Method for segmenting colorectal cancer CT image by introducing 3D UNet network of separation heavy parameter and compression excitation structure

Abstract

The objective of the invention is to provide a method for segmenting a colorectal cancer CT (Computed Tomography) image by introducing a 3DUNet network which separates heavy parameters and compresses an excitation structure (3D UNet with objective and auxiliary diagnosis) in order to solve the problem of diagnosis of colorectal cancer lesions by subjective factors, so that the objective and auxiliary diagnosis is realized, and the objective and auxiliary diagnosis can be realized, and the objective and auxiliary diagnosis can be realized, and therefore, the objective and auxiliary diagnosis can be realized, and the objective and auxiliary diagnosis can be realized, and the objective and auxiliary diagnosis can be realized, and the objective and auxiliary diagnosis can be realized, and the objective and auxiliary diagnosis can be realized. The invention provides a method for segmenting a colorectal cancer CT (Computed Tomography) image by introducing a 3D UNet network for separating heavy parameters and compressing an excitation structure. On the basis of a 3D UNet segmentation network, separation re-parameter convolution and a compression excitation structure are added into an encoder part, a decoder uses a transpose convolution mode to replace up-sampling and improve the resolution of a feature map, and the encoding-decoding part adopts jump connection for splicing; and constructing an automatic colorectal cancer focus segmentation system by using the optimized segmentation model, generating a focus segmentation result through data acquisition, data preprocessing, model reasoning and visualization processes, and assisting doctors in diagnosing focus boundaries.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning segmentation models and medical image processing in CT images, and specifically relates to a method for segmenting colorectal cancer CT images by using a 3D UNet network that introduces separation weight parameters and a compression excitation structure. Background Art

[0002] Colorectal cancer is a major public health problem worldwide. According to the latest global cancer statistics report released by the World Health Organization's International Agency for Research on Cancer (IARC) in 2022, colorectal cancer has become the third most common cancer in the world, with 1.9 million new cases worldwide by 2022, accounting for 9.6% of all cancers; at the same time, the mortality rate of colorectal cancer ranks second among all cancers. According to statistics, more than 900,000 deaths from colorectal cancer in 2022 accounted for 9.3% of the total number of deaths from all cancers. According to IARC's forecast data, it is expected that the number of new cases of colorectal cancer will exceed 35 million in 2050, an increase of 77% from 2022, which will undoubtedly bring a huge burden on medical resources, patients' families and society. For colorectal cancer, the preferred examination is colorectal endoscopy, which has the advantages of direct observation of lesions, simultaneous biopsy function and full colon examination. However, considering the overall assessment of the tumor, the basis for preoperative planning, the presence of distal metastases and the intolerance of some patients to colonoscopy, colorectal endoscopy has certain shortcomings.

[0003] Computed Tomography (CT) uses collimated X-rays, gamma rays, and highly sensitive detectors to perform continuous tomographic scans around the target area of ​​the human body. Using CT to examine colorectal cancer can avoid the discomfort caused by the intestinal preparation stage during colonoscopy and the intestinal damage that may be caused during the examination. For patients with lesions, CT examination has the following advantages: First, CT examination can provide a full picture of the lesion in the abdominal cavity, which is an important imaging basis for determining whether the lesion has invaded the surrounding tissues and organs; second, for patients who need lesion resection, CT imaging examination provides clear anatomical information for surgeons to formulate surgical plans; in addition, some patients have poor physical conditions, cardiopulmonary insufficiency, stenosis and obstruction, in which case colorectal examination is not recommended; finally, for patients in the middle and late stages, the lesions may have proximal or distal metastases, and CT images can scan multiple organs at the same time, which can effectively detect possible metastatic lesions. When doctors diagnose CT images of colorectal lesions, it is a very important task for doctors to distinguish the boundaries of lesions. The boundaries of lesions have a certain relationship with the degree of tumor invasion, benign or malignant, the formulation of surgical resection plans, and the judgment of prognosis. However, this task is usually affected by the experience of the diagnostic physician. For the same patient's images, physicians of different seniority may have inconsistent judgments on the boundaries of lesions. This subjective experience has a great impact on the diagnosis and subsequent treatment plans of colorectal cancer patients. In summary, considering the subjective factors that affect the judgment of colorectal cancer lesion boundaries in CT images by different physicians at this stage, there is an urgent need for a method that can objectively evaluate the boundaries of colorectal cancer lesions in CT images to assist physicians in judging the boundaries of colorectal cancer lesions. Summary of the invention

[0004] The purpose of the present invention is to address the problem of subjective factors in the diagnosis of colorectal cancer lesions, and to propose an objective and auxiliary diagnosis method for colorectal cancer CT image segmentation by introducing a 3D UNet network with separated reparameterization of Squeeze and Excitation (SR-SE-3D UNet). The model established by this method generates lesion segmentation results to assist physicians in diagnosing lesion boundaries.

[0005] According to a technical solution of the present invention, a method for colorectal cancer CT image segmentation is provided. The method introduces a 3D UNet network with separated heavy parameters and compressed excitation structure, comprising the following steps:

[0006] S1, data collection and grayscale correction;

[0007] S2, data preprocessing and data augmentation;

[0008] S3, model design;

[0009] S4, model training and optimization;

[0010] S5. Construction of automatic segmentation system.

[0011] This method is a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and compressed excitation structure.

[0012] According to a technical solution of the present invention, a method for colorectal cancer CT image segmentation using a 3DUNet network that introduces separation weight parameters and a compression excitation structure is provided, which specifically includes the following steps:

[0013] S1. Data collection and grayscale correction: screening of Dicom format CT image data of patients with confirmed colorectal cancer, adjusting the window width and window position of Dicom format images to optimize the display effect of tissues and organs;

[0014] S2, data preprocessing and data augmentation, converting the desensitized Dicom format CT images into high-quality NIFTI format through lossless conversion, and having senior physicians outline the ROI of the tumor area in the image, constructing data groups, and using data augmentation to expand the data groups;

[0015] S3, model design, based on the 3D UNet segmentation network, adds separation re-parameter convolution and compressed excitation structure to the encoder part, and uses transposed convolution instead of upsampling to improve the feature map resolution. The encoding-decoding part uses skip connection for splicing;

[0016] S4, model training and optimization, build an optimizer suitable for segmentation tasks, initialize hyperparameters, loss function and early stopping mechanism, and then perform model training. After training, fine-tune the model based on the evaluation indicators to optimize model performance;

[0017] S5. Automatic segmentation system construction: Use the optimized segmentation model to build an automatic segmentation system for colorectal cancer lesions. Through data collection, data preprocessing, model reasoning and visualization, it generates lesion segmentation results to assist physicians in diagnosing lesion boundaries.

[0018] In one embodiment, the method further comprises:

[0019] Before the data included in the present invention were enrolled, some patients died due to the progression of the disease, so all data were approved by the ethics committee and signed the informed consent form. Among the data of patients with colorectal cancer included, all patients' lesions were diagnosed as colorectal cancer by pathological histological examination; for patients who were not pathologically diagnosed but had typical clinical features of colorectal cancer such as blood in stool, abdominal pain, abdominal mass, and imaging features such as intestinal wall thickening and intestinal obstruction, the imaging data of patients who were highly suspected of colorectal cancer after comprehensive judgment by the multidisciplinary treatment team (MDT) were also included in the data group of this study; in the included data, all patients had to be at least 18 years old and had not undergone surgical resection, radiotherapy, chemotherapy or targeted therapy; the included colorectal cancer CT imaging data were all in original Dicom format.

[0020] In one embodiment, the method further comprises:

[0021] During the above data collection and grayscale correction, in order to facilitate doctors to accurately outline the tumor ROI area in the future, Python language combined with Pydicom and Numpy scientific databases were used to set the window width to 400HU and the window level to 40HU. The window width and window level in the Dicom format of the original colorectal cancer CT data were uniformly adjusted to enhance tissue contrast and optimize the contrast of tissues and organs, so as to facilitate doctors to observe the boundary area of ​​​​the colorectal tumor and achieve the purpose of accurately outlining the ROI area.

[0022] In one embodiment, the method further comprises:

[0023] In the above Dicom format conversion and data desensitization, the SimpleITK scientific database of Python language is used to read the Dicom format image data sequence of colorectal cancer CT images, and the ImageSeriesReader() and GetGDCMSeriesFileNames() methods are used to read the image information in Dicom format, remove the patient identity information (name, gender, date of birth and patient ID), medical information (examination date and time, examination site, diagnosis report content), equipment and examination hospital information (hospital information, equipment sequence and equipment model), and use the WriteImage method to save the image as high-quality NifTI format data to reduce the loss of image information in the image data format conversion.

[0024] In one embodiment, the method further comprises:

[0025] When the above-mentioned senior physicians outlined the tumor ROI area, they used the Label outline software written in Python to outline the tumor area in the colorectal cancer imaging data. When outlining, the outline boundary needed to fit its boundary, the tumor area needed to be completely included in the outline line, and the diagnosis and pathological changes needed to be considered. After the outlining period, the consistency of the tumor in the same case and the consistency between different annotating physicians needed to be ensured. If there was a large difference in the outline of the same tumor area by two physicians, another annotator with more years of diagnosis than the two annotating physicians would outline the tumor area.

[0026] In one embodiment, the method further comprises:

[0027] In the 3D UNet network that introduces the separation re-parameter and compression excitation structure, the tumor segmentation model structure based on colorectal cancer CT image data is based on the unique "U" structure of 3DUNet. The SR-SE-3D UNet mainly includes feature contraction (downsampling), feature expansion part (transposed convolution) and output layer, in which the feature contraction and feature expansion use the jump connection method to fuse the low-level semantic feature information of edge and texture in the shallow network with the high-level semantic features of shape and category in the deep network; in colorectal cancer CT image data, there may be small lesions and large lesions in multiple slices, so the separation re-parameter structure is added to the downsampling part. When the model infers the CT sequence, the complexity of the model in the reasoning process is reduced and the expression ability of the model in the nonlinear region is enhanced; the compression excitation module is added. When the small lesion target task is segmented, the module will recalibrate the extracted features, and assign importance to each channel based on the global feature information, so that the network can adaptively weaken the background part and enhance the target area features when segmenting the small lesion target.

[0028] In one embodiment, the method further comprises:

[0029] In the above model training and optimization, Dice loss L is used Dice , intersection loss L Iou and Focal loss L Focal Constitute the overall loss function Loss Total , where L Dice The loss function makes the model focus on the overlap between the target area's outlined labels and the model's predicted labels, avoiding the model's excessive focus on the background area. Pre is the tumor area predicted by the model, Num Pre is the number of pixels contained in it, Img True The tumor area outlined by the physician, Num True is the number of pixels contained in it, and its formula is as follows:

[0030]

[0031] Among them, L Iou The loss is based on the Jaccard coefficient, which makes the model focus on the ratio between the intersection and union of the target area outlined by the physician and the area predicted by the model. For the segmentation model of a single task in this embodiment, it can well measure the accuracy of the segmentation model, and its formula is as follows:

[0032]

[0033] Among them, L Focal The loss function is used to solve the problem that some tumor areas are small and the background areas are large. If this happens, L Focal The loss function will make the model pay more attention to the small target tumor area and reduce the attention to the background area, where α and β are hyperparameters that control the model to focus on the target area, and the initial values ​​are set to 0.25 and 2, y i，j is the true category of the pixel at position i, j (0 is the background area, 1 is the target tumor area), p i，j is the predicted category of pixel i, j, L Focal The loss function is as follows:

[0034]

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

[0036] In the above SR-SE-3D UNet colorectal cancer tumor segmentation model, the Adam optimizer is used, the initial learning rate is 0.001, the learning rate decay strategy uses the exponential decay strategy, the decay rate is set to 0.95, and the patience value in the early stopping mechanism is set to 10 Epoches. During the training process, if the loss value on the validation set does not decrease in 10 Epoches, the training is stopped and the weight parameters of 10 Epoches are saved.

[0037] In one embodiment, the method further comprises:

[0038] In the above-mentioned SR-SE-3D UNet colorectal cancer tumor segmentation model fine-tuning and model performance optimization, after the training is completed, the optimal model weight parameters are saved, and the learning rate, learning rate reduction ratio, loss function hyperparameter initial value, and Batch Size are fine-tuned by combining the transfer learning method to further enhance the generalization performance of the model.

[0039] The present invention provides a system for automatically segmenting colorectal cancer CT images using a 3D UNet network that introduces separation weight parameters and a compression excitation structure. The system uses an automatic segmentation system to introduce a 3D UNet network that introduces separation weight parameters and a compression excitation structure. The automatic segmentation system includes an optimized segmentation model established by any of the above methods. The system may also include a computer-readable storage medium, on which a computer program is stored. When the computer program is executed in a computer, the computer executes the optimized segmentation model.

[0040] The beneficial effects of the present invention are as follows: the present invention draws on the architecture of the 3D UNet segmentation network in the deep learning convolutional neural network, combines the characteristics of colorectal cancer CT image data, fully considers the size of colorectal tumors, the consistency of CT images and the label inconsistency existing in manual outlining, and solves the problem of segmenting tumor lesions in colorectal cancer CT images by adding separation weight parameters and compressed excitation structures. On the one hand, the shortcomings of the original 3D UNet network, such as slow convergence speed and low segmentation accuracy, are improved, and the speed is improved in the network inference stage, and the large convolution kernel is reduced, so that the model also has a higher segmentation ability for tumor areas of smaller volume targets; on the other hand, the problem of information interaction between sequences in CT image data is improved through the compression excitation structure, so that the model considers the channel sharing problem of the previous and next slices of the current slice when segmenting the tumor area; at the same time, the tumor area in the colorectal CT image is segmented through deep learning, an objective segmentation method, which overcomes the differences between the subjective judgments of different physicians and solves the image quality problems caused by external factors brought by different inspection equipment; finally, by building SR-SE-3D UNet into an easy-to-operate automated segmentation system, the operability of physicians in using deep learning for tumor diagnosis is greatly simplified. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A design flow chart of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compression excitation structure in an embodiment of the present invention;

[0042] Figure 2 A separation weight parameter structure diagram of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compression excitation structure in an embodiment of the present invention;

[0043] Figure 3 A diagram of a compressed excitation structure of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compressed excitation structure in an embodiment of the present invention;

[0044] Figure 4This is a SR-SE-3D UNet network structure diagram of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compression excitation structure in an embodiment of the invention;

[0045] Figure 5 This is a diagram showing the segmentation results of the method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compressed excitation structure in an embodiment of the invention. DETAILED DESCRIPTION

[0046] The following embodiments are intended to further illustrate some preferred embodiments of the present invention, but are not all embodiments. Other embodiments based on the present invention made by professionals in the field without creative work all belong to the scope of protection of the present invention. The present invention will be further described below in conjunction with the accompanying drawings.

[0047] The inspection methods without specific conditions in the embodiments are usually carried out under conventional conditions or according to the conditions recommended by the manufacturer. Figure 1 .

[0048] Figure 1 FIG. 1 is a flow chart of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compression excitation structure in an embodiment of the present invention. Figure 1 The present invention mainly includes five main parts S1 to S5, including:

[0049] Step S1, data collection and grayscale correction, screening the Dicom format CT image data of patients with confirmed colorectal cancer, adjusting the window width and window level of the Dicom format image to optimize the display effect of tissues and organs.

[0050] Step S2, data preprocessing and data augmentation, convert the desensitized Dicom format CT images into high-quality NIFTI format through lossless conversion, and have senior physicians outline the ROI of the tumor area in the image, build data groups, and expand the data groups using data augmentation.

[0051] Specifically, in data augmentation, the data outlined by the doctor are subjected to noise processing (Gaussian noise, salt and pepper noise), image blur processing (mean blur, Gaussian blur), affine transformation (translation, rotation, scaling), and elastic deformation operations. To ensure that the data is consistent with the situation when CT images are taken in a real clinical environment, the rotation angle is between ±10° and ±30° in the plane, ensuring that the model learns the complex situations that arise during shooting without distorting the shape of the organs.

[0052] Furthermore, after data augmentation, the CT data in NIFTI format are converted into a high-quality PNG model read-in format, and the mask labels in NII format are converted into a high-quality JPG image format after enhancement. Both are normalized on the Z axis. First, the mean value u of each slice pixel in the CT image is calculated, where n is the number of pixels in the CT image sequence, and x is the pixel size. i is the pixel value in the i-th image sample, and the calculation formula is:

[0053]

[0054] The standard deviation σ of the sequence is calculated by u, and the calculation formula is:

[0055]

[0056] Use the calculated mean u and standard deviation σ to normalize the CT image sequence on the Z axis. Nor is the normalized image, and the calculation formula is:

[0057]

[0058] In the label file, the pixel value of the tumor target area in the normalized image is 1, which is displayed as white in the image; the pixel value of the background part of the non-tumor area is 0, which is displayed as black in the image.

[0059] Step S3, model design, based on the 3D UNet segmentation network, adds separation re-parameter convolution and compressed excitation structure to the encoder part, the decoder uses transposed convolution instead of upsampling to improve the feature map resolution, and the encoding-decoding part uses jump connection for splicing.

[0060] Specifically, in the SR-SE-3D UNet network, the downsampling is located in the contraction part, and its purpose is to select important features in the downsampling process to reduce the data dimension and the amount of calculation and the number of parameters, while improving the receptive field of the model so that the network model can focus on the characteristics of the macroscopic area in the CT image; the transposed convolution is located in the expansion part, and its purpose is to restore the spatial feature information in the contraction process so that the network can learn the high-level semantic features of the tumor area (tumor shape and position), so as to better complete the segmentation task.

[0061] Step S4, model training and optimization, build an optimizer suitable for the segmentation task, initialize the hyperparameters, loss function and early stopping mechanism, and then perform model training. After training, fine-tune the model based on the evaluation indicators to optimize the model performance.

[0062] Step S5, automatic segmentation system construction, using the optimized segmentation model to build a colorectal cancer lesion automatic segmentation system, through data collection, data preprocessing, model reasoning and visualization process, generate lesion segmentation results, to assist doctors in diagnosing lesion boundaries.

[0063] Specifically, after the SR-SE-3D UNet segmentation model is optimized, the colorectal cancer CT image data preprocessing module is constructed in combination with the model construction process, including data desensitization, image format conversion, Z-axis normalization, and input data size adjustment operations; the model reasoning and prediction module includes SR-SE-3D UNet model structure construction, optimal model weight parameter loading, CT image data loading, and model reasoning and prediction; the model result output module includes outputting prediction results and restoring the predicted label map size using the bilinear interpolation method; the label processing module includes label post-processing and label and CT image fusion modules.

[0064] Furthermore, in the label processing module, a color overlay method, a boundary drawing method, and a transparency blending method are provided to fuse with the colorectal cancer CT image data. Among them, the color overlay method converts the label predicted by the model into a color representation and then overlays it on the original CT image; the boundary drawing method draws the boundary of the label predicted by the model on the original CT image, so that the diagnostic physician can intuitively see the outline of the model segmentation area; the transparency blending method adjusts the transparency of the tumor area label predicted by the model, and then blends it with the original CT image to facilitate the physician to observe the lesion condition and make a diagnosis.

[0065] Figure 2 This is a diagram of the separation weight parameter structure of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compressed excitation structure in an embodiment of the present invention.

[0066] In the separation re-parameter structure, the input feature map enters the separation re-parameter module and undergoes four-way parallel convolution operations. First, through the 3D convolution operation, the spatial features of the input feature map in three dimensions of depth, height and width are captured; secondly, through the 3D dilation operation, the edge of the captured CT image features is dilated. This operation can fill the tiny depressions at the boundary of the target area without losing the details of the tumor target area, making its edge smoother, while expanding the receptive field and enhancing the continuity of the feature space; finally, through the batch normalization operation, the variable offset within the feature map is reduced, making the network training more efficient; the features obtained after the four-way parallel operation are concatenated with the input features to obtain the output features through 5-feature map splicing, completing the separation re-parameter structure operation.

[0067] Furthermore, the four-way parallel operations are all completed by 3D convolution, 3D dilation and batch normalization operations. From left to right, the first path is a 3D convolution with a convolution size of 7 and a voxel cube with a dilation parameter of 1 as the structural element; the second path is a 3D convolution with a convolution size of 5 and a voxel cube with a dilation parameter of 1 as the structural element; the third path is a 3D convolution with a convolution size of 3 and a voxel cube with a dilation parameter of 2 as the structural element; the fourth path is a 3D convolution with a convolution size of 3 and a voxel cube with a dilation parameter of 3 as the structural element.

[0068] Figure 3 This is a diagram of the compressed excitation structure of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compressed excitation structure in an embodiment of the present invention.

[0069] In the compressed excitation structure, the input feature map first passes through the separation heavy parameter structure, so that the input features are spliced ​​with the input features after four parallel feature extractions, and then enter the lightweight SE module after batch normalization. The channel features are adaptively weighted to enhance the representation of the target features and strengthen the network's ability to extract key channel features from CT images. The features input to the SE module are then subjected to a deep sparse clustering operation, with the aim of mapping the previously extracted high-dimensional features to a low-dimensional space through a clustering method, thereby effectively reducing the noise and redundant information in the features, so that the features obtained can better reflect the essential characteristics of the target task. The GELU activation function is then used to weight the features obtained by deep sparse clustering, and its weight is determined by the cumulative distribution function of the standard normal distribution. A deep sparse clustering and batch normalization operation is then performed to splice the features obtained by batch normalization with the input features to complete the operation of the compressed excitation module.

[0070] Furthermore, in the compression excitation module of the above SR-SE-3D UNet, since the data to be segmented is a three-dimensional sequence data of colorectal cancer CT images, which contains multiple slice channels, but the tumor lesions only exist in some slices, the SE module with separation weight parameters is designed to help sort out the importance of slices at different positions in the sequence. The compression part is mainly composed of the global average pooling (GAP) operation, which compresses the global spatial feature information of the sequence, and then the channel information is statistically calculated through the GAP operation, which is calculated as follows:

[0071]

[0072] Among them, F sq is the compression function, z c is the pixel value of the cth channel obtained after the GAP operation, x c，i，jis the plastic limit of the cth channel at position (i, j) in the CT image feature map. H and W are the height and width of the input CT image feature, respectively. After the compression operation, the channel statistical information is stimulated. The stimulation operation is divided into two steps for calculation. The first step is the dimension reduction full connection layer operation as follows:

[0073] s1=δ(W1z+b1)

[0074] Among them, δ is the activation function. In this embodiment, δ(x)=max(0,x), W1 is the weight matrix of the first fully connected layer, which is used to perform a linear transformation on the input z feature, and b1 is a bias vector, which is used to shift the output in the linear transformation to enhance the generalization performance of the model. The two parts together constitute the importance weight of the s1 channel, and then s1 is subjected to a dimension-upgraded fully connected layer operation, and the calculation formula is as follows:

[0075] s=F ex (W2, s1, b2) = σ(W2s1 + b2)

[0076] Wherein, σ is the activation function. In this embodiment, W2 is the weight matrix of the second fully connected layer, b2 is the bias term, and this operation converts the s1 feature importance weight back to the original dimensional space s by upgrading the dimension. ex is the activation function. Finally, s is used to concatenate and fuse with the original input features. The calculation formula is:

[0077]

[0078] in, Indicates the concatenation operation of the original channel feature and the feature channel with weights, F scale is the output feature map.

[0079] Furthermore, using GELU instead of ReLu activation function after deep sparse clustering can provide smoother linear changes and alleviate the gradient vanishing problem caused by feature coefficients after sparse clustering. In order to reduce the initial calculation, the approximate calculation formula is:

[0080]

[0081] Where x is the eigenvalue of the input and tanh is the hyperbolic tangent function.

[0082] Figure 4 This is a SR-SE-3D UNet network structure diagram of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and compression excitation structure in an embodiment of the invention.

[0083] In the tumor segmentation network of SR-SE-3D UNet colorectal cancer CT images, the processed CT image is input and divided into two branches after convolution layer, batch normalization and ReLu activation function. One branch enters the compression excitation module for feature extraction and downsampling operation, and the other branch performs ReLu activation function on the original CT image, the feature data after convolution and downsampling, and then splices it with the input feature through jump connection. The features after each layer of splicing are spliced ​​with the features extracted from the next layer after transposed convolution operation to complete the fusion of low-level and high-level semantic information. Finally, the fused features are subjected to two convolutions, batch normalization and ReLu activation function operations to obtain the final feature map. Finally, the probability of each pixel in the feature map is judged through convolution and Sigmoid activation function, and the tumor area segmentation mask image predicted by the model is generated through the label post-processing module.

[0084] Figure 5 This is a diagram showing the segmentation results of the method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compressed excitation structure in an embodiment of the invention.

[0085] Furthermore, the CT image data of a patient diagnosed with colorectal cancer by pathological tissue biopsy was used for testing. The patient was male, 38 years old, with the primary lesion located in the rectum. The pathological diagnosis was rectal discoid-type moderately differentiated adenocarcinoma, with a tumor volume of 4.5x4x1.5cm, infiltrating into the outer fibers of the adventitia and adipose tissue, with visible vascular tumor thrombus and nerve involvement. The primary lesion was 6 cm away from the rectum, and no neoadjuvant chemotherapy was performed. The patient had a history of smoking and drinking, a BMI index of 19.6, and liver metastasis. The TNM stage was T4N0M0, CEA was 22.9, CA125 was 8.5, and CA199 was 2.

[0086] in, Figure 5 The left side shows the label drawn by the doctor, and the right side shows the test result of the SR-SE-3D UNet network. By zooming in on the tumor area, it can be seen that SR-SE-3D UNet shows good segmentation performance. The boundary of the target tumor area is consistent with the gold standard boundary drawn by the doctor. In the complex edge contours of continuous tumor lesions, detailed features of different sizes can be captured more accurately and clearly outlined.

[0087] For the specific calculation of a method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compression excitation structure, please refer to the above description of data collection and grayscale correction, data preprocessing and data augmentation, model design, model training and optimization, and automatic segmentation system construction in the invention method, which will not be repeated here. All or part of the above method for colorectal cancer CT image segmentation using a 3D UNet network that introduces separation weight parameters and a compression excitation structure can be implemented by software to facilitate the calculation of the corresponding operations in each step.

[0088] The process embodiments described above are merely illustrative, wherein the segmented organs, lesion types, imaging modalities, and segmentation networks may also be of other different structures, and the loss function, training method, and method for constructing a segmentation system during training may or may not be the methods used herein. Parts thereof may be selected according to actual needs to achieve the purpose of the present embodiment. Those of ordinary skill in the art may understand and implement the present invention without creative labor.

[0089] Through the description of the above implementation methods, those skilled in the art can clearly understand the various embodiments, and can implement the proposed methods with the help of different languages ​​and deep learning frameworks. Based on this understanding, the essence of the above technical solution or the contribution to the shared technology can be presented in the form of software or a full-stack Web model. The model can be embedded in the software or deployed in a Web network database, such as MySQL, SQL Server, Oracle, DB2, etc., and contains several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the above methods of each embodiment or some parts of the embodiment.

[0090] The above embodiments are intended to further illustrate some preferred embodiments of the present invention, but not all embodiments. Other embodiments based on the present invention made by professionals in this field without creative work are all within the scope of protection of the present invention.

Claims

1. A method for segmenting colorectal cancer CT images, characterized in that: The method introduces a 3D UNet network with separated heavy parameters and compressed excitation structure, comprising the following steps: S1, data collection and grayscale correction; S2, data preprocessing and data augmentation; S3, model design; S4, model training and optimization; S5. Construction of automatic segmentation system.

2. A method for colorectal cancer CT image segmentation according to claim 1, characterized in that: The steps also include: In the step S1, CT image data in Dicom format of patients with confirmed colorectal cancer are screened, and the window width and window level of the Dicom format images are adjusted; In the step S2, the desensitized Dicom format CT image is converted into a high-quality NIFTI format by lossless conversion, the ROI of the tumor area in the image is delineated, a data group is constructed, and the data group is expanded by data augmentation; In the S3 step, based on the 3D UNet segmentation network, the separation re-parameterized convolution and the compressed excitation structure are added to the encoder part, the decoder uses the transposed convolution method instead of upsampling to improve the feature map resolution, and the encoding-decoding part uses jump connections for splicing; In the step S4, an optimizer suitable for the segmentation task is constructed, hyperparameters, loss function and early stopping mechanism are initialized, and then model training is performed. After training, the model is fine-tuned in combination with the evaluation index; In the step S5, the optimized segmentation model is used to construct an automatic segmentation system for colorectal cancer lesions.

3. A method for colorectal cancer CT image segmentation according to claim 2, characterized in that: In step S1, Python language is used in combination with Pydicom and Numpy scientific databases to uniformly adjust the window width and window position in the Dicom format of the original colorectal cancer CT data to enhance tissue contrast and optimize the contrast of tissues and organs.

4. A method for colorectal cancer CT image segmentation according to claim 3, characterized in that: In the step S2, the Dicom format conversion and data desensitization are performed by using the SimpleITK scientific database of the Python language to read the Dicom format image data sequence of the colorectal cancer CT image, using the ImageSeriesReader() and GetGDCMSeriesFileNames() methods to read the image information in the Dicom format, removing the patient identity information, medical information, equipment and examination hospital information, and using the WriteImage method to save the image as NifTI format data; When outlining the ROI of the tumor area in step S2, the Label outlining software written in Python is used to outline the tumor area in the colorectal cancer image data. When outlining, the outline boundary needs to be fitted to its boundary, the tumor area needs to be completely included in the outline line, and the trial and pathological changes need to be considered. After the outlining period, the consistency of the tumor in the same case and the consistency between different annotating physicians need to be ensured. If there is a large difference in the outlining of the same tumor area by two physicians, another annotator with more years of diagnosis than the two annotating physicians will outline the tumor area.

5. A method for colorectal cancer CT image segmentation according to claim 4, characterized in that: The 3DUNet network with separation weight parameters and compression excitation structure is introduced in step S2. The tumor segmentation model structure based on colorectal cancer CT image data is based on the "U"-shaped structure unique to 3DUNet. SR-SE-3DUNet includes feature contraction (downsampling), feature expansion part (transposed convolution) and output layer. The feature contraction and feature expansion use jump connections to fuse the low-level semantic feature information of edges and textures in the shallow network with the high-level semantic features of shapes and categories in the deep network. A separation weight parameter structure is added to the downsampling part. A compression excitation module is added. When segmenting small lesion target tasks, the compression excitation module will recalibrate the extracted features and assign importance to each channel based on the global feature information.

6. A method for colorectal cancer CT image segmentation according to claim 5, characterized in that: In the model training and optimization, Dice loss L is used. Dice , intersection loss L Iou and Focal loss L Focal Constitute the overall loss function Loss Total , the L Dice The loss function makes the model focus on the overlap between the target area outline label and the model prediction label. Pre is the tumor area predicted by the model, Num Pre is the number of pixels contained in it, Img True is the delineated tumor area, Num True is the number of pixels contained in it, and its formula is as follows: Among them, L Iou The loss is based on the Jaccard coefficient.

7. A method for colorectal cancer CT image segmentation according to claim 6, characterized in that: For a single-task segmentation model, the accuracy of the segmentation model is measured as follows: L Focal The loss function is used to solve the problem that some tumor areas are small and the background areas are large. If this happens, L Focal The loss function will make the model pay more attention to the small target tumor area and reduce the attention to the background area, where α and β are hyperparameters that control the model to focus on the target area, and the initial values ​​are set to 0.25 and 2. i，j is the true category of the pixel at position i, j (0 is the background area, 1 is the target tumor area), p i，j is the predicted category of pixel i, j, L Focal The loss function is as follows:

8. The method for colorectal cancer CT image segmentation according to claim 7, characterized in that: The colorectal cancer tumor segmentation model of the SR-SE-3D UNet network uses the Adam optimizer, the initial learning rate is 0.001, the learning rate decay strategy uses the exponential decay strategy, the decay rate is set to 0.95, and the patience value in the early stopping mechanism is set to 10 Epoches. During the training process, if the loss value on the validation set does not decrease in 10 Epoches, the training is stopped and the weight parameters of 10 Epoches are saved.

9. A method for colorectal cancer CT image segmentation according to claim 8, characterized in that: After the training is completed, the optimal model weight parameters are saved, and the learning rate, learning rate reduction ratio, loss function hyperparameter initial value, and Batch Size are fine-tuned by combining the transfer learning method.

10. A 3D UNet network for automatic segmentation of colorectal cancer CT images by introducing separation weight parameters and compression excitation structure, characterized in that: The automatic segmentation system introduces a 3D UNet network with separated heavy parameters and compressed excitation structure, and the automatic segmentation system comprises the segmentation model optimized in step S5 established by any one of the methods of claims 1 to 9.

Images