Image segmentation method and apparatus

Abstract

This application relates to an image segmentation method and apparatus. The convolutional blocks in the image segmentation model are improved by avoiding direct 3D convolution processing. Instead, the traditional 3D convolution is split into pseudo-1D and pseudo-2D convolutions. The pseudo-2D convolutions extract features from sliced ​​images, while the pseudo-1D convolutions extract relationships between slices. This reduces the computational load on model parameters and lowers the hardware requirements. Simultaneously, the addition of feature gain blocks prevents the introduction of unnecessary low-level feature information that might result from repeated feature utilization. Furthermore, the receptive fields of multiple dilated convolutions enrich spatial information, further improving segmentation accuracy. This method allows the model to run smoothly on devices with low computational power without affecting computational accuracy, thereby reducing the cost of image segmentation processing.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image segmentation method and apparatus. Background Technology

[0002] Currently, in clinical practice, doctors can manually punctuate the prostate gland on MRI images to outline its contours. With the development of computer vision, deep learning can be used to automatically outline and segment the prostate gland on MRI images, but the segmentation accuracy and computational deployment still do not meet the needs of doctors.

[0003] Manual punctuation methods suffer from inconsistencies due to differing perspectives among doctors with varying experience and observations of organ contours. Deep learning-based methods often utilize large-parameter models, multinomial representations, and large datasets for automatic prostate segmentation. However, these methods require significant computational and data resources. Large datasets can lead to slow model deployment, or even make the model unusable on low-end hardware, resulting in limited clinical value and failing to meet the needs of diverse application scenarios.

[0004] However, current image segmentation models require a large amount of data processing, have high hardware computing requirements, and incur high operating costs. Summary of the Invention

[0005] Therefore, it is necessary to provide an image segmentation method and apparatus that can reduce the cost of image segmentation processing in response to the above-mentioned technical problems.

[0006] Firstly, this application provides an image segmentation method. The method includes:

[0007] Obtain the image to be segmented; the image to be segmented includes the target object to be segmented.

[0008] Obtain the image segmentation model; the image segmentation model is trained using the training set of the target object, including feature gain blocks and multiple convolutional blocks. The feature gain blocks include standard convolutions and multiple dilated convolutions. Each convolutional block includes pseudo-one-dimensional convolutions and pseudo-two-dimensional convolutions. The two dimensions of the pseudo-one-dimensional convolution have a size of 1, and the one dimension of the pseudo-two-dimensional convolution has a size of 1. The three dimensions with a size of 1 are all different.

[0009] The image features of the image to be segmented are processed by convolutional blocks and feature gain blocks in the image segmentation model to obtain an initial segmented image;

[0010] Obtain the image probability distribution for the target object, and correct the initial segmented image based on the image probability distribution to obtain the target segmented image of the target object.

[0011] In one embodiment, obtaining an image segmentation model includes:

[0012] Obtain labeled real images corresponding to sample medical images, and obtain training and test sets based on the labeled real images corresponding to sample medical images; the sample medical images include the target objects to be segmented, and the labeled real images include the target objects that have been segmented in the sample medical images.

[0013] The neural network model to be trained is obtained by inputting sample medical images from the training set into the neural network model to be trained, and obtaining the training prediction probability map output by the neural network model to be trained; the neural network model to be trained includes feature gain blocks and multiple convolutional blocks.

[0014] The image probability distribution is obtained from the training set, and the training prediction probability map is corrected based on the image probability distribution to obtain the corrected training prediction probability map.

[0015] Based on the corrected training prediction probability map and the labeled real image, obtain the first loss value, and based on the training prediction probability map and the labeled real image, obtain the second loss value.

[0016] Based on the first loss value and the second loss value, adjust the neural node parameters of the feature gain block and each convolutional block in the neural network model to be trained to obtain the trained neural network model.

[0017] The trained neural network model was tested using a test set. If the test conditions were met, the trained neural network model was used as an image segmentation model.

[0018] In one embodiment, the trained neural network model is tested using a test set. If the test conditions are met, the trained neural network model is used as an image segmentation model, including:

[0019] The sample medical images in the test set are input into the trained neural network model to obtain the test prediction probability map output by the trained neural network model.

[0020] The test prediction probability map is corrected based on the image probability distribution to obtain the corrected test prediction probability map.

[0021] The corrected test prediction probability map and the real label image are compared to obtain the index value to complete a test; the index value is used to characterize the degree of overlap between the corrected test prediction probability map and the real label image.

[0022] Multiple tests are conducted on the multiple trained neural network models obtained from multiple training sessions. The maximum index value is obtained from the index values ​​obtained in each test, and the target test prediction probability map corresponding to the maximum index value is determined.

[0023] The trained neural network model that outputs the target test prediction probability map is used as the image segmentation model.

[0024] In one embodiment, each convolutional block processes image features of the image to be segmented, including:

[0025] Multiple first feature maps are obtained based on the first image features of the input convolutional block;

[0026] Based on the amount of feature information contained in each first feature map, obtain the channel weight corresponding to each first feature map;

[0027] Based on channel weights, different first feature maps are combined to obtain at least one set of first feature maps. Pseudo-one-dimensional convolution processing is then performed on at least one set of first feature maps to obtain second image features.

[0028] Multiple second feature maps are obtained based on the second image features;

[0029] Based on channel weights, different second feature maps are combined to obtain at least one set of second feature maps. Pseudo-two-dimensional convolution processing is then performed on at least one set of second feature maps to obtain the image features output by the convolution block.

[0030] In one embodiment, the feature gain block processes image features of the image to be segmented, including:

[0031] Each dilated convolution is used to process the third image features of the input feature gain block to obtain intermediate image features with the same number of dilated convolutions; the dilation coefficient of each dilated convolution is different.

[0032] For an image feature set consisting of a third image feature and an intermediate image feature, two image features are selected from the image feature set multiple times, and the two selected image features are spliced ​​together each time to obtain multiple spliced ​​image features;

[0033] Multiple standard convolutions of different sizes are used to process the features of each stitched image separately, resulting in multiple sub-image features;

[0034] Based on the self-learning weights, the features of multiple sub-images are weighted and summed to obtain the image features output by the feature gain block.

[0035] In one embodiment, the method further includes:

[0036] Obtain the image coordinate information of the target segmented image relative to the image coordinate system;

[0037] Based on the transformation relationship between the image coordinate system and the world coordinate system, the image coordinate information is transformed to obtain the world coordinate information of the target segmentation image relative to the world coordinate system;

[0038] Based on the target segmented image and world coordinate information, an image segmentation file for the image to be segmented is generated.

[0039] In one embodiment, the method further includes:

[0040] For the segmentation contour of the target object in the target segmentation image, obtain the contour coordinate points in the segmentation contour;

[0041] Based on the variation range of the tangent vector between different contour coordinate points on the segmented contour, the contour coordinate points are filtered to obtain approximate coordinate points;

[0042] Obtain the Euclidean distance between every two adjacent approximate coordinate points. Based on the Euclidean distance between every two adjacent approximate coordinate points, filter the approximate coordinate points to obtain the target coordinate points. The target coordinate points are used to characterize the segmentation contour of the target object in the target segmentation image.

[0043] Secondly, this application also provides an image segmentation apparatus. The apparatus includes:

[0044] The image acquisition module is used to acquire the image to be segmented; the image to be segmented includes the target object to be segmented.

[0045] The model acquisition module is used to acquire the image segmentation model. The image segmentation model is trained using the training set of the target object and includes a feature gain block and multiple convolutional blocks. The feature gain block includes standard convolution and multiple dilated convolutions. Each convolutional block includes pseudo-one-dimensional convolution and pseudo-two-dimensional convolution. The two dimensions of the pseudo-one-dimensional convolution have a size of 1, and the one dimension of the pseudo-two-dimensional convolution has a size of 1. The three dimensions with a size of 1 are different from each other.

[0046] The model processing module is used to process the image features of the image to be segmented through the convolutional blocks and feature gain blocks in the image segmentation model to obtain the initial segmented image;

[0047] The model output module is used to obtain the image probability distribution of the target object, and correct the initial segmented image based on the image probability distribution to obtain the target segmented image of the target object.

[0048] Thirdly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0049] Obtain the image to be segmented; the image to be segmented includes the target object to be segmented.

[0050] Obtain the image segmentation model; the image segmentation model is trained using the training set of the target object, including feature gain blocks and multiple convolutional blocks. The feature gain blocks include standard convolutions and multiple dilated convolutions. Each convolutional block includes pseudo-one-dimensional convolutions and pseudo-two-dimensional convolutions. The two dimensions of the pseudo-one-dimensional convolution have a size of 1, and the one dimension of the pseudo-two-dimensional convolution has a size of 1. The three dimensions with a size of 1 are all different.

[0051] The image features of the image to be segmented are processed by convolutional blocks and feature gain blocks in the image segmentation model to obtain an initial segmented image;

[0052] Obtain the image probability distribution for the target object, and correct the initial segmented image based on the image probability distribution to obtain the target segmented image of the target object.

[0053] The aforementioned image segmentation method and apparatus improve the convolutional blocks in the image segmentation model, avoiding direct 3D convolution processing. Instead, it splits the traditional 3D convolution into pseudo-1D and pseudo-2D convolutions. The pseudo-2D convolutions extract features from sliced ​​images, while the pseudo-1D convolutions extract relationships between slices. This reduces the computational load on model parameters and lowers the computational requirements on hardware. Furthermore, the addition of feature gain blocks prevents the introduction of unnecessary low-level feature information that might result from repeated feature utilization. The receptive fields of multiple dilated convolutions enrich spatial information, further improving segmentation accuracy. This allows the model to run smoothly on devices with low computational power without affecting computational accuracy, thereby reducing the cost of image segmentation processing. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating an image segmentation method in one embodiment;

[0055] Figure 2 This is a schematic diagram illustrating the principle of dilated convolution in one embodiment;

[0056] Figure 3 This is a schematic diagram of the traditional 3D convolution principle in one embodiment;

[0057] Figure 4 This is a schematic diagram illustrating the principles of pseudo-1D convolution and pseudo-2D convolution in one embodiment;

[0058] Figure 5 This is a schematic diagram of the connection relationship between convolutional blocks in one embodiment;

[0059] Figure 6 This is a schematic diagram of the image segmentation model processing flow in one embodiment;

[0060] Figure 7 This is a schematic diagram of the image segmentation model training process in one embodiment;

[0061] Figure 8 This is a schematic diagram of the real image of the label and the corrected test prediction probability map in one embodiment;

[0062] Figure 9 This is a schematic diagram of the processing flow of a convolutional block in one embodiment;

[0063] Figure 10 This is a schematic diagram of grouped convolution of a convolutional block in one embodiment;

[0064] Figure 11 This is a schematic diagram of 3D convolution in one embodiment;

[0065] Figure 12 This is a schematic diagram of 3D convolution decoupling in one embodiment;

[0066] Figure 13 This is a schematic diagram of 3D decoupled weighted grouped convolution in one embodiment;

[0067] Figure 14 This is a schematic diagram of the feature gain block processing flow in one embodiment;

[0068] Figure 15 This is a schematic diagram of coordinate system transformation in one embodiment;

[0069] Figure 16 This is a schematic diagram illustrating the effect of target segmentation image in one embodiment;

[0070] Figure 17 This is a schematic diagram of the target coordinates in one embodiment;

[0071] Figure 18 This is a structural block diagram of an image segmentation device in one embodiment. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0073] In one embodiment, such as Figure 1 As shown, an image segmentation method is provided. This embodiment illustrates the application of this method to a computer device. It is understood that the computer device can specifically be a terminal or a server. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart medical devices. The server can be a standalone server or a server cluster composed of multiple servers. In this embodiment, the method includes the following steps:

[0074] Step 102: Obtain the image to be segmented; the image to be segmented includes the target object to be segmented.

[0075] The target object refers to the organ, body part, or body area that requires medical examination in the case, including but not limited to the brain, heart, blood vessels, liver, kidneys, gallbladder, pancreas, thyroid gland, prostate, uterus, and adnexa of humans or animals. The image to be segmented refers to three-dimensional medical images, including but not limited to MRI images.

[0076] Optionally, after obtaining the image to be segmented, the image can be preprocessed to ensure that the preprocessed segmented image meets the input requirements of the model. Preprocessing includes, but is not limited to, image enhancement, size adjustment, orientation normalization, and filtering for noise reduction.

[0077] Step 104: Obtain the image segmentation model. The image segmentation model is trained using the training set of the target object and includes a feature gain block and multiple convolutional blocks. The feature gain block includes standard convolution and multiple dilated convolutions. Each convolutional block includes pseudo-one-dimensional convolution and pseudo-two-dimensional convolution. The two dimensions of the pseudo-one-dimensional convolution have a size of 1, and the one dimension of the pseudo-two-dimensional convolution has a size of 1. The three dimensions with a size of 1 are different from each other.

[0078] In this training set, the sample images and the images to be segmented contain the same type of target objects. A pseudo-one-dimensional convolution with two dimensions of size 1 is equivalent to a pseudo-1D convolution, and a pseudo-two-dimensional convolution with one dimension of size 1 is equivalent to a pseudo-2D convolution.

[0079] The principle of dilated convolution is as follows: Figure 2 As shown, taking 2D convolution as an example, the principle of 3D convolution is the same as that of 2D. The one on the left is ordinary convolution. The standard convolution size is 3×3, and its field of view on the feature map is only a 3×3 area. However, the dilated convolution with a dilation factor of 2 has a feature capture range of 5×5 on the feature map, which can capture information of different ranges without increasing the amount of computation.

[0080] For each convolutional block, traditional 3D convolution, such as... Figure 3 As shown, due to the high computational complexity of 3D convolution, and the inability of ordinary 2D convolution to meet the relevant tasks of 3D medical imaging, 3D convolution can be decoupled into a combination of pseudo-2D and pseudo-1D convolutions to reduce computational complexity. For example, taking the decoupling of a 3×3×3 convolution as an example, convolution is first performed on the channels (interlayer) and then on the spatial (slice) convolution, that is, the 3D convolution is split into pseudo-1D convolution and pseudo-2D convolution, as shown below. Figure 4 As shown, a 1×1×3 convolution is performed first, followed by a 3×3×1 convolution.

[0081] Optionally, in the image segmentation model, each convolutional block can adopt, for example... Figure 5 The skip connection shown prevents the introduction of ineffective information due to excessive features when utilizing earlier layers. It employs an odd number of convolutional blocks, concatenating the output of the first block with the output of the last, the second with the second-to-last, and so on. Utilizing the shallowest information at the very beginning allows for better extraction of information from the periphery of regions, as shallow outputs in neural network models clearly describe the target, while deeper layers only contain semantic information. Feature gain blocks can be connected between the two middle convolutional blocks.

[0082] Step 106: The image features of the image to be segmented are processed by the convolutional blocks and feature gain blocks in the image segmentation model to obtain the initial segmented image.

[0083] Optional, such as Figure 6 As shown, after the model to be segmented is processed and transformed according to the model input requirements, it is input into the image segmentation model. After passing through multiple convolutional blocks and a feature gain block, the model outputs the initial segmentation result, i.e., the initial segmented image.

[0084] Step 108: Obtain the image probability distribution for the target object, and correct the initial segmentation image according to the image probability distribution to obtain the target segmentation image of the target object.

[0085] Optional, such as Figure 6 As shown, the initial segmentation image is corrected by using the image probability distribution obtained from the training set. Since the image distribution of the same type of target object is similar, this is used to further improve the final segmentation result, improve the segmentation accuracy of the target object, and obtain a target segmentation image with higher segmentation contour accuracy.

[0086] In one feasible implementation, the coordinates of the segmentation contours in the target segmentation image can be extracted. By using the transformation relationship between the image coordinate system and the world coordinate system, the actual coordinate position of the target object in the world coordinate system can be obtained.

[0087] In the above image segmentation method, the following steps are taken: First, an image to be segmented is obtained; this image includes the target object to be segmented. Second, an image segmentation model is obtained; this model is trained using a training set for the target object and includes feature gain blocks and multiple convolutional blocks. The feature gain blocks include standard convolutions and multiple dilated convolutions. Each convolutional block includes pseudo-one-dimensional convolutions and pseudo-two-dimensional convolutions. The two dimensions of the pseudo-one-dimensional convolution have a size of 1, and the one dimension of the pseudo-two-dimensional convolution has a size of 1. The three dimensions with a size of 1 are distinct. Third, the image features of the image to be segmented are processed using the convolutional blocks and feature gain blocks in the image segmentation model to obtain an initial segmented image. Fourth, the image probability distribution for the target object is obtained, and the initial segmented image is corrected based on the image probability distribution to obtain the target segmented image for the target object. The convolutional blocks in the image segmentation model have been improved. Instead of directly performing 3D convolution, the traditional 3D convolution is split into pseudo-1D and pseudo-2D convolution. The pseudo-2D convolution is used to extract features from sliced ​​images, while the pseudo-1D convolution is used to extract relationships between slices. This reduces the computational cost of model parameters and the computational requirements of hardware. At the same time, the addition of feature gain blocks can prevent the introduction of unnecessary low-level feature information that may be caused by repeated feature utilization. The receptive fields of multiple dilated convolutions are used to enrich spatial information and further improve segmentation accuracy. The model can run normally on devices with low computing power without affecting the accuracy of computation, thereby reducing the cost of image segmentation processing.

[0088] In one embodiment, obtaining an image segmentation model includes: obtaining labeled ground truth images corresponding to sample medical images; obtaining a training set and a test set based on the labeled ground truth images corresponding to the sample medical images; the sample medical images include target objects to be segmented, and the labeled ground truth images include target objects that have already been segmented in the sample medical images; obtaining a neural network model to be trained; inputting sample medical images from the training set into the neural network model to be trained to obtain a training prediction probability map output by the neural network model to be trained; the neural network model to be trained includes feature gain blocks and multiple convolutional blocks; obtaining an image probability distribution based on the training set; correcting the training prediction probability map based on the image probability distribution to obtain a corrected training prediction probability map; obtaining a first loss value based on the corrected training prediction probability map and the labeled ground truth images, and obtaining a second loss value based on the training prediction probability map and the labeled ground truth images; adjusting the neural node parameters of the feature gain blocks and each convolutional block in the neural network model to be trained based on the first loss value and the second loss value to obtain a trained neural network model.

[0089] Furthermore, the trained neural network model is tested using a test set. Under the condition that the testing conditions are met, the trained neural network model is used as the image segmentation model. This includes: inputting sample medical images from the test set into the trained neural network model to obtain the test prediction probability map output by the trained neural network model; correcting the test prediction probability map according to the image probability distribution to obtain the corrected test prediction probability map; comparing the corrected test prediction probability map with the labeled real image to obtain an index value, thus completing one test; the index value is used to characterize the overlap between the corrected test prediction probability map and the labeled real image; performing multiple tests on multiple trained neural network models obtained from multiple training iterations, obtaining the maximum index value from the index values ​​obtained in each test, and determining the target test prediction probability map corresponding to the maximum index value; and using the trained neural network model that outputs the target test prediction probability map as the image segmentation model.

[0090] Optional, such as Figure 7 As shown, sample medical images containing the target object are selected. The target object region is labeled for each sample medical image, resulting in a corresponding labeled ground truth image for each sample medical image. A dataset with a one-to-one correspondence between the sample medical images and the labeled ground truth images is generated. 90% of the dataset is used as the training set, and the remaining 10% as the test set. The sample medical images and their corresponding labeled ground truth images in the training set are read. The sample medical images undergo size, intensity, and orientation normalization and filtering. These, along with the corresponding labeled ground truth images, are converted into input features and input into the neural network model to be trained. The neural network model outputs a training prediction probability map of the training set. Utilizing the consistency of the probability distribution of MRI images of the same organ (equivalent to the same target object), the image probability distribution of the training set is introduced. A threshold is applied to the training prediction probability map and the image probability distribution of the training set. If the probability distribution is greater than the threshold, no correction is needed; if it is less than the threshold, correction is performed, resulting in a corrected training prediction probability map. The corrected training prediction probability map is compared with the labeled ground truth image to obtain a first loss value, and a second loss value is also obtained by comparing the training prediction probability map with the labeled ground truth image. The first and second loss values ​​are added together to obtain the final loss value for this training iteration. The final loss value is fed back to update the neural node parameters of the neural network model, thus completing one training iteration and obtaining the trained neural network model.

[0091] After each preset number of training iterations, sample medical images and corresponding labeled real images are read from the test set. The sample medical images are normalized and filtered for size, intensity, and orientation, and then converted into input features along with the corresponding labeled real images, which are then fed into the neural network model to be tested. The neural network model outputs a test prediction probability map for the test set. Using the image probability distribution from the training set, the test prediction probability map is corrected to obtain a corrected test prediction probability map. The overlap between the corrected test prediction probability map and the labeled real images is compared to obtain the index value for one test. During multiple iterations of training, multiple tests are performed, and the model parameters with the optimal index value are retained.

[0092] In one feasible implementation, the labeled real images corresponding to the sample medical images in the test set are as follows: Figure 8 As shown in the label, the corrected test prediction probability map is as follows. Figure 8 As shown in the `predict` table, the index is used to determine the quality of the model. Using multiple indicators for evaluation provides a better assessment and prevents a model judged by only one indicator from exhibiting significant bias on certain images. The index calculation methods include:

[0093] Index = A*Dice + B*Jaccard + C*(Accuracy + Precision + TruePrecision) / 3, where ABC are adjustable values ​​and satisfy A+B+C=1. In this example, the optimal combination is A=0.5, B=0.4, and C=0.1.

[0094] Inter = sum(label * predict), where Inter is the number of correctly predicted pixels.

[0095] union=sum(label)+sum(predict)-sum(label*predict)

[0096] Accuracy=Inter / (height*weight)

[0097] Precision=Inter / sum(predict)

[0098] TruePrecision=Inter / sum(label)

[0099] Jaccard = Inter / union

[0100] Dice=2*Inter / (sum(predict)+sum(label)

[0101] In this embodiment, the consistency of the probability distribution of MRI images of the same organ is utilized to introduce the probability distribution of the training set to correct the situation where low-prediction-value pixels may contain targets, thereby improving the segmentation accuracy of the tissue.

[0102] In one embodiment, each convolutional block processes the image features of the image to be segmented, including: obtaining multiple first feature maps based on the first image features of the input convolutional block; obtaining the channel weights corresponding to each first feature map based on the amount of feature information contained in each first feature map; combining different first feature maps based on the channel weights to obtain at least one set of first feature maps; performing pseudo-one-dimensional convolution processing on the at least one set of first feature maps to obtain second image features; obtaining multiple second feature maps based on the second image features; combining different second feature maps based on the channel weights to obtain at least one set of second feature maps; performing pseudo-two-dimensional convolution processing on the at least one set of second feature maps to obtain the image features output by the convolutional block.

[0103] The first image feature is equivalent to the input image feature of each convolutional block, which may be, but is not limited to, 3D image features.

[0104] Optionally, the processing procedure for each convolutional block is as follows: Figure 9 As shown, the output feature maps are sorted by weight, forming four groups: (highest-medium-high), (medium-high-low), (low-second-low), and (second-lowest-lowest). Then, pseudo-one-dimensional and pseudo-two-dimensional convolutions are performed on the grouped features. Using a grouped convolution with four groups implements weighted grouping convolution. Grouping features according to their weights emphasizes feature maps with high information content, thus highlighting advantageous information. Simultaneously, grouped convolution further reduces complexity. This structure is a convolutional block in a neural network model; the overall network structure is composed of multiple such convolutional blocks.

[0105] Compared to ordinary convolution, each convolutional block uses a grouped convolution processing method. For example... Figure 10 As shown, a regular convolution has six feature maps, and after the convolution operation, it also produces six feature maps. The number of standard convolutions will determine the number of output feature maps, and each standard convolution operates on all six input feature maps. Its parameter count is: w × h × 6 × 6. A grouped convolution, on the other hand, takes six feature maps as input, and after a convolution operation with two groups, it also produces six feature maps. However, when using grouped convolution, the input feature maps are grouped, and each standard convolution only operates on the feature maps grouped to it. These feature maps are then stacked together, resulting in a parameter count of: w × h × 3 × 2 × 3. This also allows for better computation of the convolution after weight grouping. Here, w represents the feature map width, and h represents the feature map height.

[0106] Specifically, the differences between the convolutional block structure in this embodiment and traditional three-dimensional convolution (also known as 3D convolution) are as follows:

[0107] like Figure 11 As shown, the left side represents the 3D input features of a prostate MRI image, which consists of four slices. Each slice generates multiple stacked feature layers, resulting in an image feature block. The feature is calculated as: w × h × 4 × d, where wh is the image's width and height, 4 is the number of slices, and d is the number of feature layers generated by each slice. The right side shows the 3D convolution process used to extract information from this input feature. A standard 3×3×3 convolution is analogous to multiplying elements of a 3x3 Rubik's Cube on a larger cube and summing the results to obtain the value of the center element. This results in a very large computational load (the input and output dimensions are the same, i.e., d is the same), which is w × h × 4 × 3 × 3 × 3 × 4 × d.

[0108] like Figure 12 As shown, the left side represents the 3D input features of a prostate MRI image, consisting of four slices. Each slice generates multiple stacked feature layers, resulting in an image feature block. The feature size is w×h×4×d, where wh is the image's width and height, 4 is the number of slices, and d is the number of feature layers generated per slice. The middle section uses pseudo-1D convolution (standard convolution 1×1×3) to extract the input features, essentially extracting them along a line. This method considers the relationships between slices, and its computational complexity is w×h×4×1×1×3×4×d. The third section uses pseudo-2D convolution (standard convolution 3×3×1) to extract the input features, essentially extracting them along a plane. This method operates on the feature layers generated from the image, capturing information from each slice and extracting its features. Its computational complexity is also w×h×4×3×3×1×4×d. The overall computational cost of this process is w×h×4×1×1×3×4×d + w×h×4×3×3×1×4×d, which is lower than the computational cost of 3D convolution. Therefore, it adopts... Figure 11 The concatenated convolution operation block within the dashed box replaces the traditional 3D convolution.

[0109] Furthermore, such as Figure 13As shown, the left side represents the 3D input image feature F1 of a prostate MRI image, which consists of four slices. Each slice generates multiple stacked image feature blocks with multiple feature layers; that is, this feature is: w×h×4×d, where wh is the width and height of the image, 4 is the number of slices, and d is the number of feature layers generated by each slice. Global average pooling is performed to extract the weights of each feature layer, and each feature layer is represented by a value with dimensions of 1×1×4×d. The F1 feature layer is grouped according to its weight values ​​(highest-medium-high), (medium-high-low), (low-second-low), (second-lowest-lowest), and then pseudo-1D grouped convolution is performed on the grouped F1 for extraction. The F2 feature layer is then grouped according to its weight values ​​(highest-medium-high), (medium-high-low), (low-second-lowest), (second-lowest-lowest), and then pseudo-2D grouped convolution is performed on the grouped F2 for extraction, resulting in the image feature F3 output by the convolution block. Grouping the feature layers according to channel weights can highlight the advantageous information and filter out the interference of low information, and can also reduce the amount of computation more significantly.

[0110] In this embodiment, the convolutional blocks in the image segmentation model are improved. Instead of directly performing 3D convolution processing, the traditional 3D convolution is split into pseudo-1D convolution and pseudo-2D convolution. Pseudo-2D convolution is used to extract features from sliced ​​images, and pseudo-1D convolution is used to extract relationships between slices. This reduces the computational load of model parameters and the computational requirements of hardware.

[0111] In one embodiment, the feature gain block processes the image features of the image to be segmented, including: processing the third image features input to the feature gain block using each dilated convolution to obtain intermediate image features with the same number of dilated convolutions; the dilation coefficients of each dilated convolution are different; for the image feature set composed of the third image features and intermediate image features, selecting two image features from the image feature set multiple times, and concatenating the two selected image features each time to obtain multiple concatenated image features; processing each concatenated image feature using multiple standard convolutions of different sizes to obtain multiple sub-image features; and summing the weights of the multiple sub-image features according to self-learning weights to obtain the image features output by the feature gain block.

[0112] Optional, such as Figure 14As shown, the feature gain block obtains S2, S3, and S4 from the input feature S through dilated convolutions with dilation coefficients d = 2, 3, and 4, respectively. These S4 and S4 are then concatenated in reverse order pairwise to form concat[s,s4], concat[s2,s3], concat[s3,s2], and concat[s4,s]. The concatenated features are then subjected to another convolution extraction, using 1×1×1 and 3×3×3 to obtain pairwise distinct features S1', S2', S3', and S4'. These features are weighted and summed to obtain S'. The weights w[w1,w2,w3,w4] are learned by the network itself. These self-learned weights, which are the four weight values ​​assigned by the network in the current part, can adapt to changes in the data. By pairwise fusing features extracted from a single feature under different receptive fields obtained through different dilated convolutions, spatial feature information can be enriched without containing redundant information. This is because the features are extracted using different standard convolutions after concatenation. Finally, the four features are assigned values ​​using learned weights and then summed to obtain the output feature.

[0113] In this embodiment, adding a feature gain block to the image segmentation model can prevent the introduction of unnecessary low-level feature information that may be caused by repeated feature utilization, and enrich the spatial information by utilizing the receptive field of multiple dilated convolutions, thereby further improving the segmentation accuracy.

[0114] In one embodiment, the method further includes: obtaining image coordinate information of the target segmented image relative to the image coordinate system; transforming the image coordinate information according to the transformation relationship between the image coordinate system and the world coordinate system to obtain world coordinate information of the target segmented image relative to the world coordinate system; and generating an image segmentation file of the image to be segmented based on the target segmented image and the world coordinate information.

[0115] Optional, such as Figure 15 As shown, on the left, X represents the horizontal coordinate of a pixel on the slice, from 0, 1, 2...X; Y represents the vertical coordinate of a pixel on the slice, from 0, 1, 2...Y; Z can be seen as the number of slices in the patient image, from 0, 1, 2...Z. On the right is DICOM (Digital Imaging and Communications in Medicine), which saves information about organs and tissues in world space when recording images, namely the voxel spacing Spacing[x,y,z] (the actual distance between pixels). The voxel spacing tells us the actual distance between two points in world coordinates; the original coordinates Origin[x,y,z] (the position of the origin of the image coordinate system relative to the origin of the world coordinate system) tell us the actual distance of each point relative to the origin of the world coordinate system.

[0116] Image coordinates need to be converted to world coordinates before they can be used by the software to display the segmentation of the prostate. The software will save the converted coordinates as an XML file for the software to read. By reading the Spacing[x,y,z] and Origin[x,y,z] information of the input image, the image coordinates can be converted to world coordinates. The conversion relationship W is:

[0117] WX=Spacing[0]*X+Origin[0];

[0118] WY=Spacing[1]*Y+Origin[1];

[0119] WZ=Spacing[2]*Z+Origin[2].

[0120] In this embodiment, the coordinates of the segmentation contours in the target segmentation image can be extracted. By using the transformation relationship between the image coordinate system and the world coordinate system, the actual coordinate position of the target object in the world coordinate system can be obtained.

[0121] In one embodiment, the method further includes: obtaining contour coordinate points in the segmentation contour of the target object in the target segmentation image; filtering the contour coordinate points according to the magnitude of the change of the tangent vector between different contour coordinate points on the segmentation contour to obtain approximate coordinate points; obtaining the Euclidean distance between every two adjacent approximate coordinate points; filtering the approximate coordinate points according to the Euclidean distance between every two adjacent approximate coordinate points to obtain target coordinate points; the target coordinate points are used to characterize the segmentation contour of the target object in the target segmentation image.

[0122] Optionally, taking the prostate as the target object as an example, the image to be segmented and the target segmentation image are as follows: Figure 16 As shown, the left side displays two MRI images of the prostate, and the right side displays the corresponding segmented image of the prostate. Figure 17 As shown, the left side shows the contour coordinates of the obtained segmented contour. Then, the curve approximation method is used. That is, if the change of the tangent vector between a point and the surrounding points is not large, this point can be used to represent all the surrounding points. Then, the entire curve is traversed and the approximated coordinates are retained. At the same time, this method may have some points that are too close to each other and are not detected. Therefore, Euclidean distance is added. When the Euclidean distance between two points is greater than a threshold, it is retained. Finally, the target coordinates of the contour are obtained in the right figure.

[0123] In this embodiment, the segmentation contour can be represented by as few coordinate points as possible, which facilitates subsequent modifications.

[0124] In one embodiment, a method for training an image segmentation model includes:

[0125] Obtain labeled real images corresponding to sample medical images, and obtain training and test sets based on the labeled real images corresponding to sample medical images; the sample medical images include the target objects to be segmented, and the labeled real images include the target objects that have been segmented in the sample medical images.

[0126] The neural network model to be trained is obtained by inputting sample medical images from the training set into the neural network model to be trained, and obtaining the training prediction probability map output by the neural network model to be trained; the neural network model to be trained includes feature gain blocks and multiple convolutional blocks.

[0127] The image probability distribution is obtained from the training set, and the training prediction probability map is corrected based on the image probability distribution to obtain the corrected training prediction probability map.

[0128] Based on the corrected training prediction probability map and the labeled real image, obtain the first loss value, and based on the training prediction probability map and the labeled real image, obtain the second loss value.

[0129] Based on the first loss value and the second loss value, adjust the neural node parameters of the feature gain block and each convolutional block in the neural network model to be trained to obtain the trained neural network model.

[0130] The sample medical images in the test set are input into the trained neural network model to obtain the test prediction probability map output by the trained neural network model.

[0131] The test prediction probability map is corrected based on the image probability distribution to obtain the corrected test prediction probability map.

[0132] The corrected test prediction probability map and the real label image are compared to obtain the index value to complete a test; the index value is used to characterize the degree of overlap between the corrected test prediction probability map and the real label image.

[0133] Multiple tests are conducted on the multiple trained neural network models obtained from multiple training sessions. The maximum index value is obtained from the index values ​​obtained in each test, and the target test prediction probability map corresponding to the maximum index value is determined.

[0134] The trained neural network model that outputs the target test prediction probability map is used as the image segmentation model.

[0135] In one embodiment, an image segmentation method includes:

[0136] The process involves: acquiring an image to be segmented, including the target object; acquiring an image segmentation model trained on a training set for the target object, comprising feature gain blocks and multiple convolutional blocks, including standard convolutions and multiple dilated convolutions, each convolutional block including pseudo-one-dimensional and pseudo-two-dimensional convolutions, where the two dimensions of the pseudo-one-dimensional convolution are of size 1, and the one dimension of the pseudo-two-dimensional convolution is of size 1, with the three dimensions of size 1 being distinct; processing the image features of the image to be segmented using the convolutional blocks and feature gain blocks in the image segmentation model to obtain an initial segmented image; acquiring the image probability distribution for the target object, and correcting the initial segmented image based on the image probability distribution to obtain the target segmented image for the target object.

[0137] Each convolutional block processes the image features of the image to be segmented, including: obtaining multiple first feature maps based on the first image features of the input convolutional block; obtaining the channel weights corresponding to each first feature map based on the amount of feature information contained in each first feature map; combining different first feature maps based on the channel weights to obtain at least one set of first feature maps; performing pseudo-one-dimensional convolution processing on at least one set of first feature maps to obtain second image features; obtaining multiple second feature maps based on the second image features; combining different second feature maps based on the channel weights to obtain at least one set of second feature maps; performing pseudo-two-dimensional convolution processing on at least one set of second feature maps to obtain the image features output by the convolutional block.

[0138] The feature gain block processes the image features of the image to be segmented, including: processing the third image features input to the feature gain block using each dilated convolution to obtain intermediate image features with the same number of dilated convolutions; the dilation coefficients of each dilated convolution are different; for the image feature set composed of the third image features and intermediate image features, two image features are selected from the image feature set multiple times, and the two selected image features are concatenated each time to obtain multiple concatenated image features; multiple standard convolutions of different sizes are used to process each concatenated image feature to obtain multiple sub-image features; and the multiple sub-image features are weighted and summed according to self-learning weights to obtain the image features output by the feature gain block.

[0139] Obtain the image coordinate information of the target segmented image relative to the image coordinate system; transform the image coordinate information according to the transformation relationship between the image coordinate system and the world coordinate system to obtain the world coordinate information of the target segmented image relative to the world coordinate system; generate the image segmentation file of the image to be segmented based on the target segmented image and the world coordinate information.

[0140] For the segmentation contour of the target object in the target segmentation image, obtain the contour coordinate points in the segmentation contour; according to the change amplitude of the tangent vector between different contour coordinate points on the segmentation contour, filter the contour coordinate points to obtain approximate coordinate points; obtain the Euclidean distance between every two adjacent approximate coordinate points, and filter the approximate coordinate points according to the Euclidean distance between every two adjacent approximate coordinate points to obtain the target coordinate points; the target coordinate points are used to represent the segmentation contour of the target object in the target segmentation image.

[0141] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0142] Based on the same inventive concept, this application also provides an image segmentation apparatus for implementing the image segmentation method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more image segmentation apparatus embodiments provided below can be found in the limitations of the image segmentation method described above, and will not be repeated here.

[0143] In one embodiment, such as Figure 18 As shown, an image segmentation device 1800 is provided, including: an image acquisition module 1801, a model acquisition module 1802, a model processing module 1803, and a model output module 1804, wherein:

[0144] Image acquisition module 1801 is used to acquire an image to be segmented; the image to be segmented includes the target object to be segmented.

[0145] The model acquisition module 1802 is used to acquire the image segmentation model. The image segmentation model is trained using the training set of the target object and includes a feature gain block and multiple convolutional blocks. The feature gain block includes standard convolution and multiple dilated convolutions. Each convolutional block includes pseudo-one-dimensional convolution and pseudo-two-dimensional convolution. The two dimensions of the pseudo-one-dimensional convolution have a size of 1, and the one dimension of the pseudo-two-dimensional convolution has a size of 1. The three dimensions with a size of 1 are different from each other.

[0146] The model processing module 1803 is used to process the image features of the image to be segmented through the convolutional blocks and feature gain blocks in the image segmentation model to obtain the initial segmented image;

[0147] The model output module 1804 is used to obtain the image probability distribution of the target object, correct the initial segmentation image according to the image probability distribution, and obtain the target segmentation image of the target object.

[0148] In one embodiment, the model acquisition module 1802 is further configured to acquire labeled real images corresponding to sample medical images, and acquire training and test sets based on the labeled real images corresponding to sample medical images; the sample medical images include target objects to be segmented, and the labeled real images include target objects already segmented in the sample medical images; acquire a neural network model to be trained, input the sample medical images in the training set into the neural network model to be trained, and obtain a training prediction probability map output by the neural network model to be trained; the neural network model to be trained includes feature gain blocks and multiple convolutional blocks; acquire image probability distribution based on the training set, and correct the training prediction probability map based on the image probability distribution to obtain a corrected training prediction probability map; acquire a first loss value based on the corrected training prediction probability map and the labeled real images, and acquire a second loss value based on the training prediction probability map and the labeled real images; adjust the neural node parameters of the feature gain blocks and each convolutional block in the neural network model to be trained based on the first loss value and the second loss value to obtain a trained neural network model; test the trained neural network model using a test set, and if the test conditions are met, use the trained neural network model as an image segmentation model.

[0149] In one embodiment, the model acquisition module 1802 is further configured to input sample medical images from the test set into the trained neural network model to obtain a test prediction probability map output by the trained neural network model; correct the test prediction probability map according to the image probability distribution to obtain a corrected test prediction probability map; compare the corrected test prediction probability map with the labeled real image to obtain an index value to complete one test; the index value is used to characterize the overlap between the corrected test prediction probability map and the labeled real image; perform multiple tests on multiple trained neural network models obtained from multiple trainings, obtain the maximum index value from the index values ​​obtained in each test, and determine the target test prediction probability map corresponding to the maximum index value; and use the trained neural network model that outputs the target test prediction probability map as an image segmentation model.

[0150] In one embodiment, the model processing module 1803 is further configured to obtain multiple first feature maps based on the first image features of the input convolutional block; obtain the channel weights corresponding to each first feature map based on the amount of feature information contained in each first feature map; combine different first feature maps based on the channel weights to obtain at least one set of first feature maps; perform pseudo-one-dimensional convolution processing on the at least one set of first feature maps to obtain second image features; obtain multiple second feature maps based on the second image features; combine different second feature maps based on the channel weights to obtain at least one set of second feature maps; perform pseudo-two-dimensional convolution processing on the at least one set of second feature maps to obtain image features output by the convolutional block.

[0151] In one embodiment, the model processing module 1803 is further configured to process the third image features input to the feature gain block using each dilated convolution to obtain intermediate image features with the same number of dilated convolutions; the dilation coefficients of each dilated convolution are different; for the image feature set composed of the third image features and intermediate image features, two image features are selected from the image feature set multiple times, and the two selected image features are concatenated each time to obtain multiple concatenated image features; multiple standard convolutions of different sizes are used to process each concatenated image feature to obtain multiple sub-image features; according to the self-learning weights, the weights of the multiple sub-image features are summed to obtain the image features output by the feature gain block.

[0152] In one embodiment, the model output module 1804 is further configured to obtain image coordinate information of the target segmented image relative to the image coordinate system; transform the image coordinate information according to the transformation relationship between the image coordinate system and the world coordinate system to obtain the world coordinate information of the target segmented image relative to the world coordinate system; and generate an image segmentation file of the image to be segmented based on the target segmented image and the world coordinate information.

[0153] In one embodiment, the model output module 1804 is further configured to: obtain contour coordinate points in the segmentation contour of the target object in the target segmentation image; filter the contour coordinate points according to the magnitude of the change of the tangent vector between different contour coordinate points on the segmentation contour to obtain approximate coordinate points; obtain the Euclidean distance between every two adjacent approximate coordinate points; filter the approximate coordinate points according to the Euclidean distance between every two adjacent approximate coordinate points to obtain target coordinate points; the target coordinate points are used to characterize the segmentation contour of the target object in the target segmentation image.

[0154] Each module in the aforementioned image segmentation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0155] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0156] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An image segmentation method characterized by, The method includes: Obtain the image to be segmented; the image to be segmented includes the target object to be segmented; An image segmentation model is obtained; the image segmentation model is trained using the training set of the target object, and includes a feature gain block and multiple convolutional blocks. The feature gain block includes standard convolution and multiple dilated convolutions. Each convolutional block includes pseudo-one-dimensional convolution and pseudo-two-dimensional convolution. The two dimensions of the pseudo-one-dimensional convolution have a size of 1, and the one dimension of the pseudo-two-dimensional convolution has a size of 1. The three dimensions with a size of 1 are different from each other. The pseudo-two-dimensional convolution realizes feature extraction of the sliced ​​image, and the pseudo-one-dimensional convolution realizes relationship extraction between slices. The image features of the image to be segmented are processed by the convolutional blocks and feature gain blocks in the image segmentation model to obtain an initial segmented image; Obtain the image probability distribution for the target object, and correct the initial segmentation image according to the image probability distribution to obtain the target segmentation image of the target object; The feature gain block processes the image features of the image to be segmented, including: Each dilated convolution is used to process the third image feature input to the feature gain block to obtain the same number of intermediate image features as the number of dilated convolutions; the dilation coefficient of each dilated convolution is different; the feature gain block is connected between the two middle convolution blocks; and the third image feature is the output of the convolution block preceding the two middle convolution blocks in the image segmentation model. For an image feature set consisting of a third image feature and an intermediate image feature, two image features are selected sequentially from the image feature set, and the two selected image features are spliced ​​together each time to obtain multiple spliced ​​image features; Multiple standard convolutions of different sizes are used to process the features of each stitched image separately, resulting in multiple sub-image features; Based on the self-learning weights, the features of multiple sub-images are weighted and summed to obtain the image features output by the feature gain block, wherein the image features output by the feature gain block are used as the input of the next convolutional block of the two middle convolutional blocks in the image segmentation model.

2. The method of claim 1, wherein, The acquisition of the image segmentation model includes: Obtain labeled real images corresponding to sample medical images, and obtain the training set and test set based on the labeled real images corresponding to the sample medical images; the sample medical images include the target objects to be segmented, and the labeled real images include the target objects that have been segmented in the sample medical images; A neural network model to be trained is obtained, and the sample medical images in the training set are input into the neural network model to be trained to obtain the training prediction probability map output by the neural network model to be trained; the neural network model to be trained includes the feature gain block and multiple convolutional blocks; The image probability distribution is obtained from the training set, and the training prediction probability map is corrected based on the image probability distribution to obtain the corrected training prediction probability map. A first loss value is obtained based on the corrected training prediction probability map and the label real image, and a second loss value is obtained based on the training prediction probability map and the label real image. Based on the first loss value and the second loss value, the neural node parameters of the feature gain block and each convolutional block in the neural network model to be trained are adjusted to obtain the trained neural network model. The trained neural network model is tested using the test set, and if the test conditions are met, the trained neural network model is used as the image segmentation model.

3. The method of claim 2, wherein, The step of testing the trained neural network model using the test set, and using the trained neural network model as the image segmentation model if the test conditions are met, includes: The sample medical images in the test set are input into the trained neural network model to obtain the test prediction probability map output by the trained neural network model; The test prediction probability map is corrected based on the image probability distribution to obtain the corrected test prediction probability map; The corrected test prediction probability map and the real label image are compared to obtain an index value to complete one test; the index value is used to characterize the degree of overlap between the corrected test prediction probability map and the real label image. Multiple tests are performed on multiple trained neural network models obtained from multiple training sessions. The maximum index value is obtained from the index values ​​obtained in each test, and the target test prediction probability map corresponding to the maximum index value is determined. The trained neural network model that outputs the target test prediction probability map is used as the image segmentation model.

4. The method of claim 1, wherein, Each convolutional block processes the image features of the image to be segmented, including: Multiple first feature maps are obtained based on the first image features of the input convolutional block; Based on the amount of feature information contained in each first feature map, obtain the channel weight corresponding to each first feature map; Based on the channel weights, different first feature maps are combined to obtain at least one set of first feature maps. The at least one set of first feature maps is then subjected to pseudo-one-dimensional convolution processing to obtain the second image features. Multiple second feature maps are obtained based on the second image features; Based on the channel weights, different second feature maps are combined to obtain at least one set of second feature maps. Pseudo-two-dimensional convolution processing is then performed on the at least one set of second feature maps to obtain the image features output by the convolution block.

5. The method of claim 1, wherein, The method further includes: Obtain the image coordinate information of the target segmented image relative to the image coordinate system; Based on the transformation relationship between the image coordinate system and the world coordinate system, the image coordinate information is transformed to obtain the world coordinate information of the target segmented image relative to the world coordinate system; Based on the target segmented image and the world coordinate information, an image segmentation file for the image to be segmented is generated.

6. The method of claim 1, wherein, The method further includes: For the segmentation contour of the target object in the target segmentation image, obtain the contour coordinate points in the segmentation contour; Based on the variation amplitude of the tangent vector between different contour coordinate points on the segmented contour, the contour coordinate points are filtered to obtain approximate coordinate points; The Euclidean distance between any two adjacent approximate coordinate points is obtained. Based on the Euclidean distance between any two adjacent approximate coordinate points, the approximate coordinate points are filtered to obtain target coordinate points. The target coordinate points are used to characterize the segmentation contour of the target object in the target segmentation image.

7. An image segmentation apparatus characterized by comprising: The device includes: An image acquisition module is used to acquire an image to be segmented; the image to be segmented includes the target object to be segmented. The model acquisition module is used to acquire an image segmentation model. The image segmentation model is trained using the training set of the target object and includes a feature gain block and multiple convolutional blocks. The feature gain block includes standard convolution and multiple dilated convolutions. Each convolutional block includes pseudo-one-dimensional convolution and pseudo-two-dimensional convolution. The two dimensions of the pseudo-one-dimensional convolution have a size of 1, and the one dimension of the pseudo-two-dimensional convolution has a size of 1. The three dimensions with a size of 1 are different from each other. The pseudo-two-dimensional convolution realizes feature extraction of the sliced ​​image, and the pseudo-one-dimensional convolution realizes relationship extraction between slices. The model processing module is used to process the image features of the image to be segmented through the convolutional blocks and feature gain blocks in the image segmentation model to obtain an initial segmented image; The model output module is used to obtain the image probability distribution for the target object, and to correct the initial segmentation image according to the image probability distribution to obtain the target segmentation image of the target object; The model processing module is further configured to process the third image feature input to the feature gain block using each dilated convolution to obtain intermediate image features with the same number of dilated convolutions; the dilation coefficients of each dilated convolution are different, the feature gain block is connected between the two middle convolution blocks, and the third image feature is the output of the convolution block preceding the two middle convolution blocks in the image segmentation model; for the image feature set composed of the third image feature and the intermediate image feature, two image features are sequentially selected from the image feature set, and the two selected image features are concatenated each time to obtain multiple concatenated image features; multiple standard convolutions of different sizes are used to process each concatenated image feature to obtain multiple sub-image features; according to the self-learning weights, the multiple sub-image features are weighted and summed to obtain the image feature output by the feature gain block, wherein the image feature output by the feature gain block is used as the input of the convolution block following the two middle convolution blocks in the image segmentation model.

8. The apparatus of claim 7, wherein, The model processing module is further configured to obtain multiple first feature maps based on the first image features of the input convolutional block; obtain the channel weights corresponding to each first feature map based on the amount of feature information contained in each first feature map; combine different first feature maps based on the channel weights to obtain at least one set of first feature maps; perform pseudo-one-dimensional convolution processing on the at least one set of first feature maps to obtain second image features; obtain multiple second feature maps based on the second image features; combine different second feature maps based on the channel weights to obtain at least one set of second feature maps; perform pseudo-two-dimensional convolution processing on the at least one set of second feature maps to obtain image features output by the convolutional block.

9. The apparatus of claim 7, wherein, The model acquisition module is specifically used for: Obtain labeled real images corresponding to sample medical images, and obtain the training set and test set based on the labeled real images corresponding to the sample medical images; the sample medical images include the target objects to be segmented, and the labeled real images include the target objects that have been segmented in the sample medical images; A neural network model to be trained is obtained, and the sample medical images in the training set are input into the neural network model to be trained to obtain the training prediction probability map output by the neural network model to be trained; the neural network model to be trained includes the feature gain block and multiple convolutional blocks; The image probability distribution is obtained from the training set, and the training prediction probability map is corrected based on the image probability distribution to obtain the corrected training prediction probability map. A first loss value is obtained based on the corrected training prediction probability map and the label real image, and a second loss value is obtained based on the training prediction probability map and the label real image. Based on the first loss value and the second loss value, the neural node parameters of the feature gain block and each convolutional block in the neural network model to be trained are adjusted to obtain the trained neural network model. The trained neural network model is tested using the test set, and if the test conditions are met, the trained neural network model is used as the image segmentation model.

10. The apparatus of claim 9, wherein, The model acquisition module is specifically used for: The sample medical images in the test set are input into the trained neural network model to obtain the test prediction probability map output by the trained neural network model; The test prediction probability map is corrected based on the image probability distribution to obtain the corrected test prediction probability map; The corrected test prediction probability map is compared with the real image of the label to obtain the index value, thus completing one test; The index value is used to characterize the degree of overlap between the corrected test prediction probability map and the label real image; Multiple tests are performed on multiple trained neural network models obtained from multiple training sessions. The maximum index value is obtained from the index values ​​obtained in each test, and the target test prediction probability map corresponding to the maximum index value is determined. The trained neural network model that outputs the target test prediction probability map is used as the image segmentation model.

Images