Training method and device of lesion segmentation model, and lesion segmentation method and device

Abstract

The application provides a training method and device of a lesion segmentation model, a lesion segmentation method and device. The training method comprises: inputting a two-dimensional CT image containing a target lesion into a pre-trained two-dimensional lesion segmentation model, and outputting a two-dimensional lesion segmentation result of the target lesion; according to position information and size information of the target lesion contained in the two-dimensional lesion segmentation result, marking a voxel unit where the target lesion is located in a three-dimensional CT image to obtain a target three-dimensional CT image with a target mark; inputting the target three-dimensional CT image into a three-dimensional lesion segmentation model, taking the voxel unit where the target mark is located as a region of interest, and outputting a three-dimensional segmentation prediction result of the target lesion; and training the three-dimensional lesion segmentation model based on a segmentation loss between the three-dimensional segmentation prediction result and the target mark until the three-dimensional lesion segmentation model converges. In this way, the application effectively improves the lesion segmentation efficiency of the target lesion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a lesion segmentation model training method and device, and a lesion segmentation method and device. BACKGROUND

[0002] Lesions refer to parts of the body that have undergone pathological changes, and their types are widely and complexly distributed. For example, liver tumors occurring in the liver and lung nodules occurring in the lungs are all lesions that need to be studied in the medical field.

[0003] Taking liver tumors, which are common lesions, as an example, lesion segmentation of liver tumors is of great significance for assisting doctors in diagnosis and treatment. At present, when performing lesion segmentation on liver tumors, the 3D-slicer platform (an open-source medical image analysis platform) is mainly used to perform contour delineation on liver tumors through human-computer interaction. However, this lesion segmentation method is time-consuming and laborious, resulting in low lesion segmentation efficiency. SUMMARY

[0004] Therefore, the present application aims to provide a lesion segmentation model training method and device, and a lesion segmentation method and device, which combine a 2D lesion segmentation model with a 3D lesion segmentation model, thereby effectively improving the lesion segmentation efficiency of a target lesion on the basis of automatically segmenting the target lesion.

[0005] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, together with the accompanying drawings.

[0006] In a first aspect, the embodiments of the present application provide a lesion segmentation model training method, which comprises:

[0007] inputting a 2D CT image containing a target lesion into a pre-trained 2D lesion segmentation model, and outputting a 2D lesion segmentation result for the target lesion;

[0008] labeling a voxel unit in which the target lesion is located in a 3D CT image according to position information and size information of the target lesion contained in the 2D lesion segmentation result, and obtaining a target 3D CT image with a target label;

[0009] inputting the target 3D CT image into a 3D lesion segmentation model, taking a voxel unit in which the target label is located as a region of interest, and outputting a 3D segmentation prediction result for the target lesion;

[0010] train the three-dimensional lesion segmentation model based on a segmentation loss between the three-dimensional segmentation prediction result and the target label until the three-dimensional lesion segmentation model converges

[0011] In a second aspect, the embodiments of the present application provide a lesion segmentation method, which comprises:

[0012] cutting a three-dimensional CT image containing a target lesion according to a reference plane in any view in space to obtain a two-dimensional slice of the three-dimensional CT image under the reference plane;

[0013] inputting the two-dimensional slice as a two-dimensional CT image containing the target lesion into a pre-trained two-dimensional lesion segmentation model to output a two-dimensional lesion segmentation result for the target lesion;

[0014] determining a region of interest in the three-dimensional CT image based on the two-dimensional lesion segmentation result, and inputting the three-dimensional CT image in which the region of interest is determined into a pre-trained three-dimensional lesion segmentation model to output an initial three-dimensional lesion segmentation result for the target lesion; wherein the three-dimensional lesion segmentation model is obtained after being trained according to any of the training methods described above.

[0015] In a third aspect, the embodiments of the present application provide a training device of a lesion segmentation model, which comprises:

[0016] a first prediction module configured to input a two-dimensional CT image containing a target lesion into a pre-trained two-dimensional lesion segmentation model to output a two-dimensional lesion segmentation result for the target lesion;

[0017] a labeling module configured to label a voxel unit in which the target lesion is located in a three-dimensional CT image according to position information and size information of the target lesion contained in the two-dimensional lesion segmentation result to obtain a target three-dimensional CT image with a target label;

[0018] a second prediction module configured to input the target three-dimensional CT image into a three-dimensional lesion segmentation model, take a voxel unit in which the target label is located as a region of interest, and output a three-dimensional segmentation prediction result for the target lesion;

[0019] a model training module configured to train the three-dimensional lesion segmentation model based on a segmentation loss between the three-dimensional segmentation prediction result and the target label until the three-dimensional lesion segmentation model converges.

[0020] In a fourth aspect, the embodiments of the present application provide a lesion segmentation device, which comprises:

[0021] a cutting module configured to cut a three-dimensional CT image containing a target lesion according to a reference plane in any view in a space to obtain a two-dimensional slice of the three-dimensional CT image under the reference plane;

[0022] a first segmentation module configured to input the two-dimensional slice as a two-dimensional CT image containing the target lesion into a pre-trained two-dimensional lesion segmentation model to output a two-dimensional lesion segmentation result for the target lesion;

[0023] a second segmentation module configured to determine a region of interest in the three-dimensional CT image based on the two-dimensional lesion segmentation result, and input the three-dimensional CT image in which the region of interest is determined into a pre-trained three-dimensional lesion segmentation model to output an initial three-dimensional lesion segmentation result for the target lesion, wherein the three-dimensional lesion segmentation model is obtained by training according to the training method of any of the above.

[0024] In a fifth aspect, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the training method of the lesion segmentation model of any of the above or the steps of the lesion segmentation method of any of the above when executing the computer program.

[0025] In a sixth aspect, an embodiment of the present application provides a computer readable storage medium, and the computer readable storage medium stores a computer program, and the computer program is executable on a processor to execute the steps of the training method of the lesion segmentation model of any of the above or the steps of the lesion segmentation method of any of the above.

[0026] The technical scheme provided by the embodiments of the present application can include the following beneficial effects:

[0027] The training method and device of the lesion segmentation model, the lesion segmentation method and device provided by the embodiments of the present application can quickly locate the position and size of the target lesion based on the 2D lesion segmentation model by combining the 2D lesion segmentation model and the 3D lesion segmentation model, and then obtain the ROI of the 3D lesion segmentation model by means of the lesion segmentation result output by the 2D lesion segmentation model, so as to avoid a large number of useless sliding window operations (such as a large number of sliding window operations on non-ROI) on the basis of obtaining a fine lesion segmentation result by the 3D lesion segmentation model, thereby effectively improving the lesion segmentation efficiency for the target lesion on the basis of realizing automatic cutting of the target lesion. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those of ordinary skill in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0029] Figure 1 A flowchart of a method for training a lesion segmentation model is shown;

[0030] Figure 2 A flowchart of another method for obtaining a two-dimensional CT image containing a target lesion is shown;

[0031] Figure 3 A flowchart of a method for obtaining a two-dimensional lesion segmentation result of a target lesion under multiple different viewing angles is shown;

[0032] Figure 4 A flowchart of a method for labeling a target lesion contained in a three-dimensional CT image is shown;

[0033] Figure 5 A flowchart of a method for obtaining a multi-channel target three-dimensional CT image is shown;

[0034] Figure 6 A flowchart of a method for determining a segmentation loss between a three-dimensional segmentation prediction result and a target label is shown;

[0035] Figure 7 A flowchart of a lesion segmentation method is shown;

[0036] Figure 8 A structural diagram of a training device of a lesion segmentation model is shown;

[0037] Figure 9 A structural diagram of a lesion segmentation device is shown;

[0038] Figure 10 A structural diagram of a computer device 1000 is shown. DETAILED DESCRIPTION

[0039] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of description and illustration, and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn according to the actual proportions. The flowcharts used in the present application show the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowcharts can not be implemented in sequence, and the steps without logical context relationship can be reversed in sequence or implemented simultaneously. In addition, one or more other operations can be added to the flowcharts or one or more operations can be removed from the flowcharts under the guidance of the content of the present application.

[0040] In addition, the described embodiments are only some of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0041] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0042] Taking liver tumors in common lesions as an example, at present, when performing lesion segmentation on liver tumors, mainly by means of a 3D-slicer platform (an open-source medical image analysis platform), the liver tumors are contoured through a man-machine interactive mode; but this lesion segmentation mode is time-consuming and laborious, resulting in low lesion segmentation efficiency.

[0043] Based on this, the embodiments of the present application provide a lesion segmentation model training method and device and a lesion segmentation method and device, through a combination of a 2D lesion segmentation model and a 3D lesion segmentation model, the position and size of the target lesion can be quickly located based on the 2D lesion segmentation model, and then the ROI (region of interest) of the 3D lesion segmentation model is obtained by means of the lesion segmentation result output by the 2D lesion segmentation model, so as to avoid a large number of useless sliding window operations (such as a large number of sliding window operations on non-interest regions) on the basis of obtaining a fine lesion segmentation result by means of the 3D lesion segmentation model, thereby effectively improving the lesion segmentation efficiency of the target lesion on the basis of realizing automatic cutting of the target lesion.

[0044] It should be noted that the training method for the lesion segmentation model provided in this application embodiment is applicable to the training device for the lesion segmentation model, and the lesion segmentation method provided in this application embodiment is also applicable to the lesion segmentation device; both the above-mentioned training device and the above-mentioned lesion segmentation device can be integrated into a computer device.

[0045] Specifically, the aforementioned computer equipment can be terminal devices, such as mobile phones, tablets, laptops, desktop computers, etc.; the aforementioned computer equipment can also be servers, which can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers that provide basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, but are not limited to these.

[0046] To facilitate understanding of the embodiments of this application, the following provides a detailed description of a training method and apparatus for a lesion segmentation model, as well as a lesion segmentation method and apparatus.

[0047] Reference Figure 1 As shown, Figure 1 This illustration shows a flowchart of a training method for a lesion segmentation model provided in an embodiment of this application. The training method includes steps S101-S104; specifically:

[0048] S101, input the two-dimensional CT image containing the target lesion into the pre-trained two-dimensional lesion segmentation model, and output the two-dimensional lesion segmentation result for the target lesion.

[0049] Here, the target lesion can be a liver tumor in the liver area, a lung nodule in the lung area, or a kidney tumor in the kidney area, etc. This application embodiment does not limit the specific lesion type to which the above-mentioned target lesion belongs.

[0050] Specifically, considering that the embodiments of the present application mainly perform lesion segmentation on the target lesion by combining a 2D lesion segmentation model (i.e., a two-dimensional lesion segmentation model) with a 3D lesion segmentation model (i.e., a three-dimensional lesion segmentation model), that is, a three-dimensional CT (Computed Tomography, electronic computed tomography) image containing the target lesion needs to be used in the subsequent 3D lesion segmentation model part; based on this, in a preferred implementation, the two-dimensional CT image mentioned in the above step S101 can be a two-dimensional slice (i.e., a two-dimensional CT picture) obtained by intercepting the three-dimensional CT image containing the target lesion through any one reference plane in space (such as the xy plane where the x axis and the y axis in the space coordinate system, the yz plane where the z axis and the y axis in the space coordinate system, etc.).

[0051] It should be noted that since the target lesion has a certain spatial volume in the three-dimensional CT image, when the three-dimensional CT image is intercepted through any one reference plane in space, multiple two-dimensional CT images containing the target lesion can be intercepted; based on this, the number of images of the two-dimensional CT image input into the two-dimensional lesion segmentation model is not limited by the embodiments of the present application (equivalent to that the two-dimensional lesion segmentation model can output the two-dimensional lesion segmentation result of the target lesion in one two-dimensional CT image, and the two-dimensional lesion segmentation model can also output the two-dimensional lesion segmentation result of the target lesion in multiple two-dimensional CT images).

[0052] S102, according to the position information and the size information of the target lesion contained in the two-dimensional lesion segmentation result, mark the voxel unit where the target lesion is located in the three-dimensional CT image to obtain a target three-dimensional CT image with a target mark.

[0053] Here, based on the fact that the target lesion has a certain spatial volume in the three-dimensional CT image, the above voxel unit is equivalent to a basic image unit in the three-dimensional CT image, and its concept can be analogous to a pixel point in the two-dimensional CT image. The specific unit volume value represented by each voxel unit is not limited by the embodiments of the present application.

[0054] Specifically, when the number of the two-dimensional CT images input into the two-dimensional lesion segmentation model is one, the obtained two-dimensional lesion segmentation result is the image region where the target lesion is located in the two-dimensional CT image (equivalent to containing the position information and size information of the target lesion, which can be in the form of 0-1 marking, which represents the distinction between the image region belonging to the target lesion and the image region not belonging to the target lesion in the two-dimensional CT image); at this time, by means of the position information and size information (i.e. the shape and size of the target lesion and other size-related information) of the target lesion contained in the two-dimensional lesion segmentation result, the target marker can be used to mark the voxel unit where the target lesion is located in the three-dimensional CT image, and the target three-dimensional CT image with the target marker is obtained.

[0055] Specifically, when the number of the two-dimensional CT images input into the two-dimensional lesion segmentation model is multiple, multiple two-dimensional lesion segmentation results are obtained, and each two-dimensional lesion segmentation result is the image region where the target lesion is located in each two-dimensional CT image; since the shape of the target lesion is usually irregular, the image region where the target lesion is located in each two-dimensional CT image can be different; at this time, unlike the case of obtaining a single two-dimensional lesion segmentation result, when multiple two-dimensional lesion segmentation results are obtained, the obtained multiple two-dimensional lesion segmentation results can be spliced in space to obtain a splicing result composed of multiple two-dimensional lesion segmentation results, so that according to the position information and size information of the target lesion contained in the splicing result, the target marker can also be used to mark the voxel unit where the target lesion is located in the three-dimensional CT image, and the target three-dimensional CT image with the target marker is obtained.

[0056] S103, inputting the target three-dimensional CT image into the three-dimensional lesion segmentation model, taking the voxel unit where the target marker is located as the region of interest, and outputting a three-dimensional segmentation prediction result for the target lesion.

[0057] Here, since the target three-dimensional CT image and the two-dimensional CT image in the above step S101 are both for the same target lesion, by combining the two-dimensional lesion segmentation model and the three-dimensional lesion segmentation model, when training the three-dimensional lesion segmentation model, the ROI region (i.e. the above-mentioned region of interest) of the three-dimensional lesion segmentation model is obtained by means of the two-dimensional lesion segmentation result output by the two-dimensional lesion segmentation model (i.e. the process of obtaining the target three-dimensional CT image based on the two-dimensional lesion segmentation result), so that the three-dimensional lesion segmentation model can be trained to focus on lesion segmentation for each ROI region, thereby avoiding a large number of useless sliding window operations (such as a large number of sliding window operations for non-interest regions unrelated to the target lesion), and facilitating to improve the training efficiency of the three-dimensional lesion segmentation model.

[0058] S104, training the three-dimensional lesion segmentation model based on the segmentation loss between the three-dimensional segmentation prediction result and the target label until the three-dimensional lesion segmentation model converges.

[0059] Specifically, in the process of training the three-dimensional lesion segmentation model, since the target label is a label of the target lesion in the three-dimensional CT image, that is, it has been determined that it belongs to the lesion rather than the false positive, therefore, based on the segmentation loss between the three-dimensional segmentation prediction result and the target label, the three-dimensional lesion segmentation model can also be trained to remove the false positive function, and further improve the segmentation result accuracy of the three-dimensional lesion segmentation model.

[0060] It should be noted that when calculating the above segmentation loss, a cross-entropy loss function can be used, or other loss functions such as a focalloss function can be used. The specific calculation method of the above segmentation loss is not limited by the embodiments of the present application.

[0061] Here, the specific implementation process of each step in the training method of the above lesion segmentation model is described in detail as follows:

[0062] For the two-dimensional CT image described in step S101, in an optional embodiment, it can be obtained by the above method of intercepting the three-dimensional CT image through any one reference plane in space (equivalent to directly intercepting the three-dimensional CT image without any processing).

[0063] On this basis, considering that different biological tissues have different CT value ranges in CT images (including two-dimensional CT images and three-dimensional CT images) (where the unit of CT value is hu, and the CT value can also be referred to as hu value), in order to highlight the target lesion contained in the CT image and enhance the effective information for segmenting the target lesion, in another optional embodiment, as shown in Figure 2 Figure 2 Another flowchart of a method for obtaining a two-dimensional CT image containing a target lesion is shown, the method comprising steps S201-S203; specifically:

[0064] S201, according to a plurality of different first CT value ranges, performing windowing processing on the three-dimensional CT image to obtain a corresponding windowing processing result under each first CT value range.

[0065] ​Here, each set of first CT value range corresponds to a set of first window width and first window level; wherein, the first window width and the first window level are determined according to the CT value range corresponding to the target tissue where the target lesion is located; for example, taking the target lesion as a liver tumor as an example, the target tissue is the liver part, at this time, according to the CT value range commonly used to observe the lesions in the liver part, a plurality of different first window width and first window level can be selected, and each set of first window width and first window level constitutes a set of first CT value range.

[0066] Specifically, taking the first window width as K1 and the first window level as N1 as an example, the first CT value range constituted by the first window width K1 and the first window level N1 is: At this time, according to the first CT value range, the specific steps of windowing processing of the three-dimensional CT image are:

[0067] (1) Retain the voxel units in the three-dimensional CT image whose CT value is in the first CT value range (equivalent to retaining the gray scale image of this part);

[0068] (2) Set the voxel units in the three-dimensional CT image whose CT value is less than the minimum value (i.e. ) of the first CT value range to all black (or all white);

[0069] (3) Set the voxel units in the three-dimensional CT image whose CT value is greater than the maximum value (i.e. ) of the first CT value range to all white (or all black).

[0070] It should be noted that the specific range of the first CT value range (i.e. the specific CT value range determined by the first window width and the first window level) depends on the target tissue where the target lesion is located, and the specific number of the first CT value range can be set according to the actual lesion segmentation requirement (such as two groups or three groups), and the specific range and the specific number of the first CT value range are not limited by the embodiments of the present application.

[0071] S202, store the corresponding windowing processing result of each first CT value range into an image channel of the three-dimensional CT image to obtain a multi-channel three-dimensional CT image containing the target lesion.

[0072] Here, after the three-dimensional CT image is windowed based on each set of first CT value range, a set of windowing processing results corresponding to each set of first CT value range can be obtained, wherein each windowing processing result can represent a three-dimensional CT image after one windowing processing, and the target lesion in each windowing processing result is more obvious than the target lesion in the three-dimensional CT image without processing.

[0073] Based on this, the result of each windowing processing is stored in an image channel of the three-dimensional CT image, which is equivalent to adding the windowing processing result stored in each image channel as effective information of the segmented target lesion (equivalent to the specific display information of the target lesion under different first CT value ranges) to the original three-dimensional CT image on the basis of the original information dimension of the three-dimensional CT image (i.e. three dimensions of length, width and height of the image), thereby obtaining a multi-channel three-dimensional CT image more conducive to segmenting the target lesion (compared to the original three-dimensional CT image).

[0074] It should be noted that the specific number of image channels of the multi-channel three-dimensional CT image is determined according to the specific number of first CT value ranges in step S201, and the specific number of image channels of the multi-channel three-dimensional CT image is not limited in the embodiments of the present application.

[0075] S203, according to the reference plane under any viewing angle in space, cutting the multi-channel three-dimensional CT image to obtain a two-dimensional slice of the multi-channel three-dimensional CT image under the reference plane, and taking the two-dimensional slice as the two-dimensional CT image.

[0076] Here, considering that the two-dimensional lesion segmentation result of the target lesion is mainly used to determine the position and size of the target lesion in the three-dimensional CT image (so as to provide the ROI region for the three-dimensional lesion segmentation model), therefore, in a preferred embodiment, the multi-channel three-dimensional CT image can be cut based on the reference planes under different viewing angles in space to obtain two-dimensional slices of the multi-channel three-dimensional CT image under different reference planes (such as two-dimensional slices of the multi-channel three-dimensional CT image under the main viewing angle, the overhead angle, the left viewing angle and the right viewing angle), so as to be able to determine the voxel unit of the target lesion in the three-dimensional CT image more accurately according to the two-dimensional lesion segmentation result of the target lesion in each two-dimensional slice by the two-dimensional lesion segmentation model (which is equivalent to the two-dimensional slices under different reference planes clearly showing the volume information of the target lesion in space).

[0077] Specifically, taking the x-axis, y-axis and Z-axis in the spatial coordinate system as an example, the plane where the y-axis and the Z-axis are located can be taken as the first reference plane under the main viewing angle in space, the plane where the x-axis and the y-axis are located can be taken as the second reference plane under the overhead angle in space, and the plane where the x-axis and the z-axis are located can be taken as the third reference plane under the left viewing angle in space; the specific selection method of the reference plane is not limited in the embodiments of the present application.

[0078] On the basis of obtaining the two-dimensional CT image by the method described in steps S201-S203, for the specific implementation process of step S101, as shown in Figure 3 Figure 3 ​A flowchart of a method for obtaining a two-dimensional lesion segmentation result of a target lesion under multiple different viewing angles is shown, and the method comprises steps S301-S303. Specifically,

[0079] S301, obtaining two-dimensional slices of the multi-channel three-dimensional CT image under multiple different reference planes.

[0080] Here, at least two-dimensional slices of the multi-channel three-dimensional CT image under the main viewing angle, the overhead viewing angle and the left viewing angle are needed for the purpose of obtaining the volume information of the target lesion in space. Based on this, in the above step S301, the multiple different reference planes at least include: a first reference plane under the main viewing angle in the space, a second reference plane under the overhead viewing angle in the space, and a third reference plane under the left viewing angle in the space.

[0081] Specifically, the specific implementation of step S301 and the specific determination method of the reference plane under different viewing angles can refer to the specific implementation process of the preceding steps S201-S203, and the repeated parts will not be described here.

[0082] It should be noted that, on the basis of the above first reference plane, the second reference plane and the third reference plane, two-dimensional slices of the multi-channel three-dimensional CT image under other reference planes (such as a reference plane under the right viewing angle in space, etc.) can also be obtained, and the present application does not make any limitation on this.

[0083] S302, inputting the obtained multiple two-dimensional slices into the pre-trained two-dimensional lesion segmentation model, and outputting a lesion segmentation result for each two-dimensional slice.

[0084] Here, the lesion segmentation result represents the prediction result of the image region where the target lesion is located in each two-dimensional slice. For example, the lesion segmentation result can be a labeling result of the two-dimensional slice based on 0-1 labeling, wherein the image region labeled as 1 in the lesion segmentation result represents the prediction result of the image region where the target lesion is located in the two-dimensional slice.

[0085] In the present application, the two-dimensional lesion segmentation model is pre-trained, wherein when the two-dimensional lesion segmentation model is pre-trained, a deep learning model with Unet structure can be used as an original model, wherein a residual connection network Resnet34 can be used in the encoder part of the original model, and the weight coefficients in the two-dimensional lesion segmentation model can be pre-trained by loading Imagenet (an open source database).

[0086] It should be noted that during the training process of the two-dimensional lesion segmentation model, two-dimensional CT images of the target lesion under multiple perspectives can also be used as training samples, so that the two-dimensional lesion segmentation model can learn the lesion features of the target lesion under different perspectives, which is beneficial to enhancing the robustness of the two-dimensional lesion segmentation model.

[0087] S303, the lesion segmentation results of the multiple two-dimensional slices obtained are used as the two-dimensional lesion segmentation results of the target lesion under different perspectives in the space.

[0088] Specifically, when a 3D CT image is cropped using the first reference plane from the main viewpoint, a 2D slice of the 3D CT image from the main viewpoint can be obtained. This 2D slice is then input into a pre-trained 2D lesion segmentation model, which outputs the 2D lesion segmentation result of the target lesion from the main viewpoint. Similarly, when a 3D CT image is cropped using the second reference plane from the top viewpoint, the 2D lesion segmentation result of the target lesion from the top viewpoint can also be obtained. Thus, based on the 2D lesion segmentation results of the target lesion from different viewpoints, the specific spatial range of the target lesion in the 3D CT image can be more clearly defined.

[0089] Regarding the specific implementation process of step S102 above, when the two-dimensional lesion segmentation results of the target lesion under multiple different perspectives are obtained according to the methods described in steps S301-S303 above, as follows: Figure 4 As shown, Figure 4 This invention provides a flowchart illustrating a method for marking target lesions in a three-dimensional CT image, comprising steps S401-S403; specifically:

[0090] S401, based on the reference plane corresponding to each of the two-dimensional slices in the space, the lesion segmentation results of each of the two-dimensional slices are stitched together to obtain a three-dimensional stitched result composed of the lesion segmentation results of multiple two-dimensional slices.

[0091] Specifically, taking the first reference plane, second reference plane, and third reference plane given in step S301 above as examples, for multiple two-dimensional slices a1 intercepted by the first reference plane, multiple two-dimensional slices a2 intercepted by the second reference plane, and multiple two-dimensional slices a3 intercepted by the third reference plane, the lesion segmentation results of multiple two-dimensional slices a1 are placed in the main view in space, the lesion segmentation results of multiple two-dimensional slices a2 are placed in the top view in space, and the lesion segmentation results of multiple two-dimensional slices a3 are placed in the left view in space. By stitching together the lesion segmentation results of the two-dimensional slices under the above multiple different viewpoints, a three-dimensional stitched result composed of the lesion segmentation results of these two-dimensional slices can be obtained.

[0092] S402, determine a target voxel unit in which the target lesion is located in the three-dimensional CT image based on the position information and the size information of the target lesion in the three-dimensional splicing result.

[0093] Here, since the above-mentioned three-dimensional splicing result is obtained based on the two-dimensional lesion segmentation result of the target lesion in different two-dimensional slices, therefore, based on the image region in which the target lesion is located in each two-dimensional lesion segmentation result, the position information and the size information of the target lesion in the three-dimensional splicing result can be directly determined.

[0094] Specifically, the determination manner of the above-mentioned target voxel unit can be determined based on the adaptive data chunking manner, that is, based on the position information and the size information of the target lesion in the three-dimensional splicing result, data chunking is performed on the three-dimensional CT image, and the data chunking result obtained is the target voxel unit in which the target lesion is located in the three-dimensional CT image.

[0095] S403, label the determined target voxel unit with the target label to obtain the target three-dimensional CT image with the target label.

[0096] It should be noted that the above-mentioned target label is only used to distinguish and represent the voxel unit belonging to the target lesion and the voxel unit not belonging to the target lesion in the three-dimensional CT image, and the target label can be a 0-1 label (for example, using the number 1 to label the voxel unit belonging to the target lesion in the three-dimensional CT image, and using the number 0 to label the voxel unit not belonging to the target lesion in the three-dimensional CT image), or an alphabetical label (for example, using the letter a to label the voxel unit belonging to the target lesion in the three-dimensional CT image, and using the letter b to label the voxel unit not belonging to the target lesion in the three-dimensional CT image), or a color label (for example, using red to label the voxel unit belonging to the target lesion in the three-dimensional CT image, and using gray to label the voxel unit not belonging to the target lesion in the three-dimensional CT image); the specific existence form of the above-mentioned target label is not limited in the embodiments of the present application.

[0097] For the specific implementation process of the above-mentioned step S103, similar to the manner of obtaining the multi-channel three-dimensional CT image through windowing processing in the above-mentioned steps S201-S203, in addition to directly inputting the target three-dimensional CT image with the target label as input data into the three-dimensional lesion segmentation model, in another optional implementation manner, the multi-channel target three-dimensional CT image obtained through windowing processing can also be used as the input data of the three-dimensional lesion segmentation model by performing windowing processing on the target three-dimensional CT image, so as to enhance the effective information for lesion segmentation on the basis of the original target three-dimensional CT image.

[0098] Specifically, asFigure 5 As shown, Figure 5 A flowchart of a method for obtaining a multi-channel target three-dimensional CT image is shown, and the method comprises steps S501-S502:

[0099] S501, performing windowing processing on the target three-dimensional CT image according to a plurality of different second CT value ranges, to obtain a corresponding windowing processing result in each second CT value range.

[0100] Here, each second CT value range corresponds to a second window width and a second window level; the second window width and the second window level are determined according to the CT value range corresponding to the target tissue where the target lesion is located.

[0101] It should be noted that the specific range and the specific number of groups of the second CT value range can be the same as the specific range and the specific number of groups of the first CT value range in the above step S201, or can be different from the specific range and the specific number of groups of the first CT value range in the above step S201, and the embodiments of the present application do not make any limitation on this.

[0102] S502, storing the corresponding windowing processing result in each second CT value range into one image channel of the target three-dimensional CT image, obtaining a multi-channel target three-dimensional CT image, and inputting the obtained multi-channel target three-dimensional CT image into the three-dimensional lesion segmentation model.

[0103] Specifically, in a preferred embodiment, the specific number of groups of the first CT value range in the above step S201 can be set to three groups, and the specific number of groups of the second CT value range in the above step S501 can be set to two groups; that is, the model input data of the two-dimensional lesion segmentation model can be set to two-dimensional slices of a three-channel three-dimensional CT image obtained based on different first window widths and first window levels, and the model input data of the three-dimensional lesion segmentation model can be set to a double-channel target three-dimensional CT image obtained based on different second window widths and second window levels.

[0104] Here, the specific implementation process of steps S501-S502 can refer to the specific implementation process of the aforementioned steps S201-S202, and the repeated parts will not be described here.

[0105] Regarding the specific model structure of the three-dimensional lesion segmentation model in steps S103-S104 above, in one optional implementation, a deep learning model with a ResUnet structure can be used; in another preferred implementation, based on the ResUnet structure, the three-dimensional lesion segmentation model can also include multiple layers of image pyramids, wherein each layer of image pyramid is used to process a target three-dimensional CT image at a specific resolution; by sampling the input target three-dimensional CT image at different scales through multiple layers of image pyramids, lesion segmentation features at different scales can be fused at the feature level, thereby maximizing the increase of global image context information, fusing lesion segmentation features at different scales, and achieving the effect of improving the lesion segmentation performance of the three-dimensional lesion segmentation model.

[0106] Based on this, such as Figure 6 As shown, Figure 6 This illustration shows a flowchart of a method for determining the segmentation loss between a 3D segmentation prediction result and a target label, according to an embodiment of this application. The method includes steps S601-S603, specifically:

[0107] S601, for each image pyramid layer, the target 3D CT image at the specific resolution used for processing by the image pyramid layer and the image feature extraction results output by the upper image pyramid layer are used as the target input data of the image pyramid layer.

[0108] It should be noted that the embodiments of this application do not limit the specific number of layers in the image pyramid.

[0109] Specifically, taking a three-layer image pyramid as an example, the input data of the first layer image pyramid (equivalent to the outermost image pyramid of the three-dimensional lesion segmentation model) is the target three-dimensional CT image at the first resolution (equivalent to the model input data of the three-dimensional lesion segmentation model); the input data of the second layer image pyramid is the target three-dimensional CT image at the second resolution and the lesion segmentation feature extraction result output by the upper layer image pyramid (i.e., the first layer image pyramid); the input data of the third layer image pyramid is the target three-dimensional CT image at the third resolution and the lesion segmentation feature extraction result output by the upper layer image pyramid (i.e., the second layer image pyramid); among them, the first resolution, the second resolution, and the third resolution are all different.

[0110] S602, the target input data is input into the image pyramid of this layer, and the segmentation prediction result of the target lesion contained in the target input data is output. Based on the segmentation prediction result and the target label, the prediction loss corresponding to the image pyramid of this layer is calculated.

[0111] Here, in each layer of the image pyramid, the target input data input into the layer of the image pyramid is subjected to a lesion segmentation prediction for the target lesion, and a segmentation prediction result for the target lesion of the layer of the image pyramid is output.

[0112] Based on this, still taking the three-layer image pyramid as an example, if the segmentation prediction result output by the first layer of the image pyramid is denoted as g1, the segmentation prediction result output by the second layer of the image pyramid is denoted as g2, and the segmentation prediction result output by the third layer of the image pyramid is denoted as g3, then through the loss function in each layer of the image pyramid, the prediction loss L1 between g1 and the target label in the first layer of the image pyramid, the prediction loss L2 between g2 and the target label in the second layer of the image pyramid, and the prediction loss L3 between g3 and the target label in the third layer of the image pyramid can be respectively calculated.

[0113] It should be noted that the loss functions in each layer of the image pyramid can be the same or different, and the embodiments of the present application do not make any limitation in this regard.

[0114] S603, based on the prediction loss corresponding to each layer of the image pyramid and the weight coefficient corresponding to each layer of the image pyramid, the weighted sum result calculated by the weighted sum is taken as the segmentation loss.

[0115] Here, the weight coefficient corresponding to each layer of the image pyramid is determined according to the specific resolution corresponding to the layer of the image pyramid; wherein the lower the corresponding specific resolution, the lower the weight coefficient corresponding to the layer of the image pyramid.

[0116] Exemplary, taking the first layer of the image pyramid for processing a target three-dimensional CT image under a first resolution and the second layer of the image pyramid for processing a target three-dimensional CT image under a second resolution as an example, when the first resolution is less than the second resolution, it can be determined that the weight coefficient W1 corresponding to the first layer of the image pyramid is less than the weight coefficient W2 corresponding to the second layer of the image pyramid.

[0117] Specifically, still taking the three-layer image pyramid as an example, the segmentation loss can be calculated by the following formula:

[0118] L 总 = L1 x w1 + L2 x w2 + L3 x w3;

[0119] Wherein, L 总 represents the segmentation loss;

[0120] L1 represents the prediction loss of the first layer of the image pyramid, and w1 represents the loss coefficient corresponding to the first layer of the image pyramid.

[0121] L2 represents the prediction loss of the second-layer image pyramid, and w2 represents the loss coefficient corresponding to the second-layer image pyramid.

[0122] L3 represents the prediction loss of the third-layer image pyramid, and w3 represents the loss coefficient corresponding to the third-layer image pyramid.

[0123] In addition to the training method for the lesion segmentation model described above, in the embodiments of this application, after training the lesion segmentation model (two-dimensional lesion segmentation model and three-dimensional lesion segmentation model), in the application stage of the three-dimensional lesion segmentation model, the embodiments of this application also provide the following lesion segmentation method, specifically:

[0124] Reference Figure 7 As shown, Figure 7 This illustration shows a flowchart of a lesion segmentation method provided in an embodiment of this application. The method includes steps S701-S703; specifically:

[0125] S701, based on a reference plane at any viewpoint in space, the three-dimensional CT image containing the target lesion is cropped to obtain a two-dimensional slice of the three-dimensional CT image under the reference plane.

[0126] Here, the specific implementation of step S701 is the same as that of step S203 above, and the repeated parts will not be described again.

[0127] S702, the two-dimensional slice is input into a pre-trained two-dimensional lesion segmentation model as a two-dimensional CT image containing the target lesion, and the two-dimensional lesion segmentation result for the target lesion is output.

[0128] Here, the specific implementation of step S702 is the same as that of step S101 above, and the repeated parts will not be described again.

[0129] It should be noted that the two-dimensional lesion segmentation model used in the training phase of the three-dimensional lesion segmentation model is also pre-trained. Based on this, in the application phase of the three-dimensional lesion segmentation model, the number of two-dimensional CT images input into the two-dimensional lesion segmentation model (equivalent to the number of two-dimensional slices obtained by cropping in step S701 above) is not limited in this embodiment.

[0130] S703, based on the two-dimensional lesion segmentation result, determine the region of interest in the three-dimensional CT image, and input the three-dimensional CT image with the determined region of interest into the pre-trained three-dimensional lesion segmentation model to output the initial three-dimensional lesion segmentation result for the target lesion.

[0131] Here, the three-dimensional lesion segmentation model is obtained after training according to any of the optional embodiments of the training method of the lesion segmentation model, and the specific training method is not described here again.

[0132] In the embodiments of the present application, after obtaining the initial three-dimensional lesion segmentation result, considering that the lesion segmentation is usually a separate segmentation for each target lesion, based on this, after performing the above step S703, the lesion volume filtering method shown in the following step 1-step 2 can be used to filter the initial three-dimensional lesion segmentation result with too small volume, to further improve the segmentation accuracy of the target lesion, specifically:

[0133] Step 1, performing connected domain analysis on the initial three-dimensional lesion segmentation result, and determining a plurality of three-dimensional connected domains belonging to the target lesion from the initial three-dimensional lesion segmentation result.

[0134] Step 2, according to the lesion volume threshold corresponding to the target lesion, screening a target three-dimensional connected domain with a volume greater than or equal to the lesion volume threshold from the plurality of three-dimensional connected domains as the final three-dimensional lesion segmentation result of the target lesion.

[0135] It should be noted that, considering that different types of lesions correspond to different volumes, the above lesion volume threshold can be determined according to the specific lesion type to which the target lesion actually segmented belongs, and the specific value of the above lesion volume threshold is not limited in the embodiments of the present application.

[0136] The training method of the lesion segmentation model and the lesion segmentation method provided in the embodiments of the present application combine the 2D lesion segmentation model and the 3D lesion segmentation model, and based on the 2D lesion segmentation model, the position and size of the target lesion can be quickly located, and then the ROI (region of interest) of the 3D lesion segmentation model is obtained by means of the lesion segmentation result output by the 2D lesion segmentation model, so that on the basis of obtaining the fine lesion segmentation result by the 3D lesion segmentation model, a large number of useless sliding window operations (such as a large number of sliding window operations for non-interest regions) can be avoided, thereby effectively improving the lesion segmentation efficiency of the target lesion on the basis of realizing the automatic cutting of the target lesion.

[0137] Based on the same inventive concept, the training device of the lesion segmentation model corresponding to the training method of the lesion segmentation model in the above embodiments is also provided in the embodiments of the present application. Since the training device in the embodiments of the present application has similar principles to solve the problem as the training method in the above embodiments of the present application, the implementation of the training device can refer to the implementation of the foregoing training method, and the repeated parts will not be described here again.

[0138] Referring toFigure 8 As shown, Figure 8 A structure diagram of a training device of a lesion segmentation model is shown; the training device comprises:

[0139] The first prediction module 801 is configured to input a two-dimensional CT image containing a target lesion into a pre-trained two-dimensional lesion segmentation model, and output a two-dimensional lesion segmentation result for the target lesion;

[0140] The marking module 802 is configured to mark a voxel unit where the target lesion is located in a three-dimensional CT image according to position information and size information of the target lesion contained in the two-dimensional lesion segmentation result, and obtain a target three-dimensional CT image with a target mark;

[0141] The second prediction module 803 is configured to input the target three-dimensional CT image into a three-dimensional lesion segmentation model, take a voxel unit where the target mark is located as a region of interest, and output a three-dimensional segmentation prediction result for the target lesion;

[0142] The model training module 804 is configured to train the three-dimensional lesion segmentation model based on a segmentation loss between the three-dimensional segmentation prediction result and the target mark, until the three-dimensional lesion segmentation model converges.

[0143] In an optional implementation, the first prediction module 801 is configured to obtain the two-dimensional CT image containing the target lesion by the following method:

[0144] According to a plurality of different first CT value ranges, the three-dimensional CT image is subjected to windowing processing to obtain a corresponding windowing processing result under each of the first CT value ranges; wherein each of the first CT value ranges corresponds to a first window width and a first window level; the first window width and the first window level are determined according to a corresponding CT value range of a target tissue where the target lesion is located;

[0145] The corresponding windowing processing result under each of the first CT value ranges is stored in an image channel of the three-dimensional CT image to obtain a multi-channel three-dimensional CT image containing the target lesion;

[0146] According to a reference plane under any viewing angle in space, the multi-channel three-dimensional CT image is cut to obtain a two-dimensional slice of the multi-channel three-dimensional CT image under the reference plane, and the two-dimensional slice is taken as the two-dimensional CT image.

[0147] In an optional implementation, the first prediction module 801 is specifically configured to:

[0148] obtaining two-dimensional slices of the multi-channel three-dimensional CT image under a plurality of different reference planes, wherein the plurality of different reference planes at least include: a first reference plane under a main viewing angle in the space, a second reference plane under a top-down viewing angle in the space, and a third reference plane under a left-side viewing angle in the space;

[0149] inputting the obtained plurality of two-dimensional slices into the pre-trained two-dimensional lesion segmentation model, and outputting to obtain a lesion segmentation result for each two-dimensional slice;

[0150] taking the obtained lesion segmentation results of the plurality of two-dimensional slices as two-dimensional lesion segmentation results of the target lesion under different viewing angles in the space.

[0151] In an optional implementation, the marking module 802 is specifically configured to:

[0152] splicing the lesion segmentation result of each two-dimensional slice based on the corresponding reference plane of each two-dimensional slice in the space, to obtain a three-dimensional splicing result composed of the lesion segmentation results of the plurality of two-dimensional slices;

[0153] determining a target voxel unit in which the target lesion is located in the three-dimensional CT image based on the position information and the size information of the target lesion in the three-dimensional splicing result;

[0154] marking the determined target voxel unit with the target marker, to obtain the target three-dimensional CT image with the target marker.

[0155] In an optional implementation, when the target three-dimensional CT image is input into the three-dimensional lesion segmentation model, the second prediction module 803 is further configured to:

[0156] performing windowing processing on the target three-dimensional CT image according to a plurality of different second CT value ranges, to obtain a corresponding windowing processing result under each second CT value range; wherein each second CT value range corresponds to a second window width and a second window level; the second window width and the second window level are determined according to a corresponding CT value range of a target tissue in which the target lesion is located;

[0157] storing the corresponding windowing processing result under each second CT value range into an image channel of the target three-dimensional CT image, to obtain a multi-channel target three-dimensional CT image, and inputting the obtained multi-channel target three-dimensional CT image into the three-dimensional lesion segmentation model.

[0158] In one optional implementation, the three-dimensional lesion segmentation model includes multiple image pyramids, each image pyramid being used to process a target three-dimensional CT image at a specific resolution; the model training module 804 is used to determine the segmentation loss between the three-dimensional segmentation prediction result and the target label using the following method:

[0159] For each image pyramid layer, the target 3D CT image at the specific resolution used for processing by that image pyramid layer and the image feature extraction results output by the upper image pyramid layer are used as the target input data for that image pyramid layer.

[0160] The target input data is input into the image pyramid of this layer, and the segmentation prediction result of the target lesion contained in the target input data is output. Based on the segmentation prediction result and the target label, the prediction loss corresponding to the image pyramid of this layer is calculated.

[0161] Based on the prediction loss corresponding to each layer of the image pyramid and the weight coefficients corresponding to each layer of the image pyramid, the weighted summation result is used as the segmentation loss by weighted summation; wherein, the weight coefficients corresponding to each layer of the image pyramid are determined according to the specific resolution corresponding to that layer of the image pyramid.

[0162] Based on the same inventive concept, this application also provides a lesion segmentation device corresponding to the above-mentioned lesion segmentation method. Since the principle of the lesion segmentation device in the embodiments of this application is similar to that of the above-mentioned lesion segmentation method in the embodiments of this application, the implementation of the lesion segmentation device can refer to the implementation of the lesion segmentation method, and the repeated parts will not be described again.

[0163] Reference Figure 9 As shown, Figure 9 This illustration shows a schematic diagram of a lesion segmentation device provided in an embodiment of this application. The lesion segmentation device includes:

[0164] The cropping module 901 is used to crop a three-dimensional CT image containing a target lesion according to a reference plane at any viewpoint in space, to obtain a two-dimensional slice of the three-dimensional CT image under the reference plane.

[0165] The first segmentation module 902 is used to input the two-dimensional slice as a two-dimensional CT image containing the target lesion into a pre-trained two-dimensional lesion segmentation model, and output a two-dimensional lesion segmentation result for the target lesion.

[0166] The second segmentation module 903 is configured to determine a region of interest in the three-dimensional CT image based on the two-dimensional lesion segmentation result, input the three-dimensional CT image in which the region of interest is determined into a pre-trained three-dimensional lesion segmentation model, and output an initial three-dimensional lesion segmentation result for the target lesion.

[0167] In an optional implementation, after the initial three-dimensional lesion segmentation result for the target lesion is output, the second segmentation module 903 is further configured to:

[0168] perform connected domain analysis on the initial three-dimensional lesion segmentation result, and determine a plurality of three-dimensional connected domains belonging to the target lesion from the initial three-dimensional lesion segmentation result.

[0169] According to a lesion volume threshold corresponding to the target lesion, a target three-dimensional connected domain with a volume greater than or equal to the lesion volume threshold is selected from the plurality of three-dimensional connected domains as a final three-dimensional lesion segmentation result of the target lesion.

[0170] As shown in Figure 10 The embodiments of the present application provide a computer device 1000 for performing the steps of the training method of the lesion segmentation model or the steps of the lesion segmentation method, and the device includes a memory 1001, a processor 1002, and a computer program stored in the memory 1001 and executable on the processor 1002, wherein the processor 1002 executes the computer program to implement the steps of the training method of the lesion segmentation model or the steps of the lesion segmentation method.

[0171] Specifically, the memory 1001 and the processor 1002 can be general memory and processor, which are not specifically limited here, and when the processor 1002 runs the computer program stored in the memory 1001, the steps of the training method of the lesion segmentation model or the steps of the lesion segmentation method can be executed.

[0172] Corresponding to the training method of the lesion segmentation model and the lesion segmentation method, the embodiments of the present application further provide a computer readable storage medium, and the computer readable storage medium stores a computer program, and the computer program is run by a processor to execute the steps of the training method of the lesion segmentation model or the steps of the lesion segmentation method.

[0173] Specifically, the storage medium can be a general storage medium such as a mobile disk, a hard disk, etc., and when a computer program on the storage medium is run, the steps of the training method of any of the above-mentioned lesion segmentation models can be executed, or the steps of any of the above-mentioned lesion segmentation methods can be executed.

[0174] In the embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other manners. The embodiments described above are merely schematic, for example, the division of the units is only a logical function division, and there can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, can be indirect couplings or communication connections through some interfaces, and electrical, mechanical or other forms.

[0175] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. In actual implementation, some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present application.

[0176] In addition, each functional unit in the embodiments provided by the present application can be integrated in one processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated in one unit.

[0177] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art, or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0178] It should be noted that like reference numerals and letters refer to like items throughout the accompanying drawings, and once an item is defined in one drawing, it should not be further defined and explained in subsequent drawings, and further, the terms "first", "second", "third" and the like are used only to distinguish descriptions, and cannot be understood as indicating or implying relative importance.

[0179] Finally, it should be noted that the above-described embodiments are merely specific embodiments of the present application, and are used to illustrate the technical solutions of the present application, but are not limiting, and the protection scope of the present application is not limited thereto, although the present application is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features, and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application. should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A training method for a lesion segmentation model, characterized in that, The training method comprises: inputting a two-dimensional CT image containing a target lesion into a pre-trained two-dimensional lesion segmentation model, and outputting a two-dimensional lesion segmentation result for the target lesion; according to position information and size information of the target lesion contained in the two-dimensional lesion segmentation result, marking a voxel unit in which the target lesion is located in a three-dimensional CT image, and obtaining a target three-dimensional CT image with a target mark; inputting the target three-dimensional CT image into a three-dimensional lesion segmentation model, taking the voxel unit in which the target mark is located as a region of interest, and outputting a three-dimensional segmentation prediction result for the target lesion; training the three-dimensional lesion segmentation model based on a segmentation loss between the three-dimensional segmentation prediction result and the target mark until the three-dimensional lesion segmentation model converges; wherein the three-dimensional lesion segmentation model comprises a plurality of image pyramids, each image pyramid being used to process a target three-dimensional CT image at a specific resolution; the segmentation loss between the three-dimensional segmentation prediction result and the target mark is determined by the following method: for each image pyramid, taking the target three-dimensional CT image at the specific resolution processed by the image pyramid and an image feature extraction result output by the upper image pyramid as target input data of the image pyramid; inputting the target input data into the image pyramid, outputting a segmentation prediction result for the target lesion contained in the target input data, and calculating a prediction loss corresponding to the image pyramid based on the segmentation prediction result and the target mark; based on the prediction loss corresponding to each image pyramid and a weight coefficient corresponding to each image pyramid, the weighted sum result obtained by weighted summation is taken as the segmentation loss; wherein the weight coefficient corresponding to each image pyramid is determined according to the specific resolution corresponding to the image pyramid.

2. The training method of claim 1, wherein, The two-dimensional CT image containing the target lesion is obtained by the following method: performing windowing processing on the three-dimensional CT image according to a plurality of different first CT value ranges, to obtain a corresponding windowing processing result under each first CT value range; wherein each first CT value range corresponds to a first window width and a first window level; the first window width and the first window level are determined according to a CT value range corresponding to a target tissue where the target lesion is located; storing the windowing processing result corresponding to each first CT value range into an image channel of the three-dimensional CT image, to obtain a multi-channel three-dimensional CT image containing the target lesion; according to a reference plane under any view angle in space, cutting the multi-channel three-dimensional CT image to obtain a two-dimensional slice of the multi-channel three-dimensional CT image under the reference plane, and taking the two-dimensional slice as the two-dimensional CT image.

3. The training method of claim 2, wherein, The inputting of the two-dimensional CT image containing the target lesion into the pre-trained two-dimensional lesion segmentation model and the outputting of the two-dimensional lesion segmentation result for the target lesion comprise: acquire two-dimensional slices of the multi-channel three-dimensional CT image under a plurality of different reference planes; wherein the plurality of different reference planes at least include: a first reference plane under a main viewing angle in the space, a second reference plane under a top-down viewing angle in the space, and a third reference plane under a left-side viewing angle in the space; input the acquired plurality of two-dimensional slices into the pre-trained two-dimensional lesion segmentation model, and output two-dimensional lesion segmentation results for each two-dimensional slice; the obtained two-dimensional lesion segmentation results of the plurality of two-dimensional slices are used as two-dimensional lesion segmentation results of the target lesion under different viewing angles in the space.

4. The training method of claim 3, wherein, the target three-dimensional CT image with the target marker is obtained by marking a voxel unit in which the target lesion is located in the three-dimensional CT image according to position information and size information of the target lesion contained in the two-dimensional lesion segmentation results, and includes: splicing the lesion segmentation results of each two-dimensional slice based on the corresponding reference plane of each two-dimensional slice in the space to obtain a three-dimensional splicing result composed of the lesion segmentation results of the plurality of two-dimensional slices; determining a target voxel unit in which the target lesion is located in the three-dimensional CT image based on position information and size information of the target lesion in the three-dimensional splicing result; the target three-dimensional CT image with the target marker is obtained by marking the determined target voxel unit with the target marker.

5. The training method of claim 1, wherein, the target three-dimensional CT image is input into a three-dimensional lesion segmentation model, including: performing windowing processing on the target three-dimensional CT image according to a plurality of different second CT value ranges to obtain a corresponding windowing processing result under each second CT value range; wherein each second CT value range corresponds to a second window width and a second window level; the second window width and the second window level are determined according to a corresponding CT value range of a target tissue in which the target lesion is located; storing the corresponding windowing processing result under each second CT value range into an image channel of the target three-dimensional CT image to obtain a multi-channel target three-dimensional CT image, and inputting the obtained multi-channel target three-dimensional CT image into the three-dimensional lesion segmentation model.

6. A lesion segmentation method characterized by, the lesion segmentation method includes: cutting a three-dimensional CT image containing a target lesion according to a reference plane under any viewing angle in a space to obtain a two-dimensional slice of the three-dimensional CT image under the reference plane; inputting the two-dimensional slice as a two-dimensional CT image containing the target lesion into a pre-trained two-dimensional lesion segmentation model to output a two-dimensional lesion segmentation result for the target lesion; determining a region of interest in the three-dimensional CT image based on the two-dimensional lesion segmentation result, and inputting the three-dimensional CT image in which the region of interest is determined into a pre-trained three-dimensional lesion segmentation model to output an initial three-dimensional lesion segmentation result for the target lesion; wherein the three-dimensional lesion segmentation model is obtained by training according to any one of the training methods of claims 1 to 5.

7. The lesion segmentation method of claim 6, wherein, After the output obtains the initial three-dimensional lesion segmentation result for the target lesion, the lesion segmentation method further comprises: Performing connected domain analysis on the initial three-dimensional lesion segmentation result, and determining a plurality of three-dimensional connected domains belonging to the target lesion from the initial three-dimensional lesion segmentation result; According to a lesion volume threshold corresponding to the target lesion, filtering out a target three-dimensional connected domain with a volume greater than or equal to the lesion volume threshold from the plurality of three-dimensional connected domains as a final three-dimensional lesion segmentation result of the target lesion. 8.A device for training a lesion segmentation model, comprising: The training device comprises: A first prediction module configured to input a two-dimensional CT image containing a target lesion into a pre-trained two-dimensional lesion segmentation model, and output a two-dimensional lesion segmentation result for the target lesion; A marking module configured to mark a voxel unit of the target lesion in a three-dimensional CT image according to position information and size information of the target lesion contained in the two-dimensional lesion segmentation result, and obtain a target three-dimensional CT image with a target mark; A second prediction module configured to input the target three-dimensional CT image into a three-dimensional lesion segmentation model, take a voxel unit where the target mark is located as a region of interest, and output a three-dimensional segmentation prediction result for the target lesion; A model training module configured to train the three-dimensional lesion segmentation model based on a segmentation loss between the three-dimensional segmentation prediction result and the target mark until the three-dimensional lesion segmentation model converges; The three-dimensional lesion segmentation model contains a plurality of image pyramids, and each image pyramid is configured to process a target three-dimensional CT image at a specific resolution; and the model training module is configured to determine the segmentation loss between the three-dimensional segmentation prediction result and the target mark by the following method: For each image pyramid, the target three-dimensional CT image at the specific resolution processed by the image pyramid and an image feature extraction result output by an upper image pyramid are taken as target input data of the image pyramid; The target input data is input into the image pyramid to output a segmentation prediction result for the target lesion contained in the target input data, and a prediction loss corresponding to the image pyramid is calculated based on the segmentation prediction result and the target mark; The prediction loss corresponding to each image pyramid and a weight coefficient corresponding to each image pyramid are used to calculate a weighted sum result by weighted summation, and the weighted sum result is taken as the segmentation loss; wherein the weight coefficient corresponding to each image pyramid is determined according to the specific resolution corresponding to the image pyramid.

9. A lesion segmentation apparatus characterized by comprising: The lesion segmentation device comprises: A cutting module configured to cut a three-dimensional CT image containing a target lesion according to a reference plane under any view angle in space to obtain a two-dimensional slice of the three-dimensional CT image under the reference plane; A first segmentation module configured to input the two-dimensional slice as a two-dimensional CT image containing the target lesion into a pre-trained two-dimensional lesion segmentation model, and output a two-dimensional lesion segmentation result for the target lesion; and A second segmentation module configured to input the two-dimensional slice as a two-dimensional CT image containing the target lesion into a pre-trained two-dimensional lesion segmentation model, and output a two-dimensional lesion segmentation result for the target lesion. A second segmentation module is configured to determine a region of interest in the three-dimensional CT image based on the two-dimensional lesion segmentation result, input the three-dimensional CT image in which the region of interest is determined into a pre-trained three-dimensional lesion segmentation model, and output an initial three-dimensional lesion segmentation result for the target lesion.

10. An electronic device, comprising: Comprising: A processor, a memory and a bus, the memory storing machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, the machine readable instructions are executed by the processor to execute the steps of the training method of the lesion segmentation model according to any one of claims 1 to 5, or execute the steps of the lesion segmentation method according to any one of claims 6 to 7.

11. A computer readable storage medium, characterized in that, The computer readable storage medium stores a computer program, when the computer program is run by the processor, the steps of the training method of the lesion segmentation model according to any one of claims 1 to 5 are executed, or the steps of the lesion segmentation method according to any one of claims 6 to 7 are executed.

Images