Margin-positive prediction system and method based on knowledge distillation learning

Abstract

The application discloses a kind of based on knowledge distillation learning's cut edge positive prediction system and method, comprising: database: store original original medical image data to be predicted, past clinical data and the intermediate medical image data obtained in training process. Image generation module: original medical image is trained using image generation network model, and intermediate medical image is generated to supplement the 3D information missing on original data set. Cut edge positive prediction module: make prediction to cut edge positive condition in combination with clinical data and generated intermediate medical image. Result shows module: the result is output to display. The application has the advantages that: the accuracy of prediction model is improved. The stability and reliability of model are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a cutting edge positive prediction system and method based on knowledge distillation learning. Background Technology

[0002] In recent years, with the increasing maturity of artificial intelligence technology, research on deep neural models has achieved significant results and has been applied to various aspects of life, such as e-commerce, the stock market and finance, social networks, and healthcare. Since the advent of the digital imaging era, the generation of massive amounts of data has provided more possibilities for the future development of medical imaging. Therefore, how to further analyze and mine big data in medical imaging, how to extract valuable information from high-dimensional medical image data, and how to closely integrate the development of modern medical imaging with precision medicine have become important issues for the future development of medical imaging.

[0003] Positive surgical margins are a common problem encountered during radical prostatectomy. Due to factors such as tumor size, location, and the anatomical characteristics of the prostate, incomplete tumor resection may occur, leading to positive surgical margins in the pathological specimen, which in turn affects the patient's prognosis and treatment strategy. Therefore, identifying the risk factors for positive surgical margins preoperatively and understanding their characteristics is of great importance.

[0004] In existing technologies, most methods employ machine learning to predict positive surgical margins after prostatectomy using clinical data and manually measured features in medical images. However, manually extracting and analyzing imaging data inevitably overlooks a large amount of hidden lesion information. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a knowledge distillation-based system and method for predicting positive surgical margins. It combines deep learning to comprehensively extract hidden information from patient medical images, improving the accuracy of postoperative positive surgical margin prediction and providing a reference for clinical surgical planning.

[0006] This invention discloses a cutting edge positive prediction system based on knowledge distillation learning. The system includes: a data input module, an image generation module, a cutting edge positive prediction module, and a result display module.

[0007] Database: Stores the original medical image data to be predicted, previous clinical data, and intermediate medical image data obtained during training.

[0008] Image generation module: The original medical images are trained using an image generation network model to generate intermediate medical images to supplement the missing 3D information in the original dataset.

[0009] Positive resection margin prediction module: Combines clinical data with generated intermediate medical images to predict the likelihood of a positive resection margin.

[0010] Results display module: Outputs the results to the display.

[0011] Furthermore, the image generation module includes a variational autoencoder. The variational autoencoder encodes the input medical image to obtain its latent vector representation. This latent vector is then input into the decoder to reconstruct the original image. Unlike traditional autoencoders, the variational autoencoder can generate diverse output images because it introduces latent variables into the computation, allowing the model to interpolate and manipulate within the latent space. This provides greater flexibility for medical image generation, as richer output results can be generated through manipulation within the latent space.

[0012] Furthermore, the variational autoencoder includes an encoder and a decoder. After the network generative model is trained, two adjacent medical images of the patient are input into the encoder to obtain the latent vectors corresponding to the two images. The average of the two latent vectors is then calculated to obtain the latent vector of the intermediate image. This latent vector is input into the decoder to generate an intermediate image that is associated with all adjacent medical images. In this way, any number of intermediate medical images can be generated.

[0013] Furthermore, the training formula for the variational autoencoder is as follows:

[0014]

[0015] Where z1 and z2 represent the latent vectors corresponding to two adjacent medical images of the patient, respectively. mid Let z represent the latent vector of the intermediate image, k represent the number of intermediate images generated, and n represent the nth sequential image generated. mid Input the data into the decoder to obtain the generated intermediate medical image.

[0016] Furthermore, the operation of adding random noise during decoder training is as follows:

[0017] z out =f(z) mid +n)

[0018] Among them, z out Let be the latent vector of the output, n be random noise, and f be the mapping function of the decoder. Random noise does not need to be added when generating the image.

[0019] The 3D convolutional neural network used in the edge positive prediction module employs a cross-entropy loss function during training to measure the difference between the model's prediction results and the true labels.

[0020] Furthermore, the cross-entropy loss function is expressed as:

[0021]

[0022] Where N is the number of training samples, x i It is the i-th medical image input, y i f(x) is the true label corresponding to the i-th medical image. i ;θ) is the model at input x i The following prediction results θ represents the predicted probability corresponding to the true label, and θ represents the model parameters.

[0023] Furthermore, the positive prediction module for the cutting edge uses a knowledge distillation method based on Neural Architecture Search (NAS) to optimize the 3D convolutional neural network model.

[0024] The knowledge distillation includes the following steps: constructing and training a teacher network, designing the teacher network structure and training the teacher network using an image dataset; obtaining a student network based on a neural network architecture search, and sampling the structure of the student network using the structure of the teacher network as the search space; training the student network based on knowledge distillation, and training the student network using a knowledge distillation loss function to obtain a student network that meets the target.

[0025] Furthermore, the knowledge distillation includes the steps of

[0026] First, define the teacher network as T, the student network as S, and use D... s and D t This represents a dataset representing a network of students and teachers. (Using f...) T f represents the output of the teacher network. S This represents the output of the student network. Use p t Let p represent the probability distribution of the teacher network output. s Let KD(p,q) represent the probability distribution of the student network output. Let KD(p,q) represent the Kullback-Leibler divergence using the temperature parameter T. Let NAS(D) represent the probability distribution of the student network output. t This indicates a search method based on a neural network architecture, using dataset D. t Search for student network structure.

[0027] The knowledge distillation process based on NAS is as follows:

[0028] S11. Construct and train the teacher network T using dataset D. t The teacher's network output f is obtained. T .

[0029] S12. Design the teacher network structure and train the teacher network T using an image dataset.

[0030] S13. Using a neural network architecture search method, with the structure of the teacher network as the search space, search for the student network structure S = NAS(D). t ), to obtain the student network output f S .

[0031] S14. Using the probability distribution p output by the teacher network t The probability distribution p of the student network output s To perform knowledge distillation training, i.e., to use the knowledge distillation loss function:

[0032]

[0033] Where α is a weighting parameter, N is the number of samples in the dataset, and CE(y) i ,f s (i) ) is the cross-entropy loss function, y i It is the true label of sample i.

[0034] S15. Minimize the knowledge distillation loss function L KD The student network S is trained to obtain a student network that meets the objectives.

[0035] This invention also discloses a cutting edge positive prediction method based on knowledge distillation learning, which is implemented on the basis of the above-mentioned cutting edge positive prediction system.

[0036] The method for predicting positive surgical margins includes the following steps:

[0037] S1: Acquire raw medical imaging data and previous clinical data;

[0038] S2: Use an image generation network model to train the original medical images and generate intermediate medical images to supplement the missing 3D information in the original dataset.

[0039] S3: Predict the occurrence of positive surgical margins by combining clinical data with generated intermediate medical images.

[0040] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method for predicting positive cut edges.

[0041] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting positive cutting edges.

[0042] Compared with the prior art, the advantages of the present invention are as follows:

[0043] Integrating clinical and medical imaging data: This invention not only utilizes structured clinical data to predict the probability of positive surgical margins, but also uses information from medical imaging, combining the two to improve the accuracy of the prediction model.

[0044] Variational autoencoder is used to expand medical image data: This invention uses a variational autoencoder to expand medical image data, thereby increasing the amount of data and improving the stability and reliability of the model.

[0045] Model optimization using knowledge distillation: This invention employs knowledge distillation to optimize the model, further improving the accuracy of the prediction model.

[0046] Experimental results show improved accuracy: The experimental results of this invention show that the accuracy is improved by 0.09 compared with previous methods, proving that this method has better predictive performance.

[0047] This study also analyzed the relationship between different features in medical imaging and the prediction results of positive surgical margins. The results showed that medical images of the prostate protruding into the bladder significantly improved the accuracy of predicting positive surgical margins. Body size and pelvic morphology also influenced the positive surgical margin prediction results. These findings help guide physicians to more accurately assess patient risks and develop more appropriate treatment plans. Attached Figure Description

[0048] Figure 1 This is a flowchart of the prediction method of the positive edge prediction system according to an embodiment of the present invention.

[0049] Figure 2 This is a flowchart of the intermediate image generation process in an embodiment of the present invention.

[0050] Figure 3 This is an image generation result using the original network in an embodiment of the present invention.

[0051] Figure 4 These are example images of a pair of generated images according to an embodiment of the present invention. The left image is the input image, and the right image is the output image.

[0052] Figure 5 This is a statistical graph showing the accuracy of the model in this embodiment of the invention on the validation set;

[0053] Figure 6 This is a schematic diagram of the network optimization method based on neural architecture search in an embodiment of the present invention;

[0054] Figure 7 This is a schematic diagram of the dataset image masking process in an embodiment of the present invention;

[0055] Figure 8This is a diagram showing the influence of different locations on the positive prediction effect of surgical margins in medical images according to an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0057] like Figure 1 As shown, this invention provides a method for predicting positive cut edges based on knowledge distillation learning, comprising the following steps:

[0058] S1: Acquire raw medical imaging data and previous clinical data;

[0059] S2: Use an image generation network model to train the original medical images and generate intermediate medical images to supplement the missing 3D information in the original dataset.

[0060] S3: Predict the occurrence of positive surgical margins by combining clinical data with generated intermediate medical images.

[0061] The aforementioned method for predicting positive surgical margins is implemented based on a system for predicting positive surgical margins, which comprises two main modules: an image generation module and a positive surgical margin prediction module. The main function of the image generation module is to generate medical images using an image generation network model to supplement the missing 3D information in the original dataset. The main function of the positive surgical margin prediction module is to combine clinical data with the generated images to predict the status of positive surgical margins.

[0062] The image generation module primarily utilizes an optimized variational autoencoder (VAE). A VAE is a machine learning model consisting of an encoder and a decoder. In medical image processing, a VAE transforms an image into intermediate latent vectors, from which an output image similar to the input image is generated. After training, the VAE can obtain the corresponding latent vectors of adjacent medical images from the same patient by inputting them into the encoder. The average of these latent vectors yields a series of intermediate image latent vectors, which are then used by the decoder to generate these intermediate images again. Unlike traditional autoencoders, the VAE can generate diverse output images because it introduces latent variables into the computation, allowing the model to interpolate and manipulate within the latent space. This provides greater flexibility for medical image processing, as richer output results can be generated through manipulation within the latent space.

[0063] In variational autoencoders (VAEs), continuous images surrounding medical images can be generated by fine-tuning the values ​​of the latent vectors. This is because VAEs can map similar images to similar latent vectors. Therefore, by fine-tuning these similar latent vectors, continuous images surrounding medical images can be generated, thereby achieving refined processing of medical images.

[0064] Variational autoencoders (VAEs) first encode the input medical image to obtain its latent vector representation. Then, by fine-tuning the values ​​of these latent vectors, a new latent vector is generated, which is then used by a decoder to generate the corresponding image. In this way, the values ​​of the latent vectors can be continuously fine-tuned to generate continuous images surrounding the medical image. Because VAEs can map similar images to similar latent vectors, the generated continuous images will also be very similar and can preserve the characteristics of the medical image. Fine-tuning the values ​​of the latent vectors requires certain skills and experience, as different latent vectors correspond to different image features. Therefore, in practical applications, it is necessary to select appropriate latent vectors based on the specific medical image and fine-tune them to achieve the best image generation results.

[0065] After the variational autoencoder network generative model is trained, two adjacent medical images of the patient are input into the encoder to obtain latent vectors for each image. The average of the two latent vectors yields the latent vector for the intermediate image. This latent vector is then input into the decoder to generate an intermediate image associated with all adjacent medical images. This method can generate any number of intermediate medical images. It's important to note that random noise needs to be added to the output latent vector during decoder training, but this noise is not required when generating the image.

[0066]

[0067] Where z1 and z2 represent the latent vectors corresponding to two adjacent medical images of the patient, respectively. mid Let z represent the latent vector of the intermediate image, k represent the number of intermediate images generated, and n represent the nth sequential image generated. mid Input the data into the decoder to obtain the generated intermediate medical image.

[0068] The operation of adding random noise during decoder training is represented as follows:

[0069] z out =f(z) mid +n)

[0070] Among them, z out Let n be the latent vector of the output, n be random noise, and f be the mapping function of the decoder. This noise is not needed when generating the image. Figure 2The intermediate image generation process based on the model in this embodiment is described.

[0071] The positive cut margin prediction module can help doctors diagnose diseases more accurately and develop more effective treatment plans. In the image generation section above, a generative network was used to augment the medical image dataset, resulting in new images with high similarity. These new images increase the quantity and diversity of the medical image dataset, thereby improving the performance of the machine learning algorithm. Next, these augmented datasets are used to train the positive cut margin prediction model, with the aim of improving the accuracy and reliability of medical images.

[0072] The difference between 3D and 2D convolutional neural networks lies in the fact that the convolution kernel moves and performs convolution operations across a series of spatially continuous images in three directions (height, width, and channels). Because medical images typically have high dimensionality, large data volumes, and complex structures, traditional machine learning methods often struggle to handle them. 3D convolutional neural networks, however, can fully utilize the temporal, spatial, and morphological information in medical images, improving the accuracy and robustness of the model and making medical image analysis more precise and reliable. Furthermore, due to the extensive movement in three directions, the number of parameters in a 3D convolutional neural network is significantly greater than that of a 2D convolutional neural network.

[0073] In this embodiment, the 3D convolutional neural network uses a cross-entropy loss function during training to measure the difference between the model's predictions and the true labels. Assume there are N training samples, where x... i It is the i-th medical image input, y i It is its corresponding real label, f(x) i ;\theta) is the model's response to input x i The following prediction results This represents the predicted probability corresponding to the true label. Therefore, the cross-entropy loss function can be expressed as:

[0074]

[0075] Here, θ represents the model parameters. Optimizing this loss function can make the model predict the labels of medical images more accurately.

[0076] Furthermore, a large training dataset is required to build an excellent image recognition network. In the field of computer vision, for deep learning image classification problems, it is empirically estimated that at least 1000 images are needed for each category. However, due to ethical and privacy concerns regarding medical image datasets, it is difficult to collect a sufficiently large public dataset. Because the dataset is relatively small and the 3D convolutional neural network model is large, the feature extraction capability of the prediction model is insufficient. Therefore, this embodiment designs a knowledge distillation optimization method based on Neural Architecture Search (NAS) to optimize the 3D convolutional neural network model. The main idea of ​​this optimization method is to use neural architecture search to sample a sub-model from a large model trained on a public dataset, and then perform knowledge distillation on the two models as the teacher network and the student network, respectively. This trained student model serves as the pre-training model for the cut edge positive prediction model.

[0077] Transferring rich knowledge from teachers to students is a crucial step in knowledge distillation. Typically, knowledge distillation uses a teacher-student model, where the teacher model is complex to train and can be ensembled from multiple separately trained models, while the student model has a relatively simple training structure. The teacher model's output Y is mapped to the probability value of the corresponding class via softmax, and the student model also outputs the corresponding probability value. Existing distillation methods include offline distillation, online distillation, and self-distillation. To improve knowledge transfer efficiency, further research is needed on the relationship between model complexity and distillation schemes. Importantly, besides ensuring that the model can output a corresponding Y for the input X, no restrictions are placed on the specific architecture, number of parameters, or whether ensembles are used.

[0078] Therefore, this embodiment will use a knowledge transfer method based on knowledge distillation to acquire knowledge from similar open-source datasets to improve the model's recognition performance. The knowledge distillation part based on Neural Architecture Search (NAS) used in this embodiment mainly includes the following steps: constructing and training a teacher network, designing the teacher network structure and training the teacher network using an image dataset; obtaining a student network based on neural architecture search, and sampling the structure of the student network using the structure of the teacher network as the search space; training the student network based on knowledge distillation, and training the student network using the knowledge distillation loss function to obtain a student network that meets the target.

[0079] First, define the teacher network as T, the student network as S, and use D... s and D t This represents a dataset representing a network of students and teachers. (Using f...) T f represents the output of the teacher network. SThis represents the output of the student network. Use p t Let p represent the probability distribution of the teacher network output. s Let KD(p,q) represent the probability distribution of the student network output. Let KD(p,q) represent the Kullback-Leibler divergence using the temperature parameter T. Let NAS(D) represent the probability distribution of the student network output. t This indicates a search method based on a neural network architecture, using dataset D. t Search for student network structure.

[0080] The process of knowledge distillation based on NAS can be described as follows:

[0081] 1. Construct and train the teacher network T using dataset D. t The teacher's network output f is obtained. T .

[0082] 2. Design the teacher network structure and train the teacher network T using an image dataset.

[0083] 3. Using a neural network architecture search method, with the structure of the teacher network as the search space, search for the student network structure S = NAS(D t ), to obtain the student network output f S .

[0084] 4. Using the probability distribution p output by the teacher network t The probability distribution p of the student network output s To perform knowledge distillation training, i.e., to use the knowledge distillation loss function:

[0085]

[0086] Where α is a weighting parameter, N is the number of samples in the dataset, and CE(y) i ,f s (i) ) is the cross-entropy loss function, y i It is the true label of sample i.

[0087] The above process is as follows Figure 6 As shown.

[0088] 5. By minimizing the knowledge distillation loss function L KD Train the student network $S$ to obtain a student network that meets the objectives.

[0089] The above algorithm process is shown in Algorithm 1.

[0090]

[0091]

[0092] This embodiment also aims to explore the influence of image features from different parts of a medical image on positive edges. To achieve this, experiments were designed to analyze the relationship between features from different parts of a medical image and the positive edge prediction results.

[0093] In the image generation experiment, this embodiment used a dataset containing 171 cases and 3 coronal MRI images for each patient. Of these, 85 patients had positive surgical margins post-surgery, and the remaining 86 had negative margins. The goal of this experiment was to train the generation model using images from all patients and generate a total of 100 medical images for each patient. Simultaneously, to demonstrate the effectiveness of the model used in this embodiment, comparative experiments were conducted using other commonly used image generation methods.

[0094] First, train the original VAE network and input medical image images. Observe the training results. The image generation results using the original network are as follows. Figure 3 As shown, the original network performs poorly on this experimental dataset, failing to reproduce the original image. The images generated by the improved VAE are compared with real images to observe their similarity. Visually, the similarity between the output image and the input image directly reflects the effectiveness of the image generation model. Figure 4 Here is an example of a pair of generated images: the left image is the input image, and the right image is the output image. It can be seen that, apart from a slight decrease in sharpness, the output image and the input image are highly similar.

[0095] Since the initial dataset contained three medical images for each patient, this experiment started with three images, gradually increasing the number of images during training. Other training settings remained unchanged, and the accuracy of the trained model on the validation set was statistically analyzed. Figure 5 As shown, the vertical axis represents the model's accuracy (%) on the test set, and the horizontal axis represents the number of medical images used to train the 3D image recognition model. In the model proposed in this embodiment, when the number of input images reaches approximately 66, the average accuracy on the test set no longer shows a significant improvement.

[0096] In training the positive surgical margin prediction model, the prostate MRI image dataset used in this embodiment includes coronal images of 116 patients and structured information for each patient, including eight clinical data points: age, surgical method, PSA, stage, ISUP grade, prostate volume, bladder neck circumference (MRI data), and IPPL (MRI data), as well as the corresponding positive surgical margin status. The dataset used for knowledge distillation is the MyConnectome brain MRI image dataset.

[0097] Table 1 shows the performance of the model in this embodiment compared to other well-known and efficient 3D image recognition models.

[0098]

[0099] The training results are shown in Table 1. In this dataset, the 3D image recognition model used in this embodiment has more parameters than other models, and its accuracy is superior. However, after distillation optimization, the model parameters are reduced, and its accuracy is still better than other models.

[0100] To investigate the influence of different image features in medical images on the positive surgical margin, this embodiment designs and analyzes the relationship between different features in medical images and the prediction results of positive surgical margin. First, the top of the right leg bone, the right leg bone itself, the pelvis, the top of the left leg bone, the left leg bone, body shape, organs surrounding the bladder, and the prostate protruding into the bladder in the dataset images are masked and treated as single-factor variables. For example... Figure 6 As shown, areas other than the area to be studied are shielded.

[0101] Results of positive prediction experiments on cut margins at different sites Figure 7 This indicates that among the eight factors mentioned above in medical imaging of prostate cancer patients, patient body size and pelvic morphology have a certain impact on positive surgical margins. Hypothesis testing revealed that incorporating medical imaging of the prostate protruding into the bladder significantly improved the accuracy of predicting positive surgical margins, while the morphology of surrounding organs and the bones of the left and right legs had virtually no impact on positive surgical margins. These experimental results are consistent with existing research findings.

[0102] The data from this experiment demonstrates that in medical imaging of prostate cancer patients, the patient's body shape and pelvic morphology can affect the outcome of positive surgical margins, thus requiring special attention during relevant examinations. Furthermore, the results show that focusing on the image features of the prostate protruding into the bladder in prostate cancer patients helps improve the predictive accuracy of positive surgical margins. These findings can help medical personnel make more accurate judgments in the early detection and treatment planning of prostate cancer.

[0103] Experimental results show that the proposed method for predicting positive surgical margins can improve performance in medical imaging tasks. Furthermore, the VAE generation model proposed in this embodiment demonstrates good performance and applicability in medical image synthesis and generation, providing a more reliable and efficient technical means for medical research and clinical practice. In addition, the knowledge distillation optimization method also has practical value in medical image analysis and processing, further improving the performance of deep learning algorithms in this field. This experiment is of significant importance and value for the analysis and processing of medical images, possessing both academic and practical significance, and providing more reliable and efficient technical support for medical research and clinical practice.

[0104] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a cutting edge positive prediction method, including the following steps:

[0105] S1: Acquire raw medical imaging data and previous clinical data;

[0106] S2: Use an image generation network model to train the original medical images and generate intermediate medical images to supplement the missing 3D information in the original dataset.

[0107] S3: Predict the occurrence of positive surgical margins by combining clinical data with generated intermediate medical images.

[0108] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). This computer-readable storage medium is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0109] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the cutting edge positive prediction method in the above embodiments; one or more instructions in the computer-readable storage medium are loaded by the processor and executed as follows:

[0110] S1: Acquire raw medical imaging data and previous clinical data;

[0111] S2: Use an image generation network model to train the original medical images and generate intermediate medical images to supplement the missing 3D information in the original dataset.

[0112] S3: Predict the occurrence of positive surgical margins by combining clinical data with generated intermediate medical images.

[0113] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the implementation methods of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the present invention.

Claims

1. A cutting edge positive prediction system based on knowledge distillation learning, characterized in that, include: The module includes a data input module, an image generation module, a positive edge prediction module, and a result display module. Database: Stores the original medical image data to be predicted, previous clinical data, and intermediate medical image data obtained during training. Image generation module: The original medical images are trained using an image generation network model to generate intermediate medical images to supplement the missing 3D information in the original dataset; The image generation module includes a variational autoencoder; The variational autoencoder encodes the input medical image to obtain its latent vector representation; The latent vectors are then input into the decoder to reconstruct the initial image; The variational autoencoder consists of an encoder and a decoder. After the network generative model is trained, two adjacent medical images of the patient are input into the encoder to obtain the latent vectors corresponding to the two images. The average of the two latent vectors is used to obtain the latent vector of the intermediate image. This latent vector is input into the decoder to generate an intermediate image that is associated with the adjacent medical images. Any number of intermediate medical images can be generated in this way. Positive resection margin prediction module: Combines clinical data with generated intermediate medical images to predict the occurrence of positive resection margins; Results display module: Outputs the results to the display screen; The 3D convolutional neural network used in the edge positive prediction module employs a cross-entropy loss function during training to measure the difference between the model's prediction results and the true labels. The cross-entropy loss function is expressed as: ， in, The number of training samples, This is the i-th medical image input. It is the real label corresponding to the i-th medical image. Is the model at the input? The following prediction results It is the predicted probability corresponding to the true label. Indicates model parameters; The positive prediction module for the cutting edge uses a knowledge distillation method based on Neural Architecture Search (NAS) to optimize the 3D convolutional neural network model.

2. The cutting edge positive prediction system based on knowledge distillation learning according to claim 1, characterized in that: The training formula for the variational autoencoder is as follows: ， in, and These represent the latent vectors corresponding to two adjacent medical images of the patient. The latent vector representing the intermediate image. Indicates the number of intermediate images generated. Indicates the generated first Zhang sequence images; will Input the data into the decoder to obtain the generated intermediate medical image.

3. The cutting edge positive prediction system based on knowledge distillation learning according to claim 2, characterized in that: The operation of adding random noise during decoder training is as follows: ， in, The latent vector of the output, It is random noise. This is the mapping function for the decoder; random noise is not needed when generating the image.

4. The cutting edge positive prediction system based on knowledge distillation learning according to claim 1, characterized in that: The knowledge distillation includes the following steps: constructing and training a teacher network, designing the teacher network structure and training the teacher network using an image dataset; obtaining a student network based on a neural network architecture search, and sampling the structure of the student network using the structure of the teacher network as the search space; training the student network based on knowledge distillation, and training the student network using a knowledge distillation loss function to obtain a student network that meets the target.

5. A method for predicting positive cut edges based on knowledge distillation learning, characterized in that: This method is implemented based on the positive prediction system for cut edges as described in any one of claims 1 to 4; The method for predicting positive surgical margins includes the following steps: S1: Acquire raw medical imaging data and previous clinical data; S2: Use an image generation network model to train the original medical images and generate intermediate medical images to supplement the missing 3D information in the original dataset; S3: Predict the occurrence of positive surgical margins by combining clinical data with generated intermediate medical images.

6. A computer device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the cutting edge positive prediction method of claim 5.

Images