Similarity measurement method for the classification of medical images into predetermined categories

• US 6,631,204 B1
• Filed: 02/04/2000
• Issued: 10/07/2003
• Est. Priority Date: 02/05/1999
• Status: Expired due to Fees

• Current Assignee: Yissum Research Development Company of the Hebrew University of Jerusalem Ltd. (Hebrew University of Jerusalem)
• Original Assignee: Yissum Research Development Company of the Hebrew University of Jerusalem Ltd. (Hebrew University of Jerusalem)
• Inventors: Smith, Yoav
• Primary Examiner(s): Patel, Jayanti K.
• Assistant Examiner(s): Choobin, Barry
• Application Number:
  US09/497,998
• Time in Patent Office: 1,341 Days
• Field of Search: 382/130, 382/192, 382/224, 382/145, 382/158, 382/133, 382/305, 382/240, 382/128, 128/922, 600/476, 600/459, 600/443, 600/410, 600/300, 378/21, 378/54, 359/548, 356/39, 345/418, 707/7, 707/100, 702/182
• US Class Current: 382/130
• CPC Class Codes: G06V 10/42 Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
• CPC Class Codes: Y10S 128/922 including image analysis

Abstract

A similarity measurement method for the classification of medical images into predetermined categories. A small set of pre-classified images is required to employ the method. The images can be real world images acquired using a camera, computer tomography, etc., or schematic drawings representing samples of different classes. The use of schematic drawings as a source of images allows a quick test of the method for a particular classification problem. The eigenvectors for each category are mathematically derived, and each image in each category is represented as a weighted linear combination of the eigenvectors. A test image is provided and projected onto the eigenvectors of each of the categories so as to reconstruct the test image with the eigenvectors. The RMS (root-mean-square) distance between the test image and each of the categories is measured. The smallest similarity measurement distance is selected, and the test image is classified in accordance with the selected smallest similarity measurement distance.

Claims

1. A similarity measurement method for the classification of medical images into predetermined categories, said method comprising the steps of:

providing a pre-classified library of images divided into at least two subgroups, each of said subgroups having a medical diagnostic meaning;

mathematically deriving a basic set of eigenvectors having coefficients that describe each of said subgroups, where each image in said subgroup is represented as a weighted linear combination of said basic set of eigenvector coefficients;

providing a test image;

projecting said test image onto said basic set of eigenvectors associated with each of said subgroups, to obtain a set of projection coefficients of said test image for each of said subgroups;

reconstructing, for each of said subgroups, said test image with said set of projection coefficients and said basic set of eigenvectors;

measuring the RMS (root-mean-square) distance between said reconstructed test image in each of said subgroups and said test image, said measured distance representing similarity;

selecting, from among said similarity measurement distances, the smallest similarity measurement distance; and

classifying said test image in accordance with said selected smallest similarity measurement distance.

2. The method of claim 1 wherein said selecting step comprises the steps of:

comparing eh of said similarity measurement distances with a predetermined threshold value; and

rejecting said test image if said threshold value is exceeded, thereby preserving the integrity of said library of images.

3. The method of claim 1 further comprising the step of adding said classified test image to said library of images.

4. The method of claim 3 wherein repeated performance of said step of adding said classified test image increases the probability of matching an existing image in said library with said test image.

5. The method of claim 1 wherein said library image providing step is performed by a communications system.

6. The method of claim 1 wherein said test image providing step is performed by a communications system.

7. The method of claim 6 wherein said communications system has a client-server architecture, wherein said pre-classified library of images is contained on the server and said method steps are performed in an algorithm which runs on a computer located at the server locations and said test image is provided by the client.

8. The method of claim 7 wherein said projecting, reconstructing, measuring, selecting and classifying steps are all performed automatically in an automatic diagnostic system, thereby saving medical diagnostic time.

9. The method of claim 1 wherein said projecting, reconstructing, measuring, selecting and classifying steps are all performed automatically in an automatic diagnostic system, thereby saving medical diagnostic time.

10. The method of claim 1 applied to diagnostic testing, wherein repeated performance of said step of adding said classified test image increases the accuracy of a diagnostic test performed in accordance with said method.

11. The method of claim 10 wherein said diagnostic test is for adenocarcinomas.

12. The method of claim 10 wherein said diagnostic test is based on brain PET scans for classification of psychiatric mental states.

13. The method of claim 10 wherein said diagnostic test is for predicting bone osteoporosis.

14. The method of claim 10 wherein said diagnostic test is for thallium images in cardiology.

15. The method of claim 10 wherein said diagnostic test is for use in spine surgery for predicting surgical results of spine curvature surgery.

16. The method of claim 1 wherein said test images are hand sketched illustrations provided for pre-classification for immediate testing of said method, before employing it to an actual problem.

17. A similarity measurement system for the classification of medical images into predetermined categories, said system comprising.

a pre-classified library of images divided into at least two subgroups, each of said subgroups having a medical diagnostic meaning;

means for mathematically deriving a basic set of eigenvectors having coefficients that describe each of said subgroups, where each image in said subgroup is represented as a weighted linear combination of said basic set of eigenvector coefficients;

image acquisition means for providing a test image;

means for projecting said test image onto said basic set of eigenvectors associated with each of said subgroups, to obtain a set of projection coefficients of said test image for each of said subgroups;

means for reconstructing, for each of said subgroups, said test image with said set of projection coefficients and said basic set of eigenvectors;

means for measuring the RMS (root-mean-square) distance between said reconstructed test image in each of said subgroups and said test image, said measured distance representing similarity;

means for selecting, from among said similarity measurement distances, the smallest similarity measurement distance; and

means for classifying said test image in accordance with said selected smallest similarity measurement distance; and

means for displaying said test image classification for diagnostic purposes.

18. The system of claim 17 wherein said library of images is stored in a computer.

19. The system of claim 17 wherein said stored library of images is arranged for dynamic growth based on addition of new test images after classification.

20. The system of claim 17 wherein said mathematical derivation means, said projecting means, said reconstructing means, said measuring means, said selecting means, and said classifying means are all implemented in a computer.