High-throughput adaptive sampling for whole-slide histopathology image analysis

1. A non-transitory computer-readable storage medium storing computer executable instructions that when executed by a computer control the computer to perform a method for detecting cancerous pathology in a region of tissue, the method comprising:

accessing an image of a region of tissue demonstrating cancerous pathology, where the image comprises a set of tiles;

generating a tile sampling selection, where the tile sampling selection defines a sampling subset of the set of tiles, where the number of tiles in the sampling subset is N, where N is less than the number of tiles in the set of tiles, N being an integer;

upon determining that a termination condition has not been met:

extracting the sampling subset from the image based on the tile sampling selection;

providing the sampling subset to an automated classifier;

receiving, from the automated classifier, a prediction probability based, at least in part, on the sampling subset, where the prediction probability indicates the probability of invasive pathology at a location in the image occupied by a tile;

constructing a probability map of the region of tissue based on the prediction probability or the sampling subset;

computing a gradient of the probability map; and

updating the tile sampling selection based on the gradient of the probability map;

upon determining that the termination condition has been met:

constructing a final probability map based, at least in part, on the probability map; and

controlling a computer aided diagnosis (CADx) system to generate a classification of the region of tissue based, at least in part, on the final probability map.

2. The non-transitory computer-readable storage medium of claim 1 , where the image of the region of tissue is a digitized whole slide image (WSI) of a region of tissue demonstrating estrogen receptor-positive (ER+) BCa.

3. The non-transitory computer-readable storage medium of claim 2 , where the WSI is a red-green-blue (RGB) color model WSI, where the WSI has dimensions of 80,000 pixels by 80,000 pixels.

4. The non-transitory computer-readable storage medium of claim 3 , where a tile has dimensions of 101 pixels by 101 pixels.

5. The non-transitory computer-readable storage medium of claim 1 , where generating the tile sampling selection comprises:

selecting the sampling subset of the set of tiles using a regular sampling approach, a uniform random sampling approach, or a quasi-Monte Carlo sampling approach.

6. The non-transitory computer-readable storage medium of claim 5 , where the quasi-Monte Carlo sampling approach employs a deterministic sequence.

7. The non-transitory computer-readable storage medium of claim 6 , where the deterministic sequence has a low discrepancy.

8. The non-transitory computer-readable storage medium of claim 7 , where the deterministic sequence measures a uniformity of a distribution of finite point sets.

9. The non-transitory computer-readable storage medium of claim 8 , where the deterministic sequence is a Sobol sequence or a Halton sequence.

10. The non-transitory computer-readable storage medium of claim 1 , where the termination condition comprises a number of iterations, a steady state condition, a processing time limit, a minimum epsilon value for an average magnitude gradient of the probability map, or a minimum approximation error between a first probability map and a second, different probability map.

11. The non-transitory computer-readable storage medium of claim 10 , where the number of iterations is twenty.

12. The non-transitory computer-readable storage medium of claim 1 , where the prediction probability represents the probability of invasive pathology at the location in the image occupied by a tile.

13. The non-transitory computer-readable storage medium of claim 12 , where constructing the probability map based on the prediction probability and the set of tiles comprises cubically interpolating the prediction probability to a plurality of tiles in the image.

14. The non-transitory computer-readable storage medium of claim 13 , where computing the gradient of the probability map comprises:

calculating a gradient image based on an X direction gradient and a Y direction gradient, where the X direction gradient represents a directional change of the probability map in the X direction, and where the Y direction gradient represents a directional change of the probability map in the Y direction; and

computing a gradient magnitude image based, at least in part, on the gradient image or the X direction gradient and the Y direction gradient.

15. The non-transitory computer-readable storage medium of claim 14 , where updating the tile sampling selection based on the gradient of the probability map comprises:

selecting 2N sample tiles from the set of tiles using a quasi-Monte Carlo sampling approach or a random sampling approach; and

generating an updated sampling selection by determining the N tiles that have the highest gradient magnitude from the 2N sample tiles, where the updated sampling selection comprises the N tiles that have the highest gradient magnitude from the 2N sample tiles.

16. The non-transitory computer-readable storage medium of claim 15 , where determining the N tiles from the 2N sample tiles that have the highest gradient magnitude comprises:

generating a ranked set of sample tiles based, at least in part, on the gradient magnitude image; and

selecting the N highest ranked sample tiles from the ranked set of sample tiles.

17. The non-transitory computer-readable storage medium of claim 1 , where the automated classifier is a 2-layer convolutional neural network (CNN) comprising:

a first layer comprising a 256 unit convolutional layer and a 256 unit pooling layer;

a second layer comprising a 256 unit fully-connected layer; and

a classification layer, where the classification layer is a softmax classifier, where the classification layer produces a first output activated by a logistic regression function and a second output activated by the logistic regression function.

18. The non-transitory computer-readable storage medium of claim 17 , where the convolutional layer generates a feature map by applying a 2 dimensional (2D) convolution of an input image with an 8 pixel by 8 pixel kernel.

19. The non-transitory computer-readable storage medium of claim 18 , where the pooling layer applies a spatial L2 pooling function without overlapping to the feature map, where the pooling layer applies a 2 pixel by 2 pixel pooling kernel to the feature map.

20. The non-transitory computer-readable storage medium of claim 1 , the method further comprising training the automated classifier.

21. The non-transitory computer-readable storage medium of claim 20 , where training the automated classifier comprises training a CNN model by:

accessing a set of digital WSIs of tissue demonstrating invasive pathology and non-invasive pathology, where a member of the set of digital WSIs comprises a set of tiles, where a member of the set of digital WSIs is an RGB color model WSI;

accessing a training set of tiles, where the training set includes a tile sampled from a region of invasive BCa tissue, and a tile sampled from a region of non-invasive tissue;

converting the training set of tiles from RGB color space to YUV color space;

normalizing the training set of tiles to a mean of zero and a variance of one; and

minimizing a softmax loss function using a stochastic gradient descent.

22. A non-transitory computer-readable storage medium storing computer executable instructions that when executed by a computer control the computer to perform a method, the method comprising:

accessing a radiological image of a region of tissue demonstrating invasive pathology, where the radiological image comprises a set of tiles;

selecting a sample set of tiles from the set of tiles using a gradient-based quasi-Monte Carlo sampling approach;

providing the sample set of tiles to an automated classifier;

receiving, from the automated classifier, a classification of the sample set of tiles;

generating an invasiveness probability map based, at least in part, on the classification;

computing a gradient image based, at least in part, on the invasiveness probability map; and

providing an invasiveness classification and localization based on the invasiveness probability map and the gradient image.

23. An apparatus for detecting cancerous pathology in a region of tissue, comprising:

a processor;

a memory;

an input/output interface;

a set of circuits; and

an interface to connect the processor, the memory, the input/output interface and the set of circuits, where the set of circuits includes:

an image acquisition circuit that acquires an image of a region of tissue demonstrating cancerous pathology, where the image comprises a set of tiles;

an adaptive sampling circuit that selects a sampling subset of tiles from the image where the sampling subset of tiles comprises N tiles, where N is less that the number of tiles in the set of tiles, N being an integer;

an invasiveness circuit that determines a probability of a presence of invasive pathology in a member of the sampling subset of tiles;

a probability map circuit that generates an invasiveness probability map based, at least in part, on the probability;

a probability gradient circuit that generates a gradient image based, at least in part, on the invasiveness probability map, where the gradient image identifies directional changes of the invasiveness probability map; and

a classification circuit that classifies the region of tissue based, at least in part, on the probability map.

24. The apparatus of claim 23 , where the adaptive sampling circuit selects the sampling subset of tiles using an iterative quasi-Monte Carlo (QMC) Sobol sequence sampling approach or an iterative QMC Halton sequence sampling approach.

25. The apparatus of claim 23 , where the adaptive sampling circuit selects the sampling subset of tiles using a uniform random sampling approach, or a gradient-based random sampling approach.

26. The apparatus of claim 23 , where the adaptive sampling circuit selects an updated subset of tiles from the image based on the gradient image, where the adaptive sampling circuit generates a set of 2N sample tiles using an iterative random sampling approach or an iterative quasi-Monte Carlo (QMC) sampling approach, where a member of the set of 2N sample tiles has a gradient magnitude, where the adaptive sampling circuit ranks the set of 2N sample tiles according to a gradient magnitude associated with a member of the set of 2N sample tiles, and where the adaptive sampling circuit selects the N sample tiles that have the highest gradient magnitudes from the set of 2N sample tiles according to the gradient magnitude.

27. The apparatus of claim 23 , where the invasiveness circuit determines the probability using a two-layer convolutional neural network (CNN) classifier trained using a stochastic gradient descent approach.

28. The apparatus of claim 27 , where the CNN classifier employs a (CS256-FC256) architecture comprising a softmax classification layer.

29. The apparatus of claim 23 , where the probability gradient circuit generates the gradient image based on a gradient in the X direction and a gradient in the Y direction, where the gradient in the X direction identifies a change in the probability in the X direction, and the gradient in the Y direction identifies a change in the probability in the Y direction, where a high gradient magnitude corresponds to a heterogeneous spatial distribution of invasive tissue or non-invasive tissue, and where a low gradient magnitude corresponds with a homogeneous distribution of invasive tissue or non-invasive tissue.

30. The apparatus of claim 23 , where the probability map circuit generates the invasiveness probability map by extending the probability of the presence of invasive pathology in a member of the subset of tiles to member of the set of tiles that is not a member of the subset of tiles using cubic interpolation.

31. The apparatus of claim 23 , where the image is a digitized whole slide image (WSI) of a region of tissue demonstrating estrogen receptor-positive (ER+) BCa.