A medical image segmentation method and a computer readable storage medium

Abstract

The application provides a medical image segmentation method and a computer readable storage medium, and the medical image segmentation method comprises the following steps: obtaining multiple medical images, classifying the medical images into training images, verification images and test images, and performing change processing on the training images; performing centroid extraction on the training images; calculating the polar coordinates of the pixel points of the training images in a polar coordinate system under a rectangular coordinate system with the image centroid (c x ,c y ) as the center; performing segmentation on the changed images based on an improved UNet network; calculating the rectangular coordinates of the pixel points of the segmented images in a rectangular coordinate system under a polar coordinate system with the image centroid (c x ,c y ) as the center to obtain a result image in the rectangular coordinate system. After the above technical scheme is adopted, the efficiency of automatic medical evaluation can be greatly improved, and the pressure of manually segmenting medical images can be relieved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and more particularly to a medical image segmentation method and a computer-readable storage medium. Background Technology

[0002] With the continuous advancement of computer technology, digital imaging technology has become increasingly mature and widespread, directly driving the rapid development of the medical imaging field. Medical imaging technology directly provides doctors with information about internal lesions in a patient's body by displaying images of specific areas, making it an indispensable auxiliary tool for clinical diagnosis and treatment. Medical image segmentation can effectively provide doctors with detailed information about the size, shape, and location of target objects, thus playing a crucial role in the clinical disease diagnosis and screening process.

[0003] However, when doctors manually segment and interpret medical images, they face problems such as slow speed and inaccurate segmentation results. Moreover, the accuracy of the segmentation results largely depends on the doctor's level of expertise, which greatly limits the chances of discovering and treating potential patients in remote areas or small hospitals and clinics. Therefore, automated medical image segmentation is of great significance.

[0004] The high-precision and high-speed recognition and segmentation of medical images has attracted increasing attention from researchers. The collaboration of scholars from fields such as applied mathematics, statistics, image processing, and computer science has led to the rapid development and remarkable achievements in medical image processing, evolving from traditional fully manual segmentation and recognition to semi-automatic and fully automatic segmentation processing.

[0005] Over the past decade, deep learning algorithms based on neural networks have demonstrated superior performance in numerous medical image segmentation tasks. As a typical supervised learning model, deep learning algorithms based on neural networks are highly dependent on the amount of data samples. However, in medical image segmentation tasks, on the one hand, medical images themselves are difficult to collect, and on the other hand, manually annotating images is time-consuming and labor-intensive, resulting in a scarcity of effective data, which directly affects the final model performance.

[0006] Therefore, a new medical image segmentation method is needed to compensate for the scarcity of existing image samples and greatly promote the progress of intelligent medical care. Summary of the Invention

[0007] In order to overcome the above-mentioned technical defects, the purpose of this invention is to provide a medical image segmentation method and a computer-readable storage medium, which can greatly improve the efficiency of automated medical assessment and alleviate the pressure of manual medical image segmentation.

[0008] This invention discloses a medical image segmentation method, comprising the following steps:

[0009] Multiple medical images are acquired, classified into training images, validation images, and test images, and the training images are modified to increase the sample size of the training images.

[0010] Centroid extraction is performed on the training images to obtain the image centroid (c) of the training images. x ,c y );

[0011] With image centroid (c x ,c y In a Cartesian coordinate system centered at , calculate the polar coordinates of the pixels in the training image in the polar coordinate system to obtain the transformed image in the polar coordinate system;

[0012] A segmented image is obtained by segmenting a variable image using an improved UNet network.

[0013] With the image centroid (c x ,c y In a polar coordinate system centered at , calculate the rectangular coordinates of the pixels in the segmented image in a rectangular coordinate system to obtain the resulting image in a rectangular coordinate system.

[0014] Preferably, the steps of acquiring multiple medical images, classifying the medical images into training images, validation images, and test images, and performing modification processing on the training images to increase the sample size of the training images include:

[0015] After acquiring multiple medical images, all medical images are divided into training images, validation images, and test images according to an 8:1:1 ratio.

[0016] Perform any one or any of the following changes:

[0017] Select n% of the training images, perform flipping processing, and record it as the first transformed image;

[0018] Select m% of the training image, translate the training image along a random direction by a% of the length of the random direction, and record it as the second transformed image;

[0019] Select p% of the training image and randomly shrink or enlarge b% of the original size of the training image, and record it as the third transformed image;

[0020] Select q% from the training image and rotate it around a random pixel in the training image by t°, and record it as the fourth transformed image;

[0021] Select r% of the training image to perform a grayscale vanishing operation, and record it as the fifth transformed image;

[0022] The first, second, third, fourth, and / or fifth modified images are combined into an augmented image and added to the training image.

[0023] Preferably, the centroid of the training image is extracted to obtain the image centroid (c) of the training image. x ,c y The steps include:

[0024] For the training image I, construct the following operator: M ij =∑ x,y I(x,y)x i y j Where I(x,y) represents the pixel value at pixel (x,y) in training image I, and the total pixel value M of training image I is calculated. ij ;

[0025] Calculate the image centroid (c) using the following operator. x ,c y )for:

[0026] c x =M 10 / M 00 ,c y =M 01 / M 00 .

[0027] Preferably, the centroid of the training image is extracted to obtain the image centroid (c) of the training image. x ,c y The steps include:

[0028] An image segmentation model is established based on an improved UNet network.

[0029] or

[0030] A heatmap is built based on the target object mask within the training image, with the image centroid (c) as the basis. x ,c y Using as the center point, the image centroid (c) is obtained according to a two-dimensional Gaussian distribution. x ,c y The pixel size near )

[0031] The intensity of the heatmap is calculated using a neural network with a stacked hourglass architecture, so that the center point of the heatmap with the highest intensity is the final centroid.

[0032] Preferably, the image centroid (c) x ,c y The steps for calculating the polar coordinates of pixels in the training image in a polar coordinate system, centered at a Cartesian coordinate system, to obtain the transformed image in polar coordinates include:

[0033] In a Cartesian coordinate system, for a pixel (x, y) in a training image, the following function is defined:

[0034]

[0035]

[0036] For a training image I of size H×W, the centroid (c) of the image is calculated according to the following operator. x ,c y The polar coordinates of the pixels in the training image I' centered in the polar coordinate system.

[0037]

[0038]

[0039] Preferably, with the image centroid (c) as the reference point, x ,c y The steps for calculating the rectangular coordinates of pixels in a segmented image in a Cartesian coordinate system, centered at a polar coordinate system, to obtain the resulting image in a Cartesian coordinate system include:

[0040] Based on the following operator, a value of size H×W with the image centroid (c) is calculated. x ,c y The polar coordinates of the pixels in the training image I' centered in the polar coordinate system. The Cartesian coordinates (x, y) of the pixels in the corresponding Cartesian coordinate system of the training image I:

[0041]

[0042]

[0043] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps described above.

[0044] Compared with existing technologies, the above technical solution has the following advantages:

[0045] 1. The improved UNet network based on edge information re-extraction proposed in this invention can effectively prevent information redundancy compared with traditional networks, and enables the network to learn deeper information, thereby improving model performance;

[0046] 2. Based on the characteristic that most medical image segmentation targets are single, round objects, this invention reduces the complexity of image segmentation tasks by utilizing polar coordinate transformation, thereby effectively improving image segmentation accuracy.

[0047] 3. The automated medical image segmentation method proposed in this invention can adapt well to segmentation tasks with small sample sizes, save a lot of manpower and resources, and is highly feasible;

[0048] 4. Tests show that the automated segmentation method proposed in this invention performs well on several publicly available medical image segmentation datasets, including the LiTS2017 dataset, the CVC-ClinicDB dataset, and the ISIC2018 tumor edge segmentation dataset. Its segmentation performance is superior to existing network algorithms. Attached Figure Description

[0049] Figure 1 A flowchart illustrating the medical image segmentation method in a preferred embodiment of the present invention is provided.

[0050] Figure 2a To conform to the original image of the training image in a preferred embodiment of the present invention;

[0051] Figure 2b This is a schematic diagram of centroid extraction of training images in a preferred embodiment of the present invention;

[0052] Figure 2c This is a schematic diagram of the polar coordinate transformation of the training image in a preferred embodiment of the present invention;

[0053] Figure 3a This is a schematic diagram of a lesion image in a rectangular coordinate system according to a preferred embodiment of the present invention;

[0054] Figure 3b This is a schematic diagram of a lesion image in polar coordinates according to a preferred embodiment of the present invention;

[0055] Figure 4a A schematic diagram of a lesion image segmentation mask in a rectangular coordinate system according to a preferred embodiment of the present invention;

[0056] Figure 4b A schematic diagram of a lesion image segmentation mask in polar coordinates according to a preferred embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of image segmentation mask prediction in accordance with a preferred embodiment of the present invention. Detailed Implementation

[0058] The advantages of the present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments.

[0059] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0060] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0061] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0062] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0063] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0064] In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the convenience of the description of the invention and have no specific meaning in themselves. Therefore, "module" and "part" can be used interchangeably.

[0065] See Figure 1 The diagram illustrates a flowchart of a medical image segmentation method according to a preferred embodiment of the present invention. In this embodiment, the medical image segmentation method mainly includes several steps: image preprocessing, centroid extraction, polar coordinate transformation, image segmentation, and inverse polar coordinate transformation. The image segmentation performance under polar coordinates is significantly improved compared to that under rectangular coordinates, resulting in higher accuracy in the final image segmentation result. Specifically, the image segmentation method includes the following steps:

[0066] S100: Acquire multiple medical images, classify the medical images into training images, validation images, and test images, and perform modification processing on the training images to increase the sample size of the training images.

[0067] Because medical images are not easy to collect, and the number of images in publicly available image sources is limited, the problem of insufficient image quantity often arises when using image segmentation models based on medical images. By executing step S100, various image processing techniques can be applied to the training images, such as inversion, desaturation, rotation, and scaling, thereby increasing the number of training images exponentially and thus increasing the sample size of the training images to compensate for the scarcity of existing image samples.

[0068] S200: Extract the centroid of the training image to obtain the image centroid (c) of the training image. x ,c y )

[0069] In step S200, the centroid of the training image is extracted to obtain the central divergence source that may reflect the lesion. It is understood that medical image segmentation mainly focuses on extracting the lesion region from the medical image, and the lesion region itself is significantly different from the surrounding area. Therefore, the centroid position obtained after centroid extraction of the training image is generally the centroid position of the lesion region or the centroid position of the non-lesion region. Diverging outward from this centroid position, the boundary between the lesion region and the non-lesion region can be obtained, thereby achieving image segmentation. Thus, calculating the centroid of the target object in the image segmentation task in step S200 can improve segmentation accuracy.

[0070] S300: With image centroid (c x ,c y In a Cartesian coordinate system centered at , calculate the polar coordinates of the pixels of the training image in the polar coordinate system to obtain the transformed image in the polar coordinate system.

[0071] In the next step, the image centroid (c) will be used as the reference point. x ,c yThe coordinates of each pixel in the training image centered at a Cartesian coordinate system are converted to the transformed image in a polar coordinate system. Each pixel in the transformed image is composed of polar coordinates in the converted polar coordinate system. A Cartesian coordinate system is an orthogonal coordinate system that explicitly expresses geometric shapes using algebraic formulas. The Cartesian coordinate system is divided into four quadrants: the first quadrant, the second quadrant, the third quadrant, and the fourth quadrant. Counting starts from the upper right corner and proceeds counterclockwise. A polar coordinate system is a coordinate system in a plane consisting of a pole, a polar axis, and a polar radius. A point O is chosen on the plane, called the pole. A ray Ox is drawn from O, called the polar axis. A unit length is chosen, and angles are usually defined as positive counterclockwise. Thus, the position of any point P on the plane can be determined by the length r of the line segment OP and the angle θ from Ox to OP. The ordered pair (r, θ) is called the polar coordinates of point P, denoted as P(r, θ); r is called the polar radius of point P, and θ is called the polar angle of point P.

[0072] See Figure 2a , 2b Figures 2 and 2c show the original training image, the training image after centroid extraction (centroids are shown as points), and the changes in the training image in polar coordinates, respectively. As can be seen from the above figures, after the transformation from rectangular coordinates to polar coordinates, the lesion region and non-lesion region are more clearly distinguished within the training image.

[0073] S400: Segment the changed image based on an improved UNet network to obtain a segmented image.

[0074] In this step, an improved UNet network with re-extraction of edge information is used as the base model for segmentation learning. The predicted segmentation map output by the model is also an image in polar coordinates. Since the target objects in medical image segmentation tasks are mostly circular, the segmentation performance in polar coordinates is significantly better than that in Cartesian coordinates.

[0075] like Figure 3a , 3bFigures 4a and 4b show the original images of polyp-like lesion areas in Cartesian and polar coordinates, respectively, and schematic diagrams of segmentation masks in Cartesian and polar coordinates. The improved UNet network consists of two parts: a shrinking network and an expanding network. The shrinking network is similar to the shrinking network in UNet, but with the difference that the output of each layer is first normalized and then activated by an activation function. Each downsampling layer contains two 3x3 convolutional layers, a 1x1 "shortcut," and a 2x2 pooling layer. The expanding network is similar to the expanding network in UNet. Each upsampling layer contains two 3x3 convolutional layers and a 1x1 "shortcut." Before each upsampling, the corresponding results from the shrinking network need to be merged. Similar to the shrinking network, the output of each layer in the expanding network needs to be normalized first and then activated by an activation function. Finally, a 1x1 convolutional network is added to determine the result corresponding to the feature map. See also... Figure 5 The segmented image obtained after image segmentation using the improved UNet network also contains a real mask for image segmentation, and the two are quite similar.

[0076] S500: In the image centroid (c x ,c y In a polar coordinate system centered at , calculate the rectangular coordinates of the pixels of the segmented image in a Cartesian coordinate system to obtain the resulting image in a Cartesian coordinate system.

[0077] Finally, the pixels of the segmented image are inversely transformed from polar coordinates to rectangular coordinates to restore the resulting image in the rectangular coordinate system. This resulting image contains only the segmented parts, thus obtaining the final segmentation result.

[0078] In a preferred embodiment, step S100 includes:

[0079] S110: After acquiring multiple medical images, divide all medical images into training images, validation images, and test images according to an 8:1:1 ratio, so as to serve as training samples for subsequent improvements to the UNet network.

[0080] S120: Perform any one or any of the following change operations:

[0081] S121: Select n% of the training images for flipping and record it as the first changed image. For example, if n=50, half of the training images are randomly flipped, and the flipping angle can be arbitrarily random.

[0082] S122: Select m% of the training image and translate the training image along a random direction by a% of the length of the random direction, and record it as the second transformed image. For example, m=10, a=6.5, so that the sample size of the original training image is increased by 10%.

[0083] S123: Select p% of the training image and randomly shrink or enlarge b% of the original size of the training image, and record it as the third variation image. For example, p=10, b=10, so that the sample size of the original training image is increased by 10%.

[0084] S124: Select q% of the training image and rotate it around a random pixel in the training image as the center by t°, and record it as the fourth changed image. For example, q = 10, t = 45, so that the sample size of the original training image is increased by 10%.

[0085] S125: Select r% of the training image to perform grayscale vanishing operation and record it as the fifth transformed image. For example, r=30, which increases the number of samples in the original training image by 30%.

[0086] After the above changes, the sample size of the original training images increased by 110%. This preprocessing (or image enhancement) effectively increases the sample size of the training set, providing a solution to the problem of models failing to fit well due to insufficient dataset size. It is understood that the values ​​of n, m, p, q, and r can be freely configured according to the needs of sample expansion.

[0087] S130: Combine the first, second, third, fourth, and / or fifth modified images into an augmented image and add it to the training image.

[0088] In another preferred embodiment, step S200 includes:

[0089] S210: For the training image I, construct the following operator: M ij =∑ x,y I(x,y)x i y j Where I(x,y) represents the pixel value at pixel (x,y) in training image I, and the total pixel value M of training image I is calculated. ij ;

[0090] S220: Calculate the image centroid (c) according to the following operator. x ,c y )for:

[0091] c x =M 10 / M 00 ,c y =M 01 / M 00 .

[0092] Furthermore, step S200 includes:

[0093] S230: An image segmentation model is established based on an improved UNet network. Specifically, by learning a target object mask, the image is directly segmented to obtain an initial segmentation result. Then, the centroid of the target object or lesion region is calculated according to the aforementioned steps. This invention uses an improved UNet network for edge information re-extraction as the basic model. The traditional UNet network mainly includes convolutional layers, downsampling layers, deconvolutional layers, upsampling layers, and skip connection structures to achieve the segmentation target. The parameter matrix passed to the downsampling layer and the parameter matrix passed to the skip connection structure are highly consistent, and information redundancy affects the learning ability of the network. This invention, after the downsampling layer, on the one hand, passes the parameter matrix to the next convolutional layer, and on the other hand, resamples the parameter matrix to obtain a new parameter matrix. The parameter matrix that has not passed through the downsampling layer is subtracted from the new parameter matrix. Since the new parameter matrix has passed through the downsampling layer and the upsampling layer consecutively, it will lose some edge features. After subtracting the two, a parameter matrix containing only edge information is obtained. Passing it into the skip connection structure can prevent information redundancy and extract more diverse information. Meanwhile, compared to the traditional UNet network which directly uses the concatenation operation, this invention adds an addition operation and incorporates a 3*3 convolutional layer into the connection structure for feature learning.

[0094] Or in another preferred embodiment, step S200 includes:

[0095] S240: Build a heatmap based on the target object mask within the training image, with the image centroid (c x ,c y Using as the center point, the image centroid (c) is obtained according to a two-dimensional Gaussian distribution. x ,c y The pixel size near )

[0096] S250: The intensity of the heatmap is calculated based on a neural network using a stacked hourglass architecture, so that the center point of the heatmap with the highest intensity is the final centroid.

[0097] Specifically, this segmentation model identifies images by learning heatmaps centered on the centroid of the target object. The heatmap is based on a target object mask, and the centroid of the target object is calculated as its center point according to the aforementioned steps. Simultaneously, the pixel size near the center point is obtained according to a 2D Gaussian distribution. The deep learning model employs a neural network based on a stacked hourglass architecture, using a total of 8 stacked hourglass modules, ultimately resulting in 8 separate heatmaps. The center point corresponding to the heatmap with the highest intensity is the final predicted centroid.

[0098] Understandably, for segmentation tasks where the target object is more irregular and larger, the first type of model (improved UNet network) is more accurate in finding the centroid, such as organs like the liver. For segmentation tasks where the target object is more regular and smaller, the second type of model (heatmap centered on the centroid of the target object) is more accurate in finding the centroid, such as tissues like tumors.

[0099] Further, step S300 includes:

[0100] S310: In a Cartesian coordinate system, for a pixel (x, y) of a training image, define the following function:

[0101]

[0102]

[0103] S320: For a training image I of size H×W, calculate the centroid (c) of the image according to the following operator. x ,c y The polar coordinates of the pixels in the training image I' centered in the polar coordinate system.

[0104]

[0105]

[0106] It is understandable that for the training image I' in polar coordinates, if the edge information is used to extract and improve the UNet network as the base model for segmentation learning, its detailed structure remains consistent with that in step S200. It should be noted that the matrix parameters between the two are not shared.

[0107] Furthermore, step S500 includes:

[0108] S510: Based on the following operator, calculate the image with size H×W and centroid (c x ,c y The polar coordinates of the pixels in the training image I' centered in the polar coordinate system. The Cartesian coordinates (x, y) of the pixels in the corresponding Cartesian coordinate system of the training image I:

[0109]

[0110]

[0111] The final segmentation result can be obtained by converting the segmented image obtained in step S400 from polar coordinates to rectangular coordinates.

[0112] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps described above.

[0113] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A medical image segmentation method, characterized in that, Includes the following steps: Multiple medical images are acquired, classified into training images, validation images, and test images, and the training images are modified to increase the sample size of the training images. Centroid extraction is performed on the training images to obtain the image centroids of the training images. ; With image centroid In a Cartesian coordinate system centered at the center, the polar coordinates of the pixels of the training image in the polar coordinate system are calculated to obtain the transformed image in the polar coordinate system; The modified image is segmented based on an improved UNet network to obtain a segmented image; With image centroid In the polar coordinate system centered at the center, the rectangular coordinates of the pixels of the segmented image in the rectangular coordinate system are calculated to obtain the result image in the rectangular coordinate system; Centroid extraction is performed on the training images to obtain the image centroids of the training images. The steps include: For the training image I, construct the following operator: ,in Represents the training image I pixels The pixel values ​​at each location are used to calculate the total pixel values ​​of training image I. ; Calculate the image centroid using the following operator. for: ； With image centroid The steps for calculating the polar coordinates of the pixels in the training image in a polar coordinate system, centered on a Cartesian coordinate system, to obtain the transformed image in a polar coordinate system, include: In a Cartesian coordinate system, for the pixels of the training image Define the following function: , For size The training image I is used to calculate the centroid of the image according to the following operator. Training images in polar coordinates centered polar coordinates of pixels : ； With image centroid The steps of calculating the rectangular coordinates of the pixels of the segmented image in a Cartesian coordinate system, centered on the polar coordinate system, to obtain the resulting image in the Cartesian coordinate system, include: Based on the following operator, the dimension is calculated as follows: Image centroid Training images in polar coordinates centered polar coordinates of pixels Cartesian coordinates of pixels in the corresponding Cartesian coordinate system : + + 。 2. The medical image segmentation method as described in claim 1, characterized in that, The steps of acquiring multiple medical images, classifying the medical images into training images, validation images, and test images, and performing modification processing on the training images to increase the sample size of the training images include: After acquiring multiple medical images, all medical images are divided into training images, validation images, and test images according to an 8:1:1 ratio. Perform any one or any of the following changes: Select n% of the training images, flip them, and record them as the first transformed image; Select m% of the training image and translate the training image along a random direction by a% of the length of the random direction, and record it as the second transformed image; Select p% of the training image and randomly shrink or enlarge b% of the original size of the training image, and record it as the third transformed image; Select q% from the training image and rotate it around a random pixel in the training image by t°, and record it as the fourth transformed image; Select r% from the training image to perform a grayscale vanishing operation, and record it as the fifth transformed image; The first, second, third, fourth, and / or fifth modified images are combined into an augmented image and added to the training image.

3. The medical image segmentation method as described in claim 1, characterized in that, Centroid extraction is performed on the training images to obtain the image centroids of the training images. The steps include: An image segmentation model is established based on an improved UNet network. or A heatmap is built based on the target object mask within the training image, and the image centroid is used as the basis for the heatmap. The image centroid is obtained by taking the center point as the center point and following a two-dimensional Gaussian distribution. Nearby pixel size; The intensity of the heatmap is calculated using a neural network with a stacked hourglass architecture, so that the center point of the heatmap with the highest intensity is the final centroid.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the steps as described in any one of claims 1-3.

Images