Using a first stain to train a model to predict the region stained by a second stain
A convolutional neural network predicts which regions of a tissue slice would be stained by a first stain by training a model to identify those regions based only on tissue stained by a second stain. Thereafter the first stain need not be used to mark cancerous regions on other tissue slices that are stained with the second stain. The training slice is stained with a first immunohistochemical stain and a second counterstain. A target region of an image of the training slice is identified using image analysis based on the first stain. A set of parameters for associated mathematical operations are optimized to train the model to classify pixels of the image as belonging to the target region based on the second stain but not on the first stain. The trained parameters are stored in a database and applied to other images of tissue not stained with the first stain.
1. A method comprising:
staining a slice of cancerous tissue with a first stain;
staining the slice with a second stain;
acquiring a first digital image of the slice;
identifying a likely cancerous region in the first digital image based on the first stain;
optimizing a plurality of parameters applied to associated mathematical operations to train a model based on the second stain but not on the first stain to classify individual pixels of the first digital image as belonging to the likely cancerous region; and
storing the plurality of parameters and associated mathematical operations of the model in a database for applying to a second digital image to indicate a probability that each pixel of the second digital image falls within the likely cancerous region, wherein the second digital image is acquired from cancerous tissue that is stained with the second stain but not with the first stain.
2. The method of claim 1 , further comprising:
generating a prediction image in which each pixel location of the second digital image is marked to indicate the probability that the associated pixel of the second digital image falls within the likely cancerous region.
3. The method of claim 1 , wherein the model is a convolutional neural network model, and wherein the optimizing the plurality of parameters involves training the convolutional neural network model.
4. The method of claim 1 , wherein the first stain stains epithelial cells, and the second stain stains nuclei.
5. The method of claim 1 , wherein the likely cancerous region is predominantly epithelial cells that have been stained by a cytokeratin stain.
6. The method of claim 1 , wherein the first stain is taken from the group consisting of: pan cytokeratin, cytokeratin 18, α-methylacyl coenzyme A racemase (AMACR), cluster of differentiation 3 (CD3) antibody stain, cluster of differentiation 4 (CD4) antibody stain and cluster of differentiation 68 (CD68) antibody stain.
7. The method of claim 1 , wherein the second stain is taken from the group consisting of: fluorescent 4, 6-diamidino-2-phenylindole (DAPI) and hematoxylin.
8. The method of claim 1 , wherein the first stain is an immunohistochemical stain.
9. The method of claim 1 , wherein the first stain is an immunofluorescence stain.
10. The method of claim 1 , wherein the second stain is a counterstain.
11. The method of claim 1 , wherein one of the mathematical operations is convolution filtering that applies a kernel to pixels surrounding a central pixel.
12. The method of claim 1 , wherein one of the mathematical operations is a Gaussian filter applied to pixel values through a convolution matrix.
13. A method comprising:
staining a slice of tissue from a cancer patient with a first stain;
staining the slice of tissue with a second stain;
acquiring a digital image of the slice of tissue;
identifying a target region in the digital image using image analysis based on the first stain;
optimizing a set of parameters for associated mathematical operations to train a model to classify individual pixels of the digital image as belonging to the target region based on the second stain but not on the first stain; and
generating a prediction image in which particular pixel locations of the digital image have an intensity associated with belonging to the target region.
14. The method of claim 13 , wherein one of the mathematical operations is convolution filtering that applies a kernel to neighboring pixels.
15. The method of claim 13 , wherein one of the mathematical operations is a Gaussian filter applied to pixel values through a convolution matrix.
16. The method of claim 13 , wherein the intensity of each of the particular pixel locations is proportional to a probability that each of the particular pixel locations belongs to the target region.
17. The method of claim 13 , wherein the intensity of each of the particular pixel locations indicates whether it is either more likely or less likely that each of the particular pixel locations belongs to the target region.
18. The method of claim 13 , wherein the first stain stains epithelial cells, and the second stain stains nuclei.
19. The method of claim 13 , wherein the first stain is an immunohistochemical stain.
20. The method of claim 13 , wherein the second stain is a counterstain.
21. The method of claim 13 , wherein the first stain is taken from the group consisting of: pan cytokeratin, cytokeratin 18, α-methylacyl coenzyme A racemase (AMACR), cluster of differentiation 3 (CD3) antibody stain, cluster of differentiation 4 (CD4) antibody stain and cluster of differentiation 68 (CD68) antibody stain.
22. The method of claim 13 , wherein the second stain is taken from the group consisting of: hematoxylin and fluorescent 4, 6-diamidino-2-phenylindole (DAPI).
23. A method with a training mode and a prediction mode, comprising:
in the training mode, staining a first slice of tissue from a cancer patient with a first stain, wherein the first stain has a first color;
staining the first slice of tissue with a second stain, wherein the second stain has a second color;
acquiring a first digital image of the first slice of tissue;
determining whether each pixel of the first digital image falls within a region of interest using image analysis based on the first color;
generating a set of parameters of a model that indicate a probability that each pixel of the first digital image falls within the region of interest using the second color but not the first color; and
in the prediction mode, using the set of parameters of the model to indicate the probability that each pixel of a second digital image falls within the region of interest using the second color, wherein the second digital image is acquired from a second slice of tissue that is stained with the second stain but not with the first stain.
24. The method of claim 23 , wherein the model is a convolutional neural network model, and wherein the generating the set of parameters involves training the convolutional neural network model.
25. The method of claim 23 , further comprising:
generating a prediction image in which each pixel location of the second digital image has an intensity indicative of the probability that the associated pixel of the second digital image falls within the region of interest.