An automatic delineation method for the radiotherapy target area of rectal cancer and related equipment

Abstract

The present invention provides an automatic delineation method for a rectal cancer radiotherapy target area and related equipment, which is applied to the field of data processing technology. This application obtains target user information, target target area label information, target area label information to be trained, a training data set matching the target area label information to be trained, and a preset target area automatic delineation model based on nnUNET; processes the training data set based on the target area label information to be trained to generate a preset delineation rule; processes the preset target area automatic delineation model based on nnUNET to generate a target target area automatic delineation model; processes the target user information to generate multi-modal CT image information of the target user and clinical prior knowledge information matching the target user; processes the multi-modal CT image information of the target user, the target target area label information and the clinical prior knowledge information matching the target user based on the target target area automatic delineation model to generate target target area delineation information.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to an automatic delineation method for rectal cancer radiotherapy target areas and related devices. Background Art

[0002] Radiotherapy is an important treatment method for rectal cancer. Tumor radiotherapy is a local treatment method that uses radiation to treat tumors. Approximately 70% of cancer patients need to undergo radiotherapy during the treatment of cancer, and about 40% of cancers can be cured by radiotherapy. The role and status of radiotherapy in tumor treatment have become increasingly prominent and have become one of the main means of treating malignant tumors.

[0003] The automatic delineation model for medical images of rectal cancer target areas based on deep learning can greatly shorten the time for doctors to modify the target areas during the clinical treatment implementation process. However, limited by the poor quality of CT imaging of rectal cancer target areas, the current automatic delineation models for rectal cancer target areas are often based on MRI. There are very few automatic delineation models for rectal cancer target areas based on CT images, and the delineation effect of the models is not ideal enough to meet the requirements of clinical use. In current clinical treatment of rectal cancer, the scenario of using CT-linac (radiation therapy linear accelerator guided by CT images) is quite common, and the delineation styles of rectal cancer target areas in each medical center are inconsistent. Therefore, there is an urgent need to develop an automatic multi-target delineation model for rectal cancer based on CT images to assist in precise treatment and accelerate the treatment process of patients.

[0004] It should be noted that the information disclosed in the above background art section is only used to enhance the understanding of the background of the present disclosure, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present application is to provide an automatic delineation method for rectal cancer radiotherapy target areas and related devices, which can at least overcome the problems existing in the prior art to a certain extent. By integrating and learning the user's previous images (multi-modal images such as plain CT and enhanced CT), the anatomical structure limitations of surrounding organs, and the clinical prior knowledge of this center, and fully learning the internal information of the custom tags (i.e., the target area contours that doctors need to train), the goal of greatly shortening the manual delineation time during the clinical radiotherapy implementation process is achieved. At the same time, the subjective differences caused by different experiences among doctors are reduced, and the customized needs of single-center automatic delineation are met.

[0006] Other features and advantages of the present application will become apparent through the following detailed description, or be learned in part through the practice of the present invention.

[0007] According to one aspect of the present application, there is provided an automatic delineation method for a rectal cancer radiotherapy target area, including: obtaining target user information, target target area label information, target area label information to be trained, a training data set matching the target area label information to be trained, and a preset target area automatic delineation model based on nnUNET, wherein the training data set includes multi-modal CT image information of other users and clinical prior knowledge information matching the other users, and the multi-modal CT images include plain CT images and enhanced CT images; processing the training data set matching the target area label information to be trained based on the target area label information to be trained to generate a preset delineation rule, wherein the preset delineation rule is used to characterize the delineation criteria for different plain CT images and different enhanced CT images and the anatomical structure limitation information of adjacent organs of the target area label; preprocessing the training data set matching the target area label information to be trained to generate a training data set with target feature data; processing the preset target area automatic delineation model based on nnUNET based on the training data set with target feature data and the preset delineation rule to generate a target area automatic delineation model; processing the target user information to generate multi-modal CT image information of the target user and clinical prior knowledge information matching the target user; processing the multi-modal CT image information of the target user, the target target area label information and the clinical prior knowledge information matching the target user based on the target area automatic delineation model to generate target area delineation information.

[0008] Another aspect of the present application is an automatic contouring device for rectal cancer radiotherapy target areas, characterized in that the device comprises: an acquisition module, configured to acquire target user information, target target area label information, target area label information to be trained, a training data set matching the target area label information to be trained, and a preset target area automatic contouring model based on nnUNET, wherein the training data set includes multi-modal CT image information of other users and clinical prior knowledge information matching the other users, and the multi-modal CT images include plain scan CT images and enhanced CT images; a processing module, configured to process the training data set matching the target area label information to be trained based on the target area label information to be trained to generate a preset contouring rule, wherein the preset contouring rule is used to represent the contouring criteria for different plain scan CT images and different enhanced CT images and the anatomical structure limitation information of adjacent organs of the target area label; preprocess the training data set matching the target area label information to be trained to generate a training data set with target feature data; process the preset target area automatic contouring model based on nnUNET based on the training data set with target feature data and the preset contouring rule to generate a target area automatic contouring model; process the target user information to generate multi-modal CT image information of the target user and clinical prior knowledge information matching the target user; and process the multi-modal CT image information of the target user, the target target area label information and the clinical prior knowledge information matching the target user based on the target area automatic contouring model to generate target area contouring information.

[0009] According to still another aspect of the present application, an electronic device is provided, characterized in that it comprises: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the automatic contouring method for rectal cancer radiotherapy target areas as described above by executing the executable instructions.

[0010] According to yet another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a second processor, the automatic contouring method for rectal cancer radiotherapy target areas as described above is implemented.

[0011] According to yet another aspect of the present application, a computer program product is provided, comprising a computer program, characterized in that when the computer program is executed by a third processor, the automatic contouring method for rectal cancer radiotherapy target areas as described above is implemented.

[0012] An automatic delineation method and related equipment for the radiotherapy target area of rectal cancer provided by this application obtain target user information, target area label information, target area label information to be trained, a training data set, and a preset target area automatic delineation model based on nnUNET from a server. The training data set includes multi-modal CT images and clinical prior knowledge of other users. Based on the target area label information to be trained, the training data set is processed to generate a preset delineation rule, including generating preset delineation feature points, then obtaining an inverse mapping and target interpolation through processing, and further obtaining initial delineation information, and finally generating a rule based on a preset loss function. Then, the training data set is preprocessed, and through steps such as feature extraction, data feature analysis, adjacent feature generation, sampling ratio determination, sampling feature generation, and data group construction, a training data set with target feature data is generated. Then, this data set and the preset delineation rule are used to train the preset target area automatic delineation model, including generating clinical prior knowledge and anatomical structure constraint information, dividing the training set and the validation set, extracting data groups and processing them to generate target data groups for training, and finally determining the target area automatic delineation model based on testing with the validation set.

[0013] For the target user, their information is processed to generate multi-modal CT image information and clinical prior knowledge information, including generating detection report information, generating clinical prior knowledge based on this, and extracting plain scan and enhanced CT images to generate multi-modal CT image information. Finally, based on the target area automatic delineation model, the target area label information is used to generate the area to be delineated information, the clinical prior knowledge information is used to generate delineation constraint information, the multi-modal CT image information is used to generate channel information, the initial target area delineation information is generated based on the area to be delineated information and the channel information, and then the target area delineation information is generated in combination with the delineation constraint information to assist in precise clinical radiotherapy.

[0014] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. Brief Description of the Drawings

[0015] Figure 1 The flowchart showing an automatic delineation method for the radiotherapy target area of rectal cancer provided by an embodiment of this application;

[0016] Figure 2 The structural schematic diagram showing an automatic delineation device for the radiotherapy target area of rectal cancer provided by an embodiment of this application;

[0017] Figure 3 The nnUNET network structural schematic diagram showing a target area automatic delineation model provided by an embodiment of this application. Detailed Description of the Embodiment

[0018] The preferred embodiments of the present invention will be described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not intended to limit the present invention.

[0019] The following will be described Figure 1 an automatic contouring method for a rectal cancer radiotherapy target area according to an exemplary embodiment of the present application. It should be noted that the following application scenarios are only shown for the convenience of understanding the spirit and principle of the present application, and the embodiments of the present application are not limited in this regard. On the contrary, the embodiments of the present application can be applied to any applicable scenario.

[0020] In one embodiment, the present application also proposes an automatic contouring method for a rectal cancer radiotherapy target area and related devices. Figure 1 A flowchart of an automatic contouring method for a rectal cancer radiotherapy target area according to an embodiment of the present application is schematically shown. As Figure 1 shown, this method is applied to a server and includes:

[0021] S101, obtaining target user information, target target area label information, target area label information to be trained, a training data set matching the target area label information to be trained, and a preset target area automatic contouring model based on nnUNET.

[0022] In one embodiment, the training data set includes multi-modal CT image information of other users and clinical prior knowledge information matching other users. The multi-modal CT images include plain scan CT images and enhanced CT images. Specifically, the target user information may include the user's basic identity information, such as name, age, gender, medical record number, etc., for uniquely identifying the user. For example, user Zhang San, male, 55 years old, with a medical record number of 20230801001. It may also include the user's clinical history information, such as whether there are other underlying diseases (such as hypertension, diabetes, etc.). These information may have a certain impact on the radiotherapy process and target area contouring. For example, the user has a history of hypertension and takes antihypertensive drugs for a long time. This may affect their physical reaction during radiotherapy, and thus needs to be considered when contouring the target area and formulating the radiotherapy plan.

[0023] For rectal cancer radiotherapy, the target area label information can be specific to the location of the tumor in the rectum (such as the upper, middle, or lower segment of the rectum), the size and shape of the tumor, etc. For example, the target area is a tumor in the middle segment of the rectum, with a size of approximately 3 cm × 2 cm and an irregular shape. It can also include a preliminary marking of the areas around the tumor that may be affected, such as the tissues within a certain range from the tumor margin. These areas also need to be focused on during radiotherapy to ensure as thorough a treatment of the tumor as possible and reduce the risk of recurrence. The target area label information to be trained can be annotation examples of different types of rectal cancer target areas collected from a large number of clinical cases. For example, for early rectal cancer, the target area label may focus on the primary tumor site and the tissues adjacent to it; for mid- and late-stage rectal cancer, the target area label may also include the areas of surrounding lymph node metastasis, etc. Taking a set of data to be trained as an example, it contains the target area annotations of rectal cancer cases with different stages, locations, and sizes. These diverse annotations will be used to train the model so that it can adapt to target area delineation in various situations.

[0024] The training dataset includes multi-modal CT image information of other users (i.e., previous users) and the clinical prior knowledge information matched with these users. In terms of multi-modal CT image information, the plain CT image can clearly show the anatomical structure of the rectum and surrounding tissues, while the enhanced CT images (venous phase and arterial phase) can better show the blood supply situation of the tumor, etc. For example, the plain CT of a user shows thickening of the rectal wall, and the enhanced CT venous phase image shows obvious enhancement at the tumor site, indicating rich blood supply. The clinical prior knowledge information can be the user's previous treatment situation (such as whether they have undergone surgery, chemotherapy, etc.), family medical history, etc. If there is a family history of colorectal cancer in the user's family, this may increase the genetic risk factors of cancer and can be used as a reference factor when training the model, enabling the model to learn the influence of different factors on the characteristics of the target area.

[0025] As Figure 3 shown, nnUNET is an advanced deep learning network structure, and its preset automatic target area delineation model plays a core role in this system. It has the ability to automatically extract medical image features. For example, it can accurately identify key information such as the contour of the rectum and the approximate location of the tumor from multi-modal CT images. The network structure of the model (such as in the document Figure 1as shown) includes multiple convolutional layers (such as 3x3×3.Conv, etc.), which can extract features from the image layer by layer. The downsampling and upsampling processes help capture information at different scales, and the skip connections can fuse feature information at different levels, so as to understand the image content more comprehensively. During the training process, the model continuously adjusts its internal parameters according to the input training data set (multi-modal CT images and related knowledge) to improve the accuracy of target area delineation. For example, by learning a large number of image data with accurate target area annotations, the model gradually learns to distinguish the characteristic differences between tumor tissues and normal tissues, so as to accurately delineate the target area on the new user images. In practical applications, when the multi-modal CT images and related information of the target user are input, the model will output the predicted target area delineation result based on the knowledge and patterns it has learned, providing reference for doctors and assisting doctors in formulating more accurate radiotherapy plans.

[0026] S102, process the training data set that matches the to-be-trained target area label information based on the to-be-trained target area label information to generate a preset delineation rule.

[0027] In one implementation, the to-be-trained target area label information is processed to generate preset delineation feature points, where the preset delineation feature points are used to represent the delineation feature points corresponding to the target area label. The preset delineation feature points are key feature points corresponding to the target area label. Taking the target area of rectal tumor as an example, possible feature points include the central position of the tumor, the most prominent points on the edge of the tumor, and the key points at the junction with adjacent organs. Suppose in a case, the tumor is located in the middle posterior wall of the rectum, and its central coordinates (in the CT image coordinate system) can be used as an important preset delineation feature point, which can roughly determine the center of the tumor position and help determine the overall range of the tumor subsequently. The most prominent point on the tumor edge close to the mesorectum is also a key feature point because this may be the key part where the tumor invades the surrounding tissues and needs special attention during delineation. The feature points at the adjacent boundary with the bladder are equally important, which define the boundary between the target area and the bladder and avoid misirradiating the bladder during radiotherapy. By processing these preset delineation feature points, the shape and position of the target area can be described more accurately, providing a basis for subsequent model training and delineation operations.

[0028] Process a preset contour feature point and a training data set that matches the target area label information to be trained based on a preset processing rule to generate an inverse mapping and a target interpolation. The inverse mapping and the target interpolation are obtained by processing based on the preset contour feature point and the training data set. For example, in the processing of the channel information of multimodal CT images, assume that the training data set contains plain CT and contrast-enhanced CT images of multiple users, and each image has different channel levels (such as different gray-level and other feature levels). Based on the preset contour feature point (such as the tumor center position mentioned above), find the matching feature point in the plain CT image and the corresponding matching feature point in the contrast-enhanced CT image. Then calculate the affine transformation matrix through these matching feature points, and further obtain the inverse mapping and the target interpolation. The inverse mapping can accurately correspond the feature points in different modality images in terms of spatial position. For example, map a certain feature point in the plain CT image to the corresponding position in the contrast-enhanced CT image to better integrate multimodal image information. The target interpolation is used to estimate the data value at an unknown position between known feature points. For example, between two adjacent tumor edge feature points, more continuous tumor boundary information can be obtained through the target interpolation, supplementing the details of the image data, enabling the model to more accurately understand the shape and scope of the target area. Especially for irregularly shaped tumor target areas, the target interpolation can provide a more accurate boundary description and improve the accuracy of contouring.

[0029] Process the training data set that matches the target area label information to be trained based on the inverse mapping and the target interpolation to obtain initial contour information, where the initial contour information includes the organ image located at the position of the target area label. The initial contour information obtained by processing the training data set based on the inverse mapping and the target interpolation includes the organ image located at the position of the target area label. Taking the rectal cancer target area as an example, the initial contour information will show the general images of the rectal tumor and its surrounding organs (such as male organs like the bladder and seminal vesicles or female organs like the uterus and vagina). In the initial contour image, it can be seen that the rectal tumor presents a certain shape (such as round, irregular, etc.), and its density shows different performances in the plain CT and contrast-enhanced CT images, and the contours of the surrounding organs are also clearly visible. For example, in male users, the relative position relationship between the seminal vesicles and the rectal tumor is reflected in the initial contour information, which helps doctors and the model to further judge whether the tumor has invaded the surrounding organs. The initial contour information provides a basic framework for subsequent accurate contouring. The model can gradually optimize and refine the target area contour according to these initial information, combined with more knowledge and algorithms, to make it more in line with the actual pathological conditions.

[0030] Process the initial delineation information based on a preset loss function to generate a preset delineation rule, which is used to represent the delineation criteria for different non-contrast CT images and contrast-enhanced CT images and the anatomical structure constraint information of adjacent organs of the target region label. The preset delineation rule covers the criteria for delineating the target region in different non-contrast CT images and contrast-enhanced CT images and the anatomical structure constraint information of adjacent organs of the target region label. For example, in non-contrast CT images, there is a certain range of delineation criteria for features such as the thickness and density of the rectal wall. The normal rectal wall may appear as a uniform medium-density shadow on non-contrast CT, with a thickness within a certain range. If local thickening or density change is found and meets certain numerical criteria (such as the thickness exceeding a certain threshold or the density being higher than normal tissue to a certain extent), it may indicate a lesion area that needs to be delineated. For contrast-enhanced CT images (arterial phase and venous phase), the enhancement characteristics of tumor tissues are different at different times. For example, it may rapidly enhance to a high density in the arterial phase, and the enhancement degree may change in the venous phase. According to these enhancement feature criteria, the range of the tumor target region is accurately delineated. At the same time, the anatomical structures of adjacent organs of the target region label, such as the bladder, prostate (in men) or uterus, vagina (in women), etc., limit the boundaries of target region delineation. For example, when delineating the target region of rectal tumors, it cannot exceed the boundaries of the normal tissues around the rectum, and adjacent organs should be avoided from being damaged to ensure the safety and effectiveness of radiotherapy. These rules are summarized based on a large amount of clinical experience and anatomical knowledge. By incorporating these rules into the training process, the model learns the correct delineation criteria and boundary constraints.

[0031] The preset loss function (such as a hybrid loss function, including the cross-entropy function and the Dice loss function) processes the initial delineation information to generate a preset delineation rule. The cross-entropy function mainly measures the classification difference between the model's prediction results and the true labels (such as the accurate target region delineation marked by doctors). In this scenario, it is to judge whether the model's classification of the target region and non-target region is accurate. For example, if the model misclassifies normal tissue as the target region or vice versa, the cross-entropy function will give a large loss value, prompting the model to adjust its parameters to improve classification accuracy. The Dice loss function focuses on the prediction errors of the model at the boundary pixels and evaluates the delineation accuracy of the target region boundary. For example, whether the delineation of the tumor edge is accurate, whether there is over-delineation or under-delineation. By continuously adjusting the model parameters to minimize the loss function value, an optimized preset delineation rule is obtained. These rules not only include the accurate definition of the target region itself but also consider various factors such as the anatomical relationship with adjacent organs, making the finally generated preset delineation rule more scientific and reasonable, and capable of accurately guiding the automatic target region delineation process in practical applications, improving the accuracy and safety of radiotherapy.

[0032] This method includes a calculation formula for obtaining a target interpolation, and the calculation formula is: ; where is the weight of the cross - entropy loss function, is the weight of the Dice loss function;

[0033] where CE represents the cross - entropy loss function, y represents the true label, and p represents the predicted probability of the model;

[0034] where Dice represents the loss function, y represents the true label, p represents the predicted probability of the model, is a very small constant to avoid the case of denominator being zero.

[0035] This formula is used to obtain the target interpolation and plays an important role in the automatic delineation model of the rectal cancer target area based on deep learning. To better understand the formula, we assume a simple example scenario: there is a small rectal cancer user image dataset containing plain - scan CT and contrast - enhanced CT images of 10 users, and we want to train a model to accurately delineate the tumor target area.

[0036] For the cross - entropy loss function , assuming that in this dataset, for a certain pixel point, its true label (indicating whether the pixel point belongs to the tumor target area, 1 means yes, 0 means no) is 1, and the probability p that the model predicts this pixel point belongs to the tumor target area is 0.8. Then according to the formula calculation:

[0037] First, calculate .

[0038] Then calculate .

[0039] So the cross - entropy loss value of this pixel point is 0.223. If such calculations are performed on each pixel point of the entire image and summed (i.e., the operation), the cross - entropy loss value of the entire image is obtained. This value reflects the degree of difference between the model prediction and the true label in classification. The smaller the value, the more accurate the model classification. During the training process, the model will continuously adjust the parameters according to this loss value to reduce the cross - entropy loss and improve the classification accuracy of the tumor target area and non - target area.

[0040] For the Dice loss function Assume that in a certain region (including multiple pixel points) in the image, the sum of the true labels y (pixel points belonging to the tumor target area are 1, otherwise 0) is (indicating that there are 50 pixel points belonging to the true tumor target area in this region), the sum of the model predicted probabilities is (the model predicts that there are 60 pixel points belonging to the tumor target area in this region), and (There are 45 pixel points where the true label is 1 and the model prediction probability is also high), let (a very small constant).

[0041] First, calculate the numerator part .

[0042] Then, calculate the denominator part .

[0043] Then the Dice coefficient of this region is 1 - (90.001 / 110.001) = 0.182. The closer the Dice coefficient is to 1, the higher the overlap between the target area predicted by the model and the true target area, and the more accurate the boundary delineation. During the training process, the model will adjust the parameters to increase the Dice coefficient and reduce the Dice loss, thereby optimizing the boundary delineation effect of the target area.

[0044] Standard interpolation calculation formula Example. Suppose during the model training process, we set the weight of the cross-entropy loss function , according to the cross-entropy loss value (assumed to be CE = 0.3) and Dice loss value (assumed to be Dice = 0.2) calculated previously.

[0045] First, calculate . Then calculate . So the target interpolation L = 0.15 + 0.1 = 0.25. The value of this target interpolation will be used in model training to adjust the model parameters to balance the influence of the cross-entropy loss and Dice loss on the model, so that the model can achieve better results in both classification accuracy and boundary delineation accuracy. In actual training, such calculations and parameter adjustments will be performed on each sample (image) of the entire dataset, and the model will be continuously iteratively optimized until the model performance reaches a satisfactory level, so as to accurately automatically delineate the target area of rectal cancer.

[0046] In another implementation, the training dataset matching the target area label information to be trained is processed to generate the channel information of the multi-modal CT image. Among them, the channel information of the multi-modal CT image contains the feature information of the same CT image at different channel levels. Based on the preset delineation feature points, the channel information of the multi-modal CT image is processed to generate the matching feature points of the plain CT image and the matching feature points of the enhanced CT image. The matching feature points of the plain CT image and the matching feature points of the enhanced CT image are processed to generate an affine transformation matrix, and the affine transformation matrix is processed to generate an inverse mapping and a target interpolation.

[0047] The method includes a calculation formula for calculating the target interpolation, and the calculation formula is:

[0048]

[0049]

[0050] Among them, are the values corresponding to adjacent feature points, respectively representing the weights for the x, y, and z directions.

[0051] Suppose we have a rectal cancer training dataset containing 10 users, and each user has plain CT and contrast-enhanced CT (venous phase, arterial phase) images. For the CT images of one of the users, the channel information of its multi-modal CT images can be understood as follows: The plain CT image itself can be regarded as one channel, which contains the basic anatomical structure information of the rectum and surrounding tissues when no contrast agent is injected, such as the thickness and density of the rectal wall and the morphology of surrounding organs. The contrast-enhanced CT venous phase image is another channel, which shows the distribution of the contrast agent in tissues during the venous phase. Tumor tissues may have specific enhancement patterns at this stage, which helps to more clearly identify the tumor range. The contrast-enhanced CT arterial phase image is the third channel, which can provide information on tumor blood supply in the early arterial phase. Combining with the venous phase image can help to more comprehensively understand the vascular characteristics of the tumor. The feature information at these different channel levels together constitutes the channel information of the multi-modal CT image, providing a rich data basis for subsequent processing and model training.

[0052] Based on the preset delineated feature points (such as the center point and edge points of the tumor, etc.), the images of the above-mentioned user are processed. Suppose the preset delineated feature point is the center point of the tumor. In the plain CT image, a point corresponding to this center point in terms of spatial position is found through an image recognition algorithm, and this point is the matching feature point of the plain CT image. Similarly, in the contrast-enhanced CT venous phase and arterial phase images, points corresponding to the preset delineated feature points are also found, that is, the matching feature points of the contrast-enhanced CT images. For example, in the plain CT image, the pixel coordinates corresponding to the tumor center point are, in the contrast-enhanced CT venous phase image are, and in the contrast-enhanced CT arterial phase image are. These matching feature points will be used for subsequent calculation of the affine transformation matrix to establish the spatial correspondence relationship between different modality images.

[0053] The preset delineation feature points are key feature points corresponding to the target area labels and are of great significance for the automatic delineation of rectal cancer tumors. For example, for the target area of rectal tumors, its central position can be used as an important preset delineation feature point, which can roughly determine the position center of the tumor in three-dimensional space and provide a basic positioning reference for subsequent delineation. In addition, the most prominent points on the tumor edge are also key feature points because this may be the most obvious part where the tumor invades surrounding tissues. Accurately identifying these points helps to determine the approximate scope of the tumor. There are also key points at the junctions with adjacent organs, such as the feature points at the adjacent boundaries with the bladder, prostate (in men) or uterus, vagina (in women), etc. These points define the boundaries between the target area and adjacent organs and are crucial for avoiding misirradiation of adjacent organs during radiotherapy.

[0054] First, in multi-modal CT images (plain CT and enhanced CT venous phase, arterial phase), based on these preset delineation feature points, through image recognition and analysis algorithms, find their corresponding positions in different modal images. For example, determine the coordinates of the tumor center position point in the plain CT image , and then find the coordinates of the points at the same anatomical position corresponding to it in the enhanced CT venous phase and arterial phase images and so on. Then, use these feature points matched in different modal images to calculate the affine transformation matrix. This matrix can describe the spatial transformation relationship between different modal images, including rotation, translation, scaling, etc. Through the affine transformation matrix, the feature points in different modal images can be accurately corresponded in spatial position. For example, accurately map a certain feature point in the plain CT image to the corresponding position in the enhanced CT image to achieve multi-modal images.

[0055] Based on the calculated affine transformation matrix and the target interpolation (calculated through a preset calculation formula using information such as adjacent feature points and weights), process the training data set to obtain the initial delineation information. The initial delineation information includes the organ images located at the positions of the target area labels. For example, a rough image of the rectal tumor and its surrounding organs (such as male organs like the bladder, seminal vesicles, etc. or female organs like the uterus, vagina, etc.) can be seen. In the initial delineation image, the shape of the rectal tumor (such as round, irregular, etc.) can be presented, its density has different manifestations in the plain CT and enhanced CT images, and the contours of the surrounding organs are also clearly visible.

[0056] Based on the initial delineation information, it is input into a deep learning model (such as a preset target region automatic delineation model based on nnUNET) for training. During the training process, the model processes the initial delineation information according to the preset loss functions (including cross-entropy function and Dice loss function, etc.), and continuously adjusts the model parameters. The cross-entropy function measures the classification difference between the model prediction result and the true label (such as the accurate target region delineation marked by a doctor), prompting the model to accurately distinguish the tumor target region and the non-target region. The Dice loss function focuses on the prediction errors at the boundary pixels and optimizes the delineation accuracy of the tumor edge. Through continuous iterative training to minimize the loss function value, the model gradually learns the accurate tumor delineation pattern, and thus can accurately perform automatic tumor delineation on new user images according to the preset delineation feature points and related processing procedures, output more accurate target region delineation results, and assist doctors in radiotherapy plan formulation.

[0057] Use the matching feature points of the plain CT image and the matching feature points of the enhanced CT image obtained previously to calculate the affine transformation matrix. Suppose we have selected three non-collinear matching feature points in the plain CT image and the corresponding three matching feature points in the enhanced CT image (taking the venous phase as an example). Through a series of mathematical calculations (involving matrix operations in linear algebra), an affine transformation matrix A can be obtained. This matrix can describe the transformation relationship from the plain CT image space to the enhanced CT image space (taking the venous phase as an example), including operations such as rotation, translation, and scaling. The same method can be used to calculate the affine transformation matrix between the plain CT image and the enhanced CT arterial phase image, and these matrices will be used for subsequent inverse mapping and target interpolation calculations to achieve accurate registration and information fusion between different modality images.

[0058] For the inverse mapping, assume that we have obtained the affine transformation matrix from the plain CT image to the enhanced CT venous phase image. If there is a point in the enhanced CT venous phase image, the corresponding point in the plain CT image can be obtained through inverse mapping calculation (using mathematical operations such as matrix inversion). This helps to accurately map the information in the enhanced CT image back to the plain CT image space and achieve the integration of multi-modal image information.

[0059] Starting from the formula of the affine transformation , we want to calculate the inverse mapping, that is, given find P.

[0060] First, rewrite the formula as , and then multiply both sides by on the left to get .

[0061] Calculate , for a 3x3 matrix its inverse matrix where is the determinant of A, is the cofactor of is the minor of (i.e., the determinant of the matrix after removing the i-th row and j-th column).

[0062] Finally, substitute into to obtain the calculation formula for the inverse mapping, that is, for a given point after affine transformation, its corresponding point P in the original space can be calculated. For example, for the x coordinate: Similarly, the calculation formulas for the x and z coordinates can be obtained. In actual calculations, according to specific image data and calculation accuracy requirements, matrix operations need to be optimized and numerically processed to ensure the accuracy and stability of the calculation results. At the same time, this formula can also be extended according to actual situations, such as considering more image dimensions or processing more transformation parameters, etc.

[0063] For target interpolation, assume we have two adjacent feature points and , and their corresponding values (such as CT values) are respectively . For the point to be interpolated, first calculate the direction weight , and similarly calculate the y-direction weight and the z-direction weight . Then calculate the interpolation result V according to the target interpolation calculation formula . For example, the CT value of is calculated to be Substitute it into the formula to calculate V = 53.2. Target interpolation can estimate the data values at unknown positions between known feature points, making the image data more continuous and complete, helping the model to more accurately understand and process image information, and improving the accuracy of target delineation.

[0064] S103. Preprocess the training data set that matches the target area label information to be trained to generate a training data set with target feature data.

[0065] In one implementation, feature extraction is performed on a training data set that matches the target region label information to be trained, generating a feature data set. Suppose our training data set contains multi-modal CT image data (plain CT and enhanced CT venous phase, arterial phase) of 50 rectal cancer patients and the corresponding target region label information. For the image data of each patient, we perform feature extraction through specific image processing algorithms. For example, the thickness feature and texture feature (such as gray level co-occurrence matrix feature, etc.) of the rectal wall are extracted from the plain CT image, and the enhancement feature of the tumor (such as enhancement degree, enhancement pattern, etc.) is extracted from the enhanced CT image. Combining these extracted features generates a feature data set, which contains rich information and can be used for subsequent model training and analysis.

[0066] Obtain any number of data features in the feature data set. In the generated feature data set, we randomly obtain 10 data features. These data features can be a certain feature value of different users. For example, one data feature may be the average thickness value of the rectal wall in the plain CT image of user A, and another data feature may be the maximum enhancement degree value of the tumor in the enhanced CT venous phase image of user B, etc. Obtaining these different data features is to further analyze the relationship and distribution between them.

[0067] Generate adjacent features based on the distance between any number of data features and other equal numbers of data features of the same category, where the adjacent features include a preset number of any number of data features. Calculate the distance between these 10 data features and other data features of the same category (such as the rectal wall thickness feature in the plain CT image of other users, etc.). The distance here can use a suitable distance metric method such as Euclidean distance. Suppose we set the preset number to 3. According to the distance calculation result, find the 3 other data features that are closest to each data feature, and these constitute the adjacent features. For example, for the data feature of the rectal wall thickness feature in the plain CT image of user A, find the rectal wall thickness features in the plain CT images of two other users (user C and user D), which are the closest to the feature of user A, and these three features constitute a group of adjacent features.

[0068] Determine the sampling ratio based on the number of data features in the training dataset, and determine the sampling rate based on the sampling ratio. Statistically calculate the total number of data features in the training set, assumed to be 200. Determine the sampling ratio according to a preset rule or algorithm. For example, if we want to select approximately 1 / 5 of the data features for subsequent operations, then the sampling ratio is 0.2. Based on this sampling ratio, calculate the sampling rate. If a simple equal-ratio sampling method is used, the sampling rate may be 0.2 (i.e., select 1 out of every 5 data features). However, in practical applications, more complex sampling strategies may be adopted according to the data distribution and specific requirements, such as stratified sampling, etc., and the calculation of the sampling rate will be different at this time.

[0069] Sample adjacent features based on the sampling rate to generate a preset number of sampled features. According to the calculated sampling rate (assumed to be 0.2), sample adjacent features. Since the preset number of adjacent features we set earlier is 3, select data features from each group of adjacent features according to the sampling rate. For example, for a certain group of adjacent features (including the rectal wall thickness features of user A, user C, and user D), sample at a sampling rate of 0.2, and maybe select 1 of the features (assume the feature of user A is selected). Perform such sampling operations on all groups of adjacent features, and finally generate a preset number (the number determined according to the sampling result here, which may be less than or equal to the initially set preset number of adjacent features 3, assume 2 sampled features are finally generated) of sampled features.

[0070] Based on any data feature and each sampling feature, multiple groups of data sets are generated, where each data set contains a preset number of data samples, and at least one data sample includes target feature data, and the target feature data is used to characterize that there is an abnormality in the image area matching the target area label information. For any one of the previously randomly obtained 10 data features (for example, the data feature of the maximum enhancement degree value of the tumor in the enhanced CT venous phase image of user B), it is combined with each sampling feature (assuming that the previously generated 2 sampling features are from a certain feature of user E and user F) to generate multiple groups of data sets. Each data set contains a preset number (assuming the preset number is 5) of data samples. Among these data sets, at least one data sample includes target feature data. For example, one of the data sets may be [the maximum enhancement degree value of the tumor of user B, the sampling feature value of user E, the sampling feature value of user F, the normal tissue feature value of user G, the feature value of user H containing the abnormal area (the feature value of user H here is the target feature data because the corresponding image area has an abnormality, such as the tumor invading the surrounding tissues, etc.)]. By generating multiple groups of data sets in this way, it can be used for subsequent model training or analysis, helping the model learn the relationships between data features and the abnormal feature patterns related to the target area label information, thereby improving the accuracy of automatic delineation of the rectal cancer target area.

[0071] S104. Process the preset target area automatic delineation model based on nnUNET according to the training data set with target feature data and the preset delineation rules to generate a target target area automatic delineation model.

[0072] In one embodiment, the preset delineation rules are processed to generate preset clinical prior knowledge information and preset anatomical structure restriction information, wherein the preset clinical prior knowledge information is used to characterize the clinical prior knowledge information of different users, and the preset anatomical structure restriction information is used to characterize the regional information of different organs. For the preset clinical prior knowledge information, assuming that in the treatment of rectal cancer, the clinical prior knowledge information of different users includes whether the user has a family history of cancer, whether other cancer-related treatments (such as chemotherapy, surgery, etc.) have been received before, and the user's basic physical condition (such as whether he suffers from other chronic diseases). For example, user A has a family history of rectal cancer and has undergone intestinal surgery before, and user B has no family history but suffers from diabetes. These different information constitute the preset clinical prior knowledge information, which can reflect the potential risks of different users during radiotherapy and possible responses to treatment. For the preset anatomical structure restriction information, taking the rectum and its surrounding organs as an example, the anatomical position relationship between the rectum and organs such as the bladder, prostate (male) or uterus, vagina (female) is important restriction information. Clarifying the position of the rectum in the pelvic cavity and its boundaries with surrounding organs can avoid mistakenly including normal organs in the target area or missing areas invaded by tumors when outlining the target area. For example, under normal circumstances, there is a certain tissue interval between the rectum and the bladder. When outlining the target area of ​​the rectal tumor, it is necessary to ensure that it does not exceed this interval to prevent damage to the bladder.

[0073] The training dataset with target feature data is divided into a training set for training the preset target area automatic delineation model and a validation set for testing the preset target area automatic delineation model based on a preset ratio. Suppose we have a training dataset with target feature data (such as tumor area features, surrounding tissue abnormality features, etc.) containing 100 users with rectal cancer. Divide it into a training set and a validation set according to a preset ratio (such as the common 80:20 ratio). Then the training set will contain data from 80 users, and the validation set will contain data from 20 users. The purpose of this division is to use most of the data (training set) to train the preset target area automatic delineation model so that it can learn the features and rules in the data, and then use a smaller but independent part of the data (validation set) to test the performance of the model and evaluate the accuracy and generalization ability of the model on unseen data.

[0074] Extract multiple groups of data sets from the training set. Each group of data sets contains a preset number of data samples, and at least one data sample includes target feature data. Extract data sets from the training set. Assume that each group of data sets contains 5 data samples, and at least one data sample includes target feature data. For example, for one of the groups of data sets, data sample 1 may be the texture features of the tumor area in the plain CT image of user C (this is the target feature data), data sample 2 is the density features of the tissue around the tumor in the enhanced CT image of user D, data sample 3 is the clinical prior knowledge information of user E (such as having received chemotherapy before), data sample 4 is the thickness features of the normal rectal wall of user F, and data sample 5 is the enhancement features of the tumor in the enhanced CT image of user G (this is also the target feature data). By extracting multiple groups of data sets in such a combined way, it can comprehensively reflect various aspects such as the imaging features and clinical information of different users in the training set, providing rich input data for subsequent model training.

[0075] Process the multiple groups of data sets based on the preset clinical prior knowledge information and the preset anatomical structure constraint information to generate target data sets. Process the multiple groups of data sets based on the previously generated preset clinical prior knowledge information (such as the medical history and treatment history of users) and the preset anatomical structure constraint information (such as the boundary relationship of organs). For example, if the preset clinical prior knowledge information indicates that a user has a history of intestinal surgery, it may affect the anatomical structure and tissue features of the intestine. When processing the data set, relevant data samples (such as imaging feature data) will be adjusted or weighted according to this information. For the preset anatomical structure constraint information, if the boundary relationship between the rectum and the bladder is known, when dealing with the imaging feature data in the area close to the bladder in the data set, the data will be processed according to this boundary constraint to ensure that the data conforms to the actual anatomical situation. After such processing, target data sets that are more in line with the actual clinical and anatomical requirements are generated, enabling the model to learn more accurate and useful information.

[0076] Train a preset target region automatic delineation model based on a target data set to generate a trained target region automatic delineation model. Process the trained target region automatic delineation model based on a validation set to generate test results. If the data samples containing target feature data in the test results indicate that there are abnormalities in the image regions matching the target region label information, then use the trained target region automatic delineation model as the target target region automatic delineation model. Use the target data set to train a preset target region automatic delineation model (such as a model based on nnUNET). During the training process, the model will learn the feature patterns in the data set and adjust its internal parameters to improve the accuracy of target region delineation. After multiple iterative trainings, a trained target region automatic delineation model is obtained. Then, input the data in the validation set into the trained model for processing to generate test results. Suppose there is the imaging data of user Xin in the validation set. After the model makes a target region delineation prediction on it, the test results may be a delineation contour and related probability values and other information. If in this test result, it is found that there are abnormalities in the image regions matching the target region label information (such as inaccurate tumor delineation range, omission of some tumor invasion regions, etc.), then the reasons need to be further analyzed. It may be that the model training is insufficient, the processing of the data set is not accurate enough, etc. If there are no such problems in the test results and the model performs well in other evaluation metrics (such as Dice coefficient, Hausdorff distance, etc.), then this trained target region automatic delineation model can be used as the target target region automatic delineation model for the actual automatic delineation task of rectal cancer radiotherapy target regions to assist doctors in performing more accurate radiotherapy treatment.

[0077] S105. Process the target user information to generate multi-modal CT image information of the target user and clinical prior knowledge information matching the target user.

[0078] In one implementation, the target user information is processed to generate the detection report information of the user. The detection report information includes the region where the primary focus of the user's rectal cancer is located, the covered region of the primary focus, the target area delineation influence parameters of the detection center, and the physical examination information of the user. Assume the target user is user Li. For the region where the primary focus of the user's rectal cancer is located, after a series of examinations (such as colonoscopy, digital rectal examination, imaging examination, etc.), it is determined that the primary focus is located in the middle segment of the rectum. The covered region of the primary focus is evaluated through imaging examinations such as CT and MRI. Assume that it is found that the tumor involves a part of the rectal wall but does not penetrate completely, and its covered region is mainly concentrated on a certain side wall and part of the posterior wall in the middle segment of the rectum. Regarding the target area delineation influence parameters of the detection center, when performing radiotherapy, this detection center is accustomed to using the dose irradiation range as a reference for the target area delineation range, and according to past experience, for tumors located in the middle segment of the rectum, a certain range of the delineated area will be appropriately enlarged to ensure sufficient irradiation dose. The user's physical examination information includes basic indicators such as height, weight, blood routine, liver and kidney functions, etc. For example, the height is 170 cm, the weight is 65 kg, the indicators in the blood routine are basically normal, and there are no obvious abnormalities in the liver and kidney functions. User Li did not undergo surgical treatment before radiotherapy but received one course of chemotherapy, and all these information will be incorporated into the detection report information.

[0079] Based on the user's detection report information, clinical prior knowledge information matching the target user is generated. Based on the above detection report information, clinical prior knowledge information matching Li is generated. Since the user's primary focus is in the middle segment of the rectum and no surgery has been performed, this means that the shape and position of the tumor may be relatively complex, and more accurate target area localization is required during radiotherapy to avoid damaging surrounding normal tissues. The user has received chemotherapy, which may have a certain impact on the morphology, blood supply, etc. of the tumor tissue, and the changes after chemotherapy need to be considered during target area delineation. The habit of this detection center using the dose irradiation range as a reference will also affect the formation of clinical prior knowledge. In subsequent model processing, the strategy for target area delineation of Li will be adjusted according to this characteristic to adapt to the treatment specifications of the detection center.

[0080] Perform image extraction processing on the target user information to generate non-contrast CT images, contrast-enhanced CT venous phase images, and contrast-enhanced CT arterial phase images of the target user. Perform image extraction processing on the target user information of Li Mou. When performing non-contrast CT examination, a series of non-contrast CT images are obtained, which clearly show the anatomical structures of the rectum and surrounding tissues, such as the thickness of the rectal wall and the distribution of surrounding adipose tissue. In the contrast-enhanced CT venous phase examination, after injecting contrast agent, scanning is performed at specific time points to obtain contrast-enhanced CT venous phase images. At this time, it can be seen that the tumor tissue has different degrees of enhancement compared with the surrounding normal tissues, which helps to more accurately identify the scope and boundary of the tumor. The contrast-enhanced CT arterial phase images can reflect the blood supply of the tumor in the early arterial phase, which is helpful for judging the activity and invasiveness of the tumor. For example, in the contrast-enhanced CT arterial phase images, it is found that some areas of the tumor have obvious early enhancement, indicating that the blood supply in this area is rich and may be the area where the tumor grows actively.

[0081] Generate multi-modal CT imaging information of the target user based on the non-contrast CT images, contrast-enhanced CT venous phase images, and contrast-enhanced CT arterial phase images of the target user. Combine the non-contrast CT images, contrast-enhanced CT venous phase images, and contrast-enhanced CT arterial phase images of Li Mou to generate his multi-modal CT imaging information. In this multi-modal CT imaging information, doctors and models can comprehensively analyze the characteristics of different modal images. For example, the non-contrast CT images provide a basic anatomical structure framework, the contrast-enhanced CT venous phase images further clarify the approximate scope of the tumor (judged by the enhanced area), and the contrast-enhanced CT arterial phase images supplement the details of the tumor blood supply. Through the comprehensive interpretation of this multi-modal CT imaging information, a more comprehensive understanding of Li Mou's rectal tumor situation can be obtained, providing a more accurate basis for subsequent automatic target volume delineation, radiotherapy plan formulation, etc. For example, in the multi-modal CT imaging information, it can be seen that the tumor appears as an isodense shadow on non-contrast CT, has uneven enhancement in the contrast-enhanced CT venous phase, has obvious early enhancement in the contrast-enhanced CT arterial phase, and the relationship with surrounding tissues (such as the relative positions with the bladder, seminal vesicles, etc.) can be clearly presented in different modal images, and these information are crucial for accurately delineating the target volume.

[0082] S106, based on the target target volume automatic delineation model, process the multi-modal CT imaging information, the target target volume label information, and the clinical prior knowledge information matching the target user to generate target target volume delineation information.

[0083] In one implementation, the target region of interest automatic delineation model processes the target region of interest label information to generate the region information to be delineated. Suppose the target region of interest label information specifies that the rectal tumor and its potentially invaded regions are to be delineated. The target region of interest automatic delineation model will process this label information according to the knowledge and algorithms it has learned. For example, by analyzing the imaging features corresponding to similar target region labels in a large number of previous cases, the model determines the approximate range of the region where a tumor may exist in the new target user's image. If the model has learned that a region with a certain specific density and morphological change in the middle rectum may be the location of the tumor, then in the target user's image, once a region with similar features is found, it is marked as the region information to be delineated. For instance, in the plain CT image of the target user, a local region in the middle rectum is found to have a slightly higher density than the surrounding normal tissues and an irregular shape. The model initially determines this region as part of the region to be delineated. At the same time, by combining information such as the enhancement features of this region in the venous and arterial phases of the contrast-enhanced CT images, the region information to be delineated is further refined to more accurately cover the tumor and its potentially invaded range.

[0084] The target region of interest automatic delineation model processes the clinical prior knowledge information matching the target user to generate the delineation restriction information. For the clinical prior knowledge information matching the target user, such as the user has received chemotherapy before, the model will generate the delineation restriction information based on this information. Since chemotherapy may cause changes such as tumor tissue necrosis and fibrosis, affecting its appearance in the image, the model will consider the boundary blurring and other situations that may be caused by these changes during delineation, adjust the delineation strategy, and avoid over-delineation or under-delineation. If the clinical prior knowledge information of the user shows that the primary focus is located in the upper rectum and close to the bladder, the model will generate the corresponding delineation restriction information according to the anatomical relationship between the bladder and the upper rectum and the target region delineation habit of this detection center (such as avoiding damaging important organs such as the bladder). For example, it is stipulated that when delineating the target region, the boundary close to the bladder side should be delineated strictly according to a certain safety distance and cannot exceed the boundary between the normal tissue and the bladder to ensure the safety of the bladder during radiotherapy.

[0085] Process the multi-modal CT image information of the target user based on the automatic target region delineation model to generate the channel information of the multi-modal CT image of the target user. The automatic target region delineation model processes the multi-modal CT image information of the target user (including non-contrast CT, enhanced CT venous phase and arterial phase images). The non-contrast CT image itself serves as a channel, and the information it contains, such as the thickness and density of the rectal wall and the basic anatomical structure of the surrounding tissues, constitutes the characteristic information of this channel. The enhanced CT venous phase image serves as another channel, and the distribution of the contrast agent in the tissue during the venous phase shown in it, such as the enhancement degree and enhancement pattern of the tumor tissue, becomes the unique characteristic information of this channel. The enhanced CT arterial phase image provides information on the tumor blood supply in the early arterial stage, such as the early enhancement characteristics of the tumor and the display of the blood supply vessels, forming the characteristic information of the third channel. The combination of the characteristic information of these different channels generates the channel information of the multi-modal CT image of the target user, and the model can comprehensively understand the situation of the tumor and the surrounding tissues by analyzing this channel information.

[0086] Process the channel information of the multi-modal CT image of the target user based on the region to be delineated information to generate the initial target region delineation information. For example, within the region to be delineated, the model will focus on the characteristic manifestations of this region in different channel information. If the region to be delineated shows an area of abnormal density in the non-contrast CT channel and presents a specific enhancement pattern in the enhanced CT venous phase channel, the model will preliminarily delineate a rough target region contour on the image based on these characteristics to form the initial target region delineation information. This initial delineation may be a relatively rough contour that includes the main suspicious parts within the region to be delineated, but may still need further optimization. For example, in the initial target region delineation information, the tumor and the surrounding tissues within a certain range may be included, but the boundary may not be precise enough and needs to be adjusted according to more information later.

[0087] Process the initial target region delineation information based on the delineation restriction information to generate the target target region delineation information. For example, according to the aforementioned delineation restriction information to avoid damaging the bladder, precisely adjust the boundary of the initial target region delineation information near the bladder side to ensure that it does not exceed the safe range. If the clinical prior knowledge indicates that the boundary of the tumor may shrink after chemotherapy, then on the basis of the initial target region delineation information, appropriately narrow the target region range to remove the false positive regions that may be caused by chemotherapy. After such a series of processes, finally generate the accurate target target region delineation information, which will provide the doctor with an accurate target region range for formulating radiotherapy plans, improving the accuracy and effectiveness of radiotherapy, and minimizing the damage to the surrounding normal tissues to the greatest extent.

[0088] In this application, the server obtains target user information, target target area label information, target area label information to be trained, a training data set, and a preset target area automatic delineation model based on nnUNET. The training data set includes multi-modal CT images of other users and clinical prior knowledge. Based on the target area label information to be trained, the training data set is processed to generate a preset delineation rule, including generating preset delineation feature points, then obtaining an inverse mapping and target interpolation through processing, and further obtaining initial delineation information. Finally, a rule is generated based on a preset loss function. Then, the training data set is preprocessed, and through steps such as feature extraction, data feature analysis, adjacent feature generation, sampling ratio determination, sampling feature generation, and data group construction, a training data set with target feature data is generated. Then, this data set and the preset delineation rule are used to train the preset target area automatic delineation model, including generating clinical prior knowledge and anatomical structure constraint information, dividing the training set and the validation set, extracting data groups and processing them to generate target data groups for training, and finally determining the target area automatic delineation model based on testing with the validation set.

[0089] For the target user, their information is processed to generate multi-modal CT image information and clinical prior knowledge information, including generating detection report information, generating clinical prior knowledge based on this, and extracting plain scan and enhanced CT images to generate multi-modal CT image information. Finally, based on the target area automatic delineation model, the target area label information is used to generate information on the area to be delineated, the clinical prior knowledge information is used to generate delineation constraint information, the multi-modal CT image information is used to generate channel information, the initial target area delineation information is generated based on the information on the area to be delineated and the channel information, and then the target area delineation information is generated by combining the delineation constraint information to assist in precise clinical radiotherapy.

[0090] In one implementation, as Figure 2 shown, this application also provides an automatic delineation device for the radiotherapy target area of rectal cancer, including:

[0091] An acquisition module 201, configured to acquire target user information, target target area label information, target area label information to be trained, a training data set matching the target area label information to be trained, and a preset target area automatic delineation model based on nnUNET, wherein the training data set includes multi-modal CT image information of other users and clinical prior knowledge information matching other users, and the multi-modal CT image includes a plain scan CT image and an enhanced CT image;

[0092] The processing module 202 is used to process the training data set matching the target area label information to be trained based on the target area label information to be trained, and generate a preset delineation rule, wherein the preset delineation rule is used to characterize the delineation standards of different plain scan CT images, different enhanced CT images and the anatomical structure restriction information of the target area label adjacent organs; pre-process the training data set matching the target area label information to be trained to generate a training data set with target feature data; process the preset target area automatic delineation model based on nnUNET based on the training data set with target feature data and the preset delineation rule to generate a target target area automatic delineation model; process the target user information to generate the target user's multimodal CT image information and the clinical prior knowledge information matching the target user; process the target user's multimodal CT image information, the target target area label information and the clinical prior knowledge information matching the target user based on the target target area automatic delineation model to generate the target target area delineation information.

[0093] The technical solution of the embodiment of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling an electronic device (which may be an air conditioner, a refrigeration device, a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method described in the embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.

[0094] The computer program product provided in the above-mentioned embodiments of the present application and the method for automatically delineating the target area for radiotherapy of rectal cancer provided in the embodiments of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application programs stored therein.

[0095] Each embodiment in the present application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the automatic delineation method, electronic device, electronic device, and readable storage medium embodiment for evaluating the target area for radiotherapy of rectal cancer, since it is basically similar to the above-mentioned automatic delineation method embodiment for radiotherapy of rectal cancer, the description is relatively simple, and the relevant parts can be referred to the partial description of the above-mentioned automatic delineation method embodiment for radiotherapy of rectal cancer.

Claims

1. A method for automatically delineating a target area for radiotherapy of rectal cancer, characterized in that: include: Obtaining target user information, target target area label information, target area label information to be trained, a training data set matching the target area label information to be trained, and a preset target area automatic delineation model based on nnUNET, wherein the training data set includes multimodal CT image information of other users and clinical prior knowledge information matching with other users, and the multimodal CT images include plain scan CT images and enhanced CT images; Based on the target area label information to be trained, a training data set matching the target area label information to be trained is processed to generate a preset delineation rule, wherein the preset delineation rule is used to characterize delineation standards of different plain scan CT images and different enhanced CT images and anatomical structure restriction information of organs adjacent to the target area label; Preprocessing a training data set that matches the target area label information to be trained to generate a training data set with target feature data; Processing the preset target area automatic delineation model based on nnUNET based on the training data set with target feature data and the preset delineation rule to generate a target area automatic delineation model; Processing the target user information to generate multimodal CT image information of the target user and clinical prior knowledge information matching the target user; Processing the multimodal CT image information of the target user, the target area label information, and the clinical prior knowledge information matching the target user based on the target area automatic delineation model to generate target area delineation information; Based on the target area label information to be trained, a training data set matching the target area label information to be trained is processed to generate a preset delineation rule, including: processing the target area label information to be trained to generate preset delineation feature points, wherein the preset delineation feature points are used to characterize the delineation feature points corresponding to the target area label; processing the preset delineation feature points and the training data set matching the target area label information to be trained based on the preset processing rules to generate an inverse mapping and a target interpolation; processing the training data set matching the target area label information to be trained based on the inverse mapping and the first target interpolation to generate initial delineation information, wherein the initial delineation information includes an organ image located at the position of the target area label; processing the initial delineation information based on a preset loss function to generate a preset delineation rule; wherein the calculation formula of the first target interpolation L is: ;in, is the weight of the cross entropy loss function, is the weight of the Dice loss function; ; Where CE represents the cross entropy loss function, y represents the true label, and p represents the predicted probability of the model; ; Where Dice represents the loss function, y represents the true label, and p represents the predicted probability of the model. is a very small constant to avoid the situation where the denominator is zero; The preset delineation feature points and the training data set matching the target area label information to be trained are processed based on preset processing rules to generate an inverse mapping and a target interpolation, and also include: processing the training data set matching the target area label information to be trained to generate channel information of a multimodal CT image, wherein the channel information of the multimodal CT image includes feature information of the same CT image at different channel levels; processing the channel information of the multimodal CT image based on the preset delineation feature points to generate plain scan CT image matching feature points and enhanced CT image matching feature points; processing the plain scan CT image matching feature points and the enhanced CT image matching feature points to generate an affine transformation matrix; processing the affine transformation matrix to generate an inverse mapping and a second target interpolation; wherein the calculation formula of the second target interpolation V is: ; (Assumption ); (Assumption ); (Assumption );in, and is the value corresponding to the adjacent feature points, , , Represent the weights for the x, y, and z directions respectively.

2. The method according to claim 1, preprocessing the training data set matching the target area label information to be trained to generate a training data set with target feature data, comprising: Extracting features from a training data set that matches the target area label information to be trained to generate a feature data set; Obtaining any number of data features in the feature data set; Generate adjacent features based on the distance between the any number of data features and other numbers of data features of the same category, wherein the adjacent features include a preset number of the any number of data features; Determine the sampling ratio based on the number of each data feature in the training data set; Based on the sampling ratio, determining a sampling ratio; Sampling the adjacent features based on the sampling ratio to generate a preset number of sampling features; Based on any data feature and each sampling feature, multiple data groups are generated, wherein each data group includes a preset number of data samples, and at least one data sample includes target feature data, and the target feature data is used to characterize the presence of an abnormality in an image area that matches the target area label information.

3. The method according to claim 2, characterized in that The preset target area automatic delineation model based on nnUNET is processed based on the training data set with target feature data and the preset delineation rule to generate the target area automatic delineation model, including: Processing the preset delineation rules to generate preset clinical prior knowledge information and preset anatomical structure restriction information, wherein the preset clinical prior knowledge information is used to characterize clinical prior knowledge information of different users, and the preset anatomical structure restriction information is used to characterize regional information of different organs; Dividing the training data set with the target feature data into a training set for training the preset target area automatic delineation model and a validation set for testing the preset target area automatic delineation model based on a preset ratio; Extracting multiple data groups from the training set, wherein each data group includes a preset number of data samples, and at least one data sample includes target feature data; Processing multiple data sets based on the preset clinical prior knowledge information and the preset anatomical structure restriction information to generate a target data set; Training the preset target area automatic delineation model based on the target data group to generate a trained target area automatic delineation model; Processing the trained target area automatic delineation model based on the validation set to generate a test result; If the data sample containing the target feature data in the test result indicates that an abnormality exists in the image region matching the target region label information, the trained target region automatic delineation model is used as the target region automatic delineation model.

4. The method according to claim 3, characterized in that The target user information is processed to generate multimodal CT image information of the target user and clinical prior knowledge information matching the target user, including: Processing the target user information to generate the user's test report information, wherein the test report information includes the area where the user's rectal cancer primary lesion is located, the coverage area of ​​the primary lesion, the target area delineation influencing parameters of the test center, and the user's physical examination information; Generate clinical prior knowledge information matching the target user based on the test report information of the user; Performing image extraction processing on the target user information to generate a plain scan CT image, an enhanced CT venous phase image, and an enhanced CT arterial phase image of the target user; Multimodal CT image information of the target user is generated based on the plain scan CT image of the target user, the enhanced CT venous phase image and the enhanced CT arterial phase image.

5. The method according to claim 4, processing the multimodal CT image information of the target user, the target area label information and the clinical prior knowledge information matching the target user based on the target area automatic delineation model to generate target area delineation information, comprising: Processing the target area label information based on the target area automatic delineation model to generate information of the area to be delineated; Processing the clinical prior knowledge information matching the target user based on the target area automatic delineation model to generate delineation restriction information; Processing the multimodal CT image information of the target user based on the target area automatic delineation model to generate channel information of the multimodal CT image of the target user; Processing the channel information of the multimodal CT image of the target user based on the information of the area to be delineated to generate initial target area delineation information; The initial target area delineation information is processed based on the delineation restriction information to generate target area delineation information.

6. An automatic delineation device for rectal cancer radiotherapy target area, characterized in that: The device comprises: An acquisition module, used to acquire target user information, target area label information, target area label information to be trained, a training data set matching the target area label information to be trained, and a preset target area automatic delineation model based on nnUNET, wherein the training data set includes multimodal CT image information of other users and clinical prior knowledge information matching with other users, and the multimodal CT images include plain scan CT images and enhanced CT images; A processing module, used to process a training data set matching the target area label information to be trained based on the target area label information to be trained, and generate a preset delineation rule, wherein the preset delineation rule is used to characterize the delineation standards of different plain scan CT images and different enhanced CT images and the anatomical structure restriction information of organs adjacent to the target area label; preprocess the training data set matching the target area label information to be trained, and generate a training data set with target feature data; based on the training data set with target feature data and the preset delineation rule, process the preset target area automatic delineation model based on nnUNET, and generate a target target area automatic delineation model; process the target user information, and generate multimodal CT image information of the target user and clinical prior knowledge information matching the target user; based on the target target area automatic delineation model, process the multimodal CT image information of the target user, the target target area label information and the clinical prior knowledge information matching the target user, and generate target target area delineation information; Based on the target area label information to be trained, a training data set matching the target area label information to be trained is processed to generate a preset delineation rule, including: processing the target area label information to be trained to generate preset delineation feature points, wherein the preset delineation feature points are used to characterize the delineation feature points corresponding to the target area label; processing the preset delineation feature points and the training data set matching the target area label information to be trained based on the preset processing rules to generate an inverse mapping and a target interpolation; processing the training data set matching the target area label information to be trained based on the inverse mapping and the first target interpolation to generate initial delineation information, wherein the initial delineation information includes an organ image located at the position of the target area label; processing the initial delineation information based on a preset loss function to generate a preset delineation rule; wherein the calculation formula of the first target interpolation L is: ;in, is the weight of the cross entropy loss function, is the weight of the Dice loss function; ; Where CE represents the cross entropy loss function, y represents the true label, and p represents the predicted probability of the model; ; Where Dice represents the loss function, y represents the true label, and p represents the predicted probability of the model. is a very small constant to avoid the situation where the denominator is zero; The preset delineation feature points and the training data set matching the target area label information to be trained are processed based on preset processing rules to generate an inverse mapping and a target interpolation, and also include: processing the training data set matching the target area label information to be trained to generate channel information of a multimodal CT image, wherein the channel information of the multimodal CT image includes feature information of the same CT image at different channel levels; processing the channel information of the multimodal CT image based on the preset delineation feature points to generate plain scan CT image matching feature points and enhanced CT image matching feature points; processing the plain scan CT image matching feature points and the enhanced CT image matching feature points to generate an affine transformation matrix; processing the affine transformation matrix to generate an inverse mapping and a second target interpolation; wherein the calculation formula of the second target interpolation V is: ; (Assumption ); (Assumption ); (Assumption );in, and is the value corresponding to the adjacent feature points, , , Represent the weights for the x, y, and z directions respectively.

7. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; Wherein, the first processor is configured to execute the method for automatically delineating a target area for radiotherapy of rectal cancer as described in any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the method for automatically delineating a target area for radiotherapy of rectal cancer as claimed in any one of claims 1 to 5 is implemented.

Images