Colorectal tumor segmentation method based on multi-branch guided network in MRI images

Abstract

The present invention provides a method for segmenting colorectal tumors in magnetic resonance imaging based on a multi-branch guided network, comprising the following steps: a data normalization step; a precise segmentation preprocessing step; a tumor determination preprocessing step; a calibration step; and a prediction result acquisition step. The present invention uses a dual-branch guided structure, with the primary branch focusing on precise segmentation of the tumor region and the secondary branch focusing on determining whether the tumor exists. The secondary branch guides the primary branch, helping to reduce the false positive rate and improve the accuracy of the segmented region. The present invention combines a confidence learning method with uncertainty analysis to reduce focus on controversial regions, allowing the model to focus more on segmenting the primary region, thereby improving the overall segmentation effect.

Description

Technical Field

[0001] The present invention relates to the field of medical image technology, and more particularly to a method for segmenting colorectal tumors in magnetic resonance imaging based on a multi-branch guided network. Background Art

[0002] Currently, with the development of deep learning, it plays a vital role in the diagnosis of colorectal cancer using magnetic resonance imaging, significantly improving the efficiency and accuracy of diagnosis through automated image analysis. Traditional diagnostic methods rely on doctors' manual analysis of images, which is time-consuming and easily affected by subjective factors. In contrast, deep learning, especially convolutional neural network (CNN) technology, can automatically extract features from a large number of medical images to achieve early detection and treatment of tumors. In the screening, staging and treatment effect evaluation of colorectal cancer, deep learning models have demonstrated excellent capabilities, not only helping doctors detect tiny lesions, but also effectively reducing the misdiagnosis rate, providing strong support for the formulation of personalized treatment plans, and thus significantly improving patients' recovery outcomes.

[0003] However, the current application of deep learning in colorectal cancer image segmentation faces several bottlenecks, primarily due to two factors. First, colorectal cancer image data suffers from class imbalance, with the majority of data coming from patients with the disease. This leads to a high false positive rate, compromising the robustness of the segmentation model. Second, due to inconsistencies in manual annotation of tumor regions, such as the transition zone, the model is easily confused in these areas, resulting in poor segmentation results. Therefore, new solutions are urgently needed to address these issues. Summary of the Invention

[0004] The present invention overcomes the shortcomings of the existing technology and provides a method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network, which focuses on judging whether a tumor exists and improves the overall segmentation effect.

[0005] The technical solutions of the present invention are as follows:

[0006] The colorectal tumor segmentation method based on the multi-branch guided network in magnetic resonance imaging includes the following steps:

[0007] S1 Data normalization step: read the patient's MRI image, normalize the acquired MRI image, and convert the normalized data into the input feature vector F input ;

[0008] S2 accurate segmentation preprocessing step: input feature vector F input Input main branch encoder main , get the multidimensional feature set of the main branch image

[0009] S3 Tumor preprocessing step: Input feature vector F input Input guide sub-branch classifier Classify sub , get the sub-branch image tumor prior feature set

[0010] S4 calibration step: the main branch image multidimensional feature set C obtained in step S2 is main and the secondary branch image tumor prior feature set C obtained in step S3 sub Input feature guidance module M guild , get the calibrated image feature set

[0011] S5: Get the prediction result step: The merged image feature set C guild Input decoder Decoder to get the segmentation prediction result Mask pre .

[0012] Furthermore, the feature guidance module M guild The specific processing steps are as follows:

[0013] S4.1. The lowest dimensional feature of the main branch feature set Input the feedforward artificial neural network model MLP to obtain the calibrated features And the calibrated features is with Eigenvectors of uniform size;

[0014] S4.2. Multidimensional feature set C of main branch image main Second Dimension and the prior feature set C of the sub-branch image tumor sub Second Dimension Calculate the corresponding feature guidance module M respectively guild The calibration attention is obtained to obtain the remaining calibration feature set S4.3, combine the calibrated image features obtained in step S4.1 and step S4.2 into a set

[0015] Furthermore, the specific processing steps of calibrating attention are as follows:

[0016] Step 4.2.1, the main branch features In the input dense convolution, i∈n represents a sampling stage; the convolution kernel is 3×3 and the convolution operation with step size of 1, 3, and 5 is performed respectively. The result is restored by the convolution kernel 1×1 and the feature size is added and merged to obtain the feature. Its characteristic size and main branch characteristics remain unchanged;

[0017] Step 4.2.2, feature The pooling operation is performed using pooling kernels of 2×2, 3×3, 5×5, and 6×6 respectively. The result is restored to the feature map size using 1×1 convolution, and the result is merged along the channel dimension to obtain a new feature. Its characteristic size and main branch characteristics remain unchanged;

[0018] Step 4.2.3: Use sub-branch features As a query, the feature is used as the key and value for attention calculation, and the attention weight is obtained. Perform dot multiplication to obtain features Step 4.2.4, the features Input the feedforward network, expand the channel four times, calculate the depth convolution and restore it to the original size to obtain the calibrated attention feature

[0019] Furthermore, the The calculation formula for attention is:

[0020]

[0021] Among them, W represents the linear operation after flattening, keeping the number of features unchanged; Q is the query vector, K is the key vector, and V is the value vector.

[0022] Furthermore, the decoder comprises: guild The highest dimension is upsampled according to the sampling stage to restore the feature size, and then the calibrated image feature set C of the corresponding dimension is used. guild The features in the last layer are merged; the features of the last layer are restored to obtain the final segmentation prediction result Mask pre .

[0023] Furthermore, the above process formula is expressed as follows:

[0024]

[0025] Where Up represents upsampling operation.

[0026] Furthermore, it also includes the S6 update step: Mask the segmentation prediction result pre and the real annotation Mask gt Input uncertain area processing module M uncertain After obtaining the segmentation loss that ignores the disputed part, backpropagation is used to update the entire model parameters.

[0027] Furthermore, the uncertainty processing module M uncertain , the specific processing includes the following steps:

[0028] Step 6.1: Count the segmentation prediction results pixel by pixel pre and the real annotation Mask gt The distribution of the number of misclassified pixels C in the dataset, for which the true label is j and the number of errors predicted as i, can be defined as follows:

[0029]

[0030] in Indicates the predicted category; y * Indicates the actual labeling result; t j represents the probability threshold, D(x) represents the distance from the tumor area; d x is the distance threshold, p represents the probability; is the number of predictions for class j when it is greater than the prediction probability threshold;

[0031] Step 6.2: Calculate the average probability Q of tumor region misclassification, where Q can be defined as:

[0032]

[0033] Where M is the number of all categories;

[0034] Step 6.3: Mark all regions with a probability higher than the average probability Q.

[0035] Step 6.4: Calculate the segmentation loss and perform backpropagation to update the model weights, ignoring the marked areas. The marked areas do not participate in the loss function calculation;

[0036] Specifically, the segmentation loss is a comprehensive weighted loss including Dice loss and cross entropy loss, and its calculation formula is:

[0037] L=λL dice +(1-λ)L ce

[0038]

[0039] In the dice loss, Pre represents the predicted area and GT represents the real labeled area, and Inter represents the intersection of the two;

[0040] The i in CE loss represents the i-th sample, c represents the classification of the sample, and y represents whether the actual classification is the current class (1 or 0). Represents the classification probability; N represents the total number of samples, and C represents the total number of categories;

[0041] Step 6.5: Perform backpropagation to calculate the gradient of each layer to update the main branch model weights.

[0042] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned segmentation method when executing the computer program.

[0043] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the segmentation method described above.

[0044] The advantages of the present invention over the prior art are:

[0045] This paper proposes a method for colorectal tumor MRI image segmentation based on a multi-branch guidance network. First, this method uses a dual-branch guidance structure: the primary branch focuses on accurate segmentation of the tumor region, while the secondary branch focuses on determining the presence or absence of the tumor. The secondary branch guides the primary branch, helping to reduce the false positive rate and improve the accuracy of the segmented region. Second, a confidence learning method based on uncertainty analysis reduces focus on controversial regions, allowing the model to focus more on segmenting the primary region, thereby improving the overall segmentation effect.

[0046] This method also feeds the segmentation prediction results and ground truth annotations into the uncertain region processing module. The segmentation loss, ignoring the disputed regions, is then back-propagated to update the model. This effectively reduces the false positive rate of model predictions and mitigates suboptimal training issues caused by disputed regions. This effectively improves the accuracy and robustness of the segmentation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of the segmentation method of the present invention;

[0048] Figure 2 It is a method architecture diagram of the present invention;

[0049] Figure 3 It is the guide module M of the present invention guild Structural diagram. DETAILED DESCRIPTION

[0050] The embodiments of the present invention are described in detail below, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not intended to limit the present invention.

[0051] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such, will not be interpreted in an idealized or overly formal sense.

[0052] The step numbers mentioned in each embodiment are for convenience of description only and do not constitute a sequence of events. The different steps in each embodiment can be combined in different sequences to achieve the objectives of the present invention. Furthermore, some structures, modules, and modeling methods not described in detail in the present invention can be implemented using conventional technical means and will not be described in detail.

[0053] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0054] like Figures 1 to 3 As shown in FIG, the method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network specifically includes the following steps:

[0055] S1 Data normalization step: read the patient's MRI image, normalize the acquired MRI image, and convert the normalized data into the input feature vector F input The details are as follows:

[0056] Obtain a patient's MRI image, using a high-resolution transverse abdominal T2 image generated using diffusion-weighted imaging technology as an example. The image dimensions are typically (C, 512, 512), where C is the number of scan slices, which may vary for different patients. Normalize the image, specifically by:

[0057] Physical normalization: By mapping the image intensity value to a uniform physical unit, the image spacing is read. If the Spacing is not 0.3 mm, it is proportionally normalized to 0.3 mm.

[0058] Z-Score normalization: Calculate the mean and variance of the image and adjust the image intensity values ​​to data with zero mean and unit variance.

[0059] After normalization, it is converted into the input feature vector F input Taking the 32-layer size as an example, the input size is (32, 1, 3, 512, 512), and its value is between (0, 1], which is converted into a tensor array to form the input feature vector F input .

[0060] S2 accurate segmentation preprocessing step: input feature vector F input Input main branch encoder main , get the multi-dimensional feature set of the main branch image

[0061] The main branch encoder Encoder main , input feature vector F input Input the image feature extraction network to extract the multidimensional feature set of the main branch image The stage represents all sampling stages. Specifically, it generally includes four stages of sampling output:

[0062] Phase 1 is a feature vector of size (32, 256, 1, 128, 128).

[0063] Phase II is a feature vector of size (32, 512, 1, 64, 64).

[0064] Phase 3 is a feature vector of size (32, 1024, 1, 32, 32).

[0065] Stage 4 is a feature vector of size (32, 2048, 1, 16, 16).

[0066] This solution uses ResNet-50 as the feature extraction network. Any deep learning network with feature extraction capabilities can be substituted. The number of network parameters and complexity should be determined based on the specific task. Four sampling stages are used, with a dimensionality reduction block input at the beginning of each stage. This solution uses a four-stage approach as an example, as it is currently the most effective method.

[0067] The specific structural order of the dimensionality reduction block is as follows: first, perform convolution with a stride of 2 and a kernel of 3; then perform batch normalization; and finally, activate with the ReLU activation function.

[0068] S3 Tumor preprocessing step: Input feature vector F input Input guide sub-branch classifier Classify sub , get the sub-branch image tumor prior feature set

[0069] Among them, the sub-branch classifier Classify sub , input feature vector F inputInput classification feature extraction network to extract multi-dimensional sub-branch image tumor prior feature set Here, stage represents all sampling stages. For the classification feature extraction network, this solution uses DenseNet-121 as the classification feature extraction network. Naturally, any classification deep learning network with feature extraction function can also be used instead. The number of network parameters and complexity are selected according to the actual task.

[0070] The specific sampling stage generally includes four stages of sampling feature output:

[0071] Phase 1 is a feature vector of size (32, 256, 1, 128, 128).

[0072] Phase II is a feature vector of size (32, 512, 1, 64, 64).

[0073] Phase 3 is a feature vector of size (32, 1024, 1, 32, 32).

[0074] Stage 4 is a feature vector of size (32, 2048, 1, 16, 16).

[0075] Then the tumor prior feature set C sub The input to the classifier yields a tumor-presence classification result. The specific classifier structure includes global pooling, a fully connected layer, and an activation function. Global pooling is performed using an average pooling operation with a kernel of 1 followed by flattening. The fully connected layer outputs the target class (where tumor-presence is a binary classification task). The activation function is a Sigmoid activation function, which converts the predicted score into a probability distribution.

[0076] The weights are updated based on the actual classification loss calculated based on whether there is a tumor.

[0077] Specifically, the loss function for classification result supervision is the cross entropy loss function, and the calculation formula is:

[0078]

[0079] Among them, M is the number of positive samples, N is the number of secondary samples, and p is the probability. i represents the number of samples, y represents the label of whether the sample is a tumor, and y i =1 means the sample is labeled as positive, y i =0 means the sample is labeled as negative.

[0080] S4 calibration step: the main branch image multidimensional feature set C obtained in step S2 is mainand the secondary branch image tumor prior feature set C obtained in step S3 sub Input feature guidance module M guild , get the calibrated image feature set

[0081] Feature-guided module M guild The specific processing steps are as follows:

[0082] S4.1. The lowest dimensional feature of the main branch feature set Input the feedforward artificial neural network model MLP to obtain the calibrated features And the calibrated features is with The feedforward artificial neural network model MLP is used to fit the nonlinear relationship between input and output (represented here as and ), increasing the model capacity.

[0083] S4.2. Set the main branch image multidimensional feature set C main Second Dimension and the prior feature set C of the sub-branch image tumor sub Second Dimension Calculate the corresponding feature guidance module M respectively guild The calibration attention is obtained to obtain the remaining calibration feature set

[0084] Calibrate attention, the specific process is as follows:

[0085] Step 4.2.1, the main branch features In the input dense convolution, i∈n represents a sampling stage (note the main branch feature here Does not include the lowest dimensional features ). Use the convolution kernel of 3×3 to perform convolution operations with step sizes of 1, 3, and 5 respectively. Use the convolution kernel of 1×1 to restore the feature size of the three step size convolution operations. Add and merge the results of the three restored feature sizes to get the feature Its characteristic size and main branch characteristics Remain unchanged.

[0086] Step 4.2.2, feature Pooling operations are performed using pooling kernels of 2×2, 3×3, 5×5, and 6×6 kernels respectively. The operation results of different pooling kernels are restored to the feature map size using 1×1 convolution. All the results of the restored feature map size are merged along the channel dimension to obtain new features. Its characteristic size and main branch characteristics Remain unchanged.

[0087] Step 4.2.3: Use sub-branch features As a query, the feature is used as the key and value for attention calculation, and the attention weight is obtained. Perform dot multiplication to obtain features The sub-branch features here Does not include the lowest dimensional features

[0088] Specifically, the The calculation formula for attention is:

[0089]

[0090] Where W represents the linear operation after flattening, maintaining the same number of features. Q is the query vector, K is the key vector, and V is the value vector. Specifically, the similarity between the query vector and the key vector is calculated to allow the model to focus on relevant parts of the input, namely the attention weight. The attention weight is then applied to the value vector to generate features that have been calibrated for attention.

[0091] Step 4.2.4, the features Input the feedforward network, expand the channel four times, calculate the depth convolution and restore it to the original size to obtain the calibrated attention feature

[0092] S4.3, combine the calibrated image features obtained in step S4.1 and step S4.2 into a set

[0093] S5: Get the prediction result step: The merged image feature set C guild Input decoder Decoder to get the segmentation prediction result Mask pre .

[0094] The decoder specifically includes: guild The highest dimension is upsampled according to the sampling stage to restore the feature size, and then the calibrated image feature set C of the corresponding dimension is used. guild The features in the last layer are merged; the features of the last layer are restored to obtain the final segmentation prediction result Mask pre The above process formula is expressed as follows:

[0095]

[0096] Where Up represents upsampling operation.

[0097] As a preference, the step S6 is also included: the segmentation prediction result Mask preand the real annotation Mask gt Input uncertain area processing module M uncertain After obtaining the segmentation loss that ignores the controversial part, back propagation updates the entire model parameters, which is equivalent to calculating the degree of one's own error to adjust the parameters of each step.

[0098] Real annotation mask gt The images are manually annotated by radiologists with more than 3 years of experience, initially reviewed by radiologists with more than 5 years of experience, and finally reviewed by radiologists with more than 10 years of experience.

[0099] The uncertainty processing module M uncertain , the specific processing includes the following steps:

[0100] Step 6.1: Count the segmentation prediction results pixel by pixel pre and the real annotation Mask gt The distribution of the number of misclassified pixels C in the dataset, for which the true label is j and the number of errors predicted as i, can be defined as follows:

[0101]

[0102] in Indicates the predicted category; y * Indicates the actual labeling result; t j represents the probability threshold, D(x) represents the distance from the tumor area; d x is the distance threshold; p represents the probability. is the number of predictions for class j when it is greater than the prediction probability threshold.

[0103] Step 6.2: Calculate the average probability Q of tumor region misclassification, where Q can be defined as:

[0104]

[0105] Where M is the number of all categories.

[0106] Step 6.3: Mark all regions with probability higher than the average Q.

[0107] Step 6.4: Calculate the segmentation loss and perform backpropagation to update the model weights, ignoring the marked areas. The marked areas do not participate in the loss function calculation.

[0108] Specifically, the segmentation loss is a comprehensive weighted loss including Dice loss and cross entropy loss, and its calculation formula is:

[0109] L=λL dice +(1-λ)L ce

[0110]

[0111] In the dice loss, Pre represents the predicted area and GT represents the true labeled area, Inter represents the intersection of the two, and λ is a constant coefficient ranging from 0 to 1, generally set to 0.6.

[0112] The i in CE loss represents the i-th sample, c represents the classification of the sample, and y represents whether the actual classification is the current class (1 or 0). represents the classification probability. N represents the total number of samples, and C represents the total number of categories.

[0113] Step 6.5: Perform backpropagation to calculate the gradient of each layer to update the main branch model weights.

[0114] S7 Model Application Step: Fix the model parameters obtained in step S6 and use them for colorectal tumor segmentation image processing. That is, for an unknown colorectal tumor image, input a high-resolution transverse abdominal T2 diffusion-weighted imaging magnetic resonance image to obtain the corresponding tumor region in the unknown colorectal tumor image.

[0115] In one embodiment, the present invention further provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the segmentation method provided in the above embodiment are implemented.

[0116] In one embodiment, the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the segmentation method provided in the above embodiment are implemented.

[0117] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, which 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-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0118] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned 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.

[0119] The descriptions of the various embodiments of the present application are relatively specific and detailed, but this should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, and such modifications and refinements should be considered within the scope of protection of the present invention. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network, characterized by: The specific steps include: S1 Data normalization step: read the patient's MRI image, normalize the acquired MRI image, and convert the normalized data into the input feature vector F input ; S2 accurate segmentation preprocessing step: input feature vector F input Input main branch encoder main , get the multidimensional feature set of the main branch image S3 Tumor preprocessing step: Input feature vector F input Input guide sub-branch classifier Classify sub , get the sub-branch image tumor prior feature set S4 calibration step: the main branch image multidimensional feature set C obtained in step S2 is main and the secondary branch image tumor prior feature set C obtained in step S3 sub Input feature guidance module M guild , get the calibrated image feature set S5: Get the prediction result step: The merged image feature set C guild Input decoder Decoder to get the segmentation prediction result Mask pre ; The feature guidance module M guild The specific processing steps are as follows: S4.

1. The lowest dimensional feature of the main branch feature set Input the feedforward artificial neural network model MLP to obtain the calibrated features And the calibrated features is with Eigenvectors of uniform size; S4.

2. Set the main branch image multidimensional feature set C main Second Dimension and the prior feature set C of the sub-branch image tumor sub Second Dimension Calculate the corresponding feature guidance module M respectively guild The calibration attention is obtained to obtain the remaining calibration feature set S4.3, combine the calibrated image features obtained in step S4.1 and step S4.2 into a set 2. The method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network according to claim 1, characterized in that: The specific processing steps of calibrating attention are as follows: Step 4.2.1, the main branch features In the input dense convolution, i∈n represents a sampling stage; the convolution kernel is 3×3 and the convolution operation with step size of 1, 3, and 5 is performed respectively. The result is restored by the convolution kernel 1×1 and the feature size is added and merged to obtain the feature. Its characteristic size and main branch characteristics remain unchanged; Step 4.2.2, feature The pooling operation is performed using pooling kernels of 2×2, 3×3, 5×5, and 6×6 respectively. The result is restored to the size of the feature map using 1×1 convolution, and the result is merged along the channel dimension to obtain a new feature. Its characteristic size and main branch characteristics remain unchanged; Step 4.2.3: Use sub-branch features As a query, the feature is used as the key and value for attention calculation, and the attention weight is obtained. Perform dot multiplication to obtain features Step 4.2.4, the features Input the feedforward network, expand the channel four times, calculate the depth convolution and restore it to the original size to obtain the calibrated attention feature 3. The method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network according to claim 2, characterized in that: The The calculation formula for attention is: Among them, W represents the linear operation after flattening, keeping the number of features unchanged; Q is the query vector, K is the key vector, and V is the value vector.

4. The method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network according to claim 1, characterized in that: The decoder specifically includes: guild The highest dimension is upsampled according to the sampling stage to restore the feature size, and then the calibrated image feature set C of the corresponding dimension is used. guild The features in the last layer are merged; the features of the last layer are restored to obtain the final segmentation prediction result Mask pre .

5. The method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network according to claim 4, characterized in that: The above process formula is expressed as follows: Where Up represents upsampling operation.

6. The method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network according to claim 1, characterized in that: Also includes the S6 update step: Mask the segmentation prediction result pre and the real annotation Mask gt Input uncertain area processing module M uncertain After obtaining the segmentation loss that ignores the disputed part, backpropagation is used to update the entire model parameters.

7. The method for colorectal tumor segmentation in magnetic resonance imaging based on a multi-branch guided network according to claim 3, characterized in that: Uncertainty processing module M uncertain , the specific processing includes the following steps: Step 6.1: Count the segmentation prediction results pixel by pixel pre and the real annotation Mask gt The distribution of the number of misclassified pixels C in the dataset, for which the true label is j and the number of errors predicted as i, can be defined as follows: in Indicates the predicted category; y * Indicates the actual labeling result; t j represents the probability threshold, D(x) represents the distance from the tumor area; d x is the distance threshold, p represents the probability; is the number of predictions for class j when it is greater than the prediction probability threshold; Step 6.2: Calculate the average probability Q of misclassification of the tumor region, where Q can be defined as: Where M is the number of all categories; Step 6.3: Mark all regions with a probability higher than the average probability Q. Step 6.4: Calculate the segmentation loss and perform backpropagation to update the model weights, ignoring the marked areas. The marked areas do not participate in the loss function calculation; Specifically, the segmentation loss is a comprehensive weighted loss including Dice loss and cross entropy loss, and its calculation formula is: L=λL dice +(1-λ)L ce In the dice loss, Pre represents the predicted area and GT represents the true labeled area, and Inter represents the intersection of the two. The i in CE loss represents the i-th sample, c represents the classification of the sample, and y represents whether the actual classification is the current class (1 or 0). Represents the classification probability; N represents the total number of samples, and C represents the total number of categories; Step 6.5: Perform backpropagation to calculate the gradient of each layer to update the main branch model weights.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the segmentation method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the segmentation method according to any one of claims 1 to 7 are implemented.

Images