A prostate image segmentation algorithm combining deep learning with traditional learning

Abstract

The present invention discloses a prostate image segmentation algorithm that combines deep learning with traditional learning. The ASSP module is introduced into the U-Net model. The improved U-Net network performs coarse segmentation of the prostate image, and then fine segmentation of the prostate is performed using an improved normal vector boundary operator. The designed algorithm for fine segmentation of the prostate image based on the grayscale difference of the semicircular area improves on the original Hodge algorithm, expanding the grayscale information in the field of a single normal vector and incorporating more pixel information. A method for detecting and reconstructing ectopic sites on the prostate boundary is designed. The curvature of a single point is represented by the slope difference of the line segments adjacent to it, and the determination of whether it is an ectopic site is automatically made. This solution solves the problem of the existing technology consuming a lot of time and being inefficient in the prostate segmentation task, and has the characteristics of faster and more accurate automated segmentation, simple and convenient operation process, and clear and intuitive segmentation results.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition technology, and in particular relates to a prostate image segmentation algorithm that combines deep learning with traditional learning. Background Art

[0002] Prostate disease is one of the most common diseases worldwide, with a particularly high incidence among men, becoming a growing threat to men's health. Prostate disease treatment is a key topic in current medical research. Segmenting the prostate from ultrasound images, as well as locating and measuring it, is crucial for prostate disease treatment. Good segmentation accuracy is beneficial for subsequent treatment outcomes.

[0003] Currently, popular prostate ultrasound image segmentation methods fall into two categories. Traditional methods, such as Sobel, extract image edge information based on specific problem-specific boundary operators. Deep learning methods, exemplified by the U-net network structure and its derivatives, have gained popularity in recent years. While the former's detection effectiveness is easily affected by image quality, the latter often requires large datasets for iterative training, resulting in significant computational overhead.

[0004] Therefore, it is necessary to design a prostate image segmentation algorithm that combines deep learning with traditional learning to solve the above problems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a prostate image segmentation algorithm that combines deep learning with traditional learning. This solution solves the problem of the existing technology consuming a lot of time and being inefficient in prostate segmentation tasks. It has the characteristics of faster and more accurate automated segmentation effect, simple and convenient operation process, and clear and intuitive segmentation results.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A prostate image segmentation algorithm combining deep learning with traditional learning includes the following steps:

[0008] S1, based on the improved U-Net, performs coarse segmentation of prostate images: Image segmentation requires outputting information for each pixel, which is then restored to the original image size during the subsequent upsampling process. This processing method uses a pooling process to reduce the size of the feature map and an upsampling process to restore the image size, but this also results in information loss, making the segmentation effect less than ideal. Therefore, to address this issue, the algorithm in this paper introduces the ASPP module; the use of dilated convolution can expand the acceptable area without sacrificing the resolution of the feature map and does not increase the number of parameters. This method can effectively obtain global information of the image and has better application value than ordinary convolution.

[0009] S101, introduce ASPP module:

[0010] This module uses multiple parallel dilated convolution layers; dilated convolution is a convolution algorithm that can achieve dense feature extraction and field of view expansion. Compared with ordinary convolution, it introduces a span of r-1 zeros in the convolution kernel, thereby effectively expanding the receptive field; specifically, a dilated convolution with a convolution kernel size of k and a dilation rate of r introduces r-1 zeros between consecutive filter values, resulting in an increase in the receptive field size to:

[0011] K=k+(k-1)×(r-1)

[0012] Where K represents the receptive field size, k represents the convolution kernel size, and r represents the expansion rate;

[0013] S102, using ASPP to replace the bottom layer of the U-Net network:

[0014] The improved U-Net prostate coarse segmentation algorithm, by leveraging the advantages of the ASPP module, consists of an encoder network, a decoder network, and the ASPP network. The U-Net architecture mainly consists of two parts: the encoder and the decoder. The input image size is 512*512. The batch normalization module is added after the convolutional layers of the encoder and decoder to accelerate convergence, and the ReLU function is used for activation. The encoding module uses pooling to downsample the feature map, and the decoding module uses a transposed convolution operation to upsample the feature map. Skip connections are used to connect the upsampled and downsampled feature maps.

[0015] S103, ASPP performs multiple sampling operations on the feature map of the previous layer:

[0016] ASPP performs five sampling operations on the feature map of the previous layer, using a 1x1 convolution kernel to improve nonlinear expression capabilities and introducing a 3x3 dilated convolution to expand the receptive field. The last convolution operation introduces average pooling, feeds the feature map into a 1x1 convolution kernel, and uses bilinear interpolation to upsample it to the original size of the image. Different dilation rates correspond to feature maps of different scales, which are obtained by the pyramid model. Each branch is independent of each other, with a smaller dilation rate used to associate small-scale information and a larger dilation rate used to associate large-scale information. In the final stage of the network, the five multi-scale feature maps are spliced ​​together and fed into a 1x1 convolution layer for fusion. The final output feature map is fed into the decoding module.

[0017] S2, completes the fine segmentation of the prostate image based on the grayscale difference of the semicircular area:

[0018] S201, improved normal vector boundary operator:

[0019] After the coarse segmentation is completed, the subsequent task is to further complete the fine segmentation of the prostate image. Due to certain characteristics of the prostate image itself, the image may have defects such as artifacts, which will lead to poor contrast at the prostate boundary. Hodge has designed a method to improve the normal vector boundary operator to solve this problem of insufficient contrast:

[0020] First, according to the Hodge method, the rough contour curve of the prostate image is represented as a point set:

[0021] H={(x1,y1),(x2,y2),…,(x n ,y n )}

[0022] Create a normal vector at each point, and set the normal vector to a point set with the same density as the pixel value; randomly select a point U(x,y) on the normal vector, with the index value key, and take a total of 2m grayscale values ​​in the positive and negative directions of point U, and set each grayscale point to a point set f i =[f key+m ,f key+m-1 ,…f key-m ], where f i Represents the grayscale value of the point with index value i;

[0023] Then calculate the cumulative grayscale difference between the upper and lower regions of a point along the normal vector direction and move the point on the normal vector to form a set C key The grayscale difference formula of the grayscale difference of the cumulative grayscale value is as follows:

[0024]

[0025] Where key represents the index value of the point, C key Grayscale difference set, i represents the value range of the upper and lower region boundary operators; the index value formula is as follows:

[0026] key=argmaxC key

[0027] Among them, key represents the index value, argmax() function represents the maximum value independent variable point set, C key Grayscale difference set;

[0028] The formula can be used to determine the point with the best segmentation effect. All discretized points of the curve on the prostate image are traversed once through the method of steps S201 to S202 to obtain the approximate boundary contour of the prostate.

[0029] S202, design a segmentation method for the semicircular area operator:

[0030] After processing by the Hodge method, most of the point sets on the image are already on the true boundary, but there are still a few segmentation points that have not been located on the true boundary. Hodge only takes into account the grayscale information of a single normal vector. The noise points on the image will cover the normal vector, resulting in inconsistent segmentation points at the front and back boundaries, and the image boundary is very rough. Therefore, on this basis, this solution continues to make further improvements based on Hodge's improvements. Since Hodge only processes a single normal vector, this method fully considers the adjacent normal vectors of the segmentation point, corrects the position of the segmentation point, and designs a segmentation method for the semicircular area operator:

[0031] Set a normal vector to The two adjacent normal vectors are set to exist The index value is key, and the point is represented as Two adjacent points are and point The boundary operator formula is as follows:

[0032]

[0033] Among them, C represents the boundary operator, b represents the point The number of points in the unilateral domain, m represents the number of domain normal vectors;

[0034] S3, ectopic site detection and reconstruction:

[0035] S301, ectopic point detection: The pixel boundaries of prostate images are jagged, making it difficult to measure curvature. However, the curvature at a single point can be defined by the line connecting adjacent points. Therefore, the curvature of a single point can be determined by the slope difference between the line segments connecting two adjacent points. The specific method is as follows:

[0036] Any single point on the boundary (x t ,y t )The curvature formula is as follows:

[0037]

[0038] Where mod represents the modular operation, tan() represents the tangent function, and x t+k with x t-k Represents the horizontal coordinate x of a single point t The horizontal coordinates of two adjacent points, y t+k with y t-k Represents the single point ordinate y t The vertical coordinates of two adjacent points;

[0039] Based on the previous normal vector operator detection, the next step is to proceed. The detection process is as follows:

[0040] When the boundary operator C detects a point F i When the two adjacent points F of the point are combined i-1 and F i+1 The spatial position relationship of F determines whether the point is actually on the boundary; i , F i-1 , F i+1 Connect the lines and form an angle with the horizontal direction, ∠1, ∠2, which are straight lines F i F i-1 , straight line F i F i+1 The angle formed by the horizontal direction; through these angles, the corresponding straight line slope can be obtained, and then by calculating the rate of change of the slope of the two straight lines, the point F can be obtained. i The curvature is obtained;

[0041] Furthermore, by performing curvature statistical analysis on the boundary points, the curvature variation range is set to between 0° and 30°. When the boundary point is within the range, it is identified as a normal point, otherwise it is an abnormal point. This method can detect the abnormal points framed in the figure.

[0042] Furthermore, when detecting ectopic points, the initial detection point is extremely critical. The point with the largest grayscale difference on the normal vector boundary and its adjacent points are selected as the initial detection point.

[0043] S302, ectopic site reconstruction: Since the boundary of the prostate image is approximately smooth, flat, and continuous during segmentation, when ectopic sites have been detected, the ectopic sites are removed and reconstructed using the following method:

[0044] After the prostate image is segmented by the normal vector, its boundary is composed of 100 points. The removed point can be replaced by the midpoint between the two points before and after the point. Because the boundary is approximately smooth, gentle, and continuous, the boundary formed around the point is similar to the real boundary, and the deviation can be ignored. Using the idea of ​​calculus, the smooth and continuous curve composed of 100 points is refined into segments of straight lines. Using the mathematical knowledge midpoint formula, F i-1 , F i+1 The point F removed between the two points i Reconstructed through the midpoint; the midpoint formula is as follows:

[0045]

[0046] where X a Indicates the horizontal coordinate of the point, Y a Indicates the vertical coordinate of the point.

[0047] F i-1 , F i+1Substitute the horizontal and vertical coordinates into the formula to obtain the reconstructed point F z .

[0048] The beneficial effects of the present invention are:

[0049] 1. To achieve better segmentation results, this paper combines traditional methods with deep learning. First, the prostate is roughly segmented using an improved U-Net network, which can better detect the true boundary points and achieve rough segmentation of the prostate.

[0050] 2. Improve the normal vector contour operator. The original Hodge method has uneven boundaries. This method expands the grayscale accumulation area. Compared with a single normal vector segment, this method incorporates more pixel information, greatly reducing the number of false positives in the segmentation results, improving segmentation accuracy and the smoothness of segmentation edges, and enhancing segmentation efficiency. This not only effectively improves segmentation accuracy, but also has strong generalization capabilities.

[0051] 3. Detect and reconstruct these ectopic points on the boundary. Due to the smooth and continuous nature of the prostate boundary, this method uses slope difference to represent curvature and detect ectopic points. Considering the continuity, smoothness, and gentleness of the boundary, the midpoint between two points is used to reconstruct ectopic points, improving segmentation accuracy. The simple and convenient operation frees up medical personnel's hands, and the segmentation results are clear and intuitive. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of the invention;

[0053] Figure 2 Schematic diagram of the dilated convolution in the present invention;

[0054] Figure 3 It is a schematic diagram of the U-Net network structure in the present invention;

[0055] Figure 4 It is a schematic diagram of the network structure of the ASPP module in the present invention;

[0056] Figure 5 Schematic diagram of the improved U-Net prostate segmentation effect diagram in the present invention;

[0057] Figure 6 Schematic diagram of the segmentation method of the semicircular area operator in the present invention;

[0058] Figure 7 This is a schematic diagram of the segmentation effect diagram based on the semicircular area operator in the present invention;

[0059] Figure 8 is a schematic diagram of the slope of the boundary point in step S301 of the present invention;

[0060] Figure 9 This is a schematic diagram of the effect diagram of the point after the reconstruction in step S302 of the present invention;

[0061] Figure 10 This is a schematic diagram of the effect of traversing 100 points on the boundary to perform ectopic point detection and reconstruction in the present invention; DETAILED DESCRIPTION

[0062] Example 1:

[0063] like Figure 1 In the present invention, a prostate image segmentation algorithm combining deep learning with traditional learning is proposed, which includes the following steps:

[0064] S1, based on the improved U-Net, performs coarse segmentation of prostate images: Image segmentation requires outputting information for each pixel, which is then restored to the original image size during the subsequent upsampling process. This processing method uses a pooling process to reduce the size of the feature map and an upsampling process to restore the image size, but this also results in information loss, making the segmentation effect less than ideal. Therefore, to address this issue, the algorithm in this paper introduces the ASPP module; the use of dilated convolution can expand the acceptable area without sacrificing the resolution of the feature map and does not increase the number of parameters. This method can effectively obtain global information of the image and has better application value than ordinary convolution.

[0065] S101, introduce ASPP module:

[0066] like Figure 2 As shown in the figure, the module uses multiple parallel dilated convolution layers. Dilated convolution is a convolution algorithm that can achieve dense feature extraction and field of view expansion. Compared with ordinary convolution, it introduces a span of r-1 zeros in the convolution kernel, thereby effectively expanding the receptive field. Specifically, a dilated convolution with a convolution kernel size of k and a dilation rate of r introduces r-1 zeros between consecutive filter values, resulting in an increase in the receptive field size to:

[0067] K=k+(k-1)×(r-1)

[0068] Where K represents the receptive field size, k represents the convolution kernel size, and r represents the expansion rate;

[0069] S102, using ASPP to replace the bottom layer of the U-Net network:

[0070] like Figure 3As shown in the figure, the improved U-Net prostate rough segmentation algorithm, by borrowing the advantages of the ASPP module, the network structure mainly consists of three parts: encoder network, decoder network and ASPP; the U-Net architecture mainly consists of two parts, encoder and decoder, with an input image size of 512*512; a BN module is added after the convolutional layer of the encoder and decoder to accelerate the convergence speed, and the ReLU function is used for activation; the encoding module uses pooling to downsample the feature map, and the decoding module upsamples the feature map through the transposed convolution operation, and uses skip connections to connect the feature maps between the upsampled and downsampled ones;

[0071] S103, ASPP performs multiple sampling operations on the feature map of the previous layer:

[0072] like Figure 4 As shown in the figure, ASPP performs five sampling operations on the feature map of the previous layer, improves the nonlinear expression ability by using a 1x1 convolution kernel, and introduces a 3x3 hole convolution to expand the receptive field; the last convolution operation introduces average pooling, feeds the feature map into a 1x1 convolution kernel, and uses bilinear interpolation to upsample to the original size of the image; different expansion rates correspond to feature maps of different scales, which are obtained by the pyramid model; each branch is independent of each other, a smaller expansion rate is used to associate small-scale information, and a larger expansion rate is used to associate large-scale information; in the final stage of the network, the five multi-scale feature maps are spliced ​​together and fed into a 1x1 convolution layer for fusion, and the final output feature map is fed into the decoding module, as shown in the figure. Figure 5 As shown;

[0073] S2, completes the fine segmentation of the prostate image based on the grayscale difference of the semicircular area:

[0074] S201, improved normal vector boundary operator:

[0075] After the coarse segmentation is completed, the subsequent task is to further complete the fine segmentation of the prostate image. Due to certain characteristics of the prostate image itself, the image may have defects such as artifacts, which will lead to poor contrast at the prostate boundary. Hodge has designed a method to improve the normal vector boundary operator to solve this problem of insufficient contrast:

[0076] First, according to the Hodge method, the rough contour curve of the prostate image is represented as a point set:

[0077] H={(x1,y1),(x2,y2),…,(x n ,y n )}

[0078] Create a normal vector at each point, and set the normal vector to a point set with the same density as the pixel value; randomly select a point U(x,y) on the normal vector, with the index value key, and take a total of 2m grayscale values ​​in the positive and negative directions of point U, and set each grayscale point to a point set f i =[f key+m ,f key+m-1 ,…f key-m ], where f i Represents the grayscale value of the point with index value i;

[0079] Then calculate the cumulative grayscale difference between the upper and lower regions of a point along the normal vector direction and move the point on the normal vector to form a set C key The grayscale difference formula of the grayscale difference of the cumulative grayscale value is as follows:

[0080]

[0081] Where key represents the index value of the point, C key Grayscale difference set, i represents the value range of the upper and lower region boundary operators; the index value formula is as follows:

[0082] key=argmaxC key

[0083] Among them, key represents the index value, argmax() function represents the maximum value independent variable point set, C key Grayscale difference set;

[0084] The formula can be used to determine the point with the best segmentation effect. All discretized points of the curve on the prostate image are traversed once through the method of steps S201 to S202 to obtain the approximate boundary contour of the prostate.

[0085] S202, design a segmentation method for the semicircular area operator:

[0086] like Figure 6 As shown in the figure, after being processed by the Hodge method, most of the point sets on the image are already on the true boundary, but there are a few segmentation points that have not yet been located on the true boundary. Hodge only takes the grayscale information of a single normal vector into consideration. The noise points on the image will cover the normal vector, resulting in inconsistent segmentation points at the front and back boundaries, and the image boundary is very rough. Therefore, on this basis, this scheme continues to make further improvements based on the Hodge improvement. Since Hodge only processes a single normal vector, this method fully considers the adjacent normal vectors of the segmentation point, corrects the position of the segmentation point, and designs a segmentation method for the semicircular area operator:

[0087] Set a normal vector to The two adjacent normal vectors are set to exist The index value is key, and the point is represented as Two adjacent points are and point The boundary operator formula is as follows:

[0088]

[0089] Among them, C represents the boundary operator, b represents the point The number of unilateral domain points, m represents the number of domain normal vectors; the specific effect is as follows Figure 7 As shown;

[0090] S3, out-of-bounds detection and reconstruction: C calculates the grayscale difference between the upper and lower semicircular frames. Compared to the grayscale difference on a single normal vector calculated by Hodge, this contains much more grayscale information, which avoids the problem of a small number of boundary points not being on the true boundary due to local noise. However, the resulting image still contains a small number of out-of-bounds points, so the next step is to detect and reconstruct the out-of-bounds points.

[0091] S301, ectopic point detection: The pixel boundaries of prostate images are jagged, making it difficult to measure curvature. However, the curvature at a single point can be defined by the line connecting adjacent points. Therefore, the curvature of a single point can be determined by the slope difference between the line segments connecting two adjacent points. The specific method is as follows:

[0092] Any single point on the boundary (x t ,y t )The curvature formula is as follows:

[0093]

[0094] Where mod represents the modular operation, tan() represents the tangent function, and x t+k with x t-k Represents the horizontal coordinate x of a single point t The horizontal coordinates of two adjacent points, y t+k with y t-k Represents the single point ordinate y t The vertical coordinates of two adjacent points;

[0095] Based on the previous normal vector operator detection, the next step is to proceed. The detection process is as follows:

[0096] When the boundary operator C detects a point F i When the two adjacent points F of the point are combined i-1 and F i+1 The spatial position relationship of F determines whether the point is actually on the boundary; i, F i-1 , F i+1 Connect the lines and form an angle with the horizontal direction, such as Figure 8 ∠1 and ∠2 are straight lines F i F i-1 , straight line F i F i+1 The angle formed by the horizontal direction; through these angles, the corresponding straight line slope can be obtained, and then by calculating the rate of change of the slope of the two straight lines, the point F can be obtained. i The curvature is obtained;

[0097] Furthermore, by performing curvature statistical analysis on the boundary points, the curvature variation range is set to between 0° and 30°. When the boundary point is within the range, it is identified as a normal point, otherwise it is an abnormal point. This method can detect the abnormal points framed in the figure.

[0098] Furthermore, when detecting ectopic points, the initial detection point is extremely critical. The point with the largest grayscale difference on the normal vector boundary and its adjacent points are selected as the initial detection point.

[0099] S302, ectopic site reconstruction: Since the boundary of the prostate image is approximately smooth, flat, and continuous during segmentation, when ectopic sites have been detected, the ectopic sites are removed and reconstructed using the following method:

[0100] After the prostate image is segmented by the normal vector, its boundary is composed of 100 points. The removed point can be replaced by the midpoint between the two points before and after the point. Because the boundary is approximately smooth, gentle, and continuous, the boundary formed around the point is similar to the real boundary, and the deviation can be ignored. Using the idea of ​​calculus, the smooth and continuous curve composed of 100 points is refined into segments of straight lines. Using the mathematical knowledge midpoint formula, F i-1 , F i+1 The point F removed between the two points i Reconstructed through the midpoint; the midpoint formula is as follows:

[0101]

[0102] where X a Indicates the horizontal coordinate of the point, Y a Indicates the vertical coordinate of the point.

[0103] F i-1 , F i+1 Substitute the horizontal and vertical coordinates into the formula to obtain the reconstructed point F z Specific effects such as Figure 9 shown.

[0104] After traversing 100 points on the boundary as described above, the effect after ectopic detection and reconstruction is as follows Figure 10 shown.

[0105] Example 2:

[0106] Dataset Preparation: This method uses a dataset of 140 prostate ultrasound images provided by a hospital. These images are derived from clinical examination data. Each image has a resolution of 576 x 768 pixels, with each pixel measuring 0.138 mm x 0.138 mm. Each prostate ultrasound image was manually labeled by an experienced clinical physician, and these labeled images serve as labels.

[0107] Dataset division: The 140 data sets are randomly divided, 100 of which are used as training sets and 40 as test sets. The training set is used to train the model, and the test set is used to test the model effect.

[0108] Experimental environment: Windows 10 operating system, hardware environment CPU is Intel (R) Core i5-6300HQ @ 2.3GHz, graphics card is GeForce GTX1080 graphics card, 16GB memory and 512GB solid state drive.

[0109] Augmentation of the training dataset: Prostate image data is diverse, and appropriate data augmentation is required. 100 images can be expanded to 1,000 images by scaling, cropping, rotating, and flipping.

[0110] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The embodiments and features in the embodiments of this application may be arbitrarily combined with each other unless they conflict. The scope of protection of the present invention shall be the technical solutions described in the claims, including equivalent alternatives to the technical features of the technical solutions described in the claims. Equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A prostate image segmentation algorithm combining deep learning with traditional learning, characterized by: The following steps are involved: S1, based on the improved U-Net, completes the rough segmentation of the prostate image: S101, introduce ASPP module: This module uses multiple parallel dilated convolution layers. Dilated convolution is a convolution algorithm that can achieve dense feature extraction and expand the field of view. Compared with ordinary convolution, it introduces a span of r-1 zeros in the convolution kernel, thereby effectively expanding the receptive field. Specifically, the convolution kernel size is k , a dilated convolution with a dilation rate of r introduces r-1 zeros between consecutive filter values, resulting in an increase in the receptive field size to: ； in represents the receptive field size, k represents the convolution kernel size, r represents the expansion rate; S102, using ASPP to replace the bottom layer of the U-Net network: The improved U-Net prostate rough segmentation algorithm, by borrowing the advantages of the ASPP module, the network structure consists of three parts: encoder network, decoder network and ASPP; the U-Net architecture consists of two parts, encoder and decoder, and the input image size is 512*512; A BN module is added after the convolutional layer of the encoder and decoder to accelerate convergence, and the ReLU function is used for activation; the encoding module uses pooling to downsample the feature map, and the decoding module upsamples the feature map through the transposed convolution operation, and uses skip connections to connect the feature maps between the upsampled and downsampled ones; S103, ASPP performs multiple sampling operations on the feature map of the previous layer: The nonlinear expression capability is improved by using a 1x1 convolution kernel, and the receptive field is expanded by introducing a 3x3 dilated convolution. The last convolution operation introduces average pooling, feeds the feature map into a 1x1 convolution kernel, and uses bilinear interpolation to upsample it to the original size of the image. Different dilation rates correspond to feature maps of different scales, which are obtained by the pyramid model. Each branch is independent of each other, with a small dilation rate used to associate small-scale information and a large dilation rate used to associate large-scale information. In the final stage of the network, the five multi-scale feature maps are spliced ​​together and fed into a 1x1 convolution layer for fusion. The final output feature map is fed into the decoding module. S2, completes the fine segmentation of the prostate image based on the grayscale difference of the semicircular area: S201, improved normal vector boundary operator: First, according to the Hodge method, the rough contour curve of the prostate image is represented as a point set: ； Create a normal vector at each point, and set the normal vector to a point set with the same density as the pixel value; select a random point on the normal vector , the index value is key, in U Take a total of 2m grayscale values ​​in the positive and negative directions of the point, and set each grayscale point as a point set ,in Represents the grayscale value of the point with index value i; Then calculate the cumulative grayscale difference between the upper and lower areas of a point along the normal vector direction and move the point on the normal vector to form a set The grayscale difference formula of the grayscale difference of the cumulative grayscale value is as follows: ； Where key represents the index value of the point. Grayscale difference set, i represents the value range of the upper and lower region boundary operators; The formula for the index value is as follows: key=argmax ; Among them, key represents the index value, and the argmax() function represents the maximum independent variable point set. Grayscale difference set; The formula can be used to determine the point with the best segmentation effect. All discretized points on the curve on the prostate image are traversed once through the method of steps S201 to S202 to obtain the approximate boundary contour of the prostate. S202, design a segmentation method for the semicircular area operator: Set a normal vector to , the two adjacent normal vectors are set to 、 ;exist The index value is key, and the point is represented as , two adjacent points are and ;point The boundary operator formula is as follows: ； in, represents the boundary operator, b represents the point The number of points in the unilateral domain, m represents the number of domain normal vectors; S3, ectopic site detection and reconstruction: S301, ectopic point detection: The pixel boundaries of prostate images are jagged, making it difficult to measure curvature. However, the curvature at a single point can be defined by the line connecting adjacent points. Therefore, the curvature of a single point can be determined by the slope difference between the line segments connecting two adjacent points. The specific method is as follows: Any single point on the boundary The curvature formula is as follows: ； Where mod represents the modular operation, tan() represents the tangent function, and Represents the horizontal coordinates of a single point The horizontal coordinates of two adjacent points, and Represents the vertical coordinate of a single point The vertical coordinates of two adjacent points; Based on the previous normal vector operator detection, the next step is to proceed. The detection process is as follows: When the boundary operator C detects a point When , the two adjacent points of the point will be combined and The spatial position relationship of determines whether the point is actually on the boundary; , , Connect the lines and form an angle with the horizontal direction. , respectively, straight lines ,straight line The angle formed by the horizontal direction; through these angles, the corresponding straight line slope can be obtained, and then the point can be obtained by calculating the rate of change of the slope of the two straight lines. The curvature at S302, ectopic site reconstruction: Since the boundary of the prostate image is approximately smooth, flat, and continuous during segmentation, when ectopic sites have been detected, the ectopic sites are removed and reconstructed using the following method: After the prostate image is segmented by the normal vector, its boundary is composed of 100 points; the eliminated point can be replaced by the midpoint between the two points before and after the point. Because the boundary is approximately smooth, gentle, and continuous, the boundary formed around the point is similar to the real boundary, and the deviation can be ignored; using the idea of ​​calculus, the smooth and continuous curve composed of 100 points is refined into a straight line, and the midpoint formula of mathematical knowledge is used to , The point removed between two points Reconstructed through the midpoint; the midpoint formula is as follows: ； in represents the horizontal coordinate of the point, Indicates the vertical coordinate of the point; Will , Substitute the horizontal and vertical coordinates into the formula to obtain the reconstructed point .

Images