Signet ring cell detector and related methods

1. A method for automatically detecting signet ring cells in an image of a biopsy tissue sample, the image comprising at least two different colors, one of the at least two different colors most likely representing cell membranes, the method comprising the steps of:

in a computer process, separating a portion of the image comprising the one of the at least two different colors that most likely represents the cell membranes from a portion of the image comprising the other one of the at least two different colors;

in a computer process, identifying edges in the portion of the image comprising the one of the at least two different colors that most likely represent the cell membranes using image intensity change in a plurality of directions, wherein the identified edges are likely the cell membranes;

from the identified edges likely to be cell membranes, determining, in a computer process, points around which the cell membranes appear radially symmetric, each of the points representing the center of a possible signet ring cell;

selecting as candidate signet ring cells, in a computer process, ones of the points representing the centers of possible signet ring cells that have an adjacent nucleus with a squashed shape by, for each of the points representing the center of a possible signet ring cell:

comparing the color of each pixel of the image to an expected color of a nucleus and classifying pixels having the expected color of the nucleus as nuclear pixels;

extracting any connected group of the nuclear pixels from the image;

determining whether the extracted connected group of nuclear pixels has a nucleus with the squashed shape on a periphery thereof; and

selecting that point as a center of a candidate signet ring cell; and

applying a convolutional neural network computer process to the selected points representing the centers of the candidate signet ring cells to determine which of the candidate signet ring cells are signet ring cells.

2. The method of claim 1 , wherein the step of determining points around which the cell membranes appear radially symmetric includes the step of finding peaks of a Hough transform wherein the points representing the centers of possible signet ring cells are selected as the peaks of the transform achieving a given threshold.

3. The method of claim 1 , wherein the selecting step is performed with a classifier.

4. The method of claim 3 , wherein the classifier comprises a convolutional neural network computer process.

5. The method of claim 3 , wherein the classifier comprises a support vector machine computer process.

6. The method of claim 1 , wherein the step of applying the convolutional neural network computer process is performed with a convolutional neural network computer trained on a supply of candidate points known to be signet ring cells, against a supply of candidate points that are not signet ring cells.

7. A signet ring cell detector for automatically detecting signet ring cells in an image of a biopsy tissue sample, the image comprising at least two different colors, one of the at least two different colors most likely representing cell membranes, the detector comprising:

a processor executing instructions for:

separating a portion of the image comprising the one of the at least two different colors that most likely represents the cell membranes from a portion of the image comprising the other one of the at least two different colors;

identifying edges in the portion of the image comprising the one of the at least two different colors that most likely represent the cell membranes using image intensity change in a plurality of directions, wherein the identified edges are likely the cell membranes;

from the identified edges likely to be cell membranes, determining points around which the cell membranes appear radially symmetric, each of the points representing the center of a possible signet ring cell;

selecting as candidate signet ring cells, ones of the points representing centers of possible signet ring cells that have an adjacent nucleus with a squashed shape by, for each of the points representing the center of a possible signet ring cell:

comparing the color of each pixels of the image to an expected color of a nucleus and classifying pixels having the expected color of the nucleus as nuclear pixels;

extracting any connected group of the nuclear pixels from the image;

determining whether the extracted connected group of nuclear pixels has a nucleus with the squashed shape on a periphery thereof; and

selecting that point as a center of a candidate signet ring cell; and

applying a convolutional neural network to the selected points representing the centers of the candidate signet ring cells to determine which of the candidate signet ring cells are signet ring cells.

8. The detector of claim 7 , wherein instructions for determining points around which the cell membranes appear radially symmetric include instructions for:

finding peaks of a Hough transform wherein the points representing the centers of possible signet ring cells are selected as the peaks of the transform achieving a given threshold.

9. The detector of claim 7 , wherein the selecting of that point as the center of the candidate signet ring cell is performed with a classifier.

10. The detector of claim 9 , wherein the classifier comprises a convolutional neural network.

11. The detector of claim 9 , wherein the classifier comprises a support vector machine.

12. The detector of claim 7 , wherein the convolutional neural network comprises a convolutional neural network trained on a supply of candidate points known to be signet ring cells, against a supply of candidate points that are not signet ring cells.