A cross-modal medical image segmentation method and system

Abstract

The present invention discloses a cross-modal medical image segmentation method and system, comprising: pixel-level annotation of images in a modality A image set, and adjusting the pixel-level labels obtained on the modality A image set according to the image resolution of the modality B image set; dividing the adjusted pixel-level labels, performing hybrid learning training on clean label images and noisy label images, and optimizing the image segmentation model; automatically segmenting the modality A images; automatically generating corresponding segmentation bounding boxes based on the obtained pixel-level segmentation results; adjusting the bounding boxes to the resolution of the modality B images; judging whether the annotated bounding boxes are redundant based on the pixel-level segmentation results on the modality B images obtained after the adjusted bounding boxes, adjusting the positions of the bounding boxes that have been judged to be redundant, and obtaining adjusted segmented images. The present invention improves the efficiency and accuracy of cross-modal segmentation and can effectively improve the applicability of cross-modal segmentation.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation and relates to a cross-modal medical image segmentation method and system. Background Art

[0002] In clinical settings, clinicians often need to analyze multiple modal images of patients for diagnosis and treatment planning; therefore, in the field of medical imaging, the problem of multimodal segmentation is a focus of research by scholars. Most traditional cross-modal segmentation methods rely on unsupervised adaptive thresholds to achieve effective cross-domain segmentation. These methods are based on generative models, and their segmentation results are affected by the quality of generated images or features. Therefore, they often lack stability and flexibility when dealing with complex multimodal segmentation tasks. In general, existing methods are greatly affected by the differences between the source domain and the target domain. Therefore, when the disparity between the source domain image and the target domain image leads to poor registration or domain adaptation, the effectiveness of cross-modal segmentation is affected. Summary of the Invention

[0003] The purpose of the present invention is to solve the problem in the prior art that when the parallax between the source domain image and the target domain image leads to poor registration or domain adaptation, the effectiveness of cross-modal segmentation is affected, and to provide a cross-modal medical image segmentation method and system.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A cross-modal medical image segmentation method, comprising:

[0006] Divide the collected image data into a modality A image set and a modality B image set;

[0007] Perform pixel-level annotation on the images in the modality A image set, and adjust the pixel-level labels obtained on the modality A image set according to the image resolution of the modality B image set;

[0008] The adjusted pixel-level labels are divided into clean label images and noisy label images, and the clean label images and noisy label images are mixed for training to optimize the image segmentation model.

[0009] Automatically segment the modality A image based on the image segmentation model; automatically generate the corresponding segmentation bounding box based on the obtained pixel-level segmentation results;

[0010] Adjust the bounding box to the resolution of the modality B image; based on the pixel-level segmentation results on the modality B image obtained after the adjusted bounding box, determine whether the annotated bounding box is redundant. If so, remove the redundant initial bounding box hints;

[0011] The position of the redundantly judged bounding box is adjusted to determine whether the segmentation target is in the adjusted bounding box. If so, the position adjustment is stopped; if not, the position adjustment is continued until the preset stop condition is reached to obtain the adjusted segmented image.

[0012] A further improvement of the present invention is:

[0013] Furthermore, the clean label images and the noisy label images are subjected to hybrid learning training, including: training the identified clean label images based on supervised learning, and the loss function is: in, Lce is the cross entropy loss function, Ldsc is the Dice loss function.

[0014] Furthermore, the clean label images and the noisy label images are mixed for training, which also includes:

[0015] Randomly crop the divided sub-image to obtain two views V1 and V2, and preprocess the randomly cropped views V1 and V2;

[0016] The preprocessed views are input into the encoder and momentum encoder of the segmentation network respectively, and the corresponding features of the two views are extracted to achieve unsupervised learning by utilizing the consistency of the features, thus obtaining the feature maps F1 and F2 of the two views.

[0017] For each pair of pixels in F1 and F2, the normalized distance between them is calculated; if the normalized distance between a pair of pixels in feature map F1 and feature map F2 is lower than a preset threshold, they are considered to be a positive pair;

[0018] The sub-image obtained by division is regarded as a binary image S, and the binary image S is average pooled to obtain an updated binary image S′, and the updated binary image S′ has the same size as the feature map F1 and the feature map F2; if , then it represents the corresponding pixel in the image It may be a noise sample with an incorrect label. Contrastive learning is performed based on the binary image S, focusing only on the noise sample. The pixel-level loss of the contrastive learning algorithm is defined as: in, , Represents pixels from feature map F1 and pixels in feature map F2 The eigenvector of In view V2, Forming a set of pixels that are facing each other;

[0019] By comparing the pixels of the two views V1 and V2, the unsupervised loss Lu is obtained as: Where m represents the mth individual image in the mini-batch M; and Represent the pixel sets of views V1 and V2 respectively; and is a collection and The size of k Is the index variable used to traverse the collection Elements in M Indicates the total number of samples; Representing a collection Chinese elements k Pixel-level loss;

[0020] For pixels with noisy labels, a combination of supervised training and self-supervised training is used, and the loss function is: in, is the loss function for training the identified clean label images based on supervised learning.

[0021] Furthermore, the pixel-level segmentation result of the modality B image is obtained based on the adjusted bounding box, specifically: based on the SAM segmentation model, the modality B image is segmented using the bounding box as a prompt.

[0022] Furthermore, based on the pixel-level segmentation result on the modality B image obtained by the adjusted bounding box, it is determined whether the marked bounding box is redundant, specifically:

[0023] Define the initial bounding box, the center coordinates are , with a width of , the height is h , applying a perturbation and , get the new bounding box ;in, and ; ; ; is the disturbance coefficient;

[0024] For the new bounding box, segment the images in the modality A image set based on the pre-trained image segmentation model, obtain the segmentation results, and calculate the Dice similarity between the segmentation results;

[0025] Based on Dice similarity and set threshold Determine whether different segmentation results are similar. If the differences between different segmentation results are large, the bounding box is considered redundant.

[0026] Furthermore, the Dice similarity between the segmentation results is:

[0027] For each pair of segmentation results, calculate the Dice similarity between them; the calculation formula of the Dice coefficient is: in, and There are two segmentation results. is the size of the intersection area between them, and are the sizes of their respective foreground regions.

[0028] Furthermore, the image segmentation model is a nnU-Net segmentation model.

[0029] Furthermore, the position of the bounding box after redundant judgment is adjusted to determine whether the segmentation target is in the position-adjusted bounding box. Specifically:

[0030] Define the coordinates of the bounding box R before adjustment, and get the segmentation result based on the segmentation network and the input image X Fθ (X, R); where the lower left coordinate of the bounding box R is , upper right coordinate ,width ,high ;

[0031] Calculate the distance between the segmentation result and the border of the box in the x-axis and y-axis directions; if the distance on the x-axis is less than Or the distance on the y-axis is less than , the segmentation result is judged to be close to the boundary; if the distance on the x-axis is greater than or the distance on the y-axis is greater than , then the segmentation result is judged to be far away from the boundary;

[0032] According to the judgment result, the border position of the prompt box is adjusted in the x-axis and y-axis directions; the adjustment amplitude on the x-axis each time is Or the corresponding height adjustment on the y-axis is ; Parameters for controlling distance judgment and adjustment amplitude;

[0033] The similarity between the adjusted prompt box and the segmentation result before adjustment is calculated based on the Dice coefficient; if the Dice coefficient is greater than a preset threshold, the prompt adjustment is stopped, or if the maximum adjustment number Imax is reached, the adjustment is stopped.

[0034] A cross-modal medical image segmentation system, comprising:

[0035] a partitioning module, wherein the partitioning module partitions the acquired image data to obtain a modality A image set and a modality B image set;

[0036] A labeling module that performs pixel-level labeling on the images in the modality A image set and adjusts the pixel-level labels obtained on the modality A image set according to the image resolution of the modality B image set;

[0037] A training module, wherein the training module divides the adjusted pixel-level labels to obtain clean label images and noisy label images, respectively, and performs hybrid learning training on the clean label images and the noisy label images to optimize the image segmentation model;

[0038] A segmentation module, which automatically segments the modality A image based on an image segmentation model; and automatically generates a corresponding segmentation bounding box based on the obtained pixel-level segmentation results;

[0039] a first determination module, which adjusts the bounding box to the resolution of the modality B image; based on a pixel-level segmentation result on the modality B image obtained after the adjusted bounding box, determines whether the marked bounding box is redundant, and if so, removes the redundant initial bounding box hint;

[0040] The second judgment module adjusts the position of the bounding box that has undergone redundant judgment, and determines whether the segmentation target is in the bounding box with adjusted position. If so, the position adjustment is stopped; if not, the position adjustment is continued until a preset stop condition is reached to obtain an adjusted segmented image.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] This method achieves accurate segmentation results for modality A images through supervised training, converts the segmentation results into bounding box prompts, and uses a SAM-based network to provide prompt information, determine whether the bounding boxes are redundant, and adjust the positions of the bounding boxes that pass the redundancy judgment to complete cross-modality segmentation of modality B images. This method improves the efficiency and accuracy of cross-modality segmentation and effectively expands the applicability of cross-modality segmentation, laying the foundation for future applications in a wider range of cross-modality segmentation scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1Schematic diagram of the process of the cross-modal medical image segmentation method of the present invention;

[0045] Figure 2 Schematic diagram of the structure of the cross-modal medical image segmentation system of the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0047] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0048] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0049] In the description of the embodiments of the present invention, it should be noted that if the terms "upper," "lower," "horizontal," "inner," etc. appear, the orientation or positional relationship indicated is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the inventive product is typically placed when in use. These terms are merely for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0050] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0051] In the description of the embodiments of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0052] The present invention is described in further detail below with reference to the accompanying drawings:

[0053] See also Figure 1 The present invention discloses a cross-modal medical image segmentation method, comprising:

[0054] S101: Divide the collected image data to obtain a modality A image set and a modality B image set;

[0055] S102: Perform pixel-level annotation on the images in the modality A image set, and adjust the pixel-level labels obtained on the modality A image set according to the image resolution of the modality B image set;

[0056] S103: Divide the adjusted pixel-level labels to obtain clean label images and noisy label images, perform hybrid learning training on the clean label images and the noisy label images, and optimize the image segmentation model;

[0057] The clean label images and the noisy label images are mixed for training, including: training the identified clean label images based on supervised learning, and the loss function is: in, Lce is the cross entropy loss function, Ldsc is the Dice loss function.

[0058] Hybrid learning training of clean label images and noisy label images also includes:

[0059] Randomly crop the divided sub-image to obtain two views V1 and V2, and preprocess the randomly cropped views V1 and V2;

[0060] The preprocessed views are input into the encoder and momentum encoder of the segmentation network respectively, and the corresponding features of the two views are extracted to achieve unsupervised learning by utilizing the consistency of the features, thus obtaining the feature maps F1 and F2 of the two views.

[0061] For each pair of pixels in F1 and F2, the normalized distance between them is calculated; if the normalized distance between a pair of pixels in feature map F1 and feature map F2 is lower than a preset threshold, they are considered to be a positive pair;

[0062] The sub-image obtained by division is regarded as a binary image S, and the binary image S is average pooled to obtain an updated binary image S′, and the updated binary image S′ has the same size as the feature map F1 and the feature map F2; if , then it represents the corresponding pixel in the image It may be a noise sample with an incorrect label. Contrastive learning is performed based on the binary image S, focusing only on the noise sample. The pixel-level loss of the contrastive learning algorithm is defined as: in, , Represents pixels from feature map F1 and pixels in feature map F2 The eigenvector of In view V2, Forming a set of pixels that are facing each other;

[0063] By comparing the pixels of the two views V1 and V2, the unsupervised loss Lu is obtained as: Where m represents the mth individual image in the mini-batch M; and Represent the pixel sets of views V1 and V2 respectively; and is a collection and The size of k Is the index variable used to traverse the collection Elements in M Indicates the total number of samples; Representing a collection Chinese elements k Pixel-level loss;

[0064] For pixels with noisy labels, a combination of supervised training and self-supervised training is used, and the loss function is: in, is the loss function for training the identified clean label images based on supervised learning.

[0065] S104: Automatically segmenting the modality A image based on the image segmentation model; automatically generating a corresponding segmentation bounding box based on the obtained pixel-level segmentation result;

[0066] S105: Adjust the bounding box to the resolution of the modality B image; based on the pixel-level segmentation result on the modality B image obtained after the adjusted bounding box, determine whether the marked bounding box is redundant; if so, remove the redundant initial bounding box hint;

[0067] The pixel-level segmentation result on the modality B image is obtained based on the adjusted bounding box. Specifically, the segmentation of the modality B image is achieved based on the SAM segmentation model and the bounding box is used as a prompt.

[0068] Based on the pixel-level segmentation results on the modality B image obtained by the adjusted bounding box, determine whether the marked bounding box is redundant, specifically:

[0069] Define the initial bounding box, the center coordinates are , with a width of , the height is h , applying a perturbation and , get the new bounding box ;in, and ; ; ; is the disturbance coefficient;

[0070] For the new bounding box, segment the images in the modality A image set based on the pre-trained image segmentation model, obtain the segmentation results, and calculate the Dice similarity between the segmentation results;

[0071] Based on Dice similarity and set threshold Determine whether different segmentation results are similar. If the differences between different segmentation results are large, the bounding box is considered redundant.

[0072] The Dice similarity between the segmentation results is:

[0073] For each pair of segmentation results, calculate the Dice similarity between them; the calculation formula of the Dice coefficient is: in, and There are two segmentation results. is the size of the intersection area between them, and are the sizes of their respective foreground regions. The image segmentation model is the nnU-Net segmentation model.

[0074] S106: Adjust the position of the bounding box that has undergone redundant judgment to determine whether the segmentation target is in the adjusted bounding box. If so, stop the position adjustment; if not, continue the position adjustment until a preset stop condition is reached to obtain an adjusted segmented image.

[0075] Define the coordinates of the bounding box R before adjustment, and get the segmentation result based on the segmentation network and the input image X Fθ (X, R); where the lower left coordinate of the bounding box R is , upper right coordinate ,width ,high ;

[0076] Calculate the distance between the segmentation result and the border of the box in the x-axis and y-axis directions; if the distance on the x-axis is less than Or the distance on the y-axis is less than , the segmentation result is judged to be close to the boundary; if the distance on the x-axis is greater than or the distance on the y-axis is greater than , then the segmentation result is judged to be far away from the boundary;

[0077] According to the judgment result, the border position of the prompt box is adjusted in the x-axis and y-axis directions; the adjustment amplitude on the x-axis each time is Or the corresponding height adjustment on the y-axis is ; Parameters for controlling distance judgment and adjustment amplitude;

[0078] The similarity between the adjusted prompt box and the segmentation result before adjustment is calculated based on the Dice coefficient; if the Dice coefficient is greater than a preset threshold, the prompt adjustment is stopped, or if the maximum adjustment number Imax is reached, the adjustment is stopped.

[0079] See also Figure 2 The present invention discloses a cross-modal medical image segmentation system, comprising:

[0080] a partitioning module, wherein the partitioning module partitions the acquired image data to obtain a modality A image set and a modality B image set;

[0081] A labeling module that performs pixel-level labeling on the images in the modality A image set and adjusts the pixel-level labels obtained on the modality A image set according to the image resolution of the modality B image set;

[0082] A training module, wherein the training module divides the adjusted pixel-level labels to obtain clean label images and noisy label images, respectively, and performs hybrid learning training on the clean label images and the noisy label images to optimize the image segmentation model;

[0083] A segmentation module, which automatically segments the modality A image based on an image segmentation model; and automatically generates a corresponding segmentation bounding box based on the obtained pixel-level segmentation results;

[0084] a first determination module, which adjusts the bounding box to the resolution of the modality B image; based on a pixel-level segmentation result on the modality B image obtained after the adjusted bounding box, determines whether the marked bounding box is redundant, and if so, removes the redundant initial bounding box hint;

[0085] The second judgment module adjusts the position of the bounding box that has undergone redundant judgment, and determines whether the segmentation target is in the bounding box with adjusted position. If so, the position adjustment is stopped; if not, the position adjustment is continued until a preset stop condition is reached to obtain an adjusted segmented image.

[0086] An embodiment of the present invention provides a terminal device. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of each of the aforementioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of each module / unit in each of the aforementioned device embodiments are implemented.

[0087] The computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to accomplish the present invention.

[0088] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0089] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0090] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0091] If the module / unit integrated into the terminal device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content of the computer-readable medium can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0092] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A cross-modal medical image segmentation method, characterized in that: include: The collected image data is divided into a modality A image set and a modality B image set; Perform pixel-level annotation on the images in the modality A image set, and adjust the pixel-level labels obtained on the modality A image set according to the image resolution of the modality B image set; The adjusted pixel-level labels are divided into clean label images and noisy label images, and the clean label images and noisy label images are mixed for training to optimize the image segmentation model. Hybrid learning training is performed on clean label images and noisy label images, including: training the identified clean label images based on supervised learning; for pixels with noisy labels, a combination of supervised training and self-supervised training is used for training; Automatically segmenting the modality A image based on an image segmentation model; automatically generating a corresponding segmentation bounding box based on the obtained pixel-level segmentation result; the image segmentation model is an nnU-Net segmentation model; Adjust the bounding box to the resolution of the modality B image; based on the pixel-level segmentation results on the modality B image obtained after the adjusted bounding box, determine whether the annotated bounding box is redundant. If so, remove the redundant initial bounding box prompt; wherein, based on the SAM segmentation model, the bounding box is used as the prompt to achieve segmentation of the modality B image; The position of the redundantly judged bounding box is adjusted to determine whether the segmentation target is in the adjusted bounding box. If so, the position adjustment is stopped; if not, the position adjustment is continued until the preset stop condition is reached to obtain the adjusted segmented image.

2. The cross-modal medical image segmentation method according to claim 1, characterized in that: The hybrid learning training of the clean label image and the noisy label image includes: training the identified clean label image based on supervised learning, and the loss function is: ； in, Lce is the cross entropy loss function, Ldsc is the Dice loss function.

3. The cross-modal medical image segmentation method according to claim 2, characterized in that: The hybrid learning training of the clean label image and the noisy label image further includes: Randomly crop the divided sub-image to obtain two views V1 and V2, and preprocess the randomly cropped views V1 and V2; The preprocessed views are input into the encoder and momentum encoder of the segmentation network respectively, and the corresponding features of the two views are extracted to achieve self-supervised learning by utilizing the consistency of the features, thus obtaining the feature maps F1 and F2 of the two views. For each pair of pixels in F1 and F2, the normalized distance between them is calculated; if the normalized distance between a pair of pixels in feature map F1 and feature map F2 is lower than a preset threshold, they are considered to be a positive pair; The sub-image obtained by division is regarded as a binary image S, and the binary image S is average pooled to obtain an updated binary image S′, and the updated binary image S′ has the same size as the feature map F1 and the feature map F2; if , then it represents the corresponding pixel in the image It may be a noise sample with an incorrect label. Contrastive learning is performed based on the binary image S, focusing only on the noise sample. The pixel-level loss of the contrastive learning algorithm is defined as: ； in, , Represents pixels from feature map F1 and pixels in feature map F2 The eigenvector of In view V2, Forming a set of pixels that are facing each other; By comparing the pixels of the two views V1 and V2, the self-supervised learning loss function Lu is obtained as: ； Where m represents the mth individual image in the mini-batch M; and Represent the pixel sets of views V1 and V2 respectively; and is a collection and The size of k Is the index variable used to traverse the collection Elements in M Indicates the total number of samples; Representing a collection Chinese elements k Pixel-level loss; For pixels with noisy labels, a combination of supervised training and self-supervised training is used, and the loss function is: ； in, is the loss function for training the identified clean label images based on supervised learning.

4. The cross-modal medical image segmentation method according to claim 3, characterized in that: The pixel-level segmentation result on the modality B image obtained based on the adjusted bounding box is used to determine whether the marked bounding box is redundant, specifically: Define the initial bounding box, the center coordinates are , with a width of , the height is h , applying a perturbation and , get the new bounding box ;in, and ; ; ; is the disturbance coefficient; For the new bounding box, segment the images in the modality B image set, obtain the segmentation results, and calculate the Dice similarity between the segmentation results; Based on Dice similarity and set threshold Determine whether different segmentation results are similar. If the differences between different segmentation results are large, the bounding box is considered redundant.

5. The cross-modal medical image segmentation method according to claim 4, characterized in that: The Dice similarity between the segmentation results is specifically: For each pair of segmentation results, calculate the Dice similarity between them; the calculation formula of the Dice coefficient is: ； in, and There are two segmentation results. is the size of the intersection area between them, and are the sizes of their respective foreground regions.

6. The cross-modal medical image segmentation method according to claim 5, characterized in that: The position adjustment of the redundantly judged bounding box and the determination of whether the segmented target is within the position-adjusted bounding box are specifically as follows: Define the coordinates of the bounding box R before adjustment, and get the segmentation result based on the segmentation network and the input image X F θ(X, R); where the lower left coordinate of the bounding box R is , upper right coordinate ,width ,high ; Calculate the distance between the segmentation result and the border of the box in the x-axis and y-axis directions; if the distance on the x-axis is less than Or the distance on the y-axis is less than , the segmentation result is judged to be close to the boundary; if the distance on the x-axis is greater than or the distance on the y-axis is greater than , then the segmentation result is judged to be far away from the boundary; According to the judgment result, the border position of the prompt box is adjusted in the x-axis and y-axis directions; the adjustment amplitude on the x-axis each time is Or the corresponding height adjustment on the y-axis is ;γ is the parameter that controls distance judgment and adjustment amplitude; The similarity between the adjusted prompt box and the segmentation result before adjustment is calculated based on the Dice coefficient; if the Dice coefficient is greater than a preset threshold, the prompt adjustment is stopped, or if the maximum adjustment number Imax is reached, the adjustment is stopped.

7. A cross-modal medical image segmentation system, characterized in that: A cross-modal medical image segmentation method according to claim 1 is used, comprising: a partitioning module, wherein the partitioning module partitions the acquired image data to obtain a modality A image set and a modality B image set; A labeling module that performs pixel-level labeling on the images in the modality A image set and adjusts the pixel-level labels obtained on the modality A image set according to the image resolution of the modality B image set; A training module, wherein the training module divides the adjusted pixel-level labels to obtain clean label images and noisy label images, respectively, and performs hybrid learning training on the clean label images and the noisy label images to optimize the image segmentation model; A segmentation module, which automatically segments the modality A image based on an image segmentation model; and automatically generates a corresponding segmentation bounding box based on the obtained pixel-level segmentation results; a first determination module, which adjusts the bounding box to the resolution of the modality B image; based on a pixel-level segmentation result on the modality B image obtained after the adjusted bounding box, determines whether the marked bounding box is redundant, and if so, removes the redundant initial bounding box hint; The second judgment module adjusts the position of the bounding box that has undergone redundant judgment, and determines whether the segmentation target is in the bounding box with adjusted position. If so, the position adjustment is stopped; if not, the position adjustment is continued until a preset stop condition is reached to obtain an adjusted segmented image.

Images