Computer-aided method for automated image feature analysis and diagnosis of digitized medical images

• US 6,011,862 A
• Filed: 06/17/1998
• Issued: 01/04/2000
• Est. Priority Date: 04/25/1995
• Status: Expired due to Term

• Current Assignee: ARCH Development Corporation (University of Chicago)
• Original Assignee: ARCH Development Corporation (University of Chicago)
• Inventors: Xu, Xin-Wei, Morishita, Junji, Katsuragawa, Shigehiko, Doi, Kunio
• Primary Examiner(s): Johns, Andrew W.
• Application Number:
  US09/098,504
• Time in Patent Office: 566 Days
• Field of Search: 382/128, 382/195, 382/199, 382/132, 382/156, 382/170, 358/458, 128/922
• US Class Current: 382/132
• CPC Class Codes: G06T 7/0012 Biomedical image inspection

Abstract

A computerized method for the detection and characterization of disease in an image derived from a chest radiograph, wherein an image in the chest radiograph is processed to determine the ribcage boundary, including lung top edges, right and left ribcage edges, and right and left hemidiaphragm edges. Texture measures including RMS variations of pixel values within regions of interest are converted to relative exposures and corrected for system noise existing in the system used to produce the image. Texture and/or geometric pattern indices are produced. A histogram(s) of the produced index (indices) is produced and values of the histograms) are applied as inputs to a trained artificial neural network, which classifies the image as normal or abnormal. In one embodiment, obviously normal and obviously abnormal images are determined based on the ratio of abnormal regions of interest to the total number of regions of interest in a rule-based method, so that only difficult cases to diagnose are applied to the artificial neural network.

Claims

1. In a computerized method for the detection and characterization of disease in an image derived from a chest radiograph, the improvement comprising:

converting RMS variations of pixel data values in a regions of interest to relative exposures based on the following relationship, ##EQU6## where RMS_p, G, C correspond to the RMS variation of pixel values in a respective region of interest, the gradient of the film used, and slope of the characteristic curve of the laser scanner, respectively, and log₁₀ e is a conversion factor from natural logarithm to logarithm to base 10.

2. The method of claim 1, further comprising:

correcting the converted RMS variations of pixel data values in inverse relationship to average distributions of RMS variations due to system noise existing in a system employed to convert the radiograph image into digital pixel data values.

3. In a computerized method for the detection and characterization of disease in an image derived from a chest radiograph, the improvement comprising:

a) determining a texture index based on at least one predetermined texture measure in ROIs in the image;

b) determining a histogram of the texture index;

c) applying values of said histogram, determined in the preceding step b), selected at predetermined upper areas of the histogram as inputs to a trained artificial neural network (ANN) having plural inputs and including a multi-layer, feed-forward network with a back-propagation algorithm, and

d) classifying said image as normal or abnormal based on an output of said ANN.

4. In a computerized method for the detection and characterization of disease in an image derived from a chest radiograph, the improvement comprising:

a) determining a geometric pattern index based on at least one predetermined geometric pattern measure in ROIs in the image;

b) determining a histogram of the geometric pattern index;

c) applying values of said histogram, determined in the preceding step b), selected at predetermined upper areas of the histogram as inputs to a trained artificial neural network (ANN) having plural inputs and including a multi-layer, feed-forward network with a back-propagation algorithm, and

d) classifying said image as normal or abnormal based on an output of said ANN.

5. In a computerized method for the detection and characterization of disease in an image derived from a chest radiograph, the improvement comprising:

a) determining a texture index based on at least one predetermined texture measure in ROIs in the image;

b) determining a geometric pattern index based on at least one predetermined geometric pattern measure in the ROIs in the image;

c) determining histograms of the texture index and the geometric pattern index;

d) applying values of said histograms, determined in the preceding step c), selected at predetermined upper areas of the histograms as inputs to a trained artificial neural network (ANN) having plural inputs and including a multi-layer, feed-forward network with a back-propagation algorithm, and

e) classifying said image as normal or abnormal based on an output of said ANN.

6. In a computerized method for the detection and characterization of disease in an image derived from a chest radiograph, the improvement comprising:

a) determining a texture index based on at least one texture measure in ROIs in the image;

b) determining the number of ROIs which contain a texture index greater than a predetermined threshold index thereby identifying potentially abnormal ROIs;

c) determining the ratio of the number of ROIs determined in the preceding step b) to total number of ROIs in the image, and if the determined ratio is less than a first predetermined ratio corresponding to the minimum ratio of abnormal ROIs detected in a training set of abnormal images, classifying the image as normal, and if the determined ratio is greater than a second predetermined ratio corresponding to the maximum ratio of abnormal ROIs detected in a training set of normal images, classifying the image as abnormal;

d) for those images which have ratios determined in said preceding step c) greater than said first predetermined ratio and less than said second predetermined ratio, determining a histogram of the texture index;

e) applying values of said histogram, determined in the preceding step d), selected at predetermined upper areas of the histogram as inputs to a trained artificial neural network (ANN) having plural inputs and including a multi-layer, feed-forward network with a back-propagation algorithm, and

f) classifying said image as normal or abnormal based on an output of said ANN.

7. In a computerized method for the detection and characterization of disease in an image derived from a chest radiograph, the improvement comprising:

a) determining a geometric pattern index based on at least one predetermined geometric pattern measure in ROIs in the image;

b) determining the number of ROIs which contain a geometric pattern index greater than a predetermined threshold index thereby identifying potentially abnormal ROIs;

c) determining the ratio of the number of ROIs determined in the preceding step b) to total number of ROIs determined in the preceding step b) to total number of ROIs in the image, and if the determined ratio is less than a first predetermined ratio corresponding to the minimum ratio of abnormal ROIs detected in a training set of abnormal images, classifying the image as normal, and if the determined ratio is greater than a second predetermined ratio corresponding to the maximum ratio of abnormal ROIs detected in a training set of normal images, classifying the image as abnormal;

d) for an image which has ratios determined in said preceding step c) greater than said first predetermined ratio and less than said second predetermined ratio, determining a histogram of the geometric pattern index;

e) applying values of said histogram, determined in the preceding step, d) selected at predetermined upper areas of the histogram as inputs to a trained artificial neural network (ANN) having plural inputs and including a multi-layer, feed-forward network with a back-propagation algorithm, and

f) classifying said image as normal or abnormal based on an output of said ANN.

8. In a computerized method for the detection and characterization of disease in an image derived from a chest radiograph, the improvement comprising:

a) determining a texture index based at on at least one texture measure in ROIs in the image;

b) determining a geometric pattern index based on at least one predetermined geometric pattern measure in said ROIs in said image;

c) determining the number of ROIs which contain a texture index greater than a predetermined threshold index thereby identifying potentially abnormal ROIs,

d) determining the number of ROIs which contain a geometric pattern greater than a predetermined threshold index thereby identifying potentially abnormal ROIs;

e) determining the ratio of the number of ROIs determined in the preceding step d) to total number of ROIs in the image, and if the determined ratio is less than a first predetermined ratio corresponding to the minimum ratio of abnormal ROIs detected in a training set of abnormal images, classifying the image as normal, and if the determined ratio is greater than a second predetermined ratio corresponding to the maximum ratio of abnormal ROIs detected in a training set of normal images, classifying the image as abnormal;

f) determining the ratio of the number of ROIs determined in the preceding step e) to total number of ROIs in the image, and if the determined ratio is less than a first predetermined ratio corresponding to the minimum ratio of abnormal ROIs detected in a training set of abnormal images, classifying the image as normal, and if the determined ratio is greater than a second predetermined ratio corresponding to the maximum ratio of abnormal ROIs detected in a training set of normal images, classifying the image as abnormal;

g) for an image which has ratios determined in at least one of said preceding steps e) and f) to be greater than said first predetermined ratio and less than said second predetermined ratio, determining a histogram of at least one of the texture index and the geometric pattern index;

h) applying values of said histogram, determined in the preceding step g) selected at predetermined upper areas of the histogram as inputs to a trained artificial neural (ANN) having plural inputs and including a multi-layer, feed-forward network with a back-propagation algorithm, and

i) classifying said image as normal or abnormal based on an output of said artificial neural network.

9. The method of claim 8, wherein:

in step g), histograms of both said texture index and said geometric pattern index are determined; and

in step h), values of both said histograms are applied as inputs to said ANN.