

Classifications

• G16H50/20 Primary

  for computer-aided diagnosis, e.g. based on medical expert systems · CPC title

• G06F18/217

  Validation; Performance evaluation; Active pattern learning techniques · CPC title

• G06T7/0014

  using an image reference approach · CPC title

• G16H30/40

  for processing medical images, e.g. editing · CPC title

• G06T2207/20076

  Probabilistic image processing · CPC title