Dual encoder attention U-net

1. A method of predicting features in binary files, executable by a processor, the method comprising:

receiving a binary file;

calculating a feature value corresponding to latent space of the binary file, wherein calculating the feature value comprises calculating one or more probability maps based on one or more determined attention priors, and wherein the one or more probability maps correspond to a processed attention feature map and a hematoxylin and eosin feature map; and

generating a new binary file on the calculated feature value,

wherein calculating the one or more probability maps comprises:

translating the processed attention feature map and the hematoxylin and eosin feature map to a first intermediate value using one or more first convolution kernels; and

performing element-wise addition and non-linear transformation on the first intermediate value through a rectified linear unit and one or more second convolution kernels to generate a second intermediate value,

wherein the feature value is calculated by a dual-encoder attention-based U-net architecture comprising:

a dual-encoder u-net, comprising:

one or more input nodes,

one or more attention prior processing nodes, and

wherein the one or more input nodes are connected by a max pooling layer, and the one or more attention prior processing nodes are connected by the max pooling layer,

wherein a first input node, among the one or more input nodes, corresponds to a first attention prior processing node, among the one or more attention prior processing nodes, and

wherein the first input node and the first attention prior processing node are connected to a corresponding first attention skip module, among one or more attention skip modules, by a skip connection.

2. The method of claim 1 , wherein the dual-encoder attention-based U-net architecture comprises:

a pre-processing module; and

a post-processing module,

wherein the dual-encoder u-net further comprises:

one or more output nodes.

3. The method of claim 1 , wherein the received binary file comprises at least one from among an image, an animation, a video, and a sound.

4. The method of claim 1 , wherein calculating the feature value comprises:

determining the one or more attention priors from the received binary file; and

generating an instance mask from the one or more probability maps, wherein the instance mask corresponds to predicted features associated with the binary files.

5. The method of claim 4 , wherein the received binary file corresponds to slide images of one or more cells stained with hematoxylin and eosin dyes.

6. The method of claim 5 , wherein the one or more attention priors correspond to a hematoxylin channel extracted from the slide images.

7. The method of claim 6 , wherein determining the one or more attention priors comprises:

color-normalizing the slide images;

factorizing the color-normalized slide images into a stain color matrix and an optical density factor; and

separating the factorized color-normalized slide images into the hematoxylin channel and an eosin channel.

8. The method of claim 1 , wherein calculating the one or more probability maps further comprises:

generating one or more coefficients from the second intermediate value using a sigmoid function; and

generating one or more probability map values through element-wise multiplication of the one or more coefficients with the hematoxylin and eosin feature map and one or more third convolution kernels.

9. The method of claim 8 , further comprising back-propagating the one or more coefficients and the one or more probability map values to the one or more first, second, and third convolution kernels.

10. A computer system for predicting features in binary files, the computer system comprising:

one or more computer-readable non-transitory storage media configured to store computer program code; and

one or more computer processors configured to access said computer program code and operate as instructed by said computer program code, said computer program code including:

receiving code configured to cause the one or more computer processors to receive a binary file;

calculating code configured to cause the one or more computer processors to calculate a feature value corresponding to latent space of the binary file, wherein to calculate the feature value comprises to calculate one or more probability maps based on one or more determined attention priors, and wherein the one or more probability maps correspond to a processed attention feature map and a hematoxylin and eosin feature map; and

generating code configured to cause the one or more computer processors to generate a new binary file on the calculated feature value,

wherein the calculating code is further configured to cause the one or more computer processors to calculate the one or more probability maps by:

translating the processed attention feature map and the hematoxylin and eosin feature map to a first intermediate value using one or more first convolution kernels; and

performing element-wise addition and non-linear transformation on the first intermediate value through a rectified linear unit and one or more second convolution kernels to generate a second intermediate value,

wherein the calculating code is further configured to cause the one or more computer processors to calculate the feature value using a dual-encoder attention-based U-net architecture comprising a dual-encoder u-net comprising one or more input nodes and one or more attention prior processing nodes,

wherein the one or more input nodes are connected by a max pooling layer, and the one or more attention prior processing nodes are connected by the max pooling layer,

wherein a first input node, among the one or more input nodes, corresponds to a first attention prior processing node, among the one or more attention prior processing nodes, and

wherein the first input node and the first attention prior processing node are connected to a corresponding first attention skip module, among one or more attention skip modules, by a skip connection.

11. The computer system of claim 10 , wherein the received binary file comprises at least one from among an image, an animation, a video, and a sound.

12. The computer system of claim 10 , wherein calculating the feature value comprises:

determining the one or more attention priors from the received binary file; and

generating an instance mask from the one or more probability maps, wherein the instance mask corresponds to predicted features associated with the binary files.

13. The computer system of claim 12 , wherein the received binary file corresponds to slide images of one or more cells stained with hematoxylin and eosin dyes.

14. The computer system of claim 13 , wherein the one or more attention priors correspond to a hematoxylin channel extracted from the slide images.

15. The computer system of claim 14 , wherein determining the one or more attention priors comprises:

color-normalizing the slide images;

factorizing the color-normalized slide images into a stain color matrix and an optical density factor; and

separating the factorized color-normalized slide images into the hematoxylin channel and an eosin channel.

16. The computer system of claim 10 , wherein calculating the one or more probability maps comprises:

generating one or more coefficients from the second intermediate value using a sigmoid function; and

generating one or more probability map values through element-wise multiplication of the one or more coefficients with the hematoxylin and eosin feature map and one or more third convolution kernels.

17. A non-transitory computer readable medium having stored thereon a computer program for predicting features in binary files, the computer program configured to cause one or more computer processors to:

receive a binary file;

calculate a feature value corresponding to latent space of the binary file, wherein to calculate the feature value comprises to calculate one or more probability maps based on one or more determined attention priors, and wherein the one or more probability maps correspond to a processed attention feature map and a hematoxylin and eosin feature map;

translate the processed attention feature map and the hematoxylin and eosin feature map to a first intermediate value using one or more first convolution kernels;

perform element-wise addition and non-linear transformation on the first intermediate value through a rectified linear unit and one or more second convolution kernels to generate a second intermediate value; and

generate a new binary file on the calculated feature value,

wherein the calculating feature value comprises calculating the feature value using a dual-encoder attention-based U-net architecture comprising a dual-encoder u-net comprising one or more input nodes and one or more attention prior processing nodes,

wherein the one or more input nodes are connected by a max pooling layer, and the one or more attention prior processing nodes are connected by the max pooling layer,

wherein a first input node, among the one or more input nodes, corresponds to a first attention prior processing node, among the one or more attention prior processing nodes, and

wherein the first input node and the first attention prior processing node are connected to a corresponding first attention skip module, among one or more attention skip modules, by a skip connection.