Pathological diagnosis support device, program, method, and system

• US 7,693,334 B2
• Filed: 11/30/2005
• Issued: 04/06/2010
• Est. Priority Date: 11/30/2004
• Status: Active Grant

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Ogura, Maki, Kiyuna, Tomoharu, Kamijo, Kenichi, Okajima, Kenji, Saitou, Akira
• Primary Examiner(s): Tucker; Wesley
• Assistant Examiner(s): CHU, RANDOLPH I
• Application Number:
  US11/290,294
• Publication Number:
  US 20060115146A1
• Time in Patent Office: 1,588 Days
• Field of Search: 382128-133, 382/143, 382/159, 382/224, 382190-208
• US Class Current: 382/190
• CPC Class Codes: G06F 18/211 Selection of the most significant subset of features
• CPC Class Codes: G06V 10/771 Feature selection, e.g. selecting representative features from a multi-dimensional feature space
• CPC Class Codes: G06V 20/69 Microscopic objects, e.g. biological cells or cellular parts
• CPC Class Codes: G16H 10/20 for electronic clinical trials or questionnaires

Abstract

A pathological diagnosis support device, a pathological diagnosis support program, a pathological diagnosis support method, and a pathological diagnosis support system extract a pathological tissue for diagnosis from a pathological image and diagnose the pathological tissue. A tissue collected in a pathological inspection is stained using, for example, hematoxylin and eosin. In consideration of the state of the tissue in which a cell nucleus and its peripheral constituent items are stained in respective colors unique thereto, subimages such as a cell nucleus, a pore, cytoplasm, interstitium are extracted from the pathological image, and color information of the cell nucleus is also extracted. The subimages and the color information are stored as feature candidates so that presence or absence of a tumor and benignity or malignity of the tumor are determined.

Claims

1. A pathological diagnosis support device, comprising:

learning pattern input means for obtaining images centered on a tumor from a pathological image and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, wherein;

the feature determining means calculates a feature of each of the learning patterns corresponding to each of the feature candidates and determines as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value; and

sequentially determines, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

the category table generator means calculates each feature of each of the learning patterns using the feature set and classifies the patterns using the category table including each feature of the learning patterns and the class information; and

the feature extracting means calculates each feature of the input patterns using the feature set,

wherein the learning pattern input means and the pattern input means select, from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belongs, calculate distance between a center of distribution of the color region and each pixel belonging to the color region, assign a signal to the pixel according to the distance, detect a peak of distribution of the signals in the pathological image, and input an image centered on the peak as the learning pattern.

2. The pathological diagnosis support device in accordance with claim 1, wherein the diagnosis means diagnoses the input patterns according a result of the diagnosis and the category table.

3. The pathological diagnosis support device in accordance with claim 1, wherein

the feature determining means prepares a predetermined number of sets for a transition of the learning pattern according to a value of the feature, calculates the feature of each of the learning patterns, determines the feature candidate, distributes the learning pattern with a weight according to the feature thus determined, sequentially causes a transition of the learning pattern to one of the sets corresponding to the feature, and sequentially determines the feature candidate under a condition that information about the sets respectively containing the learning patterns and the determined feature are known;

and

the diagnosis means causes a transition of each of the input patterns according to each feature of the input patterns and a transition table sequentially having recorded a set to which the learning pattern belongs at determination of each feature of the feature set and diagnoses the input patterns according a set to which the input pattern belongs as a result of the transition.

4. The pathological diagnosis support device in accordance with claim 1, wherein

the feature determining means prepares a predetermined number of sets for a transition of the learning pattern according to a value of the feature, calculates the feature of each of the learning patterns, determines the feature candidate, distributes the learning pattern with a weight according to the feature thus determined, sequentially causes a transition of the learning pattern to one of the sets corresponding to the feature, and sequentially determines the feature candidate under a condition that information about the sets respectively containing the learning patterns and the determined feature are known;

the feature extracting means calculates the feature that indicates a probability with which the feature at an order takes a predetermined value; and

the diagnosis means calculates, according to each feature of the input patterns and a transition table sequentially having recorded a set to which the learning pattern belongs at determination of each feature of the feature set, a probability with which the input pattern includes predetermined class information and then conducts the diagnosis.

5. The pathological diagnosis support device in accordance with claim 1, wherein the feature candidates generated by the feature generator means includes a feature candidate obtained using a complex Gabor function as a feature extraction function.

6. The pathological diagnosis support device in accordance with claim 1, wherein the feature candidates generated by the feature generator means includes a feature candidate obtained using a feature extraction function obtained by normalizing a complex Gabor function.

9. The pathological diagnosis support device in accordance with claim 5, wherein the feature determining means conducts operation for each of the learning patterns using a predetermined noise parameter for each of the feature candidates.

10. The pathological diagnosis support device in accordance with claim 5, wherein the feature determining means calculates, as the feature of each of the learning patterns corresponding to each of the feature candidates, a probability with which the feature of the learning pattern takes a predetermined value.

11. The pathological diagnosis support device in accordance with claim 1, wherein when the learning patterns can be classified irrespective of the values of the features, the category table generator means substitutes a redundant term for the value of the feature at an associated position of the category table.

12. The pathological diagnosis support device in accordance with claim 1, wherein:

each of the features of the input patterns is a value of a probability with which the feature at an order takes a predetermined value; and

the diagnosis means calculates, by use of the features, a probability with which each of the feature patterns contained in the category table takes a predetermined value of class information, the probability being used for decision.

37. A pathological diagnosis support system, comprising:

an information processing terminal for keeping pathological image data including a pathological image and information unique to a patient attached to the image; and

a pathological diagnosis server for diagnosing the pathological image data, the server comprising;

a pathological diagnosis support device according to one of claims 1 to 4, 6 to 8, 10 to l4 for diagnosing the pathological image contained in the pathological image data; and

diagnosis result storage means for storing a diagnosis result from the pathological diagnosis support device together with the information unique to the patient, wherein

the information processing terminal requests the pathological diagnosis server to transmit to the terminal the diagnosis result together with the information unique to the patient, and

the pathological diagnosis server compares the information unique to the patient received from the terminal with the information unique to the patient stored together with the diagnosis result and then transmits the diagnosis result to the terminal if the information unique to the patient received from the terminal matches the information unique to the patient stored together with the diagnosis result.

38. The pathological diagnosis support system in accordance with claim 37, further comprising an accounting server for keeping amounts of use charge respectively of the pathological diagnosis support device and the information processing terminal.

39. The pathological diagnosis support system in accordance with claim 38, wherein when the diagnosis result storage means stores the diagnosis result, the accounting server accumulates an amount of use charge of the pathological diagnosis support system.

40. The pathological diagnosis support system in accordance with claim 38, wherein when the information processing terminal receives the diagnosis result, the accounting server accumulates an amount of use charge of the information processing terminal.

7. A pathological diagnosis support device comprising:

learning pattern input means for obtaining images centered on a tumor from a pathological image and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, wherein;

the feature determining means calculates a feature of each of the learning patterns corresponding to each of the feature candidates and determines as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value; and

sequentially determines, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

the category table generator means calculates each feature of each of the learning patterns using the feature set and classifies the patterns using the category table including each feature of the learning patterns and the class information; and

the feature extracting means calculates each feature of the input patterns using the feature set,

wherein the feature candidates generated by the feature generator means includes a feature function discriminating a color of the tumor, and wherein the feature determining means compares the signal of each pixel included in the learning patterns calculated by the learning pattern input means with a predetermined threshold value; and

wherein the learning pattern input means and the pattern input means select, from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belongs, calculate distance between a center of distribution of the color region and each pixel belonging to the color region, assign a signal to the pixel according to the distance, detect a peak of distribution of the signals in the pathological image, and input an image centered on the peak as the learning pattern.

8. A pathological diagnosis support device comprising:

learning pattern input means for obtaining images centered on a tumor from a pathological image and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, wherein;

the feature determining means calculates a feature of each of the learning patterns corresponding to each of the feature candidates and determines as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value; and

sequentially determines, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

the category table generator means calculates each feature of each of the learning patterns using the feature set and classifies the patterns using the category table including each feature of the learning patterns and the class information; and

the feature extracting means calculates each feature of the input patterns using the feature set,

wherein the feature candidates generated by the feature generator means includes a feature function discriminating a color of the tumor, and wherein the feature determining means compares the signal of each pixel included in the learning patterns calculated by the learning pattern input means with a mean value of signals of pixels in the proximity of the pixel; and

wherein the learning pattern input means and the pattern input means select, from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belongs, calculate distance between a center of distribution of the color region and each pixel belonging to the color region, assign a signal to the pixel according to the distance, detect a peak of distribution of the signals in the pathological image, and input an image centered on the peak as the learning pattern.

13. A pathological diagnosis support program on computer readable medium for use with a pathological diagnosis support device comprising learning pattern input means for obtaining from a pathological image, images centered on a tumor and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, said program comprising instructions that cause the pathological diagnosis support device to perform steps of;

calculating, using the feature determining means, a feature of each of the learning patterns corresponding to each of the feature candidates and determining as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value, and sequentially determining, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

calculating, using the category table generator means, each feature of each of the learning patterns using the feature set and classifying the patterns using the category table including each feature of the learning patterns and the class information;

calculating, using the feature extracting means, each feature of the input patterns using the feature set;

processing to select, from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belongs;

processing to calculate distance between a center of distribution of the color region and each pixel belonging to the color region;

processing to assign a signal to the pixel according to the distance;

processing to detect a peak of distribution of the signals in the pathological image; and

processing to input an image centered on the peak as the learning pattern.

14. The pathological diagnosis support program in accordance with claim 13, wherein the program further executes processing in which the diagnosis means diagnoses the input patterns according a result of the diagnosis and the category table.

15. The pathological diagnosis support program in accordance with claim 13, wherein executed are

processing in which the feature determining means prepares a predetermined number of sets for a transition of the learning pattern according to a value of the feature, calculates the feature of each of the learning patterns, determines the feature candidate, distributes the learning pattern with a weight according to the feature thus determined, sequentially causes a transition of the learning pattern to one of the sets corresponding to the feature, and sequentially determines the feature candidate under a condition that information about the sets respectively containing the learning patterns and the determined feature are known;

and

processing in which the diagnosis means causes a transition of each of the input patterns according to each feature of the input patterns and a transition table sequentially having recorded a set to which the learning pattern belongs at determination of each feature of the feature set and diagnoses the input patterns according a set to which the input pattern belongs as a result of the transition.

16. The pathological diagnosis support program in accordance with claim 13, wherein executed are

processing in which the feature determining means prepares a predetermined number of sets for a transition of the learning pattern according to a value of the feature, calculates the feature of each of the learning patterns, determines the feature candidate, distributes the learning pattern with a weight according to the feature thus determined, sequentially causes a transition of the learning pattern to one of the sets corresponding to the feature, and sequentially determines the feature candidate under a condition that information about the sets respectively containing the learning patterns and the determined feature are known;

processing in which the feature extracting means calculates the feature that indicates a probability with which the feature at an order takes a predetermined value; and

processing in which the diagnosis means calculates, according to each feature of the input patterns and a transition table sequentially having recorded a set to which the learning pattern belongs at determination of each feature of the feature set, a probability with which the input pattern includes predetermined class information and then conducts the diagnosis.

17. The pathological diagnosis support program in accordance with claim 13, wherein the feature candidates generated by the feature generator means includes a feature candidate obtained using a complex Gabor function as a feature extraction function.

18. The pathological diagnosis support program in accordance with claim 13, wherein the feature candidates generated by the feature generator means includes a feature candidate obtained using a feature extraction function obtained by normalizing a complex Gabor function.

21. The pathological diagnosis support program in accordance with claim 17, further comprising processing in which the feature determining means conducts operation for each of the learning patterns using a predetermined noise parameter for each of the feature candidates.

22. The pathological diagnosis support program in accordance with claim 17, further comprising processing in which the feature determining means calculates, as the feature of each of the learning patterns corresponding to each of the feature candidates, a probability with which the feature of the learning pattern takes a predetermined value.

23. The pathological diagnosis support program in accordance with claim 13, further comprising processing in which when the learning patterns can be classified irrespectively of the values of the features, the category table generator means substitutes a redundant term for the value of the feature at an associated position of the category table.

24. The pathological diagnosis support program in accordance with claim 13, wherein:

each of the features of the input patterns is a value of a probability with which the feature at an order takes a predetermined value, the program further comprising processing in which the diagnosis means calculates, by use of the features, a probability with which each of the feature patterns contained in the category table takes a predetermined value of class information, the probability being used for decision.

19. A pathological diagnosis support program, on computer readable medium for use with a pathological diagnosis support device comprising learning pattern input means for obtaining from a pathological image, images centered on a tumor and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, said program comprising instructions that cause the pathological diagnosis support device to perform steps of;

calculating, using the feature determining means, a feature of each of the learning patterns corresponding to each of the feature candidates and determining as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value, and sequentially determining, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

calculating, using the category table generator means, each feature of each of the learning patterns using the feature set and classifying the patterns using the category table including each feature of the learning patterns and the class information;

calculating, using the feature extracting means, each feature of the input patterns using the feature set;

comparing, using the feature determining means, the signal of each pixel included in the learning patterns calculated by the learning pattern input means with a predetermined threshold value;

selecting, from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belongs;

calculating distance between a center of distribution of the color region and each pixel belonging to the color region;

assigning a signal to the pixel according to the distance;

detecting a peak of distribution of the signals in the pathological image; and

inputting an image centered on the peak as the learning pattern,

wherein the feature candidates generated by the feature generator means includes a feature function discriminating a color of the tumor.

20. A pathological diagnosis support program on computer readable medium for use with a pathological diagnosis support device comprising learning pattern input means for obtaining from a pathological image, images centered on a tumor and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, said program comprising instructions that cause the pathological diagnosis support device to perform steps of;

calculating, using the feature determining means, a feature of each of the learning patterns corresponding to each of the feature candidates and determining as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value, and sequentially determining, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

calculating, using the category table generator means, each feature of each of the learning patterns using the feature set and classifying the patterns using the category table including each feature of the learning patterns and the class information;

calculating, using the feature extracting means, each feature of the input patterns using the feature set;

comparing, using the feature determining means, the signal of each pixel included in the learning patterns calculated by the learning pattern input means with a mean value of signals of pixels in the proximity of the pixel;

selecting, from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belong;

calculating distance between a center of distribution of the color region and each pixel belonging to the color region;

assigning a signal to the pixel according to the distance;

detecting a peak of distribution of the signals in the pathological image; and

inputting an image centered on the peak as the learning pattern,

wherein the feature candidates generated by the feature generating means includes a feature function discriminating a color of the tumor.

25. A pathological diagnosis support method for use with a pathological diagnosis support device comprising learning pattern input means for obtaining from a pathological image to be used for learning, images centered on a tumor and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, the method comprising;

using a processors to perform steps of

a first step in which the feature determining means calculates a feature of each of the learning patterns corresponding to each of the feature candidates and determines as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value; and

sequentially determines, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

a second step in which the category table generator means calculates each feature of each of the learning patterns using the feature set and classifies the patterns using the category table including each feature of the learning patterns and the class information;

a third step in which the feature extracting means calculates each feature of the input patterns using the feature set; and

steps to be executed by the learning pattern input means and the pattern input means, the steps including;

a step of selecting, from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belongs;

a step of calculating distance between a center of distribution of the color region and each pixel belonging to the color region;

a step of assigning a signal to the pixel according to the distance;

a step of detecting a peak of distribution of the signals in the pathological image; and

a step of inputting an image centered on the peak as the learning pattern.

26. The pathological diagnosis support method in accordance with claim 25, wherein the method further comprises a fourth step in which the diagnosis means diagnoses the input patterns according a result of the diagnosis and the category table.

27. The pathological diagnosis support method in accordance with claim 25 wherein

in the first step the feature determining means prepares a predetermined number of sets for a transition of the learning pattern according to a value of the feature, calculates the feature of each of the learning patterns, determines the feature candidate, distributes the learning pattern with a weight according to the feature thus determined, sequentially causes a transition of the learning pattern to one of the sets corresponding to the feature, and sequentially determines the feature candidate under a condition that information about the sets respectively containing the learning patterns and the determined feature are known, the method further comprising

a fourth step in which the diagnosis means causes a transition of each of the input patterns according to each feature of the input patterns and a transition table sequentially having recorded a set to which the learning pattern belongs at determination of each feature of the feature set and diagnoses the input patterns according a set to which the input pattern belongs as a result of the transition.

28. The pathological diagnosis support method in accordance with claim 25, wherein

in the first step the feature determining means prepares a predetermined number of sets for a transition of the learning pattern according to a value of the feature, calculates the feature of each of the learning patterns, determines the feature candidate, distributes the learning pattern with a weight according to the feature thus determined, sequentially causes a transition of the learning pattern to one of the sets corresponding to the feature, and sequentially determines the feature candidate under a condition that information about the sets respectively containing the learning patterns and the determined feature are known, and

in the third step the feature extracting means calculates the feature that indicates a probability with which the feature at an order takes a predetermined value, the method further comprising

a fourth step in which the diagnosis means calculates, according to each feature of the input patterns and a transition table sequentially having recorded a set to which the learning pattern belongs at determination of each feature of the feature set, a probability with which the input pattern includes predetermined class information and then conducts the diagnosis.

29. The pathological diagnosis support method in accordance with claim 25, wherein the feature candidates generated by the feature generator means includes a feature candidate obtained using a complex (labor function as a feature extraction function.

30. The pathological diagnosis support method in accordance with claim 25, wherein the feature candidates generated by the feature generator means includes a feature candidate obtained using a feature extraction function obtained by normalizing a complex Gabor function.

33. The pathological diagnosis support method in accordance with claim 29, further comprising a step in which the feature determining means conducts operation for each of the learning patterns using a predetermined noise parameter for each of the feature candidates.

34. The pathological diagnosis support method in accordance with claim 29, further comprising a step in which the feature determining means calculates, as the feature of each of the learning patterns corresponding to each of the feature candidates, a probability with which the feature of the learning pattern takes a predetermined value.

35. The pathological diagnosis support method in accordance with claim 25, further comprising a step in which when the learning patterns can be classified irrespectively of the values of the features, the category table generator means substitutes a redundant term for the value of the feature at an associated position of the category table.

36. The pathological diagnosis support method in accordance with claim 25, wherein:

each of the features of the input patterns is a value of a probability with which the feature at an order takes a predetermined value, the method further comprising a step in which the diagnosis means calculates, by use of the features, a probability with which each of the feature patterns contained in the category table takes a predetermined value of class information, the probability being used for decision.

31. A pathological diagnosis support method comprising learning pattern input means for obtaining from a pathological image to be used for learning, images centered on a tumor and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, the method comprising;

using a processors to perform steps of

a first step in which the feature determining means calculates a feature of each of the learning patterns corresponding to each of the feature candidates and determines as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value; and

sequentially determines, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

a second step in which the category table generator means calculates each feature of each of the learning patterns using the feature set and classifies the patterns using the category table including each feature of the learning patterns and the class information;

a third step in which the feature extra6ting means calculates each feature of the input patterns using the feature set;

a fourth step in which the feature determining means compares the signal of each pixel included in the learning patterns calculated by the learning pattern input means with a predetermined threshold value;

a step of selecting from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belongs;

a step of calculating distance between a center of distribution of the color region and each pixel belonging to the color region;

a step of assigning a signal to the pixel according to the distance;

a step of detecting a peak of distribution of the signals in the pathological image; and

a step of inputting an image centered on the peak as the learning pattern,

wherein the feature candidates generated by the feature generator means includes a feature function discriminating a color of the tumor.

32. A pathological diagnosis support method, comprising learning pattern input means for obtaining from a pathological image to be used for learning, images centered on a tumor and inputting thereto the images as learning patterns;

learning pattern storage means for storing and keeping the learning patterns to which class information is attached;

feature candidate generator means for generating a plurality of feature candidates;

feature determining means for determining a feature set of features suitable for diagnosis using the feature candidates generated by the feature candidate generator means;

feature storage means for storing and keeping the set of features determined by the feature determining means;

category table generator means for generating a category table;

pattern input means for obtaining, from a pathological image to be diagnosed, images centered on a tumor candidate and inputting the images as input patterns;

feature extracting means for extracting features from the input patterns; and

diagnosis means for conducting diagnosis using the features, the method comprising;

using a processors to perform steps of

a first step in which the feature determining means calculates a feature of each of the learning patterns corresponding to each of the feature candidates and determines as a first feature of the feature set, a feature candidate for which a mutual information quantity with respect to the class information of a set of the learning patterns takes a maximum value; and

sequentially determines, under a condition that the determined feature is known, as a subsequent feature of the feature set, a feature candidate for which mutual information quantity between a feature of each learning pattern corresponding to each feature candidate and the class information of an associated one of the learning patterns takes a maximum value;

a second step in which the category table generator means calculates each feature of each of the learning patterns using the feature set and classifies the patterns using the category table including each feature of the learning patterns and the class information;

a third step in which the feature extracting means calculates each feature of the input patterns using the feature set;

a fourth step in which the feature determining means compares the signal of each pixel included in the learning patterns calculated by the learning pattern input means with a mean value of signals of pixels in the proximity of the pixel;

a step of selecting, from R, G, and B values of each pixel in the pathological image stained in advance, pixels belonging to a color region to which a cell nucleus of a predetermined tumor belongs;

a step of calculating distance between a center of distribution of the color region and each pixel belonging to the color region;

a step of assigning a signal to the pixel according to the distance;

a step of detecting a peak of distribution of the signals in the pathological image; and

a step of inputting an image centered on the peak as the learning pattern,

wherein the feature candidates generated by the feature generator means includes a feature function discriminating a color of the tumor.