Method and apparatus that divides, clusters, classifies, and analyzes images of lesions using histograms and correlation coefficients

1. A similar image providing device, comprising:

a lesion region extracting unit that extracts a lesion region from a current diagnostic image by cropping the lesion region;

a local image feature extracting unit that divides the lesion region of the current diagnostic image into local images of a uniform size, and extracts local image features that characterize the respective local images;

a clustering unit that clusters sets of the local image features around respective central cluster values;

a lesion classifying unit that computes a histogram showing the numbers of each cluster set around the respective central cluster value and the values of the local image features;

a first storing unit that stores a plurality of stored diagnostic images, each stored diagnostic image having a diagnostic result that is determined in advance and a stored histogram computed in advance showing the stored cluster sets and the values of stored local image features of each stored diagnostic image;

the lesion classifying unit compares the stored histograms of the stored diagnostic images with the computed histogram of the current diagnostic image, and classifies the lesion region according to the comparison;

a second storing unit that stores stored correlation coefficients indicating probabilities of occurrence corresponding to stored topic variables and stored local image features, each stored topic variable expresses a degree of progression or a degree of seriousness of a specific lesion, the stored correlation coefficients are determined by using the stored diagnostic images;

an expected value estimating unit that acquires current correlation coefficients for the current diagnostic image based on the local images features of the current diagnostic image and the stored correlation coefficients from the second storing unit, and estimates expected values of the probabilities of occurrence of the respective topic variables of the current diagnostic image by maximizing the probabilities of occurrence of the local image features;

the first storing unit stores expected values of the probabilities of occurrence of the stored diagnostic image;

a comparing section that compares the expected values of the probability of occurrences of the respective topic variables of the current diagnostic image with the expected values of the probability of occurrences of the respective topic variables of the stored diagnostic images, and selects one of the stored diagnostic images that best approximates the estimated expected values of the probability of occurrences of the respective topic variables of the current diagnostic image.

2. The similar image providing device as claimed in claim 1 , wherein the topic variables are in a spectrum of low values to high values of the respective degrees of progression or degrees of seriousness, and the lesion region includes the plurality of topic variables.

3. The similar image providing device as claimed in claim 1 , wherein the probabilities of occurrence of the respective topic variables of the current diagnostic image is a polynomial distribution in which a Dirichlet distribution is a prior probability.

4. The similar image providing device as claimed in claim 1 , wherein the probabilities of occurrence corresponding to stored topic variables and stored local image features is a polynomial distribution in which a Dirichlet distribution is a prior probability.

5. The similar image providing device of claim 1 , wherein

the first storing unit further stores, in association with the stored diagnostic images, diagnostic results data that express diagnostic results of the stored diagnostic images, and

the comparing section further provides, together with the stored diagnostic images, information that is based on the diagnostic results data corresponding to the stored diagnostic images.

6. A non-transitory computer readable storage medium storing a computer program causing a computer to execute a similar image providing method, comprising:

extracting a lesion region from a current diagnostic image by cropping the lesion region;

dividing the lesion region of the current diagnostic image into local images of a uniform size, and extracts local image features that characterize the respective local images;

clustering sets of the local image features around respective central cluster values;

computing a histogram showing the numbers of each cluster set around the respective central cluster value and the values of the local image features;

storing a plurality of stored diagnostic images, each stored diagnostic image having a diagnostic result that is determined in advance and a stored histogram computed in advance showing the stored cluster sets and the values of stored local image features of each stored diagnostic image;

comparing the stored histograms of the stored diagnostic images with the computed histogram of the current diagnostic image, and classifies the lesion region according to the comparison;

storing stored correlation coefficients indicating probabilities of occurrence corresponding to stored topic variables and stored local image features, each stored topic variable expresses a degree of progression or a degree of seriousness of a specific lesion, the stored correlation coefficients are determined by using the stored diagnostic images;

acquiring current correlation coefficients for the current diagnostic image based on the local images features of the current diagnostic image and the stored correlation coefficients from the second storing unit, and estimates expected values of the probabilities of occurrence of the respective topic variables of the current diagnostic image by maximizing the probabilities of occurrence of the local image features;

storing expected values of the probabilities of occurrence of the stored diagnostic image;

comparing the expected values of the probability of occurrences of the respective topic variables of the current diagnostic image with the expected values of the probability of occurrences of the respective topic variables of the stored diagnostic images, and selects one of the stored diagnostic images that best approximates the estimated expected values of the probability of occurrences of the respective topic variables of the current diagnostic image.

7. A similar image providing method, comprising:

extracting a lesion region from a current diagnostic image by cropping the lesion region;

dividing the lesion region of the current diagnostic image into local images of a uniform size, and extracts local image features that characterize the respective local images;

clustering sets of the local image features around respective central cluster values;

computing a histogram showing the numbers of each cluster set around the respective central cluster value and the values of the local image features;

storing a plurality of stored diagnostic images, each stored diagnostic image having a diagnostic result that is determined in advance and a stored histogram computed in advance showing the stored cluster sets and the values of stored local image features of each stored diagnostic image;

comparing the stored histograms of the stored diagnostic images with the computed histogram of the current diagnostic image, and classifies the lesion region according to the comparison;

storing stored correlation coefficients indicating probabilities of occurrence corresponding to stored topic variables and stored local image features, each stored topic variable expresses a degree of progression or a degree of seriousness of a specific lesion, the stored correlation coefficients are determined by using the stored diagnostic images;

acquiring current correlation coefficients for the current diagnostic image based on the local images features of the current diagnostic image and the stored correlation coefficients from the second storing unit, and estimates expected values of the probabilities of occurrence of the respective topic variables of the current diagnostic image by maximizing the probabilities of occurrence of the local image features;

storing expected values of the probabilities of occurrence of the stored diagnostic image;

comparing the expected values of the probability of occurrences of the respective topic variables of the current diagnostic image with the expected values of the probability of occurrences of the respective topic variables of the stored diagnostic images, and selects one of the stored diagnostic images that best approximates the estimated expected values of the probability of occurrences of the respective topic variables of the current diagnostic image.