Digital pathology using an artificial neural network
Various example embodiments pertain to processing images that depict tissue samples using a neural network algorithm. The neural network algorithm includes multiple encoder branches that are copies of each other that share the same parameters. The encoder branches can, accordingly, be referred to as Siamese copies of each other.
1. A computer-implemented method, comprising:
obtaining an input mage, the input image depicting a tissue sample;
determining multiple tiles of the input image;
processing each tile of the multiple tiles in a respective encoder branch of a neural network algorithm, to obtain a respective latent feature data structure; and
using a decoder branch of the neural network algorithm, the using including,
aggregating the latent feature data structures of the multiple tiles, to obtain a merged latent feature data structure, and
processing the merged latent feature data structure to infer at least one semantic histopathology feature associated with the tissue sample,
wherein the multiple encoder branches processing the multiple tiles share the same parameters.
2. A computer-implemented method, comprising:
obtaining a training input image, the training input image depicting a tissue sample;
obtaining a ground-truth label for at least one semantic histopathology feature associated with the training input image;
determining multiple tiles of the training input image; and
performing an end-to-end training of a neural network algorithm based on the ground-truth label, the neural network algorithm comprising multiple encoder branches and a decoder branch, each encoder branch processing a respective tile of the multiple tiles, the decoder branch processing a merged latent feature data structure obtained from aggregating latent feature data structures obtained from the multiple encoder branches,
wherein the multiple encoder branches share the same parameters.
3. The computer-implemented method of claim 1 ,
wherein the multiple encoder branches comprise convolutional layers.
4. The computer-implemented method of claim 1 ,
wherein the multiple tiles of the input image are determined using a selection process, and
the selection process comprises at least one of randomization of positions of the multiple tiles or thresholding of a contrast of the multiple tiles.
5. The computer-implemented method of claim 1 ,
wherein the at least one semantic histopathology feature is selected from a group consisting of: a Gleason score, cancer grading, cancer stage estimation, clinical pathway prediction, sub-tumor classification, metastasis evaluation, and microsatellite instability or stability.
6. The computer-implemented method of claim 1 ,
wherein the decoder branch comprises at least one of
an averaging layer,
a max-pooling layer,
a 1-D convolution layer,
a fully-connected layer, or
a layer implementing a matrix product with a weight kernel.
7. The method of claim 1 ,
wherein the merged latent feature data structure is processed using a fully connected layer of the decoder branch.
8. The method of claim 1 ,
wherein the multiple tiles cover an area fraction of the input image, the area fraction being in a range of 0.5% to 20%.
9. A device comprising:
at least one processor configured to,
obtain an input image, the input image depicting a tissue sample,
determine multiple tiles of the input image,
process each tile of the multiple tiles in a respective encoder branch of a neural network algorithm, to obtain a respective latent feature data structure, and
use a decoder branch of the neural network algorithm to,
aggregate the latent feature data structures of the multiple tiles to obtain a merged latent feature data structure, and
process the merged latent feature data structure to infer at least one semantic histopathology feature associated with the tissue sample,
wherein the multiple encoder branches share the same parameters.
10. A device comprising:
at least one processor configured to,
obtain a training input image, the training input image depicting a tissue sample,
obtain a ground-truth label for at least one semantic histopathology feature associated with the training input image,
determine multiple tiles of the training input image,
perform an end-to-end training of a neural network algorithm based on the ground-truth label, the neural network algorithm comprising multiple encoder branches and a decoder branch, each encoder branch configured to process a respective tile of the multiple tiles, the decoder branch configured to process a merged latent feature data structure obtained from aggregating latent feature data structures obtained from the multiple encoder branches, wherein the multiple encoder branches share the same parameters.
11. The computer-implemented method of claim 2 ,
wherein the multiple encoder branches comprise convolutional layers.
12. The computer-implemented method of claim 2 ,
wherein the multiple tiles of the training input image are determined using a selection process, and
the selection process comprises at least one of randomization of positions of the multiple tiles or thresholding of a contrast of the multiple tiles.
13. The computer-implemented method of claim 2 ,
wherein the at least one semantic histopathology feature is selected from a group consisting of: a Gleason score, cancer grading, cancer stage estimation, clinical pathway prediction, sub-tumor classification, metastasis evaluation, and microsatellite instability or stability.
14. The device of claim 9 ,
wherein the multiple encoder branches comprise convolutional layers.
15. The device of claim 9 ,
wherein the decoder branch comprises at least one of
an averaging layer,
a max-pooling layer,
a 1-D convolution layer,
a fully-connected layer, or
a layer implementing a matrix product with a weight kernel.
16. The device of claim 10 ,
wherein the multiple encoder branches comprise convolutional lavers.
17. The device of claim 10 ,
wherein the decoder branch comprises at least one of
an averaging layer,
a max-pooling layer,
a 1-D convolution layer,
a fully-connected layer, or
a layer implementing a matrix product with a weight kernel.