CN102663446A - Building method of bag-of-word model of medical focus image - Google Patents

Abstract

The present invention relates to identifying medical lesion images, and in particular to a method for constructing a bag-of-words model of medical lesion images. The method first divides the medical lesion image into two parts, the lesion area and the lesion boundary area, and obtains the lesion area and the lesion boundary area respectively. The bag of words, and then construct the word bag model of the combination of the bag of words in the lesion area and the bag of words in the boundary area of the lesion. Compared with the general bag-of-words model, the bag-of-words model obtained by the construction method of the present invention increases the relative spatial position information of the local features of the lesion area, thus helping to improve the accuracy of clinical diagnosis.

Description

A construction method of bag-of-words model for medical lesion images

technical field

The invention relates to identifying medical lesion images, in particular to a bag-of-words model for identifying medical lesion images.

Background technique

The basic idea of content-based image retrieval (Content-Based Image Retrieval, CBIR) is to extract the visual features of images for image expression, and use image expression to retrieve images. CBIR technology can provide support and assistance for image data management, clinical diagnosis, medical teaching, etc. In particular, the retrieval of similar lesions in medical images can improve the reliability of clinical diagnosis and the integrity of relevant information.

Generally, the CBIR system stores the image to be queried and the image expression of the image to be queried in the database. When the user provides a query image, the CBIR system extracts the visual features of the image for image expression, and the image of the image to be queried in the database Expressions are compared to return images similar to the query features. The visual features commonly used for image expression in the CBIR system include: color, texture, shape, edge, etc. Texture feature is one of the important features of an image, which can describe the smoothness, sparseness, regularity and other characteristics of the image. In medical images, texture features are widely used in the image expression of medical image retrieval because the extracted texture features can reflect the details hidden in the image structure that cannot be observed by human eyes. Several traditional texture feature extraction methods include: gray level co-occurrence matrix, wavelet transform, Gabor transform and so on. The gray level co-occurrence matrix represents the spatial dependence of the pixel gray level in the consistent texture mode. However, it does not fully capture the image characteristics of the local gray level. For larger parts, the effect of extracting texture features by this method is not ideal. . Wavelet transform can provide a clear mathematical framework for multi-scale requirements. Therefore, wavelet transform can effectively extract multi-scale features of texture images. However, in wavelet transform, the problem of filter bank selection remains unsolved, resulting in impact on texture analysis. The Gabor transform is the most suitable for human vision and plays an important role in image analysis. However, its transformation window size is fixed and it is not sensitive to subtle changes in texture direction and frequency, which cannot meet our practical needs. The bag of words (Bag of Visual Words, BoW) model extracts the local details of the entire image, and then quantifies the image, capturing the subtle changes and overall features of the image.

However, because the general bag-of-words model does not take into account the relative spatial position relationship between features when extracting local features, the generated bag-of-words model lacks effective spatial information and reduces the description ability of the bag-of-words model. In order to solve this problem, some researchers proposed a method of combining the spatial pyramid model and the word bag model for the classification of natural images (S.Lazebnik, C.Schmid, and J.Ponce, "Beyond bags of features: spatial pyramid matching for recognizing natural scene categories, "IEEE Conference on Computer Vision and Pattern Recognition, pp.2169-2178, 2006). This method uses the extraction method of the spatial pyramid model to subdivide the image into multiple sub-regions, then construct the bag-of-words model of each sub-region, and finally combine the bag-of-words models of all sub-regions into a large bag-of-words model. In addition, some researchers proposed to add the spatial coordinates of features to the feature vector of the bag-of-words model to improve the classification and retrieval performance of medical X-ray films (U.Avni, H.Greenspan, E.Konen, M.Sharon, and J . Goldberger, "X-ray classification and retrieval on the organ and pathology level, using patch-based visualwords, "IEEE Transactions on Medical Imaging, vol.30, pp.733-746, 2011). Because medical lesion images are generally gray-scale images, the texture information inside the lesion tissue is different from that of nearby normal tissues, and the texture information of different types of lesions is also different. In addition, the tissue structure adjacent to the lesion is also different. For example, in brain tumor images, gliomas often have a rim of edema; meningiomas are adjacent to the skull, gray matter, and cerebrospinal fluid; and pituitary tumors are often seen near the sphenoid sinus, optic chiasm, and internal carotid arteries. Therefore, although the above-mentioned prior art can well express the texture features in the lesion area, the accuracy of distinguishing different types of lesions is not enough due to the lack of grayscale change information of the lesion border area and the transition from the lesion area to the lesion border area ideal.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for constructing a bag-of-words model of a medical lesion image. The bag-of-words model obtained by the method can reflect the spatial position of the lesion area, which helps to improve the accuracy of clinical diagnosis.

The purpose of the present invention can be achieved through the following technical measures:

A method for constructing a bag-of-words model of a medical lesion image, the method is composed of the following steps:

(1) Read the medical lesion images in the database that have outlined the outline of the lesion. The read medical lesion images must contain at least 50 images of each lesion type, and each medical lesion image is processed as follows:

(1.1) First perform one-dimensional Gaussian smoothing on the lesion contour line, and then take the pixel point where the lesion contour line intersects with the normal of the lesion contour line as the starting point, and take a certain number of points along the lesion contour normal to the inside of the lesion and outside the lesion The same pixel points are respectively arranged in a column clockwise or counterclockwise with the pixel points located in the normal direction of each lesion outline, and the pixel points located in the same position in the normal direction of each lesion outline are row, arranged in the order from the inside of the lesion outline to the outside of the lesion outline or from the outside of the lesion outline to the inside of the lesion outline, to obtain the transformed image of the border area of the medical lesion image; then, the interval within the transformed image is 0 to 5 pixel points Expand outward from the center of the pixels, obtain a series of small squares with a pixel array size of 5×5, 7×7 or 9×9, rotate each column of pixels in the small squares by 90° counterclockwise, and then Arrange a queue in order from top to bottom, and then replace the pixels located in the transformed image in each queue with the corresponding gray value, and the gray value of the pixel located outside the transformed image Values are respectively assigned zero to obtain the gray value vector A1 of the lesion border area;

(1.2) In the lesion area, starting from the pixel located on the edge of the lesion contour line, the pixel points with an interval of 0 to 5 pixel points are expanded outward from the center, and a series of pixel point arrays with a size of 5×5 are obtained. , 7×7 or 9×9 small squares, each column of pixels in the small squares is rotated 90° counterclockwise, and then arranged in a row from top to bottom, and then each row of the queue is located at Pixels within the contour of the lesion are replaced with the corresponding gray values, and the gray values of the pixels outside the contour of the lesion are respectively assigned zero to obtain the gray value vector A2 of the lesion area;

(1.3) Regulate the gray value vectors A1 of the lesion boundary regions of all medical lesion images obtained in step (1.1) into one class, and standardize the gray value vectors A2 of the lesion regions of all medical lesion images obtained in step (1.2) is another class, and then k-means clustering is performed on it respectively to obtain a lesion area dictionary and a lesion boundary area dictionary composed of words formed by the cluster center;

(2) Take any medical lesion image in the database, and construct the word bag B2 of the lesion area and the word bag B1 of the lesion border area respectively; Wherein,

(2.1) The construction method of the bag of words B2 in the lesion area is as follows: take each pixel in the lesion area as the center and expand outwards to obtain a series of small squares with the same size as the small squares described in step (1.2), and then Obtain the gray value vector A2' of the secondary operation of the lesion area of the medical lesion image in the same way as step (1.2), and then calculate each gray value vector A2' and the lesion area respectively according to the order of the row queue The Euclidean distance of each word in the dictionary, and quantize each gray value vector A2′ to the word with the smallest Euclidean distance to obtain the frequency of each word in the lesion area dictionary. The dimension vector is the word bag B2 of the lesion area of the medical lesion image;

(2.2) The construction method of the bag of words B1 of the lesion border area is: first process the lesion border area of the medical lesion image in the same way as step (1.1), to obtain the transformed image of the border area, to transform each word in the image Expand outward from the center of one pixel point to obtain a series of small squares with the same size as the small squares described in step (1.1), and then use the same method as step (1.1) to obtain the secondary calculation of the lesion boundary area of the medical lesion image The gray value vector A1' of each gray value vector A1', and then calculate the Euclidean distance between each gray value vector A1' and each word in the lesion boundary area dictionary according to the order of the row queue, and put each gray value vector A1' Quantize to the word with the smallest Euclidean distance to obtain the frequency of occurrence of each word in the lesion boundary area dictionary, and the one-dimensional vector formed by the frequency number is the word bag B1 of the lesion boundary area of the medical lesion image;

(2.3) The bag-of-words model of the medical lesion image is obtained by arranging the two vectors representing the word bag B1 and the word bag B2 end-to-end to form a vector.

In the construction method of the present invention, the size of the small square is preferably 7×7, and the k value of the k-means clustering can be selected according to the number of images, and the principle of selection is that the larger the number of images, The larger the value of k is; the smaller the number of images is, the smaller the value of k is. Generally, the value of k is 1000.

The construction method described in the present invention is applicable to the construction of bag-of-words models of MRI, CT, and ultrasonic images.

Compared with the prior art, the inventive method has the following beneficial effects:

Since the image database is composed of images that have already outlined the outline of the lesion, according to the characteristics of the lesion image, the lesion image is divided into two parts: the lesion area and the lesion boundary area, and the word bags of the lesion area and the lesion boundary area are obtained respectively, and then A bag-of-words model combining bag-of-words in the lesion area and a bag-of-words in the boundary area is constructed, which increases the relative spatial position information of the local features of the lesion area compared with the general bag-of-words model, thus helping to improve the accuracy of clinical diagnosis.

Description of drawings

Fig. 1 is the flowchart of the construction method of the bag-of-words model based on medical lesion image of the present invention;

Fig. 2 is recall-precision rate curve, and wherein, the curve that is numbered as I is the recall-precision rate curve that obtains with word bag model of the present invention; Numbering is that the curve of II is to be used in lesion area construction word The recall-precision rate curve obtained by bag B2; the curve numbered III is the recall-precision rate curve obtained by the method of gray level co-occurrence matrix; the curve numbered IV is the recall rate curve obtained by the method of wavelet transform -The precision rate curve; Numbered as the curve of V is the recall-precision rate curve that obtains with the method for Gabor transformation;

Fig. 3 is recall-recall rate curve, and wherein, the curve that is numbered as I is the recall-recall rate curve that obtains with word bag model of the present invention; Numbering is that the curve that is II adopts space pyramid method to construct word The recall-precision rate curve obtained by the bag model; the curve numbered III is the recall-precision rate curve obtained by constructing the bag-of-words model by adding the spatial coordinates of the feature in the feature vector.

Detailed ways

Example 1 (Building a bag of words model)

The database used in this embodiment consists of 800 T1-weighted enhanced MRI images of the brain containing meningioma, glioma, and pituitary tumor, and each MRI image has been delineated. The method for constructing the bag-of-words model of any MRI image in the database will be described in detail below with reference to FIG. 1 .

(1) Read the MRI images in the database, and perform the following processing on each MRI image therein:

(1.1) First perform one-dimensional Gaussian smoothing on the lesion contour line, and then take the pixel point where the lesion contour line and the normal line of the lesion contour line intersect as the starting point, and take 15 points along the lesion contour normal to the inside of the lesion and outside the lesion. Pixels, and then respectively, take the pixels located in the normal direction of each lesion outline as columns, arrange clockwise or counterclockwise, and take the pixels located at the same position in the normal direction of each lesion outline as rows, Arranged in the order from the inside of the lesion outline to the outside of the lesion outline or from the outside of the lesion outline to the inside of the lesion outline, the transformed image of the boundary area of the medical lesion image is obtained; then, the pixel points with an interval of 5 pixel points in the transformed image are Expand the center outward, obtain a series of small squares with a pixel array size of 7×7, turn each column of pixels in the small squares 90° counterclockwise, and then line up in a row from top to bottom, Then, in each row queue, the pixels located in the transformed image are replaced with the corresponding gray values, and the gray values of the pixels located outside the transformed image are respectively assigned zero to obtain the lesion boundary area. Gray value vector A1. The purpose of the one-dimensional Gaussian smoothing described therein is to avoid the change of the normal direction of the lesion outline due to noise interference on the lesion outline, and a Gaussian kernel is used to smooth the lesion outline. The method of one-dimensional Gaussian smoothing is as follows:

The one-dimensional Gaussian kernel is

${\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="G"\; filenames="HDA0000156682520000021.png,HDA0000156682520000022.png"\; state="\{\{state\}\}">G</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ;</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}}\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}<span\; class="notranslate">\; ,</span>$

Here σ represents the standard deviation and X represents the position of any point in space. Assuming that the point coordinates on the lesion contour line of a lesion are b(x, y), then the point coordinates on the lesion contour line smoothed by the Gaussian kernel are

B(x', y')=b(x', y')*G _1D (X; σ)',

where σ and X are the same as those in the Gaussian kernel expression. Therefore, the angle between the normal direction of the lesion contour line and the tangent direction of the lesion contour line is

$\mathrm{<span\; class="notranslate">\; \&theta;</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arctam</span>}<span\; class="notranslate">\; (</span>\frac{{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}{{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; .</span>$

Using the angle θ, the coordinates corresponding to the points taken in the normal direction of each lesion contour can be obtained

x ₁ =x+l×cosθ,

y ₁ =y+l×sinθ,

Where l is the distance from a point taken in the normal direction of the lesion outline to a point on the lesion outline. Since the coordinates (x ₁ , y ₁ ) are not necessarily on a certain pixel point on the image, linear interpolation is used to obtain the pixel point corresponding to the point taken in the normal direction of each lesion outline.

(1.2) In the lesion area, starting from the pixel located at the edge of the lesion outline, the pixel points with an interval of 5 pixel points are expanded outwards, and a series of small pixels with an array size of 7×7 are obtained. square, rotate each column of pixel points in the small square counterclockwise by 90°, and then arrange them in a row from top to bottom, and then arrange the pixels in each row of queues that are located in the outline of the lesion respectively Replaced by the corresponding gray value, the gray value of the pixels located outside the contour line of the lesion is respectively assigned to zero, and the gray value vector A2 of the lesion area is obtained.

(1.3) Regulate the gray value vectors A1 of the lesion boundary regions of all medical lesion images obtained in step (1.1) into one class, and standardize the gray value vectors A2 of the lesion regions of all medical lesion images obtained in step (1.2) into another category, and then carry out k-means clustering on them respectively to obtain lesion area dictionaries and lesion boundary area dictionaries which are formed by the cluster centers and each have 1000 words.

(2) Take any medical lesion image in the database, and construct the word bag B2 of the lesion area and the word bag B1 of the lesion border area respectively; Wherein,

(2.1) The construction method of the bag of words B2 in the lesion area is as follows: take each pixel in the lesion area as the center and expand outwards to obtain a series of small squares with the same size as the small squares described in step (1.2), and then Obtain the gray value vector A2' of the secondary operation of the lesion area of the medical lesion image in the same way as step (1.2), and then calculate each gray value vector A2' and the lesion area respectively according to the order of the row queue The Euclidean distance of each word in the dictionary, and quantize each gray value vector A2′ to the word with the smallest Euclidean distance to obtain the frequency of each word in the lesion area dictionary. The dimension vector is the bag of words B2 of the lesion area of the medical lesion image.

(2.2) The construction method of the bag of words B1 of the lesion border area is: first process the lesion border area of the medical lesion image in the same way as step (1.1), to obtain the transformed image of the border area, to transform each word in the image Expand outward from the center of one pixel point to obtain a series of small squares with the same size as the small squares described in step (1.1), and then use the same method as step (1.1) to obtain the secondary calculation of the lesion boundary area of the medical lesion image The gray value vector A1' of each gray value vector A1', and then calculate the Euclidean distance between each gray value vector A1' and each word in the lesion boundary area dictionary according to the order of the row queue, and put each gray value vector A1' Quantize to the word with the smallest Euclidean distance to obtain the frequency of occurrence of each word in the lesion boundary area dictionary, and the one-dimensional vector formed by the frequency number is the word bag B1 of the lesion boundary area of the medical lesion image.

(2.3) The bag-of-words model of the medical lesion image is obtained by arranging the two vectors representing the word bag B1 and the word bag B2 end-to-end to form a vector.

Example 2 (verification of effect)

1. Build a CBIR system

(1) Use the construction method described in Example 1 to construct the bag-of-words model of each picture in the database described in Example 1, and then use the bag-of-words model of the obtained 2400 MRI images to construct meningioma, glioma and pituitary CBIR system for tumor inquiry1;

(2) Using the above-mentioned 2400 MRI images, the CBIR system 2, CBIR is constructed by using the bag of words B2 of the lesion area of the present invention as the bag of words model and the texture features obtained by the gray level co-occurrence matrix, wavelet transform, and Gabor transform. System 3, CBIR System 4 and CBIR System 5.

(3) Also using the above 2400 MRI images, the CBIR system 6 and ICBIR system 7 were constructed using the bag-of-words model constructed by the spatial pyramid method and the bag-of-words model constructed by adding the spatial coordinates of the features to the feature vector.

2. Search

(1) Submit 100 brain T1-weighted enhanced MRI images of meningioma, glioma, and pituitary tumor as query images to the CBIR system 1 constructed in step 1. System 1 constructs these images according to the method described in Example 1. The bag-of-words model of the query image, and then use the distance measure method to compare the similarity between these query images and each image in the CBIR system, and then classify the images in the CBIR system according to the similarity of the corresponding query image from high to high. lower order. Set the number of images returned at one time as 1, 2, ..., n, ... 2400, get the precision rate and recall rate of each image, and then calculate the precision rate and recall rate of 300 query images The mean value just obtains the curve numbered as I among Fig. 2 or Fig. 3.

(2) Submit the above-mentioned 300 query images to CBIR system 2, CBIR system 3, CBIR system 4 and CBIR system 5 constructed in step 1 respectively, and system 2 constructs query images according to lesion area word bag B2 described in the present invention Bag-of-words model, systems 3 to 5 respectively use gray-level co-occurrence matrix, wavelet transform, and Gabor transform to extract the texture features of these query images, and then use the method of distance measurement to compare these query images with each image in CBIR systems 2 to 5 The similarity of the image, and then arrange the images in the CBIR system 2-5 according to the order of the similarity of the corresponding query image from high to low. Set the number of images returned at one time as 1, 2, ..., n, ... 2400, get the precision rate and recall rate of each image, and then calculate the precision rate and recall rate of 300 query images The mean values yield the curves numbered II, III, IV and V in FIG. 2 . Figure 2 shows that among the relevant images returned by each CBIR system to the database that is less than or equal to 90%, the precision rate obtained by using the bag-of-words model of the present invention is much higher than that obtained by other texture feature extraction methods.

(3) Also use the above 300 query images to submit to the CBIR system 6 and CBIR system 7 constructed in step 1 respectively. System 6 constructs the bag-of-words model of these query images according to the spatial pyramid method, and system 7 adds The bag-of-words model of these query images is constructed by using the spatial coordinates of features, and then the similarity between these query images and each image in CBIR systems 6 and 7 is compared using the distance measure method, and then the CBIR systems 6 and 7 are compared Images are arranged in descending order of similarity to their corresponding query images. Set the number of images returned at one time as 1, 2, ..., n, ... 2400, get the precision rate and recall rate of each image, and then calculate the precision rate and recall rate of 300 query images The mean values yield the curves numbered II and III in FIG. 3 . Fig. 3 has shown that in each CBIR system returns less than or equal to 90% of the relevant images of the database, the accuracy rate obtained by adopting the bag-of-words model of the present invention is higher than that obtained by the other two bag-of-words methods adding spatial information Rate.

Claims (3)

1. A method for constructing a bag-of-words model of a medical focus image comprises the following steps:

(1) reading medical focus images with focus outlines drawn in a database, wherein the read medical focus images contain at least 50 images of each focus type, and each medical focus image is processed as follows:

(1.1) firstly, performing one-dimensional Gaussian smoothing on a focus contour line, then respectively taking pixel points of which the number is the same as that of pixel points which are taken from the focus contour line to the inside and the outside of a focus along the normal line of the focus contour as starting points, then respectively taking the pixel points taken from the normal line direction of each focus contour as columns, arranging the pixel points clockwise or anticlockwise, taking the pixel points at the same position in the normal line direction of each focus contour as lines, and arranging the pixel points in the sequence from the inside of the focus contour to the outside of the focus contour or from the outside of the focus contour to the inside of the focus contour to obtain a transformation image of a boundary region of the medical focus image; then, taking pixel points at intervals of 0-5 pixel point distances in a transformed image as centers to expand outwards, obtaining a series of small squares with the pixel point array size of 5 × 5, 7 × 7 or 9 × 9, turning each row of pixel points in the small squares by 90 degrees anticlockwise, then arranging the pixel points in the small squares into a row of queues from top to bottom, then respectively replacing the pixel points in the transformed image in each row of queues with corresponding gray values, and respectively giving zero to the gray values of the pixel points outside the transformed image to obtain a gray value vector A1 of the focus boundary region;

(1.2) in a focus area, extending outwards by taking pixel points at the distance of 0-5 pixel points as centers from pixel points at the edge of a focus contour line, acquiring a series of small squares with the pixel point array size of 5 × 5, 7 × 7 or 9 × 9, rotating each row of pixel points in the small squares by 90 degrees anticlockwise, arranging the pixel points in a row of queues from top to bottom, respectively replacing the pixel points in the focus contour line in each row of queues with corresponding gray values, and respectively giving zero to the gray values of the pixel points outside the focus contour line to obtain a gray value vector A2 of the focus area;

(1.3) specifying gray value vectors A1 of the lesion boundary regions of all the medical lesion images obtained in the step (1.1) as one type, specifying gray value vectors A2 of the lesion regions of all the medical lesion images obtained in the step (1.2) as another type, and then respectively carrying out k-means clustering on the gray value vectors A2 to obtain a lesion region dictionary and a lesion boundary region dictionary which comprise words formed by clustering centers;

(2) taking any medical focus image in the database, and respectively constructing a word bag B2 in a focus area and a word bag B1 in a focus boundary area; wherein,

(2.1) the construction method of the bag of words B2 in the focus area comprises the following steps: expanding outwards by taking each pixel point in the focus area as the center, acquiring a series of small squares with the same size as the small squares in the step (1.2), then obtaining a gray value vector A2 ' of secondary operation of the focus area of the medical focus image by the same method in the step (1.2), then respectively calculating the Euclidean distance between each gray value vector A2 ' and each word in a focus area dictionary according to the sequence of the line queue, and quantizing each gray value vector A2 ' to the word with the minimum Euclidean distance to obtain the frequency of each word in the focus area dictionary, wherein the one-dimensional vector formed by the frequency number is the bag B2 of the focus area of the medical focus image;

(2.2) the construction method of the bag of words B1 in the lesion boundary region comprises the following steps: processing the focus boundary region of the medical focus image according to the same method of the step (1.1) to obtain a transformation image of the boundary region, expanding outwards by taking each pixel point in the transformed image as a center to obtain a series of small squares with the same size as the small squares in the step (1.1), obtaining a gray value vector A1' of secondary operation of a focus boundary region of the medical focus image by the same method in the step (1.1), then respectively calculating the Euclidean distance between each gray value vector A1' and each word in the lesion boundary region dictionary according to the sequence of the line queue, and quantizes each gray value vector A1' to the word with the minimum Euclidean distance to the word to obtain the frequency of each word in the focus boundary region dictionary, the one-dimensional vector formed by the frequency number is a word bag B1 of the lesion boundary area of the medical lesion image;

(2.3) arranging two vectors representing the bag-of-words B1 and the bag-of-words B2 end to end into a vector to obtain a bag-of-words model of the medical lesion image.

2. The method as claimed in claim 1, wherein the size of the small square is 7 × 7 pixel point array.

3. The method of claim 1 or 2, wherein the k-value of the k-means cluster is 1000.

Images