Bad identification model construction method based on small sample and bad identification method

Abstract

The present application relates to the technical field of medical image processing, in particular to a BAD identification model construction method and a BAD identification method based on a small sample. In the present application, a first data set is first constructed as a training sample for model pre-training through data augmentation, then a simulated lesion conforming to the clinical pathological rule is generated in combination with the vascular anatomy features of the PPA region to obtain a second data set, and the pre-trained model is efficiently fine-tuned using the second data set to significantly improve the model generalization ability under the condition of a small sample. At the same time, the finally constructed BAD identification model can automatically analyze the image features and output a diagnosis result. Moreover, through the two-stage training strategy, the powerful feature extraction capability of the pre-trained model is retained, and the model is well adapted to the BAD diagnosis task through a small amount of parameter adjustment, thereby effectively improving the model performance and generalization ability under the condition of a small sample data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a BAD identification model construction method based on a small sample and a BAD identification method. BACKGROUND

[0002] Branch Atheromatous Disease (BAD) is one of the important causes of ischemic stroke, and its related stroke accounts for 9.1%-18.3% of all ischemic strokes. BAD progresses rapidly, and early symptoms are prone to worsen, leading to a high rate of disability and a heavy burden on the social health system.

[0003] Due to the limitations of early neuroimaging technology, the research on BAD has been in the theoretical stage for a long time, and it is difficult to directly observe the perforator artery and its lesions. In recent years, with the development of high-resolution and high-field magnetic resonance imaging (MRI) technology, the imaging research on BAD has made significant progress. Researchers have gradually explored the pathogenesis of BAD from the initial analysis of lesion characteristics, making MRI a core tool for BAD research.

[0004] However, this diagnostic standard has been proposed for a short time, and its clinical application still faces many challenges. First, the interpretation of DWI images relies on the rich experience of doctors, and BAD screening also needs to combine biochemical indicators and other clinical information, increasing the workload of doctors. In a high-intensity work environment, the accuracy of doctors' diagnosis may be affected. Second, due to the rapid early progression of BAD, many primary hospitals lack experienced neurologists, resulting in some BAD patients failing to receive timely screening and missing the best treatment opportunity. Therefore, developing an automated BAD diagnosis method based on DWI images is of great significance for improving the early diagnosis rate and clinical accessibility of BAD.

[0005] Artificial intelligence has been widely used in the field of stroke, including onset time determination, acute phase lesion automatic segmentation, penumbra identification, auxiliary decision-making, and prognosis prediction. These functions rely on neuroimaging data and need to integrate clinical data and treatment non-imaging information. In the acute phase, clinicians need to integrate a large amount of data and make quick decisions, while trained AI can quickly provide expert-level image interpretation suggestions. AI can combine and summarize the clinical and imaging features of patients and compare them with optimized big data models, providing important support for clinical decision-making. This not only improves the efficiency and accuracy of diagnosis, but also reduces the workload of doctors, allowing them to focus more on the overall care of patients.

[0006] Due to the late introduction of the BAD diagnostic criteria, the acquisition of relevant DWI (Diffusion weighted imaging) data was limited before its establishment. This delay has made researchers face the problem of insufficient data when conducting relevant research, which cannot be fully analyzed and verified. In addition, the limited amount of data also makes it difficult to train and verify AI models, further restricting the promotion of intelligent auxiliary systems in the field of stroke. Therefore, designing an effective BAD auxiliary diagnosis system for a limited number of DWI images is a problem to be solved at present and has important practical application value. SUMMARY

[0007] Therefore, the present application provides a small sample-based BAD identification model construction method and BAD identification method to solve the problem of how to use limited DWI images for BAD auxiliary diagnosis.

[0008] In a first aspect, the present application provides a small sample-based BAD identification model construction method, which comprises: obtaining diffusion weighted imaging data of BAD patients and non-BAD patients; extracting a blood supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determining whether the blood supply area slice is a lesion slice or a non-lesion slice; constructing a first data set based on the characteristics of the BAD patients and the non-BAD patients through data augmentation; performing data augmentation and fusion based on the lesions in the lesion slices and the non-lesion slices to obtain a second data set, the second data set comprising negative control samples of non-BAD patients; pre-training a visual Transformer model using the first data set; fine-tuning the pre-trained model using the blood supply area slices of the BAD patients and the second data set to obtain a BAD identification model.

[0009] In the present application, through data augmentation, a first data set is first constructed as a training sample for model pre-training, and then a simulated lesion conforming to the clinical pathological rule is generated by combining the PPA region vascular anatomy features to obtain a second data set, and the pre-trained model is efficiently fine-tuned using the second data set, which significantly improves the model generalization ability under small sample conditions. At the same time, the finally constructed BAD identification model can automatically analyze the image features and output the diagnosis result. Moreover, through the two-stage training strategy, the powerful feature extraction capability of the pre-trained model is retained, and the model is well adapted to the BAD diagnosis task through a small amount of parameter adjustment, effectively improving the model performance and generalization ability under small sample data conditions.

[0010] In an optional implementation, the method for extracting the parametric model of the paramedian artery of the pons of the BAD patient and the non-BAD patient comprises the following steps: extracting the parametric model of the paramedian artery of the pons of the BAD patient and the non-BAD patient from the diffusion weighted imaging data; and determining whether the parametric model is a lesion slice or a non-lesion slice.

[0011] In an optional implementation, after the diffusion weighted imaging data of the BAD patient and the non-BAD patient is obtained, the method further comprises the following steps: performing denoising on the diffusion weighted imaging data by using an adaptive Gaussian filtering algorithm; reconstructing the denoised data by using linear interpolation to obtain data of a size and adaptive to the target detection model; and performing standardization processing on the reconstructed data.

[0012] In the present application, the adaptive Gaussian filtering algorithm is used for denoising, which reduces imaging artifacts and enhances key information; linear interpolation is used for reconstruction, which can unify the image size and adapt the model input; and the data is standardized, which can eliminate the signal intensity difference between different scanning devices and ensure the signal consistency between different samples.

[0013] In an optional implementation, the first data set is constructed by data augmentation based on the characteristics of the BAD patient and the non-BAD patient, comprising the following steps: obtaining the parametric model of the BAD patient and the non-BAD patient based on prior knowledge; and performing parameterized modeling by using Monte Carlo simulation based on the parametric model to simulate lesion data and obtain the first data set.

[0014] In the present application, the Monte Carlo simulation is used for parameterized modeling to synthesize diversified simulated lesion data, which provides rich lesion spatial prior information for the model.

[0015] In an optional implementation, the second data set is obtained by data augmentation and fusion based on the lesions in the lesion slice and the non-lesion slice, comprising the following steps: extracting the lesions in the lesion slice by using threshold segmentation and performing geometric transformation on the lesions; and fusing the geometrically transformed lesions and the non-lesion slice by using a Gaussian pyramid fusion method to generate negative control samples with lesions and not belonging to the BAD patient.

[0016] In the present application, the lesions are extracted and geometrically transformed, and the geometrically transformed lesions and the non-lesion slice are fused by using the Gaussian pyramid fusion method to generate negative control samples with lesions and not belonging to the BAD patient, and the model is parameter fine-tuned by using the samples, which can improve the model discrimination boundary and robustness.

[0017] In an optional implementation, the visual Transformer model is pre-trained using the first data set, including: pre-training the visual Transformer model using the first data set based on a preset loss function, the preset loss function being a weight matrix quantifying spatial positions of lesions.

[0018] In the present application, the weight matrix quantifying spatial positions of lesions is used as a loss function of the pre-training model, which improves the discrimination effect between spatial related categories at the initial stage of training, so that the model can learn more interpretable feature representations.

[0019] In a second aspect, the present application provides a BAD identification method, which is applied to the BAD identification model construction method based on small samples of the first aspect and any one of the first aspect, and the method includes: obtaining diffusion weighted imaging data to be identified; extracting a blood supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determining whether the blood supply area slice is a lesion slice or a non-lesion slice; when the blood supply area slice is a lesion slice, inputting the blood supply area slice into the BAD identification model to obtain a BAD identification result; and when the blood supply area slice is a non-lesion slice, outputting a negative identification result.

[0020] In a third aspect, the present application provides a BAD identification model construction device based on small samples, which includes: a first data acquisition module for acquiring diffusion weighted imaging data of BAD patients and non-BAD patients; a first blood supply area extraction and identification module for extracting a blood supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determining whether the blood supply area slice is a lesion slice or a non-lesion slice; a first augmentation module for constructing a first data set through data augmentation based on the characteristics of the BAD patients and the non-BAD patients; a second augmentation module for data augmentation and fusion based on the lesions in the lesion slices and the non-lesion slices to obtain a second data set, the second data set including negative control samples of the non-BAD patients; a pre-training module for pre-training a visual Transformer model using the first data set; and a model construction module for fine-tuning the pre-trained model using the blood supply area slices of the BAD patients and the second data set to obtain a BAD identification model.

[0021] In a fourth aspect, the present application provides a BAD identification device, which is applied to the BAD identification model construction method based on small samples of the first aspect and any one of the first aspect of the present application. The device comprises: a second data acquisition module, configured to acquire diffusion weighted imaging data to be identified; a second blood supply area extraction and identification module, configured to extract a blood supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determine whether the blood supply area slice is a lesion slice or a non-lesion slice; a first identification module, configured to input the blood supply area slice into the BAD identification model when the blood supply area slice is the lesion slice, and obtain a BAD identification result; and a second identification module, configured to output a negative identification result when the blood supply area slice is the non-lesion slice.

[0022] In a third aspect, the present application provides a computer device, comprising a memory and a processor, which are in communication connection with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the BAD identification model construction method based on small samples of the first aspect or any one of the corresponding embodiments thereof or the BAD identification method of the second aspect.

[0023] In a fourth aspect, the present application provides a computer readable storage medium, which stores computer instructions, and the computer instructions are used to make the computer execute the BAD identification model construction method based on small samples of the first aspect or any one of the corresponding embodiments thereof or the BAD identification method of the second aspect.

[0024] In a fifth aspect, the present application provides a computer program product, which comprises computer instructions, and the computer instructions are used to make the computer execute the BAD identification model construction method based on small samples of the first aspect or any one of the corresponding embodiments thereof or the BAD identification method of the second aspect. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0026] Figure 1 is a flowchart of the BAD identification model construction method based on small samples according to an embodiment of the present application;

[0027] Figure 2 is a flowchart of the BAD identification method according to an embodiment of the present application;

[0028] Figure 3is a schematic diagram of a BAD auxiliary diagnosis system according to an embodiment of the present application;

[0029] Figure 4 is a schematic diagram of a workflow of a BAD auxiliary diagnosis system according to an embodiment of the present application;

[0030] Figure 5 is a structural block diagram of a small-sample-based BAD recognition model construction device according to an embodiment of the present application;

[0031] Figure 6 is a structural block diagram of a BAD recognition device according to an embodiment of the present application;

[0032] Figure 7 is a schematic diagram of a hardware structure of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions and advantages of embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0034] According to an embodiment of the present application, a small-sample-based BAD recognition model construction method is provided. It should be noted that the steps shown in the flowchart can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0035] In this embodiment, a small-sample-based BAD recognition model construction method is provided, which can be used in electronic devices such as computers, mobile phones, tablet computers, etc. Figure 1 is a flowchart of a small-sample-based BAD recognition model construction method according to an embodiment of the present application, as shown in Figure 1 The flowchart includes the following steps:

[0036] In step S101, the diffusion weighted imaging data of BAD patients and non-BAD patients is obtained. The diffusion weighted imaging data can be a non-open data set obtained through communication with a Branch Atheromatous Disease (BAD) research team. Since there are few current BAD diagnoses based on imaging such as diffusion weighted imaging (DWI) data, the data of BAD patients in the diffusion weighted imaging data obtained in this embodiment is also less, that is, this application can also be understood as a small sample based BAD auxiliary diagnosis and recognition.

[0037] Specifically, the diffusion weighted imaging data obtained in this embodiment includes the BAD-Study data set and other SMART-Study data set sorted by the Union Hospital. These data sets contain DWI images and related pathological information of BAD patients and control samples. These data sets need to be sorted for subsequent use. The images in these data sets are mostly stored in DICOM (Digital Imaging and Communications in Medicine) format or NifTI (Neuroimaging Informatics Technology Initiative) format. Based on this, after obtaining the data set, the DICOM / NIfTI image can be visualized and checked using the ITK-SNAP tool to ensure data quality and extract the DWI image therefrom. Subsequently, a data reading script is written to parse the patient information in the image file and match it with the pathological labels provided by BAD-Study and SMART-Study to ensure the integrity and accuracy of the data. The pathological labels include whether the patient is a BAD patient and the corresponding patient information. Thus, through the above data set processing process, the DWI images of BAD patients and non-BAD patients and the corresponding pathological labels are obtained.

[0038] Step S102, a slice of a blood supply area of a paramedian pontine artery of the diffusion weighted imaging data is extracted, and it is judged whether the slice of the blood supply area is a lesion slice or a non-lesion slice. Specifically, in the diagnosis of BAD, the PPA (paramedian pontine arteries) region is one of the regions where BAD mainly occurs, and the position, shape and size of the lesion are also important reference for diagnosing BAD. Based on this, for the obtained diffusion weighted imaging data, the PPA blood supply region is first identified and extracted, and then the extracted region is identified as having or not having a lesion, so as to realize the extraction of the blood supply area of the obtained DWI image data and the lesion identification, and provide a data basis for the subsequent construction of the training data set.

[0039] Step S103, based on the characteristics of the BAD patients and the non-BAD patients, a first data set is constructed by data augmentation. Specifically, the characteristics of the BAD patients and the non-BAD patients can be obtained by prior knowledge, for example, the corresponding characteristics can be determined based on the description of the blood supply area characteristics of the BAD patients and the non-BAD patients by the clinicians, or the corresponding characteristics can be obtained through the Internet and the like. For the obtained characteristics, it is usually insufficient to train the model as a training set. Therefore, the data augmentation technology is used to perform a series of transformation operations on the characteristics to obtain the first data set.

[0040] Step S104, based on the lesions in the lesion slices and the non-lesion slices, data augmentation and fusion are performed to obtain a second data set, and the second data set includes negative control samples of non-BAD patients. Specifically, because the obtained diffusion weighted imaging data belongs to small sample data, the obtained lesion slices and non-lesion slices also cannot meet the demand of training samples. Based on this, the lesions in the lesion slices are extracted, and data augmentation and fusion with the non-lesion slices are performed, so as to obtain a series of negative control samples with lesion characteristics but not meeting the diagnosis standard of BAD.

[0041] Step S105, the first data set is used to pre-train the visual Transformer model.

[0042] Step S106, the blood supply area slice of the BAD patient and the second data set are used to fine-tune the pre-trained model to obtain a BAD recognition model.

[0043] Specifically, few-shot learning is an important machine learning method, which aims to train a model through very few sample data. Its significance lies in effectively dealing with data scarcity, quickly adapting to new tasks, reducing labeling costs, improving model interpretability, and promoting the wide application of AI technology. Especially in the field of medical images, it is usually expensive and time-consuming to obtain a large amount of labeled data. Few-shot learning enables the model to learn and identify new categories in the case of limited data, thereby promoting the popularization and development of technology in resource-limited environments.

[0044] When the BAD recognition model is constructed by using the few-shot learning in the embodiment, the transfer learning technology is used to significantly improve the generalization ability of the model under the condition of few samples by efficiently fine-tuning the pre-trained model. For the pre-trained model, the data in the first data set is used to pre-train the visual Transformer model, and at the same time, in order to reduce the number of parameters in the model, the Tiny-ViT model (i.e., the tiny visual Transformer model) can be used to replace the visual Transformer model. After obtaining the pre-trained model, the blood supply area slice of the BAD patient and the second data set are used to fine-tune the pre-trained model to obtain the BAD recognition model.

[0045] Therefore, through the above two-stage model training process, the powerful feature extraction ability of the pre-trained model is retained, and the model is well adapted to the BAD diagnosis task through a small amount of parameter adjustment, effectively improving the model performance and generalization ability under the condition of few sample data.

[0046] The few-shot based BAD recognition model construction method provided by the embodiment of the application constructs a first data set as a training sample for model pre-training through data augmentation, then generates simulated lesions conforming to the clinical pathological law in combination with the PPA area vascular anatomy features to obtain a second data set, and efficiently fine-tunes the pre-trained model by using the second data set, thereby significantly improving the generalization ability of the model under the condition of few samples. At the same time, the finally constructed BAD recognition model can automatically analyze the image features and output the diagnosis result.

[0047] In the embodiment, a few-shot based BAD recognition model construction method is provided, and the flow includes the following steps:

[0048] In step S201, the diffusion weighted imaging data of the BAD patient and the non-BAD patient is obtained. For details, please refer to Figure 1 The step S104 of the embodiment shown in the figure will not be repeated here.

[0049] Step S202, the diffusion weighted imaging data is denoised by adaptive Gaussian filtering algorithm; the denoised data is reconstructed by linear interpolation to obtain data of a size and a target detection model.

[0050] Specifically, for the obtained diffusion weighted imaging data, in order to improve the data quality, the adaptive Gaussian filtering algorithm is used for denoising processing, and the process can be realized by using the following formula:

[0051]

[0052] wherein, is the pixel value at the coordinate after filtering; is the pixel value at the coordinate in the original image; is a Gaussian kernel parameter, which determines the shape of the Gaussian function and further affects the degree of filtering. The larger the value is, the stronger the filtering effect is, and the smoother the image is, but more details may be lost. The Gaussian kernel parameter can be dynamically calculated according to the image signal-to-noise ratio. Specifically, the noise in the image affects the stability of the signal intensity. If the noise is more, the fluctuation of the signal intensity will be larger, that is, the standard deviation is relatively large; and the mean value of the signal intensity reflects the average signal level of the region. Therefore, the ratio of the mean value of the signal intensity to the standard deviation can reflect the signal-to-noise ratio of the image to some extent. When the calculated signal-to-noise ratio of the image is large, it means that the noise is more, and at this time, the value of the Gaussian kernel parameter can be larger, so that the image can be more fully smoothed by Gaussian filtering to remove noise; when the signal-to-noise ratio is small, it means that the noise is small, and at this time, the value of the Gaussian kernel parameter can be small, so as to avoid excessive denoising leading to loss of image details.

[0053] In addition, the obtained diffusion weighted imaging data can come from many sources and have different sizes, therefore, in order to unify the input specifications and adapt to the subsequent model, the embodiment uses linear interpolation to reshape the image. In the embodiment, bilinear interpolation is specifically used to ensure the smoothness and quality of the image after transformation in the two-dimensional plane. For the specific processing process of bilinear interpolation, related technologies can be referred to for implementation. Specifically, the size of the reconstructed image is a standard size (224x224x3), which has been verified by experiments to balance the demand for calculation efficiency and lesion detail preservation.

[0054] To ensure the signal consistency between different samples, the acquired diffusion weighted imaging data is further standardized. Specifically, Z-score standardization based on global gray value can be used to eliminate the signal intensity difference between different scanning devices. The Z-score standardization based on global gray value is a method of standardizing image gray values, that is, the Z-score standardization is performed after the mean and standard deviation of the global gray value of the image are calculated respectively.

[0055] In step S203, the PPA supply area slice of the diffusion weighted imaging data is extracted, and it is judged whether the supply area slice is a lesion slice or a non-lesion slice.

[0056] Specifically, the above step S203 includes:

[0057] In step S2031, a pre-trained target detection model is used to extract the PPA supply area slice of the diffusion weighted imaging data. The target detection model is mainly used for positioning and extracting the PPA supply area in the diffusion weighted imaging data. Specifically, the target detection model can use a model capable of target detection in related technologies, such as Faster R-CNN or YOLO series network model, etc. In this embodiment, YOLOv5m is selected as the final model by balancing the detection accuracy and calculation efficiency. In addition, before use, the model is trained using image samples labeled with the supply area to enable the model to accurately identify the supply area. After the trained target detection model identifies and extracts the supply area slice, the accuracy of the supply area slice can also be manually audited.

[0058] In step S2032, a pre-trained residual network model is used to judge whether the supply area slice is a lesion slice or a non-lesion slice. Specifically, the identification of whether the supply area slice is a lesion slice or a non-lesion slice can be realized by a binary classification network. In this embodiment, a residual network such as ResNet-18 is used to realize the binary classification function. The ResNet-18 network belongs to a neural network with a certain depth, and in the binary classification task, it can learn more complex and abstract features, such as more subtle features for images with or without lesions, thereby improving the classification performance. At the same time, the network also introduces a residual block structure, which makes the network easier to learn the identity mapping through residual connection. This connection method can alleviate the gradient vanishing problem, so that the gradient can be more effectively backpropagated during training, and the deep network is also easy to train.

[0059] In addition, before classification is performed by using the residual network model, supervised learning can be performed by using the labeled lesion information to improve the reliability of the classification. Meanwhile, for the slices classified as having lesions and the slices classified as not having lesions, a corresponding database can be constructed to provide data support for subsequent data augmentation.

[0060] In step S204, a first data set is constructed through data augmentation based on the characteristics of the BAD patients and the non-BAD patients.

[0061] Specifically, the step S204 includes:

[0062] In step S2041, the blood supply area characteristics of the BAD patients and the non-BAD patients are obtained based on prior knowledge. The blood supply area characteristics can include morphology, size range, distribution position, and the size, morphology, spatial distribution of the lesions contained therein.

[0063] In step S2042, parameterized modeling is performed by using Monte Carlo simulation based on the blood supply area characteristics to simulate lesion data, and a first data set is obtained. The Monte Carlo method is a method for solving problems by using random numbers to perform a large number of repeated sampling experiments based on probability and statistics theory. Specifically, in the parameterized modeling by using Monte Carlo simulation, an elliptical bright spot is randomly generated on a black background to simulate lesions of different morphologies, sizes, and spatial distributions. By controlling the parameters of the bright spot, this method can synthesize diversified simulated lesion data, construct a support sample set (i.e., the first data set) covering a wide range of BAD manifestations, and provide rich spatial prior information of the lesions for the model.

[0064] In step S205, data augmentation and fusion are performed based on the lesions in the slices having lesions and the slices not having lesions, and a second data set is obtained. The second data set includes negative control samples of the non-BAD patients.

[0065] Specifically, the step S205 includes:

[0066] In step S2051, the lesions in the slices having lesions are extracted by using threshold segmentation, and geometric transformation is performed on the lesions. Specifically, for the obtained slices having lesions, in addition to the lesions, other areas of the blood supply area are also included. Therefore, threshold segmentation is used to extract the lesions in the lesion slices. For example, Otsu threshold method or adaptive threshold method can be used to process the slices having lesions to generate a binary image, such as 0 representing the background and 1 representing the lesions, and then the lesion mask represented as 1 is extracted.

[0067] To realize data augmentation of the extracted lesions, geometric transformation is performed on the extracted lesions in this embodiment, such as rotation, scaling, and translation of the lesions, to obtain the lesions after different transformation modes.

[0068] Step S2052, the Gaussian pyramid fusion method is used to fuse the lesion after geometric transformation and the slice without lesion to generate a negative control sample with lesion and not belonging to the BAD patient. Among them, the Gaussian pyramid is constructed by Gaussian blur and downsampling of the image to build different resolution levels (the bottom layer is high resolution and the top layer is low resolution), so as to realize multi-scale feature fusion. Specifically, the embodiment first constructs a Gaussian pyramid for the slice without lesion, then adjusts the lesion after geometric transformation to the same size as each layer of the pyramid, and simulates the blur degree of different lesions by Gaussian blur, then in each layer of the pyramid, the slice without lesion and the lesion are superimposed and fused, and finally each layer of the pyramid image is upsampled and superimposed to generate the final negative control sample. In addition, for the generated negative control sample, it can also be determined that it does not meet the BAD diagnostic criteria in an artificial review manner or other preset standards.

[0069] Step S206, the visual Transformer model is pre-trained using the first data set. Specifically, the visual Transformer model is pre-trained using the first data set based on a preset loss function, and the preset loss function is a weight matrix quantifying the spatial position of the lesion. Among them, in the feeding area, when the lesion is close to the middle position (i.e. Mid lesion), its probability of belonging to BAD is higher, and when the lesion is on the left or right position (i.e. Left lesion and Right lesion), its probability of belonging to BAD is lower. Therefore, in this preset loss function, the punishment of Mid lesion is smaller, while the punishment of misclassification of Left lesion and Right lesion is heavier.

[0070] Thus, the embodiment can improve the discrimination effect between spatially related categories in the early training stage by using the spatial prior relationship between lesions to construct a correlation loss, so that the model can learn more interpretable feature representations.

[0071] Step S207, the pre-trained model is fine-tuned using the feeding area slice of the BAD patient and the second data set to obtain a BAD recognition model. Among them, when performing parameter fine-tuning, the embodiment freezes most of the pre-training parameters of the model, and only performs efficient parameter fine-tuning through the inserted lightweight adapter module (Adapter), and uses the feeding area slice of the BAD patient and the second data set for training. Among them, the adapter module adopts a bottleneck structure to realize the optimization and adjustment of the model parameters without increasing the computational overhead.

[0072] In the embodiment, a BAD recognition method is also provided, which is applied to the small sample based BAD recognition model construction of the above-mentioned embodiments, as shown in Figure 2 The method comprises:

[0073] In step S301, diffusion-weighted imaging data to be identified is acquired; wherein the diffusion-weighted imaging data to be identified can be diffusion-weighted imaging data of a suspected BAD patient, and the BAD identification method can be used for auxiliary diagnosis. In other embodiments, the diffusion-weighted imaging data can also be other data that needs to be identified for BAD, which is not limited in the embodiment. In addition, the acquired diffusion-weighted imaging data can also be subjected to adaptive Gaussian filtering denoising, linear interpolation reconstruction, standardization and other processing.

[0074] In step S302, a PPA blood supply area slice of the diffusion-weighted imaging data is extracted, and it is judged whether the blood supply area slice is a lesion slice or a non-lesion slice; specifically, the YOLOv5 model can be used to locate the PPA blood supply area and extract the blood supply area slice. Then the ResNet-18 classifier is used to judge whether the blood supply area slice is a lesion slice or a non-lesion slice.

[0075] In step S303, when the blood supply area slice is a lesion slice, the blood supply area slice is input into the BAD identification model to obtain a BAD identification result; specifically, for the identified lesion slice, it is input into the BAD identification model constructed in the above embodiment to obtain a BAD identification result, which includes a diagnosis result of BAD positive or BAD negative, and the model can also provide a corresponding confidence score for clinical reference.

[0076] In step S304, when the blood supply area slice is a non-lesion slice, a negative identification result is output. Specifically, for the identified non-lesion slice, a BAD negative identification result can be directly obtained.

[0077] In view of the problem of medical annotation data scarcity, a dual-modal data enhancement strategy is adopted in the present application: that is, the PPA region blood vessel anatomical features are combined to generate simulated lesions conforming to the clinical pathological rules, and a random simulation method is used to synthesize diversified training samples. In the model training stage, the transfer learning technology is used, and the pre-trained model is efficiently fine-tuned to significantly improve the model generalization ability under the condition of small sample. Finally, the intelligent diagnosis system can automatically analyze the image features and output the diagnosis result. Through multi-center clinical verification, the present application has significantly improved the detection sensitivity and specificity compared with the traditional diagnosis method, and is particularly suitable for the identification of early micro-lesions, and provides an efficient and reliable intelligent solution for the clinical diagnosis of PPA-BAD.

[0078] The present application can effectively diagnose BAD lesions under limited medical image data conditions, improve diagnosis efficiency, reduce the workload of doctors, and realize more accurate personalized diagnosis and treatment.

[0079] In the embodiment, a BAD auxiliary diagnosis system is also provided, as shown inFigure 3 As shown, the system includes:

[0080] The data preprocessing module is used to acquire diffusion-weighted imaging (DWI) data and corresponding pathological labels from the raw data, and to perform quality control and standardization on them.

[0081] The PPA region lesion detection module is used to locate the PPA blood supply area and extract the ROI using a target detection model such as YOLOv5, and to perform preliminary screening using a classification model such as ResNet-18 to determine whether there are lesions in the slice.

[0082] The data augmentation module is used to generate synthetic lesion data based on prior medical knowledge and Monte Carlo simulation to expand the training samples;

[0083] The diagnostic model training module is used to employ an efficient parameter fine-tuning method. Based on the pre-trained Tiny-ViT model, it introduces an adapter module to optimize the small-sample classification performance using augmented data. In other words, it uses augmented data to efficiently fine-tune the parameters of the pre-trained Tiny-ViT model.

[0084] The testing module is used to build a clinical diagnostic framework, evaluate the performance of the diagnostic framework using independent clinical data such as DWI imaging data, output BAD diagnostic results, and ensure clinical usability.

[0085] In one alternative implementation, such as Figure 4 As shown, the system can operate according to the following process:

[0086] Step 1, data preprocessing (corresponding to the data preprocessing module).

[0087] The data in this embodiment comes from the BAD-Study and SMART-Study compiled by Peking Union Medical College Hospital, covering DWI images and related pathological information of clinically diagnosed BAD patients and control samples. First, NTK-SNAP was used to visualize the DICOM / NIfTI images to ensure data quality, and DWI images were extracted from them. Then, a data reading script was written to parse the patient information in the image files and match it with the pathological labels provided by BAD-Study and SMART-Study to ensure data integrity and accuracy.

[0088] To optimize image quality, an adaptive Gaussian filtering algorithm was used for denoising, reducing imaging artifacts and enhancing key information. Subsequently, linear interpolation was used to reshape the images to fit the model input size, and DWI signal intensity was standardized to ensure signal consistency among different samples, providing high-quality data input for subsequent analysis.

[0089] Step 2: PPA region lesion detection (corresponding to the PPA region lesion detection module).

[0090] The PPA blood supply area is extracted from the DWI image, and lesion screening is performed to construct a high-quality blood supply area image dataset. This module first preprocesses the DWI image to improve image quality, and uses a pre-trained YOLOv5 target detection model (including YOLOv5s, YOLOv5m, YOLOv5l, etc.) for automatic positioning of the PPA blood supply area. According to the trade-off between detection accuracy and computational efficiency, YOLOv5m is preferred as the final model, and the detection results are manually reviewed to ensure the accuracy of the region positioning.

[0091] After the blood supply area is extracted, ResNet-18 is further trained for lesion screening, and the extracted PPA region slices are classified into two categories to distinguish between lesion-containing and lesion-free slices. During the screening process, the lesion information labeled by doctors is used for supervised learning to improve the reliability of the classification. Finally, the selected lesion-free slices are constructed into a lesion-free sample library to support subsequent data balancing processing and ensure the data quality of the BAD diagnosis model training.

[0092] Step 3: Data augmentation (corresponding to the data augmentation module).

[0093] To enhance the discriminative ability and generalization performance of the BAD diagnosis model in small sample scenarios, a multi-strategy data augmentation module is designed. First, based on the typical features of PPA region BAD in medical prior knowledge, parameterized modeling is performed using Monte Carlo simulation to randomly generate elliptical bright spots on a black background, simulating lesions of different shapes, sizes, and spatial distributions. By controlling the parameters of the bright spots, this method can synthesize diverse simulated lesion data, constructing a support sample set covering a wide range of BAD manifestations and providing rich spatial prior information for the model.

[0094] Subsequently, the lesion area mask is extracted from the real blood supply area image (with lesion-containing slices) using threshold segmentation, and geometric transformations (such as rotation, scaling, and translation) are performed on it. Combined with the lesion-free image, a Gaussian pyramid fusion technique is used to generate negative control samples with lesion characteristics but not meeting the BAD diagnosis criteria, to improve the discriminative boundary and robustness of the model.

[0095] Through the collaborative application of the above parameter modeling and image synthesis strategies, the diversity, complexity, and structural prior of the training samples are fully expanded, effectively improving the stability and generalization ability of the model in actual clinical environments.

[0096] Step 4: Diagnosis model training (corresponding to the diagnosis model training module).

[0097] The diagnostic model training module adopts a pre-trained Tiny-ViT model as a basic network architecture. First, the model is pre-trained using a large amount of simulated data generated by parameterized modeling, and a correlation loss is constructed using the spatial prior relationship between lesions to improve the discrimination effect between spatially related categories in the early training stage, so that the model can learn more interpretable feature representations. Then, most of the pre-training parameters of the model are frozen, and only the inserted lightweight adapter module (Adapter) is used for efficient parameter fine-tuning, and the combined data of synthetic non-BAD samples and real BAD samples are used for training. This two-stage training strategy not only retains the powerful feature extraction capability of the pre-trained model, but also makes the model well adapted to the BAD diagnosis task through a small amount of parameter adjustment, effectively improving the model performance and generalization ability under small sample data conditions. The adapter module adopts a carefully designed bottleneck structure to realize the optimization and adjustment of model parameters without increasing the computational overhead.

[0098] Step 5: Test and clinical evaluation (corresponding to the test module).

[0099] The test module integrates the trained neural network to build an end-to-end diagnosis system, which specifically includes three core processing links: first, the optimized YOLOv5 model is used to locate the PPA blood supply area and extract standardized slices, then the trained ResNet-18 classifier is used to filter the lesion-containing slices, and finally the fine-tuned Tiny-ViT model is used to output the BAD detection results. The system input is the pre-processed DWI image, which is directly output after the above automatic analysis process. The binary classification BAD diagnosis conclusion (positive / negative) is output, and the corresponding confidence score is provided for clinical reference.

[0100] The system of the present application has the following advantages: 1) It integrates the training results of all previous modules to form a coherent automatic analysis process; 2) It uses a cascading network architecture to gradually refine the analysis granularity; 3) It outputs a simple and clear BAD diagnosis conclusion to meet the actual needs of clinical practice. The system realizes efficient automatic diagnosis while ensuring analysis accuracy, providing reliable support for clinical decision-making.

[0101] In this embodiment, a small sample-based BAD recognition model construction device is also provided, which is used to implement the above embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware implementation is also possible and contemplated.

[0102] The present embodiment provides a small sample-based BAD recognition model construction device, which isFigure 5 As shown in the figure, the device comprises:

[0103] The first data acquisition module 51 is configured to acquire diffusion weighted imaging data of BAD patients and non-BAD patients.

[0104] The first blood supply area extraction and identification module 52 is configured to extract a blood supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determine whether the blood supply area slice is a lesioned slice or a non-lesioned slice.

[0105] The first augmentation module 53 is configured to construct a first data set through data augmentation based on the characteristics of the BAD patients and the non-BAD patients.

[0106] The second augmentation module 54 is configured to perform data augmentation and fusion based on the lesions in the lesioned slices and the non-lesioned slices, to obtain a second data set, and the second data set comprises negative control samples of the non-BAD patients.

[0107] The pre-training module 55 is configured to pre-train a visual Transformer model using the first data set.

[0108] The model construction module 56 is configured to fine-tune the pre-trained model using the blood supply area slices of the BAD patients and the second data set, to obtain a BAD identification model.

[0109] The embodiment also provides a BAD identification device, which is applied to the small sample based BAD identification model construction method of the above embodiment. Figure 6 As shown in the figure, the device comprises:

[0110] The second data acquisition module 61 is configured to acquire diffusion weighted imaging data to be identified.

[0111] The second blood supply area extraction and identification module 62 is configured to extract a blood supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determine whether the blood supply area slice is a lesioned slice or a non-lesioned slice.

[0112] The first identification module 63 is configured to, when the blood supply area slice is a lesioned slice, input the blood supply area slice into the BAD identification model, to obtain a BAD identification result.

[0113] The second identification module 64 is configured to, when the blood supply area slice is a non-lesioned slice, output a negative identification result.

[0114] Further function descriptions of the above modules are the same as those of the above corresponding embodiments, and will not be described here.

[0115] The embodiment also provides a computer device, which has the small sample based BAD identification model construction device shown in the above Figure 5 or the BAD identification device shown in the aboveFigure 6 The BAD identification device shown.

[0116] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 7 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 7 Take a processor 10 as an example.

[0117] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0118] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0119] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device as shown by a landing page for an app. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0120] The memory 20 can include a volatile memory, such as a random access memory, and / or a non-volatile memory, such as a flash memory, a hard disk, or a solid state disk. The memory 20 can also include a combination of the above-mentioned types of memories.

[0121] The computer device also includes a communication interface 30 for the computer device to communicate with other devices or communication networks.

[0122] The embodiments of the present application also provide a computer readable storage medium, and the method according to the embodiments of the present application can be implemented in hardware, firmware, or recorded in a storage medium, or stored in a remote storage medium or a non-transitory machine readable storage medium and downloaded to a local storage medium through network, so that the method described herein can be processed by such software on a storage medium using a general purpose computer, a special purpose processor, or programmable or special purpose hardware. The storage medium can be a disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid state disk, etc. Further, the storage medium can also include a combination of the above-mentioned types of memories. It can be understood that the computer, the processor, the microprocessor controller, or the programmable hardware includes a storage component that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the method shown in the above embodiments is implemented.

[0123] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, through the operation of the computer, the method and / or technical solutions according to the present application can be invoked or provided. Those skilled in the art should understand that the form of computer program instructions in a computer readable medium includes but is not limited to source files, executable files, installation package files, etc. Correspondingly, the way of computer program instructions executed by computer includes but is not limited to: the computer directly executes the instructions, or the computer compiles the instructions and then executes the corresponding compiled program, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer readable medium can be any available computer readable storage medium or communication medium accessible to the computer.

[0124] Although the embodiments of the present application are described in conjunction with the accompanying drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A method for constructing a small sample-based BAD identification model, characterized in that, The method comprises: obtaining diffusion weighted imaging data of BAD patients and non-BAD patients; extracting a supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determining whether the supply area slice is a lesion slice or a non-lesion slice; constructing a first data set through data augmentation based on the characteristics of the BAD patients and the non-BAD patients; performing data augmentation and fusion based on the lesions in the lesion slices and the non-lesion slices to obtain a second data set, wherein the second data set comprises negative control samples of the non-BAD patients; pre-training a visual Transformer model using the first data set; fine-tuning the pre-trained model using the supply area slices of the BAD patients and the second data set to obtain a BAD recognition model; wherein the first data set is constructed through data augmentation based on the characteristics of the BAD patients and the non-BAD patients, comprising: obtaining supply area characteristics of the BAD patients and the non-BAD patients based on prior knowledge; based on the supply area characteristics, performing parameterized modeling using Monte Carlo simulation to simulate lesions of different morphologies, sizes and spatial distributions by randomly generating elliptical bright spots on a black background to obtain the first data set.

2. The method of claim 1, wherein, extracting a supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determining whether the supply area slice is a lesion slice or a non-lesion slice, comprising: extracting the supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data using a pre-trained object detection model; using a pre-trained residual network model to determine whether the supply area slice is a lesion slice or a non-lesion slice.

3. The method of claim 2, wherein, After obtaining the diffusion weighted imaging data of the BAD patients and the non-BAD patients, the method further comprises: performing denoising on the diffusion weighted imaging data using an adaptive Gaussian filtering algorithm; reconstructing the denoised data using linear interpolation to obtain data suitable for the size and the object detection model; performing standardization processing on the reconstructed data.

4. The method of claim 1, wherein, performing data augmentation and fusion based on the lesions in the lesion slices and the non-lesion slices to obtain a second data set, comprising: extracting the lesions in the lesion slices using threshold segmentation, and performing geometric transformation on the lesions; using a Gaussian pyramid fusion method to fuse the geometrically transformed lesions and the non-lesion slices to generate negative control samples that have lesions but do not belong to the BAD patients.

5. The method of claim 1, wherein, pre-training a visual Transformer model using the first data set, comprising: based on a preset loss function, pre-training the visual Transformer model using the first data set, wherein the preset loss function is a weight matrix quantifying the spatial position of the lesions.

6. A method of BAD identification, comprising: The method comprises: obtaining diffusion weighted imaging data to be recognized; extracting a supply area slice of the paramedian artery of the pontis of the diffusion weighted imaging data, and determining whether the supply area slice is a lesion slice or a non-lesion slice; when the supply area slice is a lesion slice, inputting the supply area slice into the BAD recognition model to obtain a BAD recognition result; When the blood supply area slice is a non-lesion slice, a negative recognition result is output.

7. A small sample based BAD identification model construction device, characterized in that, The device comprises: A first data acquisition module is configured to acquire diffusion-weighted imaging data of BAD patients and non-BAD patients. A first blood supply area extraction and recognition module is configured to extract a blood supply area slice of the paramedian artery of the pontis of the diffusion-weighted imaging data and determine whether the blood supply area slice is a lesion slice or a non-lesion slice. A first augmentation module is configured to construct a first data set through data augmentation based on the characteristics of the BAD patients and the non-BAD patients. A second augmentation module is configured to perform data augmentation and fusion based on the lesions in the lesion slices and the non-lesion slices to obtain a second data set, wherein the second data set comprises negative control samples of the non-BAD patients. A pre-training module is configured to pre-train a visual Transformer model using the first data set. A model construction module is configured to fine-tune the pre-trained model using the blood supply area slices of the BAD patients and the second data set to obtain a BAD recognition model. The first data set constructed based on the characteristics of the BAD patients and the non-BAD patients through data augmentation comprises: acquiring blood supply area characteristics of the BAD patients and the non-BAD patients based on prior knowledge; based on the blood supply area characteristics, performing parameterized modeling using Monte Carlo simulation to simulate lesions of different morphologies, sizes, and spatial distributions by randomly generating elliptical bright spots on a black background, and obtaining the first data set.

8. A BAD identification device, characterized by The device is applied to the small-sample-based BAD recognition model construction method of any one of claims 1-5, and the device comprises: A second data acquisition module is configured to acquire diffusion-weighted imaging data to be recognized. A second blood supply area extraction and recognition module is configured to extract a blood supply area slice of the paramedian artery of the pontis of the diffusion-weighted imaging data and determine whether the blood supply area slice is a lesion slice or a non-lesion slice. A first recognition module is configured to input the blood supply area slice to the BAD recognition model when the blood supply area slice is a lesion slice to obtain a BAD recognition result. A second recognition module is configured to output a negative recognition result when the blood supply area slice is a non-lesion slice.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing a computer to execute the small-sample-based BAD recognition model construction method of any one of claims 1-5 or the BAD recognition method of claim 6. The computer readable storage medium stores computer instructions for causing a computer to execute the small-sample-based BAD recognition model construction method of any one of claims 1-5 or the BAD recognition method of claim 6.

Images