Granted Patent B2

US 7,627,154 · App. 11/285,568 · Granted Dec 1, 2009

Abstract

Claims (37)

1. A method for automatically classifying a digital radiographic image, comprising the following steps, carried out by a programmable digital computer, of:

acquiring the digital radiographic image comprised of a matrix of rows and columns of pixels;

segmenting the image into foreground, background, and anatomy regions;

classifying a physical size of the anatomy region using image capture information;

classifying a shape pattern of the anatomy region; and

categorizing the image based on the physical size classification and the shape pattern classification;

wherein the step of classifying a physical size comprises:

collecting image capture parameters

computing features from the foreground, background and anatomy regions;

training a classifier with the features and image capture parameters; and

performing the classification using the trained classifier; and

wherein said features and said image capture parameters include one or more of the following: a ratio of the anatomy's area to the image area, a ratio of the background area to the image area, and a ratio of the foreground area to the image area.

2. A method for automatically classifying a digital radiographic image, comprising the following steps, carried out by a programmable digital computer, of:

acquiring the digital radiographic image comprised of a matrix of rows and columns of pixels;

segmenting the image into foreground, background, and anatomy regions;

classifying a physical size of the anatomy region using image capture information;

classifying a shape pattern of the anatomy region; and

categorizing the image based on the physical size classification and the shape pattern classification:

wherein the step of classifying the shape pattern comprises the steps of;

extracting edges of the anatomy regions;

generating an edge direction histogram of the anatomy region; and

classifying the edge direction histogram using a scale and rotation invariant shape classifier; and

wherein the step of categorizing the image categorizes the images according to at least one of the following classes:

a large-size anatomy with no edge, such as the PA (Posterior-Anterior) view of abdomen, thorax spine and lumbar spine;

a large-size anatomy with a one-peak shape edge direction histogram, such as the PA view of hip and shoulder;

a large-size anatomy with a two-peak shape edge direction histogram, such as the lateral (LAT) view of chest and the PA view of pelvis;

a large-size anatomy with an other share edge direction histogram, such as the PA and LAT view of skull;

a small-size anatomy with no edge, such some PA and LAT view of knee because of the setting of collimator;

a small-size anatomy with a one-peak shape edge direction histogram, such as some PA and LAT view of femur, elbow, forearm and ankle, in which the collimation region covers parts anatomy and results in only one edge detected in the image;

a small-size anatomy with a two-peak shape edge direction histogram, such as most PA and LAT view of elbow, forearm, ankle and wrist;

a small-size anatomy with an other share edge direction histogram, such as the PA and LAT view of foot, hand and fingers.

3. The method of claim 2 , wherein the step of classifying the edge direction histogram normalizes the edge direction histogram to achieve scale invariance.

4. The method of claim 2 , wherein the step of classifying the edge direction histogram comprises the steps of:

training two or more classifiers to identify different shape patterns of the edge direction histogram to achieve rotation invariance;

performing the classification of the edge histogram of the input image using the trained classifiers; and

categorizing the shape pattern of the image by combining the classification results from different classifiers.

5. The method of claim 1 wherein said features and said image capture parameters include one or more of the following: pixel spacing, width and height of the anatomy region.

Assignments (17)