Methods and apparatus for bio-fluid specimen characterization using neural network having reduced training

Patent metadata

Field	Value
Publication number	US-11386291-B2
Application number	US-201916961222-A
Country	US
Kind code	B2
Filing date	Jan 8, 2019
Priority date	Jan 10, 2018
Publication date	Jul 12, 2022
Grant date	Jul 12, 2022

Abstract

A method of training a neural network (Convolutional Neural Network-CNN) including reduced graphical annotation input is provided. The training method can be used to train a Testing CNN that can be used for determining Hemolysis (H), Icterus (I), and/or Lipemia (L), or Normal (N) of a serum or plasma portion of a test specimen. The training method includes capturing training images of multiple specimen containers including training specimens, generating region proposals of the serum or plasma portions of the training specimens; and selecting the best matches for the location, size and shape of the region proposals for the multiple training specimens. The obtained features (network and weights) from the training CNN can be used in a testing CNN. Quality check modules and testing apparatus adapted to carry out the training method, and characterization methods using abounding box regressor are described, as are other aspects.

First claim

What is claimed is:

1. A method of training a neural network, comprising:
capturing training images of a specimen container containing training specimens at a imaging location;
generating region proposals of a serum or plasma portion for input to the neural network; and
selecting region proposals that provide best matches for serum or plasma portions of the training specimens.

2. The method of claim 1 , wherein the region proposals comprise sub-regions of a virtual grid superimposed over the training images.

3. The method of claim 2 , wherein the virtual grid comprises multiple grid elements of dimension of W in width×H in height.

4. The method of claim 2 , wherein each of the region proposals comprises a rectangular sub-region within the virtual grid.

5. The method of claim 2 , wherein the region proposals are randomly selected.

6. The method of claim 1 , wherein the selecting region proposals comprises selecting a best 2,000 or more of the region proposals that are generated.

7. The method of claim 1 , wherein the best matches for the serum or plasma portions of the training specimens are based upon intensity gradients within each of the region proposals.

8. The method of claim 1 , wherein the best matches for the serum or plasma portions of the training specimens are based upon intensity gradients at the periphery of each of the region proposals.

9. The method of claim 1 , wherein the best matches for the serum or plasma portions of the training specimens are based upon summing intensity gradients at the periphery of each of the region proposals.

10. The method of claim 1 , wherein the convolutional neural network comprises an architecture including at least two layers configured to carry out convolution and pooling, and at least two additional fully-convolution layers.

11. The method of claim 1 , wherein the convolutional neural network comprises an architecture including a loss layer with a bounding box regressor.

12. The method of claim 1 , wherein the convolutional neural network comprises an architecture including a loss layer with a bounding box regressor and a SoftMax.

13. The method of claim 1 , wherein the convolutional neural network comprises an architecture including at least three layers including convolution and pooling, and at least two fully-convolutional layers, and a loss layer with a bounding box regressor and a SoftMax.

14. The method of claim 1 , wherein the capturing training images comprises different exposures for each of multiple spectra.

15. The method of claim 1 , wherein the capturing training images comprises providing different exposure times for each spectrum of red, green, and blue.

16. The method of claim 1 , wherein the capturing training images involves capturing images from multiple viewpoints with multi-spectral, multi-exposure images for each viewpoint.

17. A method of characterizing a specimen using a trained neural network, comprising:
capturing images of a specimen container containing the specimen at an imaging location;
generating a region proposal of a serum or plasma portion for input to the neural network; and
converging the region proposal to provide a match for the serum or plasma portion of the specimen through regression to provide a validated region.

18. The method of characterizing a specimen using a trained neural network of claim 17 , comprising characterization of the validated region as containing one or more of hemolysis, Icterus, lipemia, or being normal with the trained neural network.

19. A quality check system, comprising:
an image capture device configured to capture multiple images of a specimen container containing a serum or plasma portion of a specimen; and
a computer coupled to the image capture device, the computer configured and capable of being operated to:
input image data from the multiple images to a neural network,
generate a region proposal of serum or plasma portion,
converge the region proposal through regression to provide a validated region, and
output from the neural network a classification of the validated region as being one or more of hemolytic, icteric, lipemic, and normal.

20. The quality check system of claim 19 , wherein the neural network comprises an architecture including a bounding box regressor and a SoftMax.

21. The quality check system of claim 19 , wherein the neural network comprises an architecture including at least three layers including convolution and pooling, at least two fully-convolutional layers, a bounding box regressor, and a SoftMax.

22. The quality check system of claim 19 , wherein the multiple images are captured as multi-spectral, multi-exposure images.

23. The quality check system of claim 19 , wherein the multiple images are captured from multiple viewpoints with multi-spectral, multi-exposure images for each viewpoint.

Assignees

• Siemens Healthcare Diagnostics Inc

Inventors

• Kluckner Stefan
• Chang Yao-Jen
• Ma Kai
• Singh Vivek
• Chen Terrence
• Pollack Benjamin S

Classifications

• G06V10/82 Primary

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/214

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

• G06F18/24133

  Distances to prototypes · CPC title

• G06N3/045

  Combinations of networks · CPC title