Prostate cancer tissue image classification with deep learning

1. A method for classifying prostate cancer tissue data by using a trained deep learning network, wherein said classification comprises the steps of:

(a) capturing of classification histological tissue image data from a classification prostate tissue sample that has been stained with at least two stains, said classification histological tissue image data comprising a set of pixels, said stains being light absorbent, and one said stain being absorbed primarily by the stroma, the other said stain being absorbed primarily by the nuclei;

(b) identifying at least one classification nucleus in said classification histological tissue image data;

(c) identifying a classification window of pixels around said classification nucleus;

(d) classifying at least one classification nucleus as belonging to one category by applying a trained deep learning network to said classification window around said classification nucleus in said classification histological tissue image; and

(e) classifying at least one region in the classification histological tissue image data according to the categories of the classification nuclei contained in said region,

wherein the deep learning network has been previously trained by a training method comprising the steps of:

(i) capturing of training histological tissue image data from a training prostate tissue sample that has been stained with at least two stains, said training histological tissue image data comprising a set of pixels, said stains being light absorbent, and one said stain being absorbed primarily by the stroma, the other said stain being absorbed primarily by the nuclei;

(ii) segmenting said training histological tissue data into at least one training gland, where said training gland is surrounded by stromal tissue;

(iii) labelling at least one training gland according to its category;

(iv) identifying at least one training nucleus in said training histological tissue image data;

(v) labelling said training nucleus according to the category of said training nucleus centroid's surrounding tissue;

(vi) identifying a training window of pixels around said labelled training nucleus; and

(vii) training at least one deep learning network using a multitude of said training windows as training data, to recognize the category of at least one classification nucleus in classification prostate tissue image data.

2. The method of claim 1 , wherein at least one identification of a classification nucleus or a training nucleus uses a marked point process.

3. The method of claim 1 , wherein each classification window and each training window is square, the classification windows and the training windows are of the same dimensions, and each classification window and each training window is centered around the centroid of an individual classification nucleus or training nucleus, respectively.

4. The method of claim 1 , wherein said deep learning network is a convolutional neural network.

5. The method of claim 1 , wherein the classification of said regions comprise the steps of:

deriving one density map, said density map corresponding to a portion of said classification histological tissue data that represents the stroma;

segmenting said density map into at least one gland; and

classifying said prostate gland with the category of the most prevalent category among its classification nuclei.

6. The method of claim 1 , wherein the glandular segmentation further comprises the steps of:

deriving one density map, said density map corresponding to a portion of said training histological tissue image data that represents the stroma; and

segmenting said density map into at least one training gland.

7. The method of claim 6 , wherein said density map is derived using the blind color decomposition (BCD) method.

8. The method of claim 1 , wherein said training glands have been labelled according to their categories by at least one expert, the categories comprising prostate components.

9. Image capture and analysis apparatus, comprising:

(a) a microscope adapted to capture histological tissue image data from a tissue sample that has been stained with at least two stains, the said stains being light absorbent, and one said stain being absorbed primarily by stroma, the other said stain being absorbed primarily by nuclei, said training histological tissue image data comprising a set of pixels; and

(b) image processing modules adapted to perform the steps of:

(b1) creating a deep learning training data set comprising square windows centered around nuclei in said tissue image data by:

(i) segmenting captured training histological tissue data into at least one training gland, where said training gland is surrounded by stromal tissue;

(ii) labelling at least one training gland according to its category;

(iii) identifying at least one training nucleus in said training histological tissue image data;

(iv) labelling said training nucleus according to the category of said training nucleus centroid's surrounding tissue; and

(v) identifying a square training window of pixels around said labelled training nucleus;

(b2) training a deep learning network using a multitude of said training windows as training data to recognize the category of at least one nucleus in a data set of histological tissue image; and

(b3) classifying prostate cancer tissue with the trained deep learning network.

10. Image capture and analysis apparatus according to claim 9 , wherein the trained deep learning network resides on a server in a client/server architecture.

11. Image capture and analysis apparatus according to claim 9 , wherein the trained deep learning network resides on a client in a client/server architecture.

12. The method of claim 5 , wherein said density map is derived using the blind color decomposition (BCD) method.