A method and system for precise tumor segmentation in CT images

Abstract

This invention provides a method and system for accurate tumor segmentation in CT images, relating to the field of image processing technology. The method includes: determining the region to be identified in a CT image containing a tumor using an image segmentation model; determining the boundary line, center position, and nearest reference pixels of the region to be identified, along with their probability information; and then selecting correction pixels from the nearest reference pixels to obtain a correction region, thereby updating the trained image segmentation model. According to this invention, after obtaining the region to be identified in the CT image of each slice using the image segmentation model, the boundary line of the region to be identified can be mutually verified based on the nearest pixels on the boundary line and the pixels of adjacent slices. This allows for the selection of pixels that better represent the tumor edge from the nearest pixels on the boundary line, thereby achieving more accurate segmentation of the region containing the tumor.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for precise tumor segmentation in CT images. Background Technology

[0002] In related technologies, image detection models can be used to detect the tumor region in medical images (e.g., CT images), and image segmentation models can be used to segment the contour of the tumor region. However, image segmentation models usually find all pixels belonging to the tumor region by calculating the probability of whether a pixel belongs to the tumor region, and take the edge of the region enclosed by these pixels as the contour of the tumor region. However, the edge of the tumor is relatively blurry, and the accuracy of the image segmentation model is also low. For example, there are a large number of pixels with a probability close to 0.5, making it difficult to accurately determine the edge of the tumor region. Summary of the Invention

[0003] This invention provides a method and system for precise tumor segmentation in CT images, which can solve the technical problem of blurry tumor edges and low segmentation accuracy in related technologies.

[0004] According to a first aspect of the present invention, a method for precise tumor segmentation in CT images is provided, comprising:

[0005] The image segmentation model is used to determine the undetermined region where the tumor is located in the CT image of each slice;

[0006] Obtain the boundary lines and center position of the region to be determined;

[0007] Based on the pixels in the region to be determined and its boundary line, as well as the center position, determine the nearest neighbor reference pixels of the pixels on the boundary line.

[0008] The probability information for determining whether a neighboring pixel belongs to the tumor region is obtained by processing the neighboring reference pixel using the image segmentation model.

[0009] Based on the first range of the undetermined region in the CT image, the second range of the undetermined region in the CT images of adjacent slices, and the determination probability information, the correction pixel is selected from the nearest reference pixel.

[0010] Based on the corrected pixels, determine the corrected region where the tumor is located in the CT image of each slice;

[0011] Based on the correction region, the image segmentation model is updated and trained to obtain the updated image segmentation model;

[0012] The updated image segmentation model is used to segment CT images of multiple slices to obtain image segmentation results.

[0013] According to the present invention, determining the nearest neighbor reference pixels of the pixels on the boundary line based on the region to be determined and the pixels on its boundary line, as well as the center position, includes:

[0014] For the region to be determined in the CT image of the i-th slice, connect the j-th pixel on its boundary line with the center position to obtain the line numbered (i,j);

[0015] Based on the length of the line numbered (i,j), extend the length of the line outwards towards the outside of the region to be determined to obtain the search line numbered (i,j), where the extension length is proportional to the length of the line.

[0016] Using the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice as the center, and the extended length of the search line numbered (i,j) as the radius, obtain the circular search neighborhood numbered (i,j).

[0017] The pixel located within the circular search neighborhood numbered (i,j) and on the search line numbered (i,j), and whose probability information belongs to the preset interval, is determined as the nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice.

[0018] According to the present invention, selecting correction pixels from neighboring reference pixels based on a first range of the region to be determined in a CT image, a second range of the region to be determined in CT images of adjacent slices, and the determination probability information, includes:

[0019] Obtain the center location of the second range of the region to be determined in the CT images of adjacent slices;

[0020] Based on the first range, the center position of the first range, and the center position of the second range, the projection range of the first range in adjacent tomographic CT images is determined, wherein the center position of the projection range overlaps with the center position of the second range, and the shape and size of the projection range are the same as the first range.

[0021] Obtain the projection line of the line numbered (i,j) in the CT image of the adjacent slice, wherein the projection line numbered (i,j) has one end at the center of the projection range and is parallel to the line numbered (i,j).

[0022] Extend the projection line numbered (i,j) so that the projection line numbered (i,j) intersects with both the boundary line of the projection range and the boundary line of the second range, thus obtaining the intersection line numbered (i,j).

[0023] Determine the projection search point of the nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice on the intersection line numbered (i,j).

[0024] Centered on the intersection of the line of intersection numbered (i,j) and the boundary line of the second range, and with the extended length of the search line numbered (i,j) as the radius, a reference neighborhood numbered (i,j) is set.

[0025] The pixel located within the reference neighborhood of (i,j) and on the intersection line of (i,j) is set as the reference pixel.

[0026] Based on the CT values ​​and decision probability information of neighboring reference pixels, the CT values ​​and decision probability information of projection search points, and the CT values ​​and decision probability information of comparison pixels, the selection criteria for each neighboring reference pixel relative to the CT images of each adjacent slice are determined.

[0027] The nearest neighbor reference pixel with the highest index is selected as the correction pixel.

[0028] According to the present invention, the selection index of each neighboring reference pixel relative to the CT images of each adjacent slice is determined based on the CT values ​​and determination probability information of neighboring reference pixels, the CT values ​​and determination probability information of the projection search point, and the CT values ​​and determination probability information of the comparison pixel, including:

[0029] According to the formula

[0030]

[0031] The selection index for obtaining the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. ,in, Let be the CT value of the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. This refers to the probability information of determining the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. Let be the CT value of the (k+1)th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. This refers to the probability information of determining the (k+1)th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. Let be the CT value of the k-th projection search point on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. This refers to the probability information for determining the k-th projection search point on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. Let be the CT value of the k-th reference pixel on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. This refers to the probability information for determining the k-th reference pixel on the intersection line numbered (i,j) of the CT image of the s-th slice adjacent to the i-th slice. This represents the number of faults adjacent to the i-th fault.

[0032] According to the present invention, an image segmentation model is updated and trained based on a correction region to obtain an updated image segmentation model, comprising:

[0033] Based on the correction region, the correction boundary of the region where the tumor is located;

[0034] Set first annotation information for pixels within the correction boundary, and set second annotation information for pixels outside the correction boundary;

[0035] Obtain the probability information of each pixel obtained by the image segmentation model;

[0036] Based on the determination probability information, the first annotation information and the second annotation information, as well as the nearest neighbor reference pixels and their selection index, the loss function of the image segmentation model is determined;

[0037] The image segmentation model is trained based on the loss function of the image segmentation model to obtain an updated image segmentation model.

[0038] According to the present invention, the loss function of the image segmentation model is determined based on the determination probability information, the first annotation information, the second annotation information, and the nearest neighbor reference pixels and their selection indices, including:

[0039] Based on the probability information of other pixels in the CT image besides the nearest neighbor reference pixel, as well as the first annotation information and the second annotation information, the enhancement training loss function is determined.

[0040] Based on the nearest neighbor reference pixels in the CT image, their selection index and decision probability information, as well as the first annotation information and the second annotation information, the boundary selection loss function is determined.

[0041] The loss function of the image segmentation model is determined based on the enhanced training loss function and the boundary selection loss function.

[0042] According to the present invention, a boundary selection loss function is determined based on each nearest neighbor reference pixel in a CT image, its selection index and decision probability information, as well as first annotation information and second annotation information, including:

[0043] According to the formula

[0044]

[0045] Determine the boundary and choose the loss function ,in, Let $\frac{j}{k}$ be the probability value that the k-th nearest neighbor reference pixel corresponding to the $j-th pixel on the boundary line of the region to be determined based on the first annotation information belongs to the region where the tumor is located. This refers to the probability information of the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined. The selection criterion is the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined. Let m be the probability value that the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined according to the second annotation information belongs to the region where the tumor is located, m is the number of nearest neighbor reference pixels corresponding to the j-th pixel on the boundary line of the region to be determined, and n is the number of pixels on the boundary line of the region to be determined.

[0046] According to a second aspect of the present invention, a precise tumor segmentation system for CT images is provided, comprising:

[0047] The undetermined region module uses an image segmentation model to determine the undetermined region where the tumor is located in the CT image of each slice.

[0048] The center position module obtains the boundary line of the region to be determined and the center position of the region to be determined;

[0049] The nearest neighbor reference pixel module determines the nearest neighbor reference pixels of the pixels on the boundary line based on the pixels on the region to be determined and its boundary line, as well as the center position.

[0050] The probability information determination module determines the probability information of a neighboring pixel belonging to the tumor region, obtained by the image segmentation model processing the neighboring reference pixel.

[0051] The pixel correction module selects correction pixels from nearby reference pixels based on the first range of the undetermined region in the CT image, the second range of the undetermined region in the CT images of adjacent slices, and the determination probability information.

[0052] The correction region module determines the correction region where the tumor is located in the CT image of each slice based on the correction pixels.

[0053] The update module updates and trains the image segmentation model based on the correction region to obtain the updated image segmentation model;

[0054] The image segmentation results module uses an updated image segmentation model to segment CT images from multiple slices and obtain the image segmentation results.

[0055] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0056] According to the present invention, after obtaining the region to be determined where the tumor is located in the CT image of each slice through an image segmentation model, the boundary line of the region to be determined can be mutually verified based on the nearest neighbor pixels on the boundary line and the pixels of adjacent slices. This allows for the selection of pixels that are more representative of the tumor edge from the nearest neighbor pixels on the boundary line, thereby enabling more precise segmentation of the tumor region. When determining the selection criteria, the nearest neighbor reference pixels can be verified by the projected pixels in adjacent slice CT images, and the change benchmark of the comprehensive level of pixels near the nearest neighbor reference pixels can be obtained. Then, the selection criteria are determined based on the change in the comprehensive level of the nearest neighbor reference pixels and adjacent pixels, as well as the change benchmark. Based on the selection criteria, the nearest neighbor reference pixel with the largest relative difference in comprehensive level on both sides is determined as the pixel on the boundary line, thus determining the corrected boundary line. This maximizes the relative difference in comprehensive level inside and outside the corrected boundary line, improving the contrast inside and outside the tumor region, and also improving the accuracy and objectivity of the boundary line. When updating and training an image segmentation model, the training intensity of the loss function can be adjusted by selecting an index, so that the training intensity and training value of pixels closer to the boundary line are greater, thereby rapidly improving the image segmentation model's accuracy in recognizing the boundary line and the accuracy in judging whether pixels near the boundary line are within the tumor area, thus improving training efficiency and training effect. Attached Figure Description

[0057] Figure 1 A schematic flowchart of a method for precise tumor segmentation in CT images according to an embodiment of the present invention is shown exemplarily;

[0058] Figure 2 A block diagram of a CT image tumor precision segmentation system according to an embodiment of the present invention is shown as an example. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0061] Figure 1 An exemplary flowchart of a method for precise tumor segmentation in CT images according to an embodiment of the present invention is shown, the method comprising:

[0062] Step S1: Using an image segmentation model, determine the region to be identified where the tumor is located in the CT image of each slice;

[0063] Step S2: Obtain the boundary line of the region to be determined and the center position of the region to be determined;

[0064] Step S3: Based on the pixels on the boundary line of the region to be determined and the center position, determine the nearest neighbor reference pixels of the pixels on the boundary line.

[0065] Step S4: Determine the probability information of the nearest neighbor pixels obtained by the image segmentation model from processing the nearest neighbor reference pixels to determine whether the nearest neighbor pixels belong to the tumor region.

[0066] Step S5: Based on the first range of the region to be determined in the CT image, the second range of the region to be determined in the CT images of adjacent slices, and the determination probability information, select the correction pixel from the nearest reference pixel.

[0067] Step S6: Based on the correction pixels, determine the correction region where the tumor is located in the CT image of each slice;

[0068] Step S7: Based on the correction region, update and train the image segmentation model to obtain the updated image segmentation model;

[0069] Step S8: Using the updated image segmentation model, the CT images of multiple slices are segmented to obtain the image segmentation results.

[0070] According to an embodiment of the present invention, the method for accurate tumor segmentation in CT images, after obtaining the region to be determined where the tumor is located in the CT image of each slice through the image segmentation model, can mutually verify the boundary line of the region to be determined based on the neighboring pixels on the boundary line and the pixels of adjacent slices, thereby selecting pixels that are more representative of the tumor edge from the neighboring pixels of the boundary line, and thus segmenting the region where the tumor is located more accurately.

[0071] According to an embodiment of the present invention, in step S1, the image segmentation model can be a deep learning neural network model such as the U-NET model. The CT images of each slice can be used as input to the image segmentation model. After the image segmentation model processes the CT image, it can obtain a mask image with the same size as the CT image. By superimposing the mask image with the CT image, the region to be determined where the tumor is located can be identified in the CT image. Furthermore, the probability information of each pixel in the CT image belonging to the region where the tumor is located can also be obtained. For example, if the probability information is greater than 0.5, it is determined that the pixel belongs to the region where the tumor is located. Therefore, the set of all pixels with a probability information greater than 0.5 is the region to be determined. However, tumors typically exhibit characteristics such as high CT values, indistinct edges, and irregular shapes. Image segmentation models may calculate a probability of 0.5 for pixels near their edges. For example, pixel A might have a probability of 0.51, while a nearby pixel B might have a probability of 0.49. Since these two pixels are close in location, have insignificant CT values, and similar probabilities, while pixel A can be classified as part of the undefined region based on these probabilities, this classification method may contain errors. Near the boundary line, pixels within the undefined region can easily be confused with those outside. For instance, if the image segmentation model has inherent errors, this classification method may be inaccurate; pixel A might not actually belong to the tumor region, or pixel B might actually belong to the tumor region. This results in potential errors in boundary line determination.

[0072] According to one embodiment of the present invention, in step S2, the boundary line of the region to be determined can be found, and the center position of the region to be determined can be determined. The coordinates of the center position can be the average of the coordinates of all pixels in the region to be determined. For example, the average of the ordinates of all pixels in the region to be determined can be used as the ordinate of the center position, and the average of the abscissas of all pixels in the region to be determined can be used as the abscissa of the center position, thereby determining the center position.

[0073] According to one embodiment of the present invention, in step S3, neighboring reference points near the boundary line can be selected first, that is, pixels that may serve as the boundary, so as to select more suitable pixels from these pixels as pixels on the boundary line, thereby obtaining a more accurate boundary line.

[0074] According to one embodiment of the present invention, determining the nearest neighbor reference pixel of the pixel on the boundary line based on the undetermined region and the pixels on its boundary line, as well as the center position, includes: for the undetermined region in the CT image of the i-th slice, connecting the j-th pixel on its boundary line with the center position to obtain a line numbered (i,j); extending the length of the line numbered (i,j) outward from the undetermined region according to the length of the line numbered (i,j), to obtain a search line numbered (i,j), wherein the extension length is proportional to the length of the line; taking the j-th pixel on the boundary line of the undetermined region in the CT image of the i-th slice as the center and the extension length of the search line numbered (i,j) as the radius, obtaining a circular search neighborhood numbered (i,j); determining the pixel located within the circular search neighborhood numbered (i,j) and on the search line numbered (i,j) whose probability information belongs to a preset interval as the nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the undetermined region in the CT image of the i-th slice.

[0075] According to one embodiment of the present invention, the boundary line includes multiple pixels. Connecting the j-th pixel with the center position yields a line numbered (i,j), which is the j-th line within the undetermined region of the i-th tomographic CT image. Furthermore, pixels with a probability of close to 0.5 may exist both inside and outside the boundary line; therefore, a search can be performed both inside and outside the boundary line, meaning that neighboring reference pixels can be set both inside and outside the boundary line. Thus, the aforementioned line can be extended to obtain a search line numbered (i,j), with the extension length proportional to the length of the line, for example, 0.1 times the length of the line. Further, a circular region is set with the extended length as the radius and the j-th pixel on the boundary line as the center; this is the circular search neighborhood numbered (i,j), within which pixels suitable as boundary lines can be searched. Furthermore, pixels located within the circular search neighborhood (i,j) and on the search line (i,j), with a probability information higher than a preset lower threshold, can be identified as the nearest neighbor reference pixels corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. That is, pixels that simultaneously satisfy three conditions: Condition 1 is being located within the circular search neighborhood (i,j), meaning that pixels within this circular region are relatively close to the j-th pixel on the boundary line. Therefore, a search is performed within this circular region to determine if there exists a pixel more suitable as a boundary than the j-th pixel on the boundary line. Condition 2 is being located on the search line (i,j). Since there are multiple pixels within the circular region, they may also be relatively close to other pixels on the boundary line, for example, the (j+1)-th pixel. Therefore, the search scope can be further narrowed down to the search line (i,j) to determine if there exists a pixel more suitable as a boundary than the j-th pixel on the boundary line. Condition 3 states that the probability information belongs to a preset interval (e.g., [0.45, 0.55]). This means the image segmentation model has been trained multiple times and possesses a certain level of accuracy. Therefore, pixels with low probability information are more likely to be outside the undetermined region, while pixels with high probability information are more likely to be within the undetermined region. Consequently, pixels not belonging to the preset interval are more likely not to be on the boundary line. Therefore, only pixels with probability information within the preset interval are searched to determine if there exists a pixel more suitable as a boundary than the j-th pixel on the boundary line. Pixels that simultaneously satisfy all three conditions can be used as the nearest neighbor reference pixels corresponding to the j-th pixel on the boundary line of the undetermined region in the CT image of the i-th slice. Among these nearest neighbor reference pixels, it can be determined whether there exists a pixel more suitable as a boundary than the j-th pixel on the boundary line.

[0076] According to an embodiment of the present invention, in step S4, as described above, the image segmentation model can determine the determination probability information of each pixel point, and therefore, the determination probability information of the nearest neighbor reference pixel point can be found.

[0077] According to an embodiment of the present invention, in step S5, the nearest reference pixels can be verified based on the pixels in the CT images of adjacent slices to determine whether there is a pixel among the nearest reference pixels that is more suitable as a boundary than the j-th pixel on the boundary line, thereby determining the correction pixel.

[0078] According to one embodiment of the present invention, selecting correction pixels from neighboring reference pixels based on a first range of a region to be determined in a CT image, a second range of a region to be determined in CT images of adjacent slices, and the determination probability information, includes: obtaining the center position of the second range of the region to be determined in CT images of adjacent slices; determining the projection range of the first range in CT images of adjacent slices based on the first range, the center position of the first range, and the center position of the second range, wherein the center position of the projection range overlaps with the center position of the second range, and the shape and size of the projection range are the same as the first range; obtaining the projection line numbered (i,j) of the line connecting (i,j) in CT images of adjacent slices, wherein the projection line numbered (i,j) has the center position of the projection range as one end, and the projection line numbered (i,j) is parallel to the line numbered (i,j); extending the projection line numbered (i,j) so that the projection line numbered (i,j) is... The projection line intersects with both the boundary line of the projection range and the boundary line of the second range, obtaining the intersection line numbered (i,j); the projection search point of the nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice is determined on the intersection line numbered (i,j); with the intersection line numbered (i,j) and the boundary line of the second range as the center, and the extended length of the search line numbered (i,j) as the radius, a control neighborhood numbered (i,j) is set; the pixel located in the control neighborhood numbered (i,j) and on the intersection line numbered (i,j) is set as the control pixel; based on the CT value and determination probability information of the nearest neighbor reference pixel, the CT value and determination probability information of the projection search point, and the CT value and determination probability information of the control pixel, the selection index of each nearest neighbor reference pixel relative to the CT images of each adjacent slice is determined; the nearest neighbor reference pixel with the highest selection index is determined as the correction pixel.

[0079] According to one embodiment of the present invention, the adjacent slices of the i-th slice are the (i+1)-th slice and the (i-1)-th slice. If i=1, then the only adjacent slice is the second slice; if the i-th slice is the last slice, then the only adjacent slice is the (i-1)-th slice. The center position of the second range of the region to be determined in the CT image of the adjacent slices can be determined in the same way as the method for determining the center position of the region to be determined described above, and will not be repeated here.

[0080] According to one embodiment of the present invention, by projecting a first range onto the CT image of an adjacent slice, the center of the first range can coincide with the center of a second range in the CT image of the adjacent slice. With this position as the center, a range with the same shape and size as the first range is set to obtain the projection range of the first range in the CT image of the adjacent slice.

[0081] According to one embodiment of the present invention, the center of the projection range is connected to the j-th pixel on its boundary line in the same manner as the line numbered (i,j), resulting in a projection line numbered (i,j). Since the center positions are coincident during projection, even if the tumor has an irregular shape, the projection range and the second range share a common center. A straight line extending outward from this common center can pass through the boundary line of both the projection range and the second range, i.e., it intersects with both boundary lines. If the second range is larger than the projection range, it needs to be extended to intersect with the boundary line of the second range. In this case, the extended projection line is designated as the intersection line numbered (i,j), and is extended by at least 1.1 times. If the projection range differs significantly from the second range, it can be extended further until an intersection is obtained. If the second range is smaller than the projection range, it can intersect with the boundary line of the second range without extension. In this case, the projection line is appropriately extended, for example, by 1.1 times, to obtain the intersection line numbered (i,j).

[0082] According to one embodiment of the present invention, the projection search point of the nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice can be determined on the intersection line numbered (i,j). That is, the index of the nearest neighbor reference pixel on the search line numbered (i,j) can be determined, and the pixel with the same index on the intersection line numbered (i,j) is found, which is the projection search point. The CT value and determination probability information of the projection search point can represent the status of the corresponding position of the tumor in adjacent CT images, and can be used to verify the pixels on the boundary line of the region to be determined.

[0083] According to one embodiment of the present invention, a circular region is set with the intersection point of the intersection line numbered (i,j) and the boundary line of the second range as the center, and the extended length of the search line numbered (i,j) as the radius. This circular region is the control neighborhood numbered (i,j). Pixels located within the control neighborhood numbered (i,j) and on the intersection line numbered (i,j) are set as control pixels. The CT value and determination probability information of the control pixels can represent the boundary status of the tumor region in adjacent tomographic CT images and can be used to verify the pixels on the boundary line of the region to be determined.

[0084] According to an embodiment of the present invention, the selection index of each neighboring reference pixel relative to the CT images of each adjacent slice is determined based on the CT values ​​and determination probability information of the nearest neighbor reference pixels, the CT values ​​and determination probability information of the projection search points, and the CT values ​​and determination probability information of the reference pixels. This includes: obtaining the selection index of the kth nearest neighbor reference pixel corresponding to the jth pixel on the boundary line of the region to be determined in the CT image of the i-th slice according to formula (1). ,

[0085] (1)

[0086] in, Let be the CT value of the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. This refers to the probability information of determining the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. Let be the CT value of the (k+1)th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. This refers to the probability information of determining the (k+1)th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. Let be the CT value of the k-th projection search point on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. This refers to the probability information for determining the k-th projection search point on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. Let be the CT value of the k-th reference pixel on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. This refers to the probability information for determining the k-th reference pixel on the intersection line numbered (i,j) of the CT image of the s-th slice adjacent to the i-th slice. This represents the number of faults adjacent to the i-th fault.

[0087] According to one embodiment of the present invention, as described above, the CT value of the tumor region is higher than that of the background region; therefore, the CT value can reflect the probability that a pixel belongs to the tumor region. On the other hand, the determination probability information is the probability that a pixel belongs to the tumor region calculated by the image segmentation model. Therefore, both the CT value and the determination probability information can reflect the probability that a pixel belongs to the tumor region. Multiplying the two together yields a comprehensive level of probability that a pixel belongs to the tumor region, i.e., This represents the overall probability that the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice belongs to the region where the tumor is located. This represents the overall probability that the (k+1)th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice belongs to the tumor region. Using this overall probability level to distinguish pixels inside and outside the tumor region improves the discriminative power compared to using only the probability information calculated by the image segmentation model. In other words, the overall probability levels of pixels inside and outside the tumor region differ more significantly, making them easier to distinguish. Furthermore, if both the k-th and (k+1)-th nearest neighbor reference pixels belong to the tumor region, their overall probability levels are relatively close. If both belong to the background region outside the tumor, their overall probability levels are also relatively close. Only when the k-th nearest neighbor reference pixel belongs to the tumor region and the (k+1)-th pixel belongs to the background region outside the tumor is the difference between them large. Therefore, the difference between them can be calculated. This is used to find pixels on the boundary line among multiple nearest neighbor reference points based on the difference. That is, if the overall level difference between the kth nearest neighbor reference pixel and the (k+1)th pixel is large, then the probability that the kth nearest neighbor reference pixel is on the boundary line is high.

[0088] According to one embodiment of the present invention, The overall level of the k-th projection search point is the overall level of the projection of the k-th nearest neighbor reference pixel in adjacent slices. This projection search point is located either outside or inside the second range in the CT images of adjacent slices, and can be used to describe the overall level inside or outside the tumor in the CT images of adjacent slices. If it is located inside the tumor, it can describe the condition inside the tumor, that is, the CT value condition inside the tumor and the condition of the tumor features identified by the image segmentation model. If it is located outside the tumor, it can describe the condition of the background region, that is, the CT value condition of the background region and the condition of the background region features identified by the image segmentation model. Similarly, This can represent the overall level of pixels near the boundary of the second range in CT images of adjacent slices. Therefore, This can represent the difference or change between the condition inside or outside the tumor and the condition at the tumor boundary. It can also represent the difference or change between the projection position of the k-th nearest neighbor reference pixel and the condition at the tumor boundary. This difference can be used to describe the difference between the overall level of pixels in the region near the k-th nearest neighbor reference pixel and the average level of the tumor boundary. If this difference is large, it indicates that the overall level of pixels in the region near the k-th nearest neighbor reference pixel varies greatly, and the difference between the inside / outside of the tumor and the boundary line is large. For example, even two pixels located inside or outside the tumor region may have a large difference in their overall levels. If this difference is small, it indicates that the overall level of pixels in the region near the k-th nearest neighbor reference pixel is relatively stable, and the difference between the inside / outside of the tumor and the boundary line is small. Therefore, It can represent the horizontal difference between the internal or external regions of a tumor and its boundary line. This represents the average level difference corresponding to all adjacent faults, serving as the benchmark for the change in the overall level of pixels in the region near the k-th nearest neighbor reference pixel. That is, if the benchmark is high, then... A larger value is needed to represent the k-th nearest neighbor reference pixel as a pixel on the boundary line. If the change reference is low, then even if... Even a small value may result in the greatest horizontal change between the kth nearest neighbor reference pixel and the (k+1)th nearest neighbor reference pixel, meaning that the kth nearest neighbor reference pixel is determined to be a pixel on the boundary line.

[0089] According to one embodiment of the present invention, the ratio of the overall level change of the k-th nearest neighbor reference pixel to that of the (k+1)-th nearest neighbor reference pixel and the change benchmark can be obtained as the selection index for the k-th nearest neighbor reference pixel. The nearest neighbor reference pixel with the highest selection index can be determined as the correction pixel; that is, the nearest neighbor reference pixel with the largest overall level change under its specific change benchmark is the correction pixel, i.e., the pixel on the boundary line. In other words, among multiple nearest neighbor reference pixels, the pixel with the largest relative change in overall level is found, and this pixel is the pixel on the boundary line. The relative difference in overall level is largest on both sides of this pixel, therefore this pixel can be used as a dividing point. By determining the correction pixels of all pixels on the boundary line in the same way, the correction boundary can be obtained, i.e., the more accurate boundary line after correction. In step S6, the area within the correction boundary is the correction area where the tumor is located in the CT image. By the same method, the correction area where the tumor is located in the CT image of each slice can be obtained. Furthermore, manual verification on multiple CT images revealed that the correction boundary of the above-mentioned correction area improved the segmentation accuracy by approximately 8.2% compared to the boundary obtained by segmentation using only the U-NET model.

[0090] In this way, the nearest reference pixels can be verified by projecting pixels in adjacent tomographic CT images, and the change benchmark of the comprehensive level of pixels near the nearest reference pixels can be obtained. Then, based on the change in the comprehensive level of the nearest reference pixels and adjacent pixels and the change benchmark, selection criteria are determined. Based on the selection criteria, the nearest reference pixel with the largest relative difference in comprehensive level on both sides is determined as the pixel on the boundary line, thereby determining the corrected boundary line. This maximizes the relative difference in comprehensive level inside and outside the corrected boundary line, improves the contrast inside and outside the tumor area, and also improves the accuracy and objectivity of the boundary line.

[0091] According to an embodiment of the present invention, in step S7, after obtaining a more accurate correction region, the image segmentation model can be updated and trained to further enhance the performance of the image segmentation model, so that the accuracy of the segmentation result obtained by the image segmentation model is closer to the accuracy of the corrected segmentation result.

[0092] According to one embodiment of the present invention, an image segmentation model is updated and trained based on a correction region to obtain an updated image segmentation model, including: determining the correction boundary of the tumor region based on the correction region; setting first annotation information for pixels within the correction boundary and setting second annotation information for pixels outside the correction boundary; acquiring the determination probability information of each pixel obtained by the image segmentation model; determining the loss function of the image segmentation model based on the determination probability information, the first annotation information, the second annotation information, and the nearest neighbor reference pixels and their selection index; and training the image segmentation model based on the loss function of the image segmentation model to obtain an updated image segmentation model.

[0093] According to one embodiment of the present invention, based on the more accurate correction boundary, the pixels in the CT image can be re-labeled. For example, the probability of pixels outside the correction boundary belonging to the tumor region is labeled as 0 to obtain second labeling information, and the probability of pixels inside and outside the correction boundary belonging to the tumor region is labeled as 1 to obtain first labeling information. Further, the loss function of the image segmentation model can be determined by the determination probability information calculated for each pixel by the image segmentation model and the aforementioned re-labeled probabilities.

[0094] According to one embodiment of the present invention, determining the loss function of an image segmentation model based on the determination probability information, the first annotation information, the second annotation information, and the nearest neighbor reference pixels and their selection indices includes: determining an enhancement training loss function based on the determination probability information of other pixels in the CT image besides the nearest neighbor reference pixels, and the first annotation information and the second annotation information; determining a boundary selection loss function based on each nearest neighbor reference pixel in the CT image, its selection indices, determination probability information, and the first and second annotation information; and determining the loss function of the image segmentation model based on the enhancement training loss function and the boundary selection loss function.

[0095] According to one embodiment of the present invention, if a pixel in a CT image is not a nearest neighbor reference pixel, then the pixel is relatively far from the boundary line. The cross-entropy loss function corresponding to the pixel can be determined using first or second annotation information and the pixel's decision probability information. For example, if the pixel is outside the correction boundary, the cross-entropy loss function between the pixel's decision probability information and the second annotation information can be calculated; if the pixel is inside the correction boundary, the cross-entropy loss function between the pixel segment's decision probability information and the first annotation information can be calculated. Further, the cross-entropy loss functions corresponding to all pixels other than the nearest neighbor reference pixel can be summed to obtain the enhancement training loss function.

[0096] According to one embodiment of the present invention, the boundary selection loss function corresponding to the nearest neighbor reference pixel can be further determined. Based on each nearest neighbor reference pixel in the CT image, its selection index and decision probability information, as well as the first annotation information and the second annotation information, the boundary selection loss function is determined, including: determining the boundary selection loss function according to formula (2). ,

[0097] (2)

[0098] in, Let $\frac{j}{k}$ be the probability value that the k-th nearest neighbor reference pixel corresponding to the $j-th pixel on the boundary line of the region to be determined based on the first annotation information belongs to the region where the tumor is located. This refers to the probability information of the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined. The selection criterion is the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined. Let m be the probability value that the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined according to the second annotation information belongs to the region where the tumor is located, m is the number of nearest neighbor reference pixels corresponding to the j-th pixel on the boundary line of the region to be determined, and n is the number of pixels on the boundary line of the region to be determined.

[0099] According to an embodiment of the present invention, in formula (2), if the k-th nearest neighbor reference pixel is on or within the correction boundary, then its corresponding loss function is: Its value is equal to ,in, Values ​​between 0 and 1, and The larger, the better The smaller the value, the greater the selection metric for pixels on the correction boundary, and the larger the selection metric for pixels closer to the correction boundary. Therefore, pixels closer to the boundary line correspond to... The smaller the value, the better. The smaller the overall value, the more likely it is to cause problems. The larger the overall value, the greater the training adjustment range and efficiency when the loss function is reduced during training. This improves the training strength and value of pixels near the boundary line, and to a large extent improves the accuracy of the image segmentation model in judging whether pixels near the boundary line belong to the tumor region. The closer to the boundary line, the greater the training strength, and the higher the accuracy of the image segmentation model in recognizing the boundary line.

[0100] According to one embodiment of the present invention, similarly, if the k-th nearest neighbor reference pixel is outside the correction boundary, then its corresponding loss function is: Its value is equal to Furthermore, the closer the nearest neighbor reference pixel is to the correction boundary, the larger the selection index. The smaller, The smaller, The larger the pixel size, the greater the training intensity and value of pixels near the boundary line. This significantly improves the accuracy of the image segmentation model in determining whether pixels near the boundary line belong to the tumor region. Furthermore, the closer the pixel is to the boundary line, the greater the training intensity, resulting in higher accuracy for the image segmentation model in recognizing the boundary line.

[0101] According to one embodiment of the present invention, the boundary selection loss function can be obtained by summing the loss functions corresponding to each nearest neighbor reference pixel for each pixel on the boundary line of the region to be determined. Further, the boundary selection loss function and the enhancement training loss function can be weighted and summed (each weight can be 0.5) to obtain the loss function of the image segmentation model. Through backpropagation of the loss function, the parameters of the image segmentation model can be adjusted using gradient descent to reduce the loss function. An updated image segmentation model can be obtained by training with multiple CT images.

[0102] In this way, the training intensity of the loss function can be adjusted by selecting indicators, so that the training intensity and training value of pixels closer to the boundary line are greater, thereby rapidly improving the image segmentation model's accuracy in recognizing boundary lines and the accuracy in judging whether pixels near the boundary line are within the tumor area, thus improving training efficiency and training effect.

[0103] According to an embodiment of the present invention, in step S8, other CT images can be segmented using the updated image segmentation model to obtain image segmentation results of CT images with multiple slices, that is, a more accurate segmentation result of the region where the tumor is located.

[0104] According to an embodiment of the present invention, the method for precise tumor segmentation in CT images, after obtaining the region to be determined where the tumor is located in the CT image of each slice through an image segmentation model, can mutually verify the boundary line of the region to be determined based on the nearest neighbor pixels on the boundary line and the pixels of adjacent slices. This allows for the selection of pixels that are more representative of the tumor edge from the nearest neighbor pixels on the boundary line, thereby achieving more precise segmentation of the tumor region. When determining the selection criteria, the nearest neighbor reference pixels can be verified by the projected pixels in adjacent slice CT images, and the change benchmark of the comprehensive level of the pixels near the nearest neighbor reference pixels can be obtained. Then, the selection criteria are determined based on the change in the comprehensive level of the nearest neighbor reference pixels and adjacent pixels, as well as the change benchmark. Based on the selection criteria, the nearest neighbor reference pixel with the largest relative difference in comprehensive level on both sides is determined as the pixel on the boundary line, thus determining the corrected boundary line. This maximizes the relative difference in comprehensive level inside and outside the corrected boundary line, improving the contrast inside and outside the tumor region, and also improving the accuracy and objectivity of the boundary line. When updating and training an image segmentation model, the training intensity of the loss function can be adjusted by selecting an index, so that the training intensity and training value of pixels closer to the boundary line are greater, thereby rapidly improving the image segmentation model's accuracy in recognizing the boundary line and the accuracy in judging whether pixels near the boundary line are within the tumor area, thus improving training efficiency and training effect.

[0105] Figure 2 A block diagram of a precise tumor segmentation system for CT images according to an embodiment of the present invention is shown, the system comprising:

[0106] The undetermined region module uses an image segmentation model to determine the undetermined region where the tumor is located in the CT image of each slice.

[0107] The center position module obtains the boundary line of the region to be determined and the center position of the region to be determined;

[0108] The nearest neighbor reference pixel module determines the nearest neighbor reference pixels of the pixels on the boundary line based on the pixels on the region to be determined and its boundary line, as well as the center position.

[0109] The probability information determination module determines the probability information of a neighboring pixel belonging to the tumor region, obtained by the image segmentation model processing the neighboring reference pixel.

[0110] The pixel correction module selects correction pixels from nearby reference pixels based on the first range of the undetermined region in the CT image, the second range of the undetermined region in the CT images of adjacent slices, and the determination probability information.

[0111] The correction region module determines the correction region where the tumor is located in the CT image of each slice based on the correction pixels.

[0112] The update module updates and trains the image segmentation model based on the correction region to obtain the updated image segmentation model;

[0113] The image segmentation results module uses an updated image segmentation model to segment CT images from multiple slices and obtain the image segmentation results.

[0114] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0115] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for precise tumor segmentation in CT images, characterized in that, include: The image segmentation model is used to determine the undetermined region where the tumor is located in the CT image of each slice; Obtain the boundary lines and center position of the region to be determined; Based on the pixels in the region to be determined and its boundary line, as well as the center position, determine the nearest neighbor reference pixels of the pixels on the boundary line. The probability information for determining whether a neighboring pixel belongs to the tumor region is obtained by processing the neighboring reference pixel using the image segmentation model. Based on the first range of the region to be determined in the CT image, the second range of the region to be determined in the CT images of adjacent slices, and the determination probability information, correction pixels are selected from the nearest neighbor reference pixels. This step includes: determining the selection index of each nearest neighbor reference pixel relative to the CT images of each adjacent slice based on the CT values ​​and determination probability information of the nearest neighbor reference pixels, the CT values ​​and determination probability information of the projection search point, and the CT values ​​and determination probability information of the reference pixel. Specifically, this includes: according to the formula... The selection index for obtaining the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. ,in, Let be the CT value of the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. This refers to the probability information of determining the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. Let be the CT value of the (k+1)th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. This refers to the probability information of determining the (k+1)th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice. Let be the CT value of the k-th projection search point on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. This refers to the probability information for determining the k-th projection search point on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. Let be the CT value of the k-th reference pixel on the intersection line numbered (i,j) in the CT image of the s-th slice adjacent to the i-th slice. This refers to the probability information for determining the k-th reference pixel on the intersection line numbered (i,j) of the CT image of the s-th slice adjacent to the i-th slice. This represents the number of faults adjacent to the i-th fault. The nearest neighbor reference pixel with the highest index is selected as the correction pixel. Based on the corrected pixels, determine the corrected region where the tumor is located in the CT image of each slice; Based on the correction region, the image segmentation model is updated and trained to obtain the updated image segmentation model; The updated image segmentation model is used to segment CT images of multiple slices to obtain image segmentation results.

2. The method for precise tumor segmentation in CT images according to claim 1, characterized in that, Based on the region to be determined and the pixels on its boundary line, as well as the center position, determine the nearest neighbor reference pixels of the pixels on the boundary line, including: For the region to be determined in the CT image of the i-th slice, connect the j-th pixel on its boundary line with the center position to obtain the line numbered (i,j); Based on the length of the line numbered (i,j), extend the length of the line outwards towards the outside of the region to be determined to obtain the search line numbered (i,j), where the extension length is proportional to the length of the line. Using the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice as the center, and the extended length of the search line numbered (i,j) as the radius, obtain the circular search neighborhood numbered (i,j). The pixel located within the circular search neighborhood numbered (i,j) and on the search line numbered (i,j), and whose probability information belongs to the preset interval, is determined as the nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice.

3. The method for precise tumor segmentation in CT images according to claim 2, characterized in that, Based on the first range of the region to be determined in the CT image, the second range of the region to be determined in the CT images of adjacent slices, and the determination probability information, correction pixels are selected from the nearest reference pixels, including: Obtain the center location of the second range of the region to be determined in the CT images of adjacent slices; Based on the first range, the center position of the first range, and the center position of the second range, the projection range of the first range in adjacent tomographic CT images is determined, wherein the center position of the projection range overlaps with the center position of the second range, and the shape and size of the projection range are the same as the first range. Obtain the projection line of the line numbered (i,j) in the CT image of the adjacent slice, wherein the projection line numbered (i,j) has one end at the center of the projection range and is parallel to the line numbered (i,j). Extend the projection line numbered (i,j) so that the projection line numbered (i,j) intersects with both the boundary line of the projection range and the boundary line of the second range, thus obtaining the intersection line numbered (i,j). Determine the projection search point of the nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined in the CT image of the i-th slice on the intersection line numbered (i,j). Centered on the intersection of the line of intersection numbered (i,j) and the boundary line of the second range, and with the extended length of the search line numbered (i,j) as the radius, a reference neighborhood numbered (i,j) is set. The pixel located within the reference neighborhood of (i,j) and on the intersection line of (i,j) is set as the reference pixel.

4. The method for precise tumor segmentation in CT images according to claim 3, characterized in that, Based on the correction region, the image segmentation model is updated and trained to obtain the updated image segmentation model, including: Based on the correction region, the correction boundary of the region where the tumor is located; Set first annotation information for pixels within the correction boundary, and set second annotation information for pixels outside the correction boundary; Obtain the probability information of each pixel obtained by the image segmentation model; Based on the determination probability information, the first annotation information and the second annotation information, as well as the nearest neighbor reference pixels and their selection index, the loss function of the image segmentation model is determined; The image segmentation model is trained based on the loss function of the image segmentation model to obtain an updated image segmentation model.

5. The method for precise tumor segmentation in CT images according to claim 4, characterized in that, Based on the determination probability information, the first annotation information, the second annotation information, and the nearest neighbor reference pixels and their selection indices, the loss function of the image segmentation model is determined, including: Based on the probability information of other pixels in the CT image besides the nearest neighbor reference pixel, as well as the first annotation information and the second annotation information, the enhancement training loss function is determined. Based on the nearest neighbor reference pixels in the CT image, their selection index and decision probability information, as well as the first annotation information and the second annotation information, the boundary selection loss function is determined. The loss function of the image segmentation model is determined based on the enhanced training loss function and the boundary selection loss function.

6. The method for precise tumor segmentation in CT images according to claim 5, characterized in that, Based on the nearest neighbor reference pixels in the CT image, their selection indices and decision probabilities, as well as the first and second annotation information, the boundary selection loss function is determined, including: According to the formula Determine the boundary and choose the loss function ,in, Let $\frac{j}{k}$ be the probability value that the k-th nearest neighbor reference pixel corresponding to the $j-th pixel on the boundary line of the region to be determined based on the first annotation information belongs to the region where the tumor is located. This refers to the probability information of the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined. The selection criterion is the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined. Let m be the probability value that the k-th nearest neighbor reference pixel corresponding to the j-th pixel on the boundary line of the region to be determined according to the second annotation information belongs to the region where the tumor is located, m is the number of nearest neighbor reference pixels corresponding to the j-th pixel on the boundary line of the region to be determined, and n is the number of pixels on the boundary line of the region to be determined.

7. A precise tumor segmentation system for CT images, used to perform the method as described in any one of claims 1-6, characterized in that, include: The undetermined region module uses an image segmentation model to determine the undetermined region where the tumor is located in the CT image of each slice. The center position module obtains the boundary line of the region to be determined and the center position of the region to be determined; The nearest neighbor reference pixel module determines the nearest neighbor reference pixels of the pixels on the boundary line based on the pixels on the region to be determined and its boundary line, as well as the center position. The probability information determination module determines the probability information of a neighboring pixel belonging to the tumor region, obtained by the image segmentation model processing the neighboring reference pixel. The pixel correction module selects correction pixels from nearby reference pixels based on the first range of the undetermined region in the CT image, the second range of the undetermined region in the CT images of adjacent slices, and the determination probability information. The correction region module determines the correction region where the tumor is located in the CT image of each slice based on the correction pixels. The update module updates and trains the image segmentation model based on the correction region to obtain the updated image segmentation model; The image segmentation results module uses an updated image segmentation model to segment CT images from multiple slices and obtain the image segmentation results.

Images