Deformable nnunet-based image segmentation method and apparatus

Abstract

A deformable nnUNet-based image segmentation method and apparatus. The method comprises: acquiring a 3D pelvis image and, on the basis of deformable nnUNet, identifying one or more target regions in the 3D pelvis image, the deformable nnUNet comprising, successively connected, a deformable convolution-based feature encoder, a feature fusion module and a feature decoder; calculating reconstruction parameters on the basis of the 3D pelvis image, the reconstruction parameters at least comprising the radius of a disc region, and calculating the reconstruction parameters on the basis of the area of the largest connected component of the 3D pelvis image; and, on the basis of the reconstruction parameters, obtaining segmented images of the one or more target regions. The accuracy of image segmentation and the effect of image reconstruction are improved.

Description

Image segmentation method and device based on deformable nnUNet Technical Field

[0001] The present invention relates to the field of image processing, and in particular to an image segmentation method and device based on a deformable nnUNet. Background Art

[0002] Pelvic segmentation has always been an essential step in the clinical diagnosis and surgical planning of pelvic bone diseases, and the processing of CT images is a fundamental task in this step. CT images retain the actual anatomical structure, including depth information, providing surgeons with more detailed information about the injury site. Therefore, they are often used for 3D reconstruction for subsequent surgical planning and postoperative outcome evaluation. Therefore, accurately segmenting the pelvic bones from pelvic CT images is crucial for assessing the severity of pelvic injuries and helping surgeons make correct judgments and choose the appropriate surgical method.

[0003] The primary task of pelvic bone segmentation from pelvic CT images is to obtain segmented representations of the four key joints of the pelvic bone structure: the lumbar vertebrae, sacrum, left hip, and right hip. Existing pelvic bone segmentation methods are either manual or semi-automatic, with limited accuracy when addressing image appearance variations caused by multi-site domain offsets, the presence of contrast vessels, fecal stones and chyme, fractures, low dose, metal artifacts, and other factors. With the advancement of deep learning technology in image segmentation, classification, and feature learning, utilizing computer vision, feature recognition, and fusion techniques for pelvic CT image segmentation and classification is an effective solution.

[0004] In recent years, due to the widespread and successful application of UNet in medical imaging, many variants have been derived, and nnUNet is one of the variants with excellent performance. This network focuses on designing a reasonable data processing process and fully utilizing the characteristics of the dataset to train the model. Its performance surpasses other UNet variants that make adjustments to the architecture and has been introduced into a variety of medical image processing tasks. However, when using nnUNet to segment pelvic CT images in the medical field, existing technologies require obtaining 3D pelvic CT images for training deep learning models. The training, verification, and testing of pelvic CT segmentation models require the use of clear, annotated 3D CT images for training to improve the model's ability to fit and generalize to new samples. However, pelvic CT images contain noise such as fractures, blood vessels, and metal artifacts, which reduces the accuracy of CT image segmentation.

[0005] Therefore, in the process of segmentation and reconstruction of 3D pelvic CT, due to the large volume of pelvic data, the possibility of metal and other noise, unclear boundaries, and the long-term lack of effectively labeled data sets for training, there is currently no fruitful research on the segmentation and reconstruction of the pelvis, the features learned by the model are inaccurate, and the segmentation and reconstruction effect is poor.

[0006] Summary of the Invention

[0007] The purpose of the present invention is to provide an image segmentation method and device based on a deformable nnUNet to solve the technical problem of poor image segmentation effect proposed in the above background technology.

[0008] The first aspect of the present invention provides an image segmentation method based on a deformable nnUNet, specifically comprising:

[0009] Acquire a 3D pelvic image, and identify one or more target regions in the 3D pelvic image based on a deformable nnUNet, wherein the deformable nnUNet includes a deformable convolution-based feature encoder, a feature fusion module, and a feature decoder connected in sequence;

[0010] Calculating reconstruction parameters based on the 3D pelvic image, wherein the reconstruction parameters include at least a radius of a disk region, and the reconstruction parameters are calculated based on an area of ​​a maximum connected component of the 3D pelvic image;

[0011] A segmented image of the one or more target regions is obtained based on the reconstruction parameters.

[0012] Preferably, the feature encoder based on deformable convolution includes:

[0013] Offset adjustment of the convolution kernels of multiple convolution layers. The offset of the convolution kernels of each convolution layer is not exactly the same.

[0014] The convolution module is constructed by connecting multiple offset convolution layers using residual connections.

[0015] Connect the convolution module and the difference layer to form a deformable convolution feature encoder,

[0016] A feature encoder based on deformable convolution is used to identify irregular pelvic medical features in 3D pelvic images and obtain an encoded 3D pelvic image feature map.

[0017] Preferably, the adjusting the convolution kernels of the multiple convolution layers specifically includes:

[0018] Preferably, the difference layer adjusts the output of the convolution module using the following methods (3) and (4): x(e) = B(w i ,qj ), (3) B(w i ,q j )=w1q1+w2q2+w3q3+w4q4. (4).

[0019] Preferably, the feature fusion module is composed of Ladder-ASPP, which specifically includes:

[0020] An upper global pooling layer and a lower fusion layer, wherein the upper global pooling layer includes at least a global pooling module, and the output of the global pooling module is mixed with the input image of the feature fusion module;

[0021] The lower fusion layer includes at least multiple parallel dilated convolution layers and fusion layers with variable sampling rates. The levels of each parallel dilated convolution layer are different, and the input of each parallel dilated convolution layer includes at least the input image of the input feature fusion module and the output of the previous parallel dilated convolution layer.

[0022] Preferably, the output of the upper global pooling layer is mixed with the input image of the feature fusion module after 1×1 convolution and Sigmoid, and the mixing process is weighted feature channel multiplication.

[0023] Preferably, the upper pooling layer is calculated in the following manner (Formula 5): y1=B[C1[GP S (x)]]×x, (5).

[0024] Preferably, the lower fusion layer is calculated using the following method (Formula 6):

[0025] Preferably, before identifying one or more target regions in the 3D pelvic image based on the deformable nnUNet, the method further includes training the deformable nnUNet;

[0026] The training of the deformable nnUNet specifically includes:

[0027] Establish a loss function and train the deformable nnUNet using backpropagation;

[0028] The loss function is composed of the sum of the cross entropy loss function and the dice overall loss function.

[0029] A second aspect of the present invention provides an image segmentation device based on a deformable nnUNet, comprising:

[0030] a segmentation module for acquiring a 3D pelvic image and identifying one or more target regions in the 3D pelvic image based on a deformable nnUNet, wherein the deformable nnUNet comprises a deformable convolution-based feature encoder, a feature fusion module, and a feature decoder connected in sequence;

[0031] a calculation module, configured to calculate reconstruction parameters based on the 3D pelvic image, wherein the reconstruction parameters include at least a radius of a disk region, and the reconstruction parameters are calculated based on an area of ​​a maximum connected component of the 3D pelvic image;

[0032] A post-processing module is used to obtain segmented images of the one or more target areas based on the reconstruction parameters.

[0033] The present invention provides an image segmentation method and device based on a deformable nnUNet, which utilizes a feature encoder of a deformable convolution to extract features from pelvic images. Compared with ordinary convolution, the method improves the effectiveness and accuracy of the encoder in identifying and extracting features of irregular pelvic images. At the same time, the method adaptively calculates reconstruction parameters based on the image itself, accurately removes artifacts and holes, and ensures that morphological reconstruction conforms to the actual situation of the image, thereby improving the reconstruction effect from feature information to image, thereby improving the accuracy and effect of image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] FIG1 is a flow chart of an image segmentation method based on a deformable nnUNet according to an embodiment of the present invention;

[0035] FIG2 is a schematic diagram of the structure of a deformable nnUNet model provided by an embodiment of the present invention;

[0036] FIG3 is a schematic diagram of the structure of a feature encoder based on deformable convolution according to an embodiment of the present invention;

[0037] FIG4 is a schematic diagram of the structure of a feature fusion module provided in an embodiment of the present invention;

[0038] FIG5 is a schematic diagram of a segmented image provided by an embodiment of the present invention;

[0039] FIG6 is a schematic structural diagram of an image segmentation device based on a deformable nnUNet provided by an embodiment of the present invention.

[0040] FIG7 is a schematic diagram of the overall model architecture of an image segmentation method based on a deformable nnUNet provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0042] The components of the embodiments of the present invention generally described and shown in the drawings herein may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention.

[0043] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0044] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0045] Overall model architecture of this method: This method proposes a deformable encoder-decoder network and applies it to pelvic segmentation. As shown in Figure 7, this is an enhanced nnUNet, consisting of four parts: an encoder, an intermediate processing module, a decoder, and post-processing. Stage 1 represents a low-resolution nnUNet network, and stage 2 represents a cascade of high- and low-resolution nnUNet networks. The encoder and decoder in both stages have four layers, each consisting of deformable convolutions with a residual structure. Each layer halves the length, width, and height of the input image and doubles the number of channels.

[0046] An embodiment of the present invention provides an image segmentation method based on a deformable nnUNet. As shown in FIG1 , the image segmentation method based on a deformable nnUNet includes:

[0047] Acquire a 3D pelvic image, and identify one or more target regions in the 3D pelvic image based on a deformable nnUNet, wherein the deformable nnUNet includes a deformable convolution-based feature encoder, a feature fusion module, and a feature decoder connected in sequence;

[0048] Calculating reconstruction parameters based on the 3D pelvic image, wherein the reconstruction parameters include at least a radius of a disk region, and the reconstruction parameters are calculated based on an area of ​​a maximum connected component of the 3D pelvic image;

[0049] A segmented image of the one or more target regions is obtained based on the reconstruction parameters.

[0050] The present invention provides an image segmentation method based on a deformable nnUNet, which uses a feature encoder of a deformable convolution to extract features from pelvic images. Compared with ordinary convolution, the method improves the effectiveness and accuracy of the encoder in identifying and extracting features of irregular pelvic images. At the same time, the method adaptively calculates reconstruction parameters based on the situation of the image itself, accurately removes artifacts and holes, and makes the morphological reconstruction conform to the actual situation of the image, thereby improving the reconstruction effect from feature information to image, thereby improving the accuracy and effect of image segmentation.

[0051] Specifically, the 3D pelvic image is a 3D pelvic CT image. Before identifying one or more target areas in the 3D pelvic image based on the deformable nnUNet, the method further includes training the deformable nnUNet. As shown in FIG2 , the deformable nnUNet includes a feature encoder based on a deformable convolution, a feature fusion module, and a feature decoder connected in sequence. The feature encoder based on the deformable convolution is used to encode the features of the 3D pelvic image during the downsampling process to extract the features of the 3D pelvic image. As shown in FIG3 , the feature encoder based on the deformable convolution includes: offset adjustment of the convolution kernels of multiple convolution layers, wherein the offsets of the convolution kernels of each convolution layer are not exactly the same, connecting the multiple offset convolution layers using a residual connection to form a convolution module, connecting the convolution module and the difference layer to form a deformable convolution feature encoder. The feature encoder based on the deformable convolution is used to identify irregular pelvic medical features in the 3D pelvic image and obtain an encoded 3D pelvic image feature map.

[0052] As an optional embodiment, the convolution kernels of multiple convolution layers are adjusted using the following formulas (1) and (2):

[0053] Among them, w represents the weight, e0 represents the position of the pixel, and e n Represents the position of the adjacent pixel that falls into L. And L~ is the deformation result of L. Compared with L, L~ is a pixel that contains the offset position Δe n irregular grid.

[0054] As an optional embodiment, the difference layer is a bilinear difference layer, and the difference layer adjusts the output of the convolution module using the following methods (3) and (4): x(e) = B(w i ,q j ), (3) B(w i ,q j )=w1q1+w2q2+w3q3+w4q4. (4)

[0055] Among them, and because B is two-dimensional, so wi is the corresponding weight, q j is the 4 pixel units around the irregular sampling position that participate in the calculation, B(w i， q j ) is the bilinear interpolation kernel.

[0056] In the prior art, standard square convolution blocks are used for feature extraction. However, in actual use, the feature extraction effect of images with objects of different scales or deformations at different positions is poor. The image segmentation method based on the variable nnUNet provided by the present invention adopts deformable convolution and a parallel network to learn offsets (offsets), so that the convolution kernel is offset at the sampling point and concentrated on the area where the target object is located in the image, thereby adapting the scale or receptive field size to accurately locate the image and improve the feature extraction effect of the image; further, the residual connection method is integrated to increase the amount of information. The information of the next layer includes not only the information x of the layer, but also the new information after the layer is nonlinearly transformed. Such processing makes the information increase layer by layer, avoiding the problem of information loss caused by the increase in the number of network layers. The encoder based on deformable convolution provided by the present invention adds residual deformable convolution during the downsampling process to make the feature encoder more effective and accurate in identifying the medical features of irregular pelvis. After the offset operation, the convolution layer is converted into a deformable convolution that adapts to the edge of the organ or lesion to adapt to the image. Furthermore, residual connection and bilinear difference are added to optimize the deformable convolution, making the effect of convolution on feature extraction more significant and improving the accuracy of feature extraction.

[0057] The feature fusion module is composed of Ladder-ASPP, specifically including: an upper global pooling layer and a lower fusion layer, wherein the upper global pooling layer includes at least a global pooling module, and the output of the global pooling module is mixed with the input image of the feature fusion module; the lower fusion layer includes at least multiple parallel atrous convolution layers and fusion layers with variable sampling rates, each parallel atrous convolution layer has a different level, and the input of each parallel atrous convolution layer includes at least the input image of the input feature fusion module and the output of the previous parallel atrous convolution layer. ASPP is composed of atrous convolution, which uses multiple parallel atrous convolution layers with different sampling rates. Exemplarily, as shown in FIG4 , an example Ladder-ASPP includes an upper global pooling layer. The output of the upper global pooling layer is mixed with the input image of the feature fusion module after 1×1 convolution and Sigmoid, and the mixing process is weighted feature channel multiplication. An example Ladder-ASPP also includes multiple parallel hole convolution layers with variable sampling rates for realizing feature fusion, such as multiple 3×3 convolution layers, whose expansion coefficients can be 1, 2, 5, and 7, respectively. The features extracted at each sampling rate are further processed in a separate branch and fused through concat (shown in C in FIG4 ) to generate the input of the next branch.

[0058] As an optional embodiment, the upper pooling layer calculates y1=B[C1[GP S (x)]]×x, (5)

[0059] Among them, x is the feature map obtained by the feature encoder, and then global pooling is performed using GP S (x) represents, C1 represents the weight of each feature channel through 1×1 convolution, and B is the normalization of the feature weight.

[0060] The lower fusion layer is calculated as follows (Formula 6):

[0061] Among them, let G K，D is the output of densely connected pyramid pooling, where K is the level of the pyramid and D is the dilation rate. Furthermore, in the case where ⊕ represents a cascade operation, each x is represented as:

[0062] The feature fusion module provided by the present invention constructs convolution kernels with different receptive fields using different dilation rates to acquire multi-scale object information. The features extracted for each sampling rate are further processed in separate branches and fused to generate the final result, thereby improving the processing accuracy of the feature fusion processing effect. To preserve feature information, global pooling and 1×1 convolution paths are further added to ensure that important features are not lost, thereby improving the image processing accuracy of the feature fusion module and the quality of the fused image.

[0063] Furthermore, the feature decoder includes multiple layers of decoding modules, and the input of each layer of decoding modules at least includes the output of the previous layer of decoding modules and the image information input to the feature decoder.

[0064] The training of the deformable nnUNet specifically includes: establishing a loss function and training the deformable nnUNet using back propagation; the loss function is composed of the sum of the cross entropy loss function and the dice overall loss function. Compared with the traditional dice overall loss function, which can learn detail losses but may not be able to back propagate gradients, the method provided by the present invention introduces the cross entropy loss function to alleviate the problem of gradient disappearance, so that the overall L loss It can update parameters more accurately and optimize the performance of the model.

[0065] Preferably, the loss function is specifically: L loss =L cross +L dice . (7)

[0066] After completing the deformable nnUNet pre-training, a 3D pelvic image is obtained, and one or more target regions in the 3D pelvic image are identified based on the deformable nnUNet. The one or more target regions include at least four important joints of the pelvic bone structure: lumbar vertebrae, sacrum, left hip, and right hip.

[0067] Calculating reconstruction parameters based on the 3D pelvic image, wherein the reconstruction parameters include at least a radius of a disk region, and the reconstruction parameters are calculated based on an area of ​​a maximum connected component of the 3D pelvic image;

[0068] Calculate the ratio of the largest connected component in the 3D pelvic image to the total area of ​​the 3D pelvic image, and calculate the radius of the disk area based on the ratio: r = 30 × round (R / (H × W)) + 1, (10)

[0069] Where r is the radius of the disk area, R represents the area of ​​the largest connected region in the segmentation result, and H and W represent the height and width of the input image, respectively.

[0070] For the disk area, if the ratio value is larger, a larger S will be used; conversely, if the ratio value is smaller, a smaller S will be used.

[0071] The segmented images of the one or more target areas are obtained based on the reconstruction parameters, specifically including: constructing a target disc based on the reconstruction parameters, and obtaining the segmented images of the one or more target areas based on the target disc and the one or more target areas in the 3D pelvic image. As shown in FIG5 , FIG5 shows a comparison example of reconstructed medical images after morphological processing. The reconstruction process is actually morphological reconstruction, and finally the segmented representations of the lumbar spine, sacrum, left hip and right hip are obtained. Among them, MCR represents the maximum connected region of the anatomical structure in the predicted result, SDF represents the MCR-based signed distance function, and GT represents the true segmentation result.

[0072] Morphological reconstruction enables binary image filtering while maintaining large objects unchanged, removing artifacts and holes and smoothing boundaries. Compared to fixed disk size methods, this method flexibly determines the disk size based on image characteristics. If the disk is larger, more small areas are removed. Conversely, when the reconstruction parameter values ​​are smaller, fewer areas are removed, maximizing the preservation of key image features and improving image segmentation effectiveness.

[0073] The deformable nnUNet-based image segmentation proposed in this paper can be applied to pelvic segmentation. Compared to the existing nnUNet, the deformable nnUNet provided by this invention uses deformable convolution with a residual structure to generate feature maps. The original image is spliced ​​with the output of different decoder layers to obtain a better feature representation. In addition, unlike the general pyramid pooling module, the deformable nnUNet of this invention includes a better feature fusion module Ladder-ASPP. Ladder-ASPP adopts a dense connection method, but due to the use of depthwise separable convolution, it only requires a small amount of memory. It also improves image recognizability during the morphological reconstruction process.

[0074] First, a feature encoder with residual deformable convolutions performs downsampling to extract features. Feature fusion is performed through the Ladder-ASPP module, and then input into a feature decoder using transposed convolutions to decode features and identify segmented images. Backpropagation is used to optimize performance using the cross-entropy + DICE loss function. The output segmented images are then processed using morphological post-processing. Finally, the final model is trained on the public pelvic dataset CTPlevic1K, with input 3D CT images of size 128×128×128.

[0075] The embodiment of the present invention further provides an image segmentation device based on a deformable nnUNet, as shown in FIG6 , comprising:

[0076] a segmentation module for acquiring a 3D pelvic image and identifying one or more target regions in the 3D pelvic image based on a deformable nnUNet, wherein the deformable nnUNet comprises a deformable convolution-based feature encoder, a feature fusion module, and a feature decoder connected in sequence;

[0077] a calculation module, configured to calculate reconstruction parameters based on the 3D pelvic image, wherein the reconstruction parameters include at least a radius of a disk region, and the reconstruction parameters are calculated based on an area of ​​a maximum connected component of the 3D pelvic image;

[0078] A post-processing module is used to obtain segmented images of the one or more target areas based on the reconstruction parameters.

[0079] It is not difficult to find that this embodiment is a device embodiment corresponding to the first embodiment, and this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment are still valid in this embodiment, and to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the first embodiment.

[0080] It is worth noting that all modules involved in this embodiment are logical modules. In actual applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovations of the present invention, this embodiment does not include units that are not closely related to solving the technical problems proposed by the present invention. However, this does not mean that other units do not exist in this embodiment.

[0081] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0082] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0083] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Thus, this specification may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This specification may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0085] The various embodiments in this specification are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are generally similar to the method embodiments, so the description is relatively simple. For relevant parts, refer to the description of the method embodiments.

[0086] The foregoing is merely an example of the present invention and is not intended to limit the present invention. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. An image segmentation method based on deformable nnUNet, characterized in that: The image segmentation method based on deformable nnUNet specifically includes: Acquire a 3D pelvic image, and identify one or more target regions in the 3D pelvic image based on a deformable nnUNet, wherein the deformable nnUNet includes a deformable convolution-based feature encoder, a feature fusion module, and a feature decoder connected in sequence; Calculating reconstruction parameters based on the 3D pelvic image, wherein the reconstruction parameters include at least a radius of a disk region, and the reconstruction parameters are calculated based on an area of ​​a maximum connected component of the 3D pelvic image; obtaining a segmented image of the one or more target regions based on the reconstruction parameters; The deformable convolution-based feature encoder includes: Offset adjustment of the convolution kernels of multiple convolution layers. The offset of the convolution kernels of each convolution layer is not exactly the same. The convolution module is constructed by connecting multiple offset convolution layers using residual connections. Connect the convolution module and the difference layer to form a deformable convolution feature encoder, A feature encoder based on deformable convolution is used to identify irregular pelvic medical features in 3D pelvic images and obtain an encoded 3D pelvic image feature map.

2. The image segmentation method based on deformable nnUNet according to claim 1, characterized in that The adjusting of the convolution kernels of the multiple convolution layers specifically includes: Among them, w represents the weight, e0 represents the position of the pixel, and e n represents the position of the adjacent pixels falling into L; and L ～ is the deformation result of L; compared with L, L ～ is a parameter containing the offset position Δe n irregular grid.

3. The image segmentation method based on deformable nnUNet according to claim 1, characterized in that The difference layer adjusts the output of the convolution module using the following methods (3) and (4): x(e)=B(w i ,q j ), (3) B(w i ,q j )=w1q1+w2q2+w3q3+w4q4. (4) Among them, w i is the corresponding weight, q j is the 4 pixel units around the irregular sampling position that participate in the calculation, B(w i ,q j ) is the bilinear interpolation kernel.

4. The image segmentation method based on deformable nnUNet according to claim 1, characterized in that The feature fusion module is composed of Ladder-ASPP, which specifically includes: An upper global pooling layer and a lower fusion layer, wherein the upper global pooling layer includes at least a global pooling module, and the output of the global pooling module is mixed with the input image of the feature fusion module; The lower fusion layer includes at least multiple parallel dilated convolution layers and fusion layers with variable sampling rates. The levels of each parallel dilated convolution layer are different, and the input of each parallel dilated convolution layer includes at least the input image of the input feature fusion module and the output of the previous parallel dilated convolution layer.

5. The image segmentation method based on deformable nnUNet according to claim 4, characterized in that The output of the upper global pooling layer is mixed with the input image of the feature fusion module after 1×1 convolution and Sigmoid, and the mixing process is weighted feature channel multiplication.

6. The image segmentation method based on deformable nnUNet according to claim 4, characterized in that The upper pooling layer is calculated as follows: y1=B[C1[GP S (x)]]×x, (5) Among them, x is the feature map obtained by the feature encoder, and then global pooling is performed using GP S (x) represents, C1 represents the weight of each feature channel through 1×1 convolution, and B is the normalization of the feature weight.

7. The image segmentation method based on deformable nnUNet according to claim 4, characterized in that The lower fusion layer is calculated as follows: Among them, let G K，D is the output of densely connected pyramid pooling, where K is the level of the pyramid and D is the expansion rate. Furthermore, In the case of cascade operations, each x is represented as:

8. The image segmentation method based on deformable nnUNet according to claim 1, characterized in that Before identifying one or more target regions in the 3D pelvic image based on the deformable nnUNet, the method further includes training the deformable nnUNet; The training of the deformable nnUNet specifically includes: Establish a loss function and train the deformable nnUNet using backpropagation; The loss function is composed of the sum of the cross entropy loss function and the dice overall loss function.

9. An image segmentation device based on deformable nnUNet, characterized in that: The image segmentation device based on deformable nnUNet includes: a segmentation module for acquiring a 3D pelvic image and identifying one or more target regions in the 3D pelvic image based on a deformable nnUNet, wherein the deformable nnUNet comprises a deformable convolution-based feature encoder, a feature fusion module, and a feature decoder connected in sequence; a calculation module, configured to calculate reconstruction parameters based on the 3D pelvic image, wherein the reconstruction parameters include at least a radius of a disk region, and the reconstruction parameters are calculated based on an area of ​​a maximum connected component of the 3D pelvic image; A post-processing module is used to obtain segmented images of the one or more target areas based on the reconstruction parameters.

Images