KR20020079742A - Convolution filtering of similarity data for visual display of enhanced image - Google Patents

Abstract

The similarity of the data elements is adjusted to provide an improved image of the object. Data is collected from the object using appropriate sensors and stored as a number of separate data points. Data points are compared to each other to determine how similar the data attributes of one data element are to other data elements. Similarity is assigned to each data element to indicate an attribute difference between the two data elements. The similarity for each particular data element is then adjusted based on the weighted similarity of the group of data elements adjacent to that data element itself. Adjusted analog values are stored for each data element and used to generate an image representing the object.

Description

Convolution Filtering of Similarity Data for Visible Display of High-Quality Images {CONVOLUTION FILTERING OF SIMILARITY DATA FOR VISUAL DISPLAY OF ENHANCED IMAGE}

It is often required to display an image of an object on a visual display for the user to see. The object on which the image is to be displayed may be an object of a common physical system, such as a building, a landscape at a particular point in time, or a tree or other living thing. In the medical field, it is often required to display an image of all or several parts of an animal or human body, such as the spinal cord, brain, bone, muscle, or other tissues. It is very important for the purpose of medical diagnosis to accurately identify tumors, masses or other abnormal tissues of interest. This diagnosis is often performed by medical images.

Images of the object may be collected as many as various symptoms and stored in various techniques. In the medical technology field, it is common to collect images using a variety of techniques including magnetic resonance imaging (MRI). Various types of information are collected using such medical devices. The collected and stored data is in most cases formed into images for the user to view, which may be a medical diagnostic team that includes the radiologist and the patient's primary care physician. As part of the process of generating the image data and the image obtained therefrom, a convolution filter can act on the initial data from the projector to refine the data and reduce noise.

In the medical projection field, it is very important to obtain accurate and precise images of the object. MRI is often used to find abnormal tissue in the human body. In many cases, abnormal tissue is a tumor, whether malignant or benign. It is particularly important to determine whether the tumor is malignant or benign early in the medical diagnosis process. In medical terms, where the tumor is determined to be malignant, it is very important to find out the extent of metastasis, in medical terms. If the tumor has not yet metastasized, special treatment regimens are good and are often advantageous for the human body to overcome cancer. On the other hand, when cancer has metastasized, other treatments and medical treatments are required and often treat the cancer more aggressively, increasing the likelihood that all cancerous tissues will be completely removed from the patient's body and restored to their health. Such imaging becomes especially important when cancer is present in the brain, lungs, lymph nodes, or other parts of the body where surgical examination or treatment is not easy.

For medical consuls, it is also important to minimize the false positives so that the patient is not overtreated. If the image appears to be cancerous in countless parts of the body, it is very important to make sure that healthy non-cancerous tissue does not look like cancerous tissue. False positives lead to inadequate medical diagnosis and result in different or excessively aggressive treatment than is required. Therefore, in the medical field, it is important to display an image showing all parts of the human body in which abnormal sarcomas are properly and accurately identified without any healthy breakfast being mistakenly recognized as abnormal sarcomas.

This invention belongs to the field of controlling the visual display of an object, and more particularly, to a convolution filter for a visual display device.

1 is a block diagram showing the various stages of the image acquisition display process;

2 is a block diagram showing generation and convolution processing of analogous data according to the present invention;

3 is a diagram showing an example of a Gaussian type convolution filter having a sigma of 1;

Figures 4a and 4b is a diagram illustrating the split results for the tumor in the magnetic resonance projector before and after the convolution filter according to the present invention.

According to the principles of the present invention, there is provided a method of adjusting the values of data elements stored in a representation of an object in which an image is stored in a memory. Data is collected from the object using appropriate sensors. Each of the plurality of data elements representing the various locations of the specimen is stored in memory. If each of the data elements from these sensors is called a first value, this first value represents the nature of the specimen. If multiple datasets are generated from the projection of the object, there will be multiple first values for each of the various locations of the object. The first values are combined with several data elements organized in a pattern, where data elements adjacent to each other in the pattern in which the data elements are organized represent adjacent positions in an object. In this regard, additional adjustment of the first data values may occur via convolution filters or high pass or low pass filters to achieve smoothness and noise rejection of the image result. The result of this process is a second value for each of the various data elements.

Subsequently, a new value, called a similarity, is assigned to each of the data elements, which indicates the similarity of the corresponding data element to the standard, reference set of data elements representing the properties of the specimen. Similarities are merged with several data elements organized in a pattern, where data elements are adjacent to each other in the organized pattern. The similarity of a particular data element is then adjusted based on the data element itself and the weighted similarity of the group of adjacent data elements. This procedure is known as convolution filtering and yields a controlled similarity. The adjusted analog is stored as adjusted for each data element. An image is then formed of an array of multiple visible pixels for viewing by the user, the display characteristics of which are based on the adjusted similarity of the data elements.

According to one embodiment of this invention, a convolution filter comes out of the sensor and acts on the data stored in the memory. The convolution filter preferably serves as the final stage of the multi-stage processing of the data. In the first series of steps, data from the sensors is collected and stored in memory. The data from the sensors is organized according to the various properties of the tissue under examination in each medical image. The area of interest, including at least the reference organization or standard organization, is designated by the user. Within any region of interest, the data elements of the training set are determined. The user additionally selects one or more training data elements to accurately reflect the nature of the tissue to be examined. After the selection of the training set, the remaining data elements representing the object are processed with a test sample to be compared with the selected training set. The Euclidean distance between each test sample and each element of the selected training set is calculated and a similarity is obtained that represents the shortest distance between any element of the training set and the test sample. This similarity is the relative similarity expressed numerically between the test sample and the training dataset. The analogous value is stored in a memory location to be used for generating the image.

After the analogous data is generated, the convolution filter of the present invention works. The convolution filter compares the similarity of one data element with the similarities of adjacent data elements and adjusts the similarity of the data element under inspection by applying one or more weighters. This convolutional filtering is done every element throughout. Then, an image is displayed based on the adjusted analogs that more accurately indicate the properties of the object.

According to one embodiment of this invention, the working convolution filter is a conventional 5x5 square array weighted according to a Gaussian formula with a standard deviation sigma of 1. According to another embodiment of the present invention, a convolutional filter acting in a windowing in the Fourier domains, an isotropic diffusion filtering, or adjacent pixels of an image is normally similar. It can be a probability filter that takes advantage of the fact that it belongs to an organization. A Markov field type filter that collects spatial information into similar data may be used. Other convolutions based on spatial characteristics of the human body, such as symmetry of the brain or other body parts, may be used to increase the clarity of the image and confidence in the classification of the tissue.

1 illustrates an overview of the entire process of implementing the invention. Step 12 in which image acquisition takes place is a step of initially acquiring data representing the specimen. As described under the background of the invention, the image may be any object that requires more attention. For example, according to this invention, the object may be one that can occur universally in the natural world, such as landscapes, plants, or sunsets. This invention seeks to be very sensitive to the differences between two very similar features in an object, which is very useful for projecting a single object, finding similar features of an object, and distinguishing them from different features in the same object. Do. This invention is very suitable for application when the specimen is a patient and the image is made for the purpose of medical diagnosis of the tissue of the patient's body.

Magnetic resonance projection using pattern recognition of US Pat. No. 5,311,131, and Quantization and Standardization of Magnetic Resonance Measurements of US Pat. No. 5,644,232, and Quantization and Standardization of Magnetic Resonance Measurements of US Pat. No. 5,818,231, incorporated herein by reference. Methods of acquiring and trimming images using magnetic resonance projection techniques are already known in the art. These patents describe various methods of obtaining medical images using NMR or MRI techniques and are not repeated here. These patents are provided as background information and may be studied in more detail if necessary in understanding the preferred embodiments and applicability of this invention. The image acquisition method described in these patents is suitable for use with this invention to acquire and store the image required in step 12. Other medical imaging methods may be used, including X-rays, Positron Emission Tomography, Computed Tomography, various radiation methods, ultrasound methods, or other available methods.

Therefore, when an image is required to be acquired and stored in the memory, it is desirable to further study the diagnosis image of the medical analysis and the diagnosis of cancer location that may exist in the body of the patient under examination. In order to proceed to the next step, a user, typically a medical professional or physician, selects an area of interest that requires further testing. The region of interest can be selected by using a computer mouse to mark the type of tissue desired on the screen, or by forming a box around the target position, such as a known tumor position, or by other means for the user to define the region of interest. have.

After the region of interest is determined, aggregation in the region of interest is performed as shown in step 16. The number of cohorts corresponds to the number of different tissue types in the area of interest. This allows you to divide your organization into different categories within your area of interest. If the area of interest is very small and special, there may be only one type of organization within the area of interest. Optionally, there may be more than one type of organization within the area of interest, and grouping will allow different types of organizations to be found within a given area of interest. Since the area of interest is smaller than the premise image, the number of types of tissues in the classification will be small, and one type of organization in the area of interest can be distinguished more clearly from other organizations.

According to one embodiment, the grouping performed in step 16 is automatic grouping in an unlimited form of division. A fuzzy grouping mechanism is performed that automatically determines the number of populations used for classification using Xie-Beni Fuzzy Validity Criterion. The Xzai-Benney fuzzy effect axiom is based on the simplicity of the number of groups and the differentiation of different types of organizations within a particular domain. Such fuzzy groupings are known in the art and are described, for example, in XL. No. 13, pp. 841-847 (1991) of IEEE Transactions on Patiern Analysis and Machine Intelligence . As shown in the paper by Xie and G. Beni, the literature is published. Optionally, grouping may be performed using a non-automatic method.

The user then performs inductive tissue selection in step 18, described below, in accordance with the present invention. After organizational grouping is performed, the user selects one or more populations that indicate the type of organization of the region of interest for further study. The type of tissue selected for the standard use of this invention will typically be a tumor that is suspected to be malignant. In some applications, it may be time to find out if the tumor has already been identified as a malignant tumor and whether the exact shape of the tumor and whether the cancer has spread beyond the boundaries of the initial tumor. You may want to know if cancer is present in other parts of your body.

According to this invention, a population representing a tissue known to be a particular tissue that requires additional testing is selected. In this example, the tumor is assumed to be selected. The group corresponding to the specific organization is selected. The selected population contains a training set of individual data elements that indicate the type of tissue to be further examined and studied. A comparison is then made between each component of the training set and each data element outside of the training set. Preferably, the data elements that make up the entire initial image in step 12 are compared to the training set to determine which tissue under study is found outside of the region of interest, especially if it is found elsewhere in the body. .

Any known available method may be used to compare the data elements of the training set with other sets of data elements representing the subject. Although other methods may be used, we will focus on one specific comparison method in detail.

An overview of the available methods for performing the specific tissue classification of step 20 is as follows. The corresponding data element from each sensor has one or more properties that describe the corresponding part of the object represented by the data element. Each of these properties is quantified. For example, if the acquired image is an MRJ image, then the nature of each data element is either longitudinal longitudinal T1 or transverse relaxation T2 or weighted T1 or T2 image, proton density space, or MRI as known in the art. It may include features such as other parameters that are normally measured. Therefore, each of the known data elements in the training set has a numerical value associated with each of the properties that provide a specification for the data element. In so doing, each data element will be described by several different numbers in which one value is stored for each property. In doing so, the data is diversified. The numerical values may define the position of data elements in multidimensional space and reflect the magnetic resonance properties of the tissue corresponding to those positions. That is, each parameter represents one of the dimensions for locating the object in the Euclidean geometric field. If two properties of an object are stored for each data element, the field is a two-dimensional Euclidean plane. If three parameters are stored, the data elements can be considered to be a location in the three-dimensional Euclidean field. Likewise, if four physical parameters are indicated, the object can be considered as a location in the four-dimensional Euclidean field. Therefore, each component of the training set has a location in the multidimensional Euclidean field.

The values associated with each component of the training set are compared with the values associated with every other data element of the specimen. Since the numerical values of the data elements are related to the magnetic resonance properties of the specimen for each data element, their position in the multidimensional Euclidean field is related to the magnetic resonance properties of the object. In the comparison of the training set to the test set, the closer the two data elements are to each other in the multidimensional Euclidean field, the more similar they are. As such, the distance to each other between two data elements is accurate and reliable, indicating the similarity of the magnetic resonance properties of the object represented by the data elements (data elements outside the training set are referred to as test samples). It is a measure.

The distance between the data elements of the training set and each data element outside of the training set is determined, and a similarity is assigned to that data element.

The distance from the data sample and the test sample data element of the training set may be expressed as a numerical value that provides a similarity to the test sample data element. This similarity number indicates the similarity between the test sample and the training set. This similarity number indicates the similarity between the test sample and the training set. The smaller the distance, the more similar the data elements are. As such, the distance 0 indicates that the data elements are identical to each other and are similar enough to be said to be components of the same class. Larger distances indicate less similarity. Depending on the measure used, the numerical distance value may be a terminal number such as 1 to 10 or the like. Optionally, it may have many numbers, such as 7 to 10 digits. In some embodiments it is expected that the numerical value of the distance is large enough for almost all test samples of the dataset that the log value can be taken after the distance value is obtained to normalize the distance values. The absolute distance value is not as important as the relative distance value compared to the distance for the other data elements of the test set and the distance of the training data element when compared to the other elements of the training data set.

In one embodiment, the actual similarity may be calculated as follows.

The data collected during the MRI test is a four-dimensional array ( ) Where I represents the number of rows in the array, J represents the number of columns, K represents the number of properties from the sensor, and L represents the number of anatomical slices to examine. The K properties from the sensor are those used for tissue classification. Property values for specific array elements ) Can be accessed by specifying an address consisting of row-column-sequence-slice, where i = 1... I is an argument representing an action value, and j = 1. J is a factor representing the column position, k = 1 .... K is the factor for the property from the sensor, l = 1. L is the factor for the number of splices. Each data element in the array can be thought of as occupying a spatial location in a two-dimensional, three-dimensional, quadratic, five-dimensional, or other multidimensional field. Such a data element having a volume in space may be referred to as a voxel. In the case of a two-dimensional screen image, each display element is generally referred to as a pixel.

In one embodiment of the MR classification, a set of samples in which a user represents a single tissue class ( ) Where N represents the number of samples in the training set. These training samples are identified and surveillance classification is performed using the nearest method. Traditional close proximity methods can be used at this stage, as described in the 1996 edition of Gose et al., Which classifies unknown test samples as belonging to the same class as the nearest test sample location in the training set. Parametric decision rule. Test Sample ( ) And training samples ( The similarity between) can be determined by the Euclidean term, such as Equation 1 in the K- dimensional feature space.

The distance from the test sample to each training sample of the set is measured and the maximum of this distance is selected. That is, the similarity to the test sample is based on how close it is to the nearest one of the samples of the training set of that training set. The result of the proximity classification is a three-dimensional pseudo-array containing the maximum distances at each voxel location ( i , j , l ). It contains similarity data in).

After similarities are obtained, in some embodiments it may be desirable to divide the tissues into different classes. According to this embodiment, different organizations are ranked based on how closely they are related to the training set of data elements. If they are extremely close, they may be assigned different patterns or intensities of color, indicating that they are more similar for the data elements of the training class than for other classes of data elements, even if they are not within the training set itself. In some embodiments, no specific tissue classification step 20 is performed and does not form part of this invention. Rather, it is optional that may be performed if necessary.

Figure 2 shows the processing steps for the operation of the convolution filter according to the invention. The raw data is obtained in step 12 corresponding to step 12 in FIG. Processing of the data from the sensor is performed in step 24. Many data processing techniques are known for filtering the collected data, one of which is convolution filtering, which acts on the row data itself.

In one embodiment, processing 24 and generation 25 of similarity data correspond to the steps 14-20 of FIG. 1 described above. In other embodiments, processing 24 and generation 25 of similarity data may use other techniques that are acceptable in the art.

According to this invention, convolution occurs after the generation of analogous data, thereby obtaining a unique advantage that was not possible in the prior art. After completion of step 25, similarity data is obtained and stored in the first memory. Similarity data is a representation of an object using similarities of multiple data elements corresponding to positions in the object. The similarity of each data element represents the magnetic resonance characteristic of the object in a defined region of interest compared to the magnetic resonance characteristic of another object elsewhere or in the same object using the same projection method and protocol. In one embodiment, the numerical value of the similarity describes how closely the magnetic resonance characteristics of one data element correspond to the magnetic resonance characteristics of the data elements of the training set, and the shorter the distance, the more similar.

The analogous values are stored in memory so that logically adjacent data elements are organized in a manner corresponding to adjacent positions in the object under examination. In some types of computers, these may be adjacent locations in the computer's memory. Optionally, the locations may be only logical addresses, and data elements need not be stored adjacent to each other in physical memory. In this way, data elements appear to be adjacent to each other from a functional point of view, although they can be stored strictly speaking in any desired way.

Then, convolutional filtering in step 26 is performed, wherein the similarity of the data elements is adjusted based on the similarity of adjacent data elements. An important advantage is obtained by performing convolution filtering on similarity data. Adjustment of the similarity of any data element is performed as follows. A data element will be referred to as a pixel in this example as it will be displayed on a two-dimensional computer screen, but of course, this example and the invention may be a voxel or other element in multidimensional space. The similarity of the selected pixels is controlled by a convolution kernel of a given size and shape. The convolution kernel determines the set of filter weights applied to adjacent pixels in its space and the similarities for the selected pixels. The convolutional process can be visualized by superimposing the convolutional kernel on the analogous data matrix with a central filter element of the convolutional kernel that is generally consistent with the similarity of the selected pixel. The adjusted similarity is calculated by multiplying the similarities for the selected pixel and its spatially adjacent pixels by their corresponding convolution filter weights and then summing the multiplications. In one embodiment, the convolution kernel consists of a Gaussian distributed 5x5 matrix. Optionally, the kernel may be hexagonal or circular, or any other shape that defines the control kernel to be applied to the selected pixel.

3 illustrates a method of performing convolution filtering on similar data using a 5x5 matrix Gaussian convolution filter 36 having a standard deviation and a sigma of 1. FIG. In this particular shape a weight is assigned to each pixel position superimposed by the convolution kernel 36 as shown in FIG. The weight is based on the distance from the center pixel 34, which in this example is the selected pixel whose similarity is adjusted. Each pixel position superimposed by the kernel is provided with a weight by Gaussian distribution as shown in legend 38 to the right of Gaussian kernel 36. The center pixel 34 is given the maximum weight. Adjacent pixels are given a smaller weight according to the distance from the center of the selected pixel 34.

Looking at legend 38, it can be seen that the center pixel adjacent to Gaussian kernel 36 will have a white wing, thus a weight slightly greater than 0.15. Although in some embodiments it may be possible to have a weighting factor of 0.2 or 0.5 or some other value where it is desirable to weight the primordial value of the selected pixel itself, the center pixel 34 according to one embodiment of the invention is 0.162. Has a weighted number of. For narrower Gaussian curves, sigma may be chosen lower than 1, such as 0.6, 0.8, etc., which will give greater weight to more center pixels. In addition, larger sigma may be selected, such as 1.5 or 2. If you're concerned about not detecting very small cancers, you'll smooth your data with larger sigmas.

The next four nearest pixels 40 are weighted to the next highest value. In the embodiment shown in FIG. 3 this value for each of these four pixels is about 0.1. The exact path value is shown in legend 38 on the right side, although other values are possible, but in the illustrated embodiment it has a row number of 0.0983. In short, the similarity of the center pixel 34 is in this case multiplied by a weighted number of ˜0.16. Each next closest similarity for adjacent pixels 40 is multiplied by each corresponding weighted number, in this case ˜0.1, to obtain the product for each of the four similarities for the four pixels 40. The four multiplications are then added to the multiplication of the weighted similarities for the center pixel 34 to obtain a sum. Like other pixels, such as 44, 46, 50, etc., superimposed by the convolutional kernel 36, the remaining adjacent pixels 42 have similarities multiplied by their respective weighted numbers. Of course, the weighted number becomes smaller depending on the distance that a particular pixel has from the center pixel 34. For example, the weighting factor for one pixel at position 42 according to this embodiment is ˜0.06, while the weight for another pixel 44 is ˜0.02, and another pixel 46, 50 The weights are 0.013 and 0.003, respectively. In other words, the exact value can be seen by comparing the convolution filter 30 to the legend 38.

After all multiplications of the weights multiplied by the respective similarity data for the pixels represented by the weights are calculated, all multiplications are added together to obtain a sum for all multiplications across the convolution filter. A new similarity is obtained. According to this invention, the top and bottom of the center pixel 34 are replaced by the top and bottom calculated by the sum of the multiplications of all the pixels superimposed by the convolution kernel. Now, the convolution for this pixel 34 is complete.

Then, a new pixel is selected to become the center pixel 34, where convolution will be performed. The convolutional kernel successively overlaps each pixel of the entire image. The similarity superimposed by the 5x5 Gaussian block underneath the convolutional kernel is multiplied by each corresponding weight defined by the kernel. This process is performed for each of every pixel, the multiplication of which is added and the calculated value replaces the primordial value at the center pixel. This has the effect of incorporating information from adjacent pixels into the similarity for each pixel. As explained, pixels away from the center pixel are less weighted than those closer to the center pixel. Then, in step 22, an image is displayed on the display device.

The result of convolutional filtering can be provided to the user by various acceptable techniques. Of course, the actual row numeric data can be provided in printed matter or other visible display devices. In some cases it may be useful to provide the actual similarity data itself, or it may be advantageous if the physician makes detailed decisions about a particular tissue. However, most users who have numerical printouts of analogues themselves will not be as useful as having images representing convolved analogues. Also, a better way to provide the result of the convolution filter is in the form of a monochromatic layer superimposed on the corresponding gray MR image. In so doing, the image may be gray in which the pixels have different values. The adjusted similarity data is applied to the pixels on such gray images to create a new image more clearly showing the viewer the location of the cancer or other tissue being examined.

In another option, the convolution filter data may be displayed on a multicolor display device using different colors, showing different similarities. The tissue most similar to the training set itself may be indicated in a first color, such as red. The next most similar tissue in the training set may be displayed in a color other than red. The next most similar classes after this class may be represented in a third color different from the previous two classes. In this way, an image shown using the results of the convolution filter can be displayed on the color display, where different colors of pixels can clearly indicate the difference in similarity with the training set tissue.

In FIG. 2, a more visible display may be performed on the image after the convolution filter is applied. As described in the first embodiment of the present invention, instead of providing the adjusted similarity data directly from the convolution filter 26 to the display device 22, the data is thresholded in a thresholding step 28. Can be compared to, for example, followed by a post-processing step 30 to generate a binary image. According to the threshold comparison step 28, all pixels whose similarity is above a predetermined threshold are considered to belong to the same class. If the similarity is lower than the threshold, the organization is determined not to belong to the class of interest and discarded. With the generation of binary images, all tissues determined to be of interest class are represented with the same color and characteristics as all other tissues in the interest class. Pixels that are not within the class of interest are shown to have distinctly different colors, and the colors may be gray or actual spectral colors. In this way, the binary image can have sharply defined edges approaching the edge of the binary image, whether each pixel is within or outside the class of interest, and thus more sharply contrast the areas of the tissue of the image under examination. If a gray image is used, the pixels in the class can be in color at one end of the spectrum or in black or white, and all other pixels can be represented in their respective grayscale colors. For example, if all of the data elements with similarities above the threshold become black, this overlaps the top of the MRI image to more clearly show areas that are very similar to each other. After the threshold comparison in step 28, the binary image post-processing is performed in step 30 to make the binary image smoother and filtered. Thereafter, the binary image of step 30 is provided to the display device 22 for the user to observe. In this way, a convolution filter may be applied to the data, if necessary, immediately before the threshold comparison followed by the generation of the binary image.

According to one embodiment, when the similarity of each pixel is adjusted by the convolution filter, the adjusted similarity is stored in the second memory, and the original similarity remains in the first memory and used for all subsequent convolution calculations. . In other words, when the convolution is applied to the first pixel and its similarity is adjusted, this similarity is stored in a separate memory, and the original similarity remains in the first memory. Then, the convolution kernel is applied to the next adjacent pixel whose similarity is just adjusted. The primordial luxury data stored in the first memory is used for convolution to adjust another pixel in the center of the current convolutional kernel. In that way, only the primitive value from each of the pixels is used to adjust the similarity of the center pixel. According to this method, the similarity changes only by the relationship to the original value, and no adjustment is made by any of the similarity data of the pixels which have already changed.

Optionally, if a pixel has adjusted its similarity, the adjusted similarity is used for subsequent convolutional filtering of pixels adjacent thereto. This has the effect of increasing the intensity of that particular pixel compared to others, since its similarity has already been adjusted by the pixels adjacent to it. Although not always good, the latter may be used in some embodiments where an attempt is made to make the similarity as uniform as possible.

In addition to the 5x5 matrix Gaussian distributed convolution filter as described with respect to FIG. 3, other types of convolution filtering may be used. For example, the window processor of the Fourier definition zone may act as a low pass filter. The outline of the window processing section of the Fourier definition area can be seen from the standard image processing textbooks, and thus will not be described in detail here.

Since tumors or other tissues in general in the human body are generally seen in images as a plurality of adjacent pixels, the adjacent pixels can provide information to increase confidence in the division or segmentation of the data into classes. The convolution filter of this invention has been found to greatly improve the readability of the image and to reduce false labeling while increasing appropriate labeling for tumors. Thus, while similar types of tissues can be more readily identified in the image behind the convolution filter of the present invention, sham positives are much more likely to be properly interpreted because they will be removed or reduced.

4A and 4B illustrate the improvement obtained in this invention when a convolution filter is applied. Four white arrows 50, 52, 54, 56 are shown in FIG. Each of these arrows points to a known tumor in the specimen by necropsy performed after the MRI image is generated. It is noteworthy that large tumors, referred to as arrows 50 and 52, contain dead cells that were previously part of living cancerous tissue in their central necrotic tissue.

In performing the divisions for this example, known tumor locations were used to set up the training set, and the similarities for each pixel in the object were calculated. A binary classification result was obtained by applying a threshold to those pixels having or exceeding a given threshold and labeled black in FIGS. 4A and 4B. These black pixels, identified as very similar to tumors, were superimposed on conventional MR images.

The image of FIG. 4A, which provides binary classification results performed without convolution filtering, includes a number of sham positives. For example, in the central region 60 there is a fake labeled black region, which may be considered a tumor by some physicians. Other pseudopositive bodies 62, 64 are seen at various locations in the image. More importantly, the known tumor location, referred to as arrow 54, is not labeled black, indicating that it is a false negative location in the current image. Tumors known to be in position 54 by necropsy performed after MRI were taken were not identified using conventional standard MR imaging.

Fig. 4B shows the binary shape of the grayscale image of the same MR image data shown in Fig. 4A followed by convolution filtering according to the present invention. That is, the similarity data value analysis described in connection with the present invention was performed on the data elements constituting the MR image. After the similarity for each data element corresponding to the pixels in the figure was obtained, a convolution filter was applied to the similarity data, and the similarity of each pixel was adjusted by the weight of the adjacent pixel. Then the threshold was applied and all pixels with similarities greater than or equal to the threshold were displayed in black. Pixels with similarities less than the threshold were shown in grayscale patterns according to their original magnetic resonance properties. This has a good effect of eliminating or reducing the number of sham positives, while describing the actual tumor tissue more clearly. For example, as shown in Figure 4b, tumor locations (50, 52, 56) can be seen more clearly see. Most important of all, the tumor tissue at position 54 can now be seen as the black zone indicated by arrow 54. By adjusting the value of each pixel according to the weight of the adjacent pixels, all the results reinforce the effect of adjacent tumors, including the pixels where one location 54 is now apparently seen as tumor tissue.

Convolution filters also have the good effect that labeling of multiple fake benign tumor sites is completely removed from the image. Other false positives appear to be more clearly false positives, thereby not confusing the physician under analysis. For example, the black labeling of sham positives in the bottom region 60 completely disappeared. Almost all fake positive labeling was removed throughout the image. As shown in regions 62 and 64, the labeling of sham progenitors before convolutional filtering was greatly reduced in size after using the convolution filter of the present invention. In this way, since these areas were reduced in size due to the convolution filter instead of being enlarged or maintained at the same size as occurred in the actual tumor tissue at positions 50, 52, 54, 56, the surgeon determined that these areas Instead, you can be more confident that they represent fake positives instead.

The method of the present invention obtains good results by applying a convolution filter as a post-processing step of similarity data after the classification. There are reports that a convolution filter has been used to process image data from the original sensor, but as far as the inventions of the present invention are concerned, such convolution filters have not previously been used in post-processing of analogue data to improve discrimination. Did.

From the above, although certain embodiments of the invention have been described for illustrative purposes, it will be appreciated that various adjustments may be made without departing from the spirit and scope of the invention. Therefore, the invention is not to be restricted except in light of the attached claims.

Claims (14)

In the method of adjusting data, Storing a representation of the object in the first memory using a plurality of separate data elements representing corresponding positions in the object, Assign each data point a similarity that represents the object's properties. Organize the similarities of the separate data elements into a pattern, in such a way that adjacent data elements in such a pattern indicate adjacent positions within the object, Adjust the similarity of the arbitrary data element according to the weighted similarity of the arbitrary data element and a plurality of adjacent data elements, Store the adjusted analogs of the data elements, Forming a user observable image of a plurality of distinct visible pixels having display characteristics in accordance with the adjusted similarity of the data elements. The method according to claim 1, And all data elements disposed within three of the arbitrary data elements are considered adjacent. The method according to claim 1, And all data elements disposed within two of the arbitrary data elements are considered adjacent. The method according to claim 1, Wherein the weight of one adjacent data element compared to other adjacent data elements is due to the weighting of the Gaussian pattern of the data elements. The method according to claim 1, Adjusting the value of the arbitrary pixel, Form a convolution kernel with the arbitrary pixel and a plurality of adjacent pixels, Multiplying the similarity of the pixel with adjacent pixels by the weight of each pixel in the convolution kernel to obtain first multiplications, A weighted similarity is obtained by adding the multiplications of all the pixels in the convolution filter, Storing the similarity of the arbitrary pixel as the weighted similarity. The method according to claim 5, The sum of all weighted numbers is one. The method according to claim 1, Form the data elements of the training set, Form the data elements of the testing set, Comparing each test data element with data elements of the training set to obtain a similarity for the test data element. The method according to claim 1, Storing the similarity of each data element in the first memory, And storing the adjusted similarities of the data elements in a second memory. The method according to claim 1, Organizing similarities of the separate data elements is performed by organizing a storage location of the separate data elements in the first memory. The method according to claim 1, Additionally generating a binary image of the MR data, Generating a binary image of the MR data, Setting a threshold for the similarity, Generating a data element class comprising all data elements having a superiority greater than the threshold. The method according to claim 10, Display all pixels representing data elements in any class in monochrome, And additionally indicative of pixels that are not in the any class using a different color than data elements in the any class. The method according to claim 11, Wherein said monochromatic color is one color selected from the grayscale < RTI ID = 0.0 > The method according to claim 11, Wherein said solid color selected for the class of interest is black, and all other tissues not within the class of interest are represented in a grayscale pattern with a color different from black. The method according to claim 11, Wherein said monochromatic color is selected from a visually visible color spectrum and pixels representing tissue that is not within the class of interest are displayed using other colors in the color spectrum.