Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/70—Determining position or orientation of objects or cameras
  • G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V10/00—Arrangements for image or video recognition or understanding
  • G06V10/20—Image preprocessing
  • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/30—Subject of image; Context of image processing
  • G06T2207/30004—Biomedical image processing
  • G06T2207/30008—Bone
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
  • G06V2201/03—Recognition of patterns in medical or anatomical images

Definitions

• This invention relates generally to techniques for processing radiographs, and more particularly to techniques for automatically recognizing the projection view of radiographs.
• the importance of recognizing the projection view of radiographs includes the following two aspects. Firstly, it can help automate the image rendering procedure and optimize the image display quality. According to the workflow of a Computer Radiograph (CR) system, a technologist takes radiographs ordered in an examination, and then scans each CR cassettes while manually typing in the projection view associated with the cassette. This projection view information, together with the body part information which is obtained when the examination is ordered, determine the characteristics of the radiograph and directly influence the choice of image rendering parameters. Therefore, the success of recognizing the projection view of radiograph can help eliminate the need of the radiologist input, automate the image rendering process, and expedite the workflow. Secondly, projection view recognition can also benefit image management in Picture Archiving and Communication Systems
• RSNA 2002 Computerized Method for Automated Classification of Body Parts in Digital Radiographs
• the method examines the similarity of a given image to a set of pre-determined template images by using the cross-correlation values as the similarity measure.
• the manual generation of these template images is quite time consuming, and more particularly, it is highly observer dependent, which may introduce error into the classification.
• GuId et. al. (“Comparison of Global Features for Categorization of Medical Images", SPIE medical Imaging 2004) discloses a method to evaluate a set of global features extracted from images for classification. In both methods, no preprocessing is implemented to reduce the influence of irrelevant and often distracting data.
• the unexposed regions caused by the blocking of the x-ray collimator during the exposure may result in a significant white borders surrounding the image.
• Applicants have noted that if such regions are not removed in a pre-processing step and therefore used in the computation of similarity measures, the classification results can be seriously biased.
• Recent literature focuses on natural scene image classification. Examples include QBIC (W. Niblack, et al, "The QBIC project: Querying images by content using color, texture, and shape” Proc. SPIE Storage and Retrieval for Image and Video Databases, Feb 1994), Photobook (A.Pentland, et. al.
• the object of the present invention is to provide an automated method for recognizing the projection view of radiographs.
• this objective is achieved by the following steps: correcting the orientation of the radiograph, extracting a region of interest from the radiograph, and recognizing the projection view of radiograph.
• pre-processing the input radiograph can be accomplished.
• Preprocessing an input radiograph comprises sub-sampling the original image, segmenting the image into foreground, background and anatomy, and normalizing the image intensity based on characteristics of the anatomy.
• Correcting the orientation of radiograph comprises detecting the orientation of the radiograph and reorienting the radiograph accordingly.
• Extracting a region of interest from the radiograph includes detecting the medial axis of the anatomy, determining the center, size and shape of a region of interest and locating the region of interest in the radiograph.
• Recognizing the projection view of radiograph is accomplished by classifying the radiograph with respect to all possible views and combining the classification results to determine the most likely projection view of radiograph.
• the present invention provides some advantages. For example, features of the method promote robustness. Preprocessing of radiographs helps avoid the interference from the collimation areas and other noise. In addition, features used for orientation classification are invariant to size, translation and rotation. Features of the method also promote efficiency. For example, the processes can be implemented on a sub-sampled coarse resolution image, which greatly speeds up the recognition process.
• FIGS. IA 5 IB and 1C show flow charts illustrating the automated method for recognizing the projection view of radiographs.
• FIG. 2 is a flow chart illustrating the preprocessing step.
• FIGS. 3A-3E illustrate diagrammatic views showing the results from the preprocessing step.
• FIG. 3 A displays the original image of a radiograph.
• FIGS. 3B-3D depict its corresponding foreground, background and anatomy images from the segmentation, respectively.
• FIG. 3E displays the normalized image to emphasize the anatomy.
• FIGS. 4A-4C show diagrammatic views illustrating the detection of the medial axis of an anatomy.
• FIG. 4A shows the original image.
• FIG. 4B is the Euclidean distance map calculated from the anatomy image.
• FIG. 4C displays the medial axis detected from the anatomy image.
• FIGS. 5A-5B show diagrammatic views illustrating the shape of the region of interest (ROI).
• FIG. 5 A depicts the region of interest extracted from the cervical spine radiograph.
• FIG. 5B shows the region of interest found in the hand radiograph.
• the white dot represents the center of ROI.
• FIG. 6 shows a flow chart illustrating the classification of radiograph with respect to all possible views. DETAILED DESCRIPTION OF THE INVENTION
• the present invention discloses a method for automatically recognizing the projection view of radiographs.
• a flow chart of a method in accordance with the present invention is shown in Figure IA.
• the method includes three stages: correcting the orientation of radiograph (step 11); extracting a Region of Interest (ROI) from the input radiograph(step 12); and recognizing the projection view of the radiograph (step 13).
• ROI Region of Interest
• an additional step can be applied prior to the image orientation correction.
• This additional step (step 14) is the preprocessing of radiographs, and will be more particularly described below.
• the step of recognizing the projection view of radiograph can include two steps: 1) classifying the radiograph using a set of features and pre-trained classifiers, each classifier trained to recognize one projection view (step 15); 2) determining the projection view of radiographs by combining the classification results (step 16).
• step 14 the step for preprocessing radiographs.
• the purpose of preprocessing includes three aspects: (1) minimizing the number of pixels that need to be processed, but without degrading the performance of recognition; (2) reducing the interference from collimation areas (foreground) and direct exposure areas (background), so that the orientation recognition is driven by the diagnostically useful part of image data (anatomy); and (3) generating a consistent intensity and contrast image for the subsequent processes.
• the method step for preprocessing radiographs starts with sub-sampling the original image to a small-size coarse resolution image (step 20). The sub-sampled image is then segmented into foreground, background, and anatomy regions (step 21).
• Step 22 the foreground and background regions are removed from the image (step 22), and only the anatomy region is kept for further processing. Finally, the result image is normalized based on the intensity range of the anatomy region (step 23).
• Sub-sampling the original image (step 20) can be performed by known methods used for reducing image size while preserving enough information for orientation recognition.
• a Guassian pyramid data structure is employed to generate sub-sampled images
• Image segmentation (step 21) can be accomplished by using methods known to those skilled in the art. One such segmentation method is to find two thresholds from the image histogram, then segment the image into foreground, background and anatomy regions.
• Figure 3A shows an exemplary radiograph and Figures 3B-3D show its foreground, background and anatomy images obtained from segmentation.
• the foreground and background regions are removed from the sub-sampled image (step 22). This can be accomplished by setting the pixels in these regions to a pre-defined value, with the pixels in the remaining anatomy region kept unchanged.
• the preprocessed image contains the diagnostically useful part of image data, therefore the interference from coUimation areas can be minimized and the intensity range of anatomy region can be accurately detected.
• image intensity normalization (step 23) is performed over the image in order to compensate for difference in exposure densities caused by patient variations and examination conditions.
• One technique to achieve normalization is to detect minimum and maximum brightness values from the image histogram, preferably computed from pixels in the anatomy region, then apply a linear or log transfer function to adjust the image brightness into a pre-defined range. Histogram equalization could be further performed on the image to spread out those peaks in the image histogram, so that more details in low-contrast regions in the image can be better shown.
• Figure 3E displays a resulting image after intensity normalization by using this method.
• other known techniques such as the tone scale method disclosed in U.S. Patent No. 5,633,511 issued on 1997 by Lee et al.
• the orientation correction (step 11) of a radiograph comprises detecting the orientation of the radiograph and reorienting it into the position preferred by radiologists.
• the orientation detection can be accomplished using methods known to those skilled in the art.
• One suitable method is disclosed in commonly assigned U.S. Serial No. 10/993,055, entitled “DETECTION AND CORRECTION METHOD FOR RADIOGRAPH ORIENTATION", filed on November 19, 2004 in the names of Luo et al, and which is assigned to the assignee of this application, and incorporated herein by reference. Due to the examination condition, the size and position, as well as orientation of anatomy from the same examination would be varying, hi addition, the portion of anatomy shown in the radiograph is also varied depending on the patient's situation and the setting of collimation blades. These factors may result in the different appearances of radiographs, which pose challenges to the orientation recognition.
• a Region of Interest is extracted from the radiograph.
• the ROI is intended to capture the diagnostically useful part from image data, and minimize the distraction and interference caused by the factors mentioned above.
• the projection view recognition can focus on the diagnostically important region.
• the ROI extraction method includes two steps: detecting a medial axis of anatomy in the radiograph and locating the ROI accordingly.
• the medial axis is used to describe the anatomy in radiographs. Using the medial axis is attractive in that it provides a simple description of position and orientation of anatomy in radiographs, and greatly helps limit search complexity and expedites processing.
• the medial axis can be detected by using the Euclidean distance map.
• the contour of anatomy is detected and used to calculate Euclidean distance map as shown in Figure 4B for the image shown in Figure 4A.
• the maximum ridge is detected and used as the medial axis.
• Figure 4C depicts a resultant medial axis. This method is particularly suited for the radiographs of extremities, such as elbow, knee, and wrist, which tend to have well-defined medial axes, even though the shapes can be complex.
• the medial axis can be detected by Multiscale Medial Analysis ( Morse et. al. "Multiscale Medial Analysis of Medical Images", Image and Vision Computing, VoI 12 No.6, 1994).
• Multiscale Medial Analysis Morse et. al. "Multiscale Medial Analysis of Medical Images", Image and Vision Computing, VoI 12 No.6, 1994.
• An advantage of using MMA is that it works directly on image intensities, and does not require a prior segmentation of the image or explicit determination of object boundaries.
• ROI extraction starts with searching for the center of ROI, followed by determining the size and shape of ROI based on the features of the anatomy.
• the center of ROI is dependent on the characteristics of the anatomy in the examination.
• the center of ROI is located at the center of neck, as shown in Figure 5 A.
• the center of palm can be used as the center of ROI as shown in Figure 5B.
• the size of ROI is related to the size of the anatomy in radiographs, which can be derived from the anatomy image with the help of the medial axis. According to the present invention, the size of ROI is proportional to the minimal distance from the edge of anatomy to the medial axis.
• the shape of the ROI two types are preferably employed in the present invention.
• One is a rectangle shape, the other is an adapted shape aligned with the medial axis.
• the medial axis provides the position and orientation information of anatomy in radiograph, so adapting the shape of ROI along the medial axis can help reduce the effects caused by translation and rotation and ensure that the ROI is translation- and rotation- invariant, which in turn ensures that the whole projection view recognition method is robust.
• Figures 5 A and 5B show examples of the different shapes of ROI extracted from radiographs of different anatomy objects.
• the medial axis of cervical spine is nearly straight, so the shape of ROI is rectangle.
• the medial axis may not always hold straight.
• the shape of ROI may appear as a twisted or slanted strip surrounding the medial axis, as shown in Figure 5B.
• the choice of the ROI shape is largely dependant on how it affects the performance of recognition. If a simple rectangle shape can satisfy the requirements, it will be adopted; otherwise a more complicated adapted shape will be considered.
• the ROI is further divided into N*N blocks and a set of low-level visual features are computed from sub-blocks.
• the number of sub-blocks is determined empirically to balance the trade off between the computation complexity and recognition performance.
• possible low-level features could be the gray level mean, variance, edge information, texture and other image features extracted from sub-blocks.
• the scaled i th feature component, ⁇ ,- of a feature vector, x is calculated as:
• Xj Xj — mint I rnaxi — mint where min t and maxt represent the range of the i th feature component JC,- of x over the training examples.
• FIG. 6 wherein projection view classification is accomplished by a set of pre-trained classifiers.
• Each classifier is trained to classify one projection view from all the others, and its output represents how closely the input radiograph match such projection view.
• the number of classifiers equals to the total number of projection views that an examination can possibly have.
• chest radiographs generally have four projection views, anterior-posterior view (AP), posterior-anterior view (PA), lateral view (LAT) and oblique view (OBL). Therefore four classifiers are needed to classify chest radiographs.
• AP anterior-posterior view
• PA posterior-anterior view
• LAT lateral view
• OLB oblique view
• a method is provided to create such classifiers. This method is composed of a training step and a testing step.
• a collection of training images is first obtained with known target projection view information.
• a set of extracted features from individual training images and their associated target outputs, which specify the correct or incorrect projection view are used to train a classifier.
• the classifier can be any of methods known in the art, such as a neural network and support vector machine. If the original features are not effective for classification, an optional step can be added, which computes a suitable transformation from the original features. The benefit of adding this step is to further study the characteristics of training features and derive the most discriminate features for classification. Examples of such transformations include normalization, feature extraction by principle component analysis (PCA) or independent component analysis (ICA), or a non-linear transformation to create secondary features.
• PCA principle component analysis
• ICA independent component analysis
• a set of a pre-trained classifiers can be obtained by the steps of: collecting a pre-determined number of training images with known projection view information; locating a region of interest for each of the training images; computing a set of features from the region of interest of each of the training images; associating a target output specifying the known projection view of each of the training images; computing a transformed feature set for each of the training images using principal component analysis based on all the training images; and training a classifier with the transformed feature set and target output.
• a testing step is performed on novel images to evaluate the performance of classification. If the performance cannot satisfy the pre-set requirement, the classifier may be biased by, or overfit the training data. When this happens, the classifier can be retrained until it performs best on both training and testing data.
• features described above are extracted from the ROI of a novel radiographic image (step 61).
• the final step is to determine the most likely projection view of the input radiograph by combining the outputs of the classifiers (step 65, new in Figure 6).
• Bayesian decision rule is used to combine the results from classifiers and infer the projection view of radiograph as the one with the highest confidence.
• a computer program product may include one or more storage media, for example; magnetic storage media such as magnetic disk (such as a floppy disk) or magnetic tape; optical storage media such as optical disk, optical tape, or machine readable bar code; solid-state electronic storage devices such as random access memory (RAM), or read-only memory (ROM); or any other physical device or media employed to store a computer program having instructions for controlling one or more computers to practice the method according to the present invention.
• the system of the invention can include a programmable computer having a microprocessor, computer memory, and a computer program stored in said computer memory for performing the steps of the method.
• the computer has a memory interface operatively connected to the microprocessor.
• the system includes a digital camera that has memory that is compatible with the memory interface.
• a photographic film camera and scanner can be used in place of the digital camera, if desired.
• a graphical user interface (GUI) and user input unit, such as a mouse and keyboard can be provided as part of the computer.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Multimedia (AREA)
• Image Processing (AREA)
• Image Analysis (AREA)
• Apparatus For Radiation Diagnosis (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=35987029

Family Applications (1)

Country Status (4)

Families Citing this family (3)

Family Cites Families (2)

• 2005
  • 2005-11-21 CN CNA2005800401449A patent/CN101065777A/zh active Pending
  • 2005-11-21 EP EP05825471A patent/EP1815433A1/en not_active Withdrawn
  • 2005-11-21 JP JP2007543501A patent/JP2008520396A/ja active Pending
  • 2005-11-21 WO PCT/US2005/042649 patent/WO2006058176A1/en active Application Filing

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events