A method and system for differentiating pancreatic cysts based on semi-supervised learning

Abstract

The present invention provides a method and system for differentiating pancreatic cysts based on semi-supervised learning, which solves the technical problem of difficult differentiation of pancreatic cysts in the prior art. The method includes the following steps: training a semi-supervised learning model with a training set with a small number of artificial labels to obtain a trained semi-supervised learning model; using the trained semi-supervised learning model to extract a pancreatic contour probability map of multi-phase pancreatic images; using the trained semi-supervised learning model to extract a pancreatic cyst contour probability map of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map; extracting the pancreatic cyst area of the pancreatic cyst contour probability map, and differentiating the benign and malignant of the pancreatic cysts according to the pancreatic cyst area and combining clinical knowledge-driven features. By adopting a method of step-by-step segmentation of the pancreas and pancreatic cysts, and using the pancreatic contour probability map obtained by pancreatic segmentation to guide the segmentation of pancreatic cysts, more accurate segmentation of pancreatic cysts is achieved, thereby improving the accuracy of differentiating pancreatic cysts.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and particularly relates to a method and system for differentiating pancreatic cysts based on semi-supervised learning. Background Art

[0002] Driven by deep learning algorithms, the field of computer-aided diagnosis (CAD) has developed rapidly, especially in the field of medical image segmentation. However, due to the limitations of existing CT imaging, single-phase CT scans often make it difficult to accurately locate the contours of organs, and it is even more complex to determine the contours of lesions. Since different phases can enhance different details, referring to different phases is an effective strategy to identify the boundaries of organs or lesions as completely as possible. In recent years, omics features driven by clinical knowledge can be well combined with the features extracted by deep learning to further improve the performance of disease diagnosis and are widely used in many lesion studies. Different from cysts in other organs, pancreatic cysts often cannot be accurately examined by needle biopsy because the pancreas is located deep in the abdomen and is adjacent to various organs and blood vessels. And there is a lack of reliable imaging techniques in the prior art to provide important evidence for differentiating the benign and malignant of pancreatic cysts. Summary of the Invention

[0003] In view of the above problems, embodiments of the present invention provide a method and system for differentiating pancreatic cysts based on semi-supervised learning to solve the technical problem of difficult differentiation of pancreatic cysts in the prior art.

[0004] In a first aspect, a method for differentiating pancreatic cysts based on semi-supervised learning according to an embodiment of the present invention includes the following steps:

[0005] Training a semi-supervised learning model with a training set with a small number of artificial labels to obtain a trained semi-supervised learning model;

[0006] Using the trained semi-supervised learning model to extract a pancreatic contour probability map of multi-phase pancreatic images;

[0007] Using the trained semi-supervised learning model to extract a pancreatic cyst contour probability map of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map;

[0008] Extracting the pancreatic cyst area of the pancreatic cyst contour probability map, and differentiating the benign and malignant of the pancreatic cyst according to the pancreatic cyst area and combining clinical knowledge-driven features.

[0009] In an embodiment of the present invention, it further includes:

[0010] Using the trained semi-supervised learning model to extract a pancreatic contour deformation field of multi-phase pancreatic images, discriminating the approximation degree between the pancreatic contour deformation field and the original pancreatic contour deformation field to obtain a pancreatic morphology score;

[0011] Using the trained semi-supervised learning model, extract the pancreatic cyst contour deformation field of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map, and discriminate the approximation degree between the pancreatic cyst contour deformation field and the original pancreatic cyst contour deformation field to obtain the pancreatic cyst morphology score;

[0012] Sort the pancreatic morphology scores and pancreatic cyst morphology scores from high to low, screen out a group of data with high scores in the training set to generate data with machine labels, and add the data with machine labels to the training set to iteratively train the semi-supervised learning model until all data in the training set have machine labels.

[0013] In one embodiment of the present invention, the pancreatic contour deformation field and the pancreatic contour probability map are extracted by the first segmentation network of the semi-supervised learning model; the pancreatic cyst contour deformation field and the pancreatic cyst contour probability map are extracted by the second segmentation network of the semi-supervised learning model, and the first segmentation network and the second segmentation network form a cascaded segmentation network.

[0014] In one embodiment of the present invention, the pancreatic morphology score is obtained through the first morphology discrimination network of the semi-supervised learning model; the pancreatic cyst morphology score is obtained through the second morphology discrimination network of the semi-supervised learning model; the scores for sorting and screening are the sum of the pancreatic morphology scores and the pancreatic cyst scores corresponding to each data in the training set.

[0015] In one embodiment of the present invention, extract the region in the pancreatic cyst contour probability map with a threshold greater than 0.5 to obtain the pancreatic cyst region.

[0016] In one embodiment of the present invention, the method for differentiating the benign and malignant of pancreatic cysts based on the pancreatic cyst region and combined with clinical knowledge-driven features includes:

[0017] Divide the pancreatic cyst region into several image patches, and map the information contained in each image patch to the corresponding sequence representation vector patch embedding;

[0018] Record the position representation vector position embedding corresponding to each sequence representation vector;

[0019] Input the sequence representation vector patch embedding and the position representation vector position embedding into the clinically driven Transformer and combine the clinical knowledge-driven features to differentiate the benign and malignant of pancreatic cysts.

[0020] In one embodiment of the present invention, the clinical knowledge-driven features include: the maximum diameter of the cyst, the presence or absence of mural nodules, the presence or absence of solid components, and whether the pancreatic duct is dilated.

[0021] Second aspect, a method for differentiating pancreatic cysts based on semi-supervised learning according to an embodiment of the present invention includes:

[0022] A model training module: configured to train a semi-supervised learning model through a training set with a small number of manual labels to obtain a trained semi-supervised learning model;

[0023] A first segmentation module: configured to use the trained semi-supervised learning model to extract a pancreatic contour probability map of multi-phase pancreatic images;

[0024] A second segmentation module: configured to use the trained semi-supervised learning model to extract a pancreatic cyst contour probability map of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map;

[0025] A differentiation module: configured to extract the pancreatic cyst area of the pancreatic cyst contour probability map, and differentiate the benign and malignant nature of the pancreatic cyst according to the pancreatic cyst area and combined with clinically knowledge-driven features.

[0026] In an embodiment of the present invention, it further includes:

[0027] A first discrimination module: configured to use the trained semi-supervised learning model to extract a pancreatic contour deformation field of multi-phase pancreatic images, and discriminate the approximation degree between the pancreatic contour deformation field and the original pancreatic contour deformation field to obtain a pancreatic morphology score;

[0028] A second discrimination module: configured to use the trained semi-supervised learning model to extract a pancreatic cyst contour deformation field of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map, and discriminate the approximation degree between the pancreatic cyst contour deformation field and the original pancreatic cyst contour deformation field to obtain a pancreatic cyst morphology score;

[0029] An iterative training module: configured to sort the pancreatic morphology scores and the pancreatic cyst morphology scores from high to low, and screen out a group of data with the top scores in the training set to generate data with machine labels, and add the data with machine labels to the training set to perform iterative training on the semi-supervised learning model until all data in the training set have machine labels.

[0030] In an embodiment of the present invention, the pancreatic contour deformation field and the pancreatic contour probability map are extracted through a first segmentation network of the semi-supervised learning model; the pancreatic cyst contour deformation field and the pancreatic cyst contour probability map are extracted through a second segmentation network of the semi-supervised learning model, and the first segmentation network and the second segmentation network form a cascaded segmentation network.

[0031] In an embodiment of the present invention, the pancreatic morphology score is obtained through a first morphology discrimination network of the semi-supervised learning model; the pancreatic cyst morphology score is obtained through a second morphology discrimination network of the semi-supervised learning model; the score for sorting and screening is the sum of the pancreatic morphology scores and the pancreatic cyst scores corresponding to each data in the training set.

[0032] In one embodiment of the present invention, in the identification module, an area with a threshold greater than 0.5 in the pancreatic cyst contour probability map is extracted to obtain the pancreatic cyst area.

[0033] In one embodiment of the present invention, the identification module includes:

[0034] A mapping unit: used to divide the pancreatic cyst area into several image patches, and map the information contained in each image patch to the corresponding sequence representation vector patch embedding;

[0035] A recording unit: used to record the position representation vector position embedding corresponding to each sequence representation vector;

[0036] An identification unit: used to input the sequence representation vector patch embedding and the position representation vector position embedding into a clinically driven Transformer and combine clinical knowledge to drive features to identify the benign and malignant nature of pancreatic cysts.

[0037] In one embodiment of the present invention, the clinically driven features in the identification unit include: the maximum diameter of the cyst, the presence or absence of mural nodules, the presence or absence of solid components, and whether the pancreatic duct is dilated.

[0038] In a third aspect, the present invention provides an electronic device, which includes:

[0039] A processor, a memory, and an interface for communicating with a gateway;

[0040] The memory is used to store programs and data, and the processor calls the programs stored in the memory to execute a method for identifying pancreatic cysts based on semi-supervised learning provided in any one of the first aspects.

[0041] In a fourth aspect, the present invention provides a computer-readable storage medium, and the computer-readable storage medium includes a program, and the program is used to execute a method for identifying pancreatic cysts based on semi-supervised learning provided in any one of the first aspects when being executed by a processor.

[0042] As can be seen from the above description, a method and system for differentiating pancreatic cysts based on semi-supervised learning in an embodiment of the present invention adopt a method of step-by-step segmentation of the pancreas and pancreatic cysts. First, the pancreatic image is segmented; then, the probability map of the pancreatic contour that has been segmented first is used as the attention map for the segmentation of pancreatic cysts to guide the segmentation of pancreatic cysts, which can avoid the interference of similar tissues outside the pancreas, increase the effectiveness of feature extraction in the process of the pancreatic cyst contour, and achieve more accurate segmentation of pancreatic cysts; finally, after the segmentation of pancreatic cysts is achieved, a probability map of the pancreatic cyst contour is obtained, and the cyst area in the probability map of the pancreatic cyst contour is extracted and the clinical knowledge-driven features are fully utilized, greatly improving the ability to differentiate the benign and malignant of cysts. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The flowchart of a method for differentiating pancreatic cysts based on semi-supervised learning according to an embodiment of the present invention is shown.

[0044] Figure 2 The architecture diagram of a method for differentiating pancreatic cysts based on semi-supervised learning according to an embodiment of the present invention is shown.

[0045] Figure 3 The structural diagram of a system for differentiating pancreatic cysts based on semi-supervised learning according to an embodiment of the present invention is shown.

[0046] Figure 4 The structural diagram of an electronic device according to the present invention is shown. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts belong to the scope of protection of the present invention.

[0048] As Figure 1 and in conjunction with Figure 2 shown, a method for differentiating pancreatic cysts based on semi-supervised learning according to an embodiment of the present invention includes the following steps:

[0049] S110: Train a semi-supervised learning model with a training set with a small number of artificial labels to obtain a trained semi-supervised learning model. The semi-supervised learning model uses a large amount of unlabeled data and at the same time uses a small amount of labeled data to perform subsequent steps. The training set is used in the supervised learning model to estimate the semi-supervised learning model.

[0050] S120: Extract the pancreatic contour probability map of multi-phase pancreatic images using the trained semi-supervised learning model. The multi-phase pancreatic images are obtained by means such as CT or MRI and are registered. And through S150: Extract the pancreatic contour deformation field of multi-phase pancreatic images using the trained semi-supervised learning model, and determine the approximation degree between the pancreatic contour deformation field and the original pancreatic contour deformation field to obtain the pancreatic morphology score.

[0051] Among them, the pancreatic contour deformation field and the pancreatic contour probability map are extracted by the first segmentation network of the semi-supervised learning model. The first segmentation network is a 3D encoding-decoding network, which has an encoding-decoding path corresponding to each phase of the multi-phase pancreatic image one by one, and the features between each path can interact with each other to fuse the features of the pancreatic images in each phase and improve the pancreatic segmentation effect. The first segmentation network can adopt one of 3D-Unet or VNet models, etc.

[0052] The extracted pancreatic contour probability map is the probability value corresponding to the possibility of each pixel in the multi-phase pancreatic image being the pancreas. The higher this probability value, the greater the possibility that this pixel point is the pancreas. The label of the extracted pancreatic contour deformation field can be generated by the ndimage.distance_transform_edt function. The ndimage.distance_transform_edt function is a function in the scipy library in this field, which is used for distance transformation and calculates the distance from non-zero points in the image to the nearest background point, and can calculate the distance change of the nearest background element of each foreground element in the deformation field. The pancreatic morphology score is obtained through the first morphology discrimination network of the semi-supervised learning model. The specific process is as follows: The generated pancreatic contour deformation field and the original multi-phase pancreatic images are input into the first discrimination network for discrimination. In the way of the discriminator of the generative adversarial network, the authenticity of the pancreatic contour deformation field will be judged. (If true: the label is the deformation field generated by ndimage.distance_transform_edt; if false: it is the deformation field generated by the model). Then, the pancreatic morphology score of the pancreatic contour deformation field will be output, which ranges from 0 to 1 point (the scoring mechanism is to score according to the approximation degree between the pancreatic contour deformation field and the original pancreatic contour deformation field. The closer the pancreatic contour deformation field is to the original pancreatic contour deformation field, the closer the score value is to 1).

[0053] S130: Extract the pancreatic cyst contour probability map of multi-phase pancreatic cyst images under the guidance of the trained semi-supervised learning model using the pancreatic contour probability map. Similarly, the multi-phase pancreatic cyst images are obtained by means of CT or MRI and are registered. And through S160: Extract the pancreatic cyst contour deformation field of multi-phase pancreatic cyst images under the guidance of the trained semi-supervised learning model using the pancreatic contour probability map, and discriminate the approximation degree between the pancreatic cyst contour deformation field and the original pancreatic cyst contour deformation field to obtain the pancreatic cyst morphology score.

[0054] Among them, the pancreatic cyst contour deformation field and the pancreatic cyst contour probability map are extracted by the second segmentation network of the semi-supervised learning model. Similarly, the second segmentation network is also a 3D encoder-decoder network, which has an encoder-decoder path corresponding one by one to each phase of the multi-phase pancreatic cyst images, and the features between each path can interact with each other to fuse the features of the pancreatic cyst images in each phase and improve the pancreatic cyst segmentation effect. The second segmentation network can adopt one of the 3D-Unet or VNet models, etc.

[0055] The specific segmentation process of the pancreatic cyst is as follows: The pancreatic contour probability map will be used as an auxiliary attention map for pancreatic cyst segmentation (the principle is: the higher the value of the pancreatic contour probability map, the greater the weight of the second segmentation network corresponding to this part. On the premise of knowing the pancreatic contour probability map, according to the area with a larger pancreatic contour probability map value, some position information can be provided for the cysts on the pancreas, thereby guiding the segmentation of the cysts). In this way, the second segmentation network will be more inclined to capture the features contained in the attention map, and the pancreatic contour probability map will guide the segmentation of the cysts to focus on the pancreas, which can effectively reduce the interference of similar tissues outside the pancreas and increase the effectiveness of feature extraction. Combining S120 and S130, it can be seen that the first segmentation network and the second segmentation network form a cascaded segmentation network, which can enhance the overall performance of the segmentation network and improve the learning speed, and can also maintain the original structure when the training set changes.

[0056] The pancreatic cyst morphology score is obtained through the second morphology discrimination network of the semi-supervised learning model, and its process is the same as the pancreatic morphology score mechanism, so it will not be elaborated here.

[0057] After obtaining the pancreatic morphology score and the pancreatic cyst score, execute S170: Sort the pancreatic morphology score and the pancreatic cyst morphology score from high to low and select a group of data with the top scores in the training set to generate data with machine labels, and add the data with machine labels to the training set to perform iterative training on the semi-supervised learning model until all data in the training set have machine labels.

[0058] Specifically, before training the semi-supervised learning model, only a small portion of the data in the training set is labeled, and the other data is unlabeled. First, use the labeled data to train the semi-supervised learning model. Then, score the unlabeled data. After scoring, select the top K data in terms of scores, group these K data as a set of data, and add the corresponding generated machine labels to the training set to continue training the semi-supervised learning model. During the screening process, machine labels are preferentially generated for the data with high scores according to the score levels. For a set of data with high scores, the numbers of this set can be obtained, and then use the currently trained semi-supervised learning model to predict these numbered data (i.e., predict through the ndimage.distance_transform_edt function) to generate machine labels. Finally, the data with a quantity less than one set will also generate labels (if the remaining data is less than K and cannot form a set of data, then all these remaining data will generate machine labels). It should be noted that the scores for sorting and screening are the sum of the scores of the pancreatic morphology scores and pancreatic cyst scores corresponding to each data in the training set, making the selected data more valuable for reference. After multiple iterations, all the data in the training set will have machine labels (that is, the training set has incorporated data with machine labels, and the data originally with a small number of labels will gradually generate data with machine labels after iteration).

[0059] S140: After obtaining the pancreatic cyst contour probability map, extract the pancreatic cyst region of the pancreatic cyst contour probability map. For example, extract the region where the threshold is greater than 0.5 in the pancreatic cyst contour probability map to obtain the pancreatic cyst region. Of course, the closer this threshold is to 1, the more it can reflect the true situation of the pancreatic cyst. Extracting the region where the threshold is greater than 0.5 can improve the effectiveness of extracting effective features on the premise of ensuring that important feature data is not lost. Then, based on the pancreatic cyst region and combined with clinical knowledge-driven features, identify the benign and malignant nature of the pancreatic cyst. The specific method is as follows: First, divide the extracted pancreatic cyst region into several image patches, and map the information contained in each image patch to the corresponding sequence representation vector patch embedding; then record the corresponding position representation vector position embedding for each sequence representation vector; finally, input the sequence representation vector patch embedding and the position representation vector position embedding into the clinically driven Transformer and combine clinical knowledge-driven features to identify the benign and malignant nature of the pancreatic cyst. The clinical knowledge-driven features specifically include: the maximum diameter of the cyst, the presence or absence of mural nodules, the presence or absence of solid components, and whether the pancreatic duct is dilated.

[0060] In summary, the present invention has the following beneficial effects:

[0061] 1. The first - form discrimination network and the second - form discrimination network of the present invention discriminate the deformation fields of the pancreas and cysts, and score the forms of the pancreas and cysts. Sort according to the pancreas form score + cyst form score and select the top K data (i.e., a set of data) with the highest scores to generate machine labels, and add these K data with machine labels to the current training set for training. Iterating according to this method can help better train the semi - supervised learning model, achieve better segmentation performance, can preferentially utilize more data with better forms to guide difficult samples, effectively inhibit the problem of performance degradation as the number of machine labels increases, and overcome the problem that since most medical images are 3D, the acquisition of medical image annotation data is very expensive, and generating high - quality annotations requires professional doctors to spend a lot of time on annotation.

[0062] 2. The present invention adopts a two - step strategy for organ lesion segmentation and designs a cascaded segmentation network for the pancreas and pancreatic cysts (i.e., the first segmentation network and the second segmentation network). First, the pancreas is segmented, and the deformation field strategy can be used to optimize the contour of the pancreas. Then, the pancreas contour probability map is used to guide the subsequent segmentation of pancreatic cysts. The pancreas contour probability map can guide the segmentation of pancreatic cysts to focus within the pancreas, effectively reducing the interference of similar tissues outside the pancreas and increasing the effectiveness of feature extraction. At the same time, the segmentation of pancreatic cysts can also use the deformation field strategy to optimize the contour of the cysts. It overcomes the problem that since the segmentation of lesions (pancreatic cysts) is more difficult than that of organs (the pancreas), most existing networks do not effectively utilize the prior knowledge of organ segmentation to guide the segmentation of lesions, resulting in less than ideal segmentation results of lesions.

[0063] 3. The present invention designs an end - to - end network for lesion (pancreatic cyst) segmentation and lesion (pancreatic cyst) discrimination. The pancreatic cyst segmentation adopts the idea of using organ segmentation results to guide lesion segmentation, and designs a clinically - driven Transformer for the benign and malignant discrimination of cysts. The clinically - driven Transformer not only effectively utilizes the sequence information of the image, but also makes full use of the clinically - driven features of the position information, greatly improving the ability to discriminate the benign and malignant of pancreatic cysts. It overcomes the problem that most benign and malignant discrimination networks are not trained end - to - end with the previous segmentation network, and rarely consider clinically - driven features, and mostly conduct category identification based on convolutional neural networks (CNNs).

[0064] Based on the same inventive concept, the embodiments of the present application further provide a pancreatic cyst discrimination system based on semi-supervised learning, which can be used to implement a pancreatic cyst discrimination method based on semi-supervised learning described in the above embodiments, as in the following embodiments. Since the principle of solving problems by a pancreatic cyst discrimination system based on semi-supervised learning is similar to that of a pancreatic cyst discrimination method based on semi-supervised learning, the implementation of a pancreatic cyst discrimination system based on semi-supervised learning can refer to the method implementation, and the repeated parts will not be elaborated. Hereinafter, the term "unit" or "module" may refer to a combination of software and / or hardware that can achieve a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0065] The present invention provides a multi-modal fusion pancreatic segmentation system based on deep learning, as Figure 2 shown. In Figure 2 it, the system includes:

[0066] A model training module 210: configured to train a semi-supervised learning model through a training set with a small number of manual labels to obtain a trained semi-supervised learning model;

[0067] A first segmentation module 220: configured to use the trained semi-supervised learning model to extract a pancreatic contour probability map of multi-phase pancreatic images;

[0068] A second segmentation module 230: configured to use the trained semi-supervised learning model to extract a pancreatic cyst contour probability map of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map;

[0069] A discrimination module 240: configured to extract the pancreatic cyst region of the pancreatic cyst contour probability map, and discriminate the benign and malignant nature of the pancreatic cyst according to the pancreatic cyst region and in combination with clinically knowledge-driven features.

[0070] In an embodiment of the present invention, it further includes:

[0071] A first discrimination module 250: configured to use the trained semi-supervised learning model to extract a pancreatic contour deformation field of multi-phase pancreatic images, and discriminate the approximation degree between the pancreatic contour deformation field and the original pancreatic contour deformation field to obtain a pancreatic morphology score;

[0072] A second discrimination module 260: configured to use the trained semi-supervised learning model to extract a pancreatic cyst contour deformation field of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map, and discriminate the approximation degree between the pancreatic cyst contour deformation field and the original pancreatic cyst contour deformation field to obtain a pancreatic cyst morphology score;

[0073] Iterative training module 270: It is used to sort the pancreatic morphology scores and pancreatic cyst morphology scores from high to low, screen out a set of data with the top scores in the training set to generate data with machine labels, and add the data with machine labels to the training set for semi-supervised learning model iterative training until all data in the training set have machine labels.

[0074] In an embodiment of the present invention, the pancreatic contour deformation field and the pancreatic contour probability map are extracted by the first segmentation network of the semi-supervised learning model; the pancreatic cyst contour deformation field and the pancreatic cyst contour probability map are extracted by the second segmentation network of the semi-supervised learning model, and the first segmentation network and the second segmentation network form a cascaded segmentation network.

[0075] In an embodiment of the present invention, the pancreatic morphology score is obtained through the first morphology discrimination network of the semi-supervised learning model; the pancreatic cyst morphology score is obtained through the second morphology discrimination network of the semi-supervised learning model; the scores for sorting and screening are the sum of the scores of the pancreatic morphology scores and the pancreatic cyst scores corresponding to each data in the training set.

[0076] In an embodiment of the present invention, in the discrimination module 240, the region in the pancreatic cyst contour probability map with a threshold greater than 0.5 is extracted to obtain the pancreatic cyst region.

[0077] In an embodiment of the present invention, the discrimination module includes:

[0078] Mapping unit 241: It is used to divide the pancreatic cyst region into several image patches and map the information contained in each image patch to the corresponding sequence representation vector patch embedding.

[0079] Recording unit 242: It is used to record the position representation vector position embedding corresponding to each sequence representation vector.

[0080] Discrimination unit 243: It is used to input the sequence representation vector patch embedding and the position representation vector position embedding into the clinically driven Transformer and combine the clinical knowledge-driven features to discriminate the benign and malignant of pancreatic cysts.

[0081] In an embodiment of the present invention, the clinical knowledge-driven features in the discrimination unit 243 include: the maximum diameter of the cyst, the presence or absence of mural nodules, the presence or absence of solid components, and whether the pancreatic duct is dilated.

[0082] The embodiments of the present application also provide a specific implementation manner of an electronic device capable of implementing all the steps in the above-mentioned method for discriminating pancreatic cysts based on semi-supervised learning. Refer to Figure 3 , and the electronic device 300 specifically includes the following content:

[0083] A processor 310, a memory 320, a communication unit 330, and a bus 340;

[0084] Among them, the processor 310, the memory 320, and the communication unit 330 communicate with each other through the bus 340; the communication unit 330 is used to implement information transmission between related devices such as server-side devices and terminal devices.

[0085] The processor 310 is used to call the computer program in the memory 320, and when the processor executes the computer program, all steps in a method for differentiating pancreatic cysts based on semi-supervised learning in the above-mentioned embodiment are implemented.

[0086] Those of ordinary skill in the art should understand that the memory can be, but is not limited to, a random access memory (Random Access Memory, abbreviated as RAM), a read-only memory (Read Only Memory, abbreviated as ROM), a programmable read-only memory (Programmable Read-Only Memory, abbreviated as PROM), an erasable programmable read-only memory (Erasable Programmable Read-Only Memory, abbreviated as EPROM), an electrically erasable programmable read-only memory (Electric Erasable Programmable Read-Only Memory, abbreviated as EEPROM), etc. Among them, the memory is used to store programs, and after receiving an execution instruction, the processor executes the program. Further, the software programs and modules in the above-mentioned memory may also include an operating system, which may include various software components and / or drivers for managing system tasks (such as memory management, storage device control, power management, etc.), and may communicate with various hardware or software components to provide a running environment for other software components.

[0087] The processor may be an integrated circuit chip with signal processing capabilities. The above-mentioned processor may be a general-purpose processor, including a central processing unit (Central Processing Unit, abbreviated as CPU), a network processor (Network Processor, abbreviated as NP), etc. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0088] The present application also provides a computer-readable storage medium, and the computer-readable storage medium includes a program, and the program is used to execute a method for differentiating pancreatic cysts based on semi-supervised learning provided in any one of the foregoing method embodiments when being executed by a processor.

[0089] Those of ordinary skill in the art should understand that all or part of the steps for implementing the above method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including those of the above method embodiments; and the aforementioned storage medium includes: various media such as ROM, RAM, magnetic disks, or optical discs that can store program codes, and the specific type of the medium is not limited in this application.

[0090] As described above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for differentiating pancreatic cysts based on semi-supervised learning, characterized in that, it includes the following steps: Training a semi-supervised learning model with a training set with a small number of manual labels to obtain a trained semi-supervised learning model; Using the trained semi-supervised learning model to extract the pancreatic contour probability map of multi-phase pancreatic images; Using the trained semi-supervised learning model to extract the pancreatic cyst contour probability map of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map; Extracting the pancreatic cyst area of the pancreatic cyst contour probability map, and differentiating the benign and malignant of the pancreatic cyst according to the pancreatic cyst area and combining clinical knowledge-driven features; It also includes: Using the trained semi-supervised learning model to obtain the pancreatic contour deformation field and the pancreatic cyst contour deformation field, respectively discriminating and scoring the pancreatic contour deformation field and the pancreatic cyst contour deformation field to obtain the pancreatic morphology score and the pancreatic cyst morphology score; Sorting the pancreatic morphology score and the pancreatic cyst morphology score from high to low, screening out a group of data with the highest scores in the training set to generate data with machine labels, and adding the data with machine labels to the training set to iteratively train the semi-supervised learning model until all data in the training set have machine labels.

2. A method for differentiating pancreatic cysts based on semi-supervised learning according to claim 1, characterized in that, The step of using the trained semi-supervised learning model to obtain the pancreatic contour deformation field and the pancreatic cyst contour deformation field, respectively discriminating and scoring the pancreatic contour deformation field and the pancreatic cyst contour deformation field to obtain the pancreatic morphology score and the pancreatic cyst morphology score includes: Using the trained semi-supervised learning model to extract the pancreatic contour deformation field of multi-phase pancreatic images, discriminating the approximation degree between the pancreatic contour deformation field and the original pancreatic contour deformation field to obtain the pancreatic morphology score; Using the trained semi-supervised learning model to extract the pancreatic cyst contour deformation field of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map, and discriminating the approximation degree between the pancreatic cyst contour deformation field and the original pancreatic cyst contour deformation field to obtain the pancreatic cyst morphology score.

3. A method for differentiating pancreatic cysts based on semi-supervised learning according to claim 2, characterized in that, The pancreatic contour deformation field and the pancreatic contour probability map are extracted by the first segmentation network of the semi-supervised learning model; the pancreatic cyst contour deformation field and the pancreatic cyst contour probability map are extracted by the second segmentation network of the semi-supervised learning model, and the first segmentation network and the second segmentation network form a cascaded segmentation network.

4. A method for differentiating pancreatic cysts based on semi-supervised learning according to claim 3, characterized in that, The pancreatic morphology score is obtained by the first morphology discrimination network of the semi-supervised learning model; the pancreatic cyst morphology score is obtained by the second morphology discrimination network of the semi-supervised learning model; the scores for sorting and screening are the sum of the pancreatic morphology scores and the pancreatic cyst scores corresponding to each data in the training set.

5. A method for differentiating pancreatic cysts based on semi-supervised learning according to claim 1, characterized in that, Extracting the area where the threshold in the pancreatic cyst contour probability map is greater than 0.5 to obtain the pancreatic cyst area.

6. The method for differentiating pancreatic cysts based on semi-supervised learning according to claim 5, characterized in that, the differentiating of the benign and malignant of pancreatic cysts according to the pancreatic cyst region and combined with clinically knowledge-driven features includes: dividing the pancreatic cyst region into several image patches, and mapping the information contained in each image patch to the corresponding sequence representation vector patch embedding; recording the position representation vector position embedding corresponding to each sequence representation vector; inputting the sequence representation vector patch embedding and the position representation vector position embedding into a clinically-driven Transformer and combining clinically knowledge-driven features to differentiate the benign and malignant of pancreatic cysts.

7. The method for differentiating pancreatic cysts based on semi-supervised learning according to claim 1, characterized in that, the clinically knowledge-driven features include: the maximum diameter of the cyst, the presence or absence of mural nodules, the presence or absence of solid components, and whether the pancreatic duct is dilated.

8. A system for differentiating pancreatic cysts based on semi-supervised learning, characterized in that, it includes: a model training module: used to train a semi-supervised learning model through a training set with a small number of manual labels to obtain a trained semi-supervised learning model; a first segmentation module: used to extract the pancreatic contour probability map of multi-phase pancreatic images by using the trained semi-supervised learning model; a second segmentation module: used to extract the pancreatic cyst contour probability map of multi-phase pancreatic cyst images under the guidance of the pancreatic contour probability map by using the trained semi-supervised learning model; a differentiation module: used to extract the pancreatic cyst region of the pancreatic cyst contour probability map, and differentiate the benign and malignant of pancreatic cysts according to the pancreatic cyst region and combined with clinically knowledge-driven features; it further includes: a discrimination module: used to obtain the pancreatic contour deformation field and the pancreatic cyst contour deformation field by using the trained semi-supervised learning model, respectively perform discrimination scoring on the pancreatic contour deformation field and the pancreatic cyst contour deformation field to obtain the pancreatic morphology score and the pancreatic cyst morphology score; an iterative training module: used to sort the pancreatic morphology score and the pancreatic cyst morphology score from high to low, screen out a group of data with high scores in the training set to generate data with machine labels, and add the data with machine labels to the training set to perform iterative training on the semi-supervised learning model until all data in the training set have machine labels.

9. An electronic device, characterized in that, the device includes: a processor, a memory, and an interface for communicating with a gateway; the memory is used to store programs and data, and the processor calls the programs stored in the memory to execute the method for differentiating pancreatic cysts based on semi-supervised learning according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, the computer-readable storage medium includes a program, and the program is used to execute the method for differentiating pancreatic cysts based on semi-supervised learning according to any one of claims 1 to 7 when being executed by a processor.

Images