Model training and detection method for cerebral vascular aneurysm detection

Abstract

This application discloses a model training and detection method for cerebral aneurysm detection. The training method includes: performing frequency domain transformation on NCCT images to decompose them into high-frequency components; applying spectral energy perturbation to the high-frequency components to generate at least one perturbed view, and combining it with the NCCT images to obtain multiple sets of input views; inputting CTA images into a pre-trained teacher detection model to obtain corresponding cerebral aneurysm detection pseudo-labels and intermediate-layer teacher features; inputting all input views into a student detection model to be trained to obtain cerebral aneurysm detection results and intermediate-layer student features for each set of input views; and training the student detection model based on the cerebral aneurysm detection pseudo-labels, intermediate-layer teacher features, cerebral aneurysm detection results, and intermediate-layer student features. The student detection model trained by this application can improve the detection accuracy of cerebral aneurysms in contrast-free scenarios.

Description

Technical Field

[0001] This application generally relates to the field of computer vision technology. More specifically, this application relates to a method and electronic device for training a student detection model. Furthermore, this application also relates to a method and electronic device for detecting cerebral aneurysms. Background Technology

[0002] Cerebral aneurysms are a common cerebrovascular disease in clinical practice. Rupture of aneurysms can easily lead to subarachnoid hemorrhage, resulting in extremely high rates of disability and death. Early and accurate detection of cerebral aneurysms is a crucial prerequisite for preventing aneurysm rupture and developing appropriate clinical treatment plans, and it has significant clinical importance in reducing the risk of death and disability from cerebrovascular diseases.

[0003] Currently, the gold standard for clinical diagnosis of cerebral aneurysms is computed tomography (CTA). This technique enhances vascular visualization through intravenous injection of iodine contrast agents, clearly revealing the morphology of cerebral blood vessels and aneurysm lesions. However, it has significant limitations. Iodine contrast agents pose a risk of allergic reactions and may cause nephrotoxicity in patients with renal insufficiency, thus having clearly defined contraindications. Furthermore, this examination involves a high radiation dose and is costly, making it unsuitable for large-scale population screenings and repeated examinations. Non-contrast computed tomography (NCCT) does not require contrast agent injection and offers advantages such as being non-invasive, rapid, low-cost, and low-radiation, making it suitable for screenings and for individuals contraindicated for contrast agents. However, in NCCT images, the density difference between cerebral blood vessels and surrounding brain tissue is minimal, and vascular boundaries are blurred, making it easy to miss small aneurysms. Traditional manual image interpretation heavily relies on the physician's clinical experience, resulting in a high rate of missed diagnoses. Existing deep learning-based detection models are mostly designed for vascular imaging images; when directly applied to NCCT images, their detection performance significantly degrades, failing to meet the practical needs of contrast-free clinical detection. In recent years, knowledge distillation technology has been widely applied in the field of cross-modal learning. Existing research mostly adopts a teacher-student framework, transferring knowledge from a teacher model trained on vascular imaging images to a student model trained on NCCT images. However, most existing distillation methods perform feature alignment in the spatial domain, failing to fully consider the weak information of vascular edges in NCCT images, and thus cannot effectively enhance the model's ability to perceive vascular edges and aneurysm morphology. Frequency domain image processing technology has been proven to effectively enhance image edge details, but existing solutions only apply it to general image enhancement tasks, without deep integration with the knowledge distillation framework. This prevents cross-modal feature knowledge transfer and lacks a suitable technical solution for contrast-free cerebral aneurysm detection scenarios, failing to fundamentally address the industry pain point of insufficient aneurysm detection accuracy in NCCT images.

[0004] In view of this, there is an urgent need to provide a solution for detecting cerebral aneurysms in NCCT images, so as to improve the detection accuracy of cerebral aneurysms in contrast-free scenarios and provide safe and effective detection technology support for large-scale population screening and for people who are contraindicated for contrast agents. Summary of the Invention

[0005] In order to at least address one or more of the technical problems mentioned above, this application proposes a scheme for detecting cerebral aneurysms in NCCT images in several aspects.

[0006] In a first aspect, this application provides a method for training a student detection model, characterized in that the student detection model is used for cerebral aneurysm detection in plain computed tomography (NCCT) images, and the training method includes the following operations:

[0007] Obtain a training dataset, which includes multiple sets of paired NCCT images and corresponding computed tomography angiography (CTA) images, wherein the paired NCCT images and CTA images are spatially registered images of the same subject; The NCCT image is subjected to frequency domain transformation to decompose it into high-frequency components containing information on blood vessel edges and texture details; Apply spectral energy perturbation to the high-frequency components to generate at least one perturbation view, and combine it with the NCCT image as an unperturbed view to obtain multiple sets of input views; The CTA image is input into the pre-trained teacher detection model to obtain the corresponding cerebral aneurysm detection pseudo-label and intermediate layer teacher features; Input all input views into the student detection model to be trained to obtain the cerebral aneurysm detection results and intermediate layer student features corresponding to each set of input views; Based on the pseudo-labels for cerebral aneurysm detection, intermediate-layer teacher features, cerebral aneurysm detection results, and intermediate-layer student features, a joint loss function is constructed, and the student detection model is trained using the joint loss function.

[0008] In some embodiments, the spectral energy perturbation includes at least one of random masking perturbation, random enhancement perturbation, and random phase shift perturbation; wherein, the random masking perturbation is to randomly select a preset proportion of coefficients in the high-frequency components and set them to zero; the random enhancement perturbation is to multiply the coefficients of the high-frequency components by a random scaling factor within a preset range; and the random phase shift perturbation is to randomly shift the phase of the high-frequency components.

[0009] In some embodiments, the perturbation probability and perturbation intensity of the spectral energy perturbation are adaptively adjusted based on the local gradient magnitude of the high-frequency component; and applying the spectral energy perturbation to the high-frequency component includes the following operations: For each position of the high-frequency component, generate a random number r located in the interval [0,1]. Determine whether the random number r at this location is less than its perturbation probability. ; If the random number r at that position is less than its perturbation probability If a spectral energy perturbation is applied to that location, the larger the gradient magnitude, the smaller the perturbation intensity. If the random number r at that position is greater than or equal to its perturbation probability Then no spectral energy perturbation is applied to that position.

[0010] In some embodiments, the perturbation probability at each location of the high-frequency component is determined. Includes the following operations: Calculate the gradient magnitude at each position of the high-frequency component and normalize it to the [0,1] interval; Based on the normalized gradient magnitude, the perturbation probability at each location of the high-frequency component is determined using the following formula. :

[0011] in, The normalized gradient magnitude. This represents the global disturbance intensity coefficient.

[0012] In some embodiments, the joint loss function includes detection loss, contrastive distillation loss, and frequency domain consistency loss; the expression for the joint loss function is:

[0013] in, For the joint loss function, To detect the loss, To compare distillation losses, For frequency domain consistency loss, , and This is the preset loss balance coefficient.

[0014] In some embodiments, the detection loss is used to constrain the consistency between the cerebral aneurysm detection results output by the student detection model and the cerebral aneurysm detection pseudo-labels; the expression for the detection loss is:

[0015] in, The number of perturbation views, This represents the detection result of the cerebral aneurysm corresponding to the k-th perturbation view. False label for cerebral aneurysm detection For classifying losses, The bounding box regression loss is defined as either focus loss or cross-entropy loss, while the classification loss is defined as either smoothing L1 loss or GIoU loss.

[0016] In some embodiments, the contrastive distillation loss employs InfoNCE loss to constrain the feature consistency between the intermediate-layer student features and the intermediate-layer teacher features; the expression for the contrastive distillation loss is:

[0017] in, The number of perturbation views, For temperature coefficient, For the intermediate layer student features corresponding to the k-th perturbation view, Characteristics of middle-level teachers For cosine similarity,

[0018] pos represents the set of positive samples, corresponding to the features of teachers in the intermediate layer. negative sample set This includes intermediate-layer student features and intermediate-layer teacher features corresponding to other images within the current batch.

[0019] In some embodiments, the frequency domain consistency loss is used to constrain the feature consistency of intermediate layer student features corresponding to different input views in the frequency domain; the expression for the frequency domain consistency loss is:

[0020] in, The number of perturbation views, Let be the high-frequency component of the intermediate layer student feature corresponding to the i-th perturbation view. The high-frequency component is the intermediate layer student feature corresponding to the j-th perturbation view, which is extracted by performing a frequency domain transformation on the intermediate layer student feature.

[0021] In a second aspect, this application provides a method for detecting cerebral aneurysms, characterized in that it comprises: Acquire plain computed tomography (NCCT) images of the subjects to be tested, wherein the NCCT images are CT images without contrast enhancement; The NCCT image to be tested is input into a pre-trained student detection model; wherein the student detection model is trained using the methods for training student detection models described in the embodiments of the first aspect above; The student detection model outputs the detection results of cerebral aneurysms in the NCCT images to be tested.

[0022] In a third aspect, this application provides an electronic device, characterized in that it includes a processor; and A memory storing program instructions for training a student detection model or for detecting cerebral aneurysms, wherein when the program instructions are executed by the processor, the method according to any one of the embodiments of the first aspect described above is implemented, or the method according to each of the second aspects described above is implemented.

[0023] Using the above-described scheme for detecting cerebral aneurysms in NCCT images, this embodiment of the application obtains high-frequency components carrying vessel edges and texture details by performing frequency domain transformation decomposition on NCCT images, specifically adapting to the image characteristics of weak vessel edge information in NCCT images. By applying spectral energy perturbation to the high-frequency components to generate multiple sets of input views, the student model can be guided to learn robust vessel edge features under different perturbation conditions, enhancing the model's ability to perceive aneurysm morphology and edge information. By combining the detection pseudo-labels output by the teacher model with intermediate layer features to construct a joint loss function to complete model training, cross-modal knowledge transfer is achieved from the teacher model trained on CTA images to the student model trained on NCCT images. This alleviates the detection performance degradation problem that occurs when existing detection models are transferred from CTA images to NCCT images, providing a well-adapted model foundation for cerebral aneurysm detection in contrast-free scenarios.

[0024] Furthermore, this embodiment of the application employs a student detection model trained using the aforementioned method. This model can detect cerebral aneurysms simply by inputting NCCT images without contrast enhancement, eliminating the need for iodine contrast agents. This avoids the risks of allergic reactions and nephrotoxicity associated with contrast agents, while also reducing the radiation dose and economic cost required for detection. The trained student model's excellent perception of vessel edges and aneurysm features in NCCT images improves the detection rate of cerebral aneurysms in contrast-free scenarios, reducing the reliance on physicians' clinical experience in traditional manual image interpretation. This approach is suitable for large-scale population screening and cerebral aneurysm detection needs in individuals contraindicated to contrast agents. Attached Figure Description

[0025] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts, wherein: Figure 1An exemplary flowchart of a method for training a student detection model according to an embodiment of this application is shown; Figure 2 An exemplary flowchart of a method for detecting cerebral aneurysms according to an embodiment of this application is shown; Figure 3 An exemplary structural block diagram of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] It should be understood that the terms "comprising" and "including" used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0028] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0029] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0030] The specific embodiments of this application will be described in detail below with reference to the accompanying drawings. The execution subject of the embodiments of this application is an electronic device with graphics processing capabilities and deep learning model computing capabilities, including but not limited to medical image post-processing workstations, servers, computer terminals, or embedded medical devices, and this application does not limit these.

[0031] Figure 1 An exemplary flowchart of a method 100 for training a student detection model according to an embodiment of this application is shown. It will be understood that the student detection model is used for cerebral aneurysm detection in NCCT images.

[0032] like Figure 1 As shown, in step S101, the training dataset is obtained. The training dataset includes multiple sets of paired NCCT images and corresponding CTA images. It should be noted that the paired NCCT images and CTA images are spatially registered images from the same subject.

[0033] NCCT images are uncontrast-free, non-contrast-enhanced 3D volumetric CT images of the subject's head, while CTA images are 3D volumetric CT angiography images of the head acquired after intravenous injection of iodine contrast agent in the same subject. Spatial registration is achieved by aligning the spatial coordinates of the NCCT and CTA images using a rigid registration algorithm, ensuring a one-to-one correspondence between the spatial positions of the same anatomical structure in the two sets of images. After obtaining the training dataset, the NCCT and CTA images can be preprocessed, including but not limited to grayscale normalization, skull removal, and region of interest cropping, to improve the stability and computational efficiency of model training.

[0034] Next, in step S102, the NCCT image is transformed in the frequency domain to decompose it into high-frequency components containing information about blood vessel edges and texture details.

[0035] Specifically, the frequency domain transformation adopts wavelet transform, which can be two-dimensional or three-dimensional. The wavelet basis can be Haar wavelet or Daubechies wavelet, and the decomposition level is set to 1. Through wavelet transform, in addition to decomposing and obtaining high-frequency components containing information on blood vessel edges and texture details, low-frequency approximate components reflecting the overall gray-level distribution of the image can also be obtained. Among them, the high-frequency components contain detail sub-bands in three directions: horizontal, vertical and diagonal, corresponding to key anatomical information such as the contour and edges of cerebral blood vessels.

[0036] Next, in step S103, spectral energy perturbation is applied to the high-frequency components to generate at least one perturbed view, which is then combined with the NCCT image as an unperturbed view to obtain multiple sets of input views.

[0037] Spectral energy perturbation may include, but is not limited to, at least one of random masking perturbation, random enhancement perturbation, and random phase shift perturbation. Specifically, random masking perturbation involves randomly selecting a preset proportion of coefficients in the high-frequency components and setting them to zero; this preset proportion can be set to 20% to 50%. Random enhancement perturbation involves multiplying the coefficients of the high-frequency components by a random scaling factor within a preset range; this random scaling factor can be set to a value between 0.5 and 2.0. Random phase shift perturbation involves randomly shifting the phase of the high-frequency components.

[0038] In the embodiments of this application, the perturbation probability and perturbation intensity of the spectral energy perturbation are adaptively adjusted according to the local gradient magnitude of the high-frequency component. Based on this, applying spectral energy perturbation to the high-frequency component can be performed by: first, generating a random number r within the interval [0,1] for each position of the high-frequency component; then determining whether the random number r at that position is less than its perturbation probability. If the random number r at that position is less than its perturbation probability If a spectral energy perturbation is applied to a location, the larger the gradient magnitude, the higher the probability that the location corresponds to a potential vascular edge region, and the smaller the applied perturbation intensity, in order to avoid damaging key vascular anatomical features; conversely, if the random number r at that location is greater than or equal to its perturbation probability... Then no spectral energy perturbation is applied to that position, and the original high-frequency coefficients at that position are preserved.

[0039] Among them, the perturbation probability at each position of the high-frequency component is determined. The following steps can be performed: First, calculate the gradient magnitude at each location of the high-frequency component and normalize it to the [0,1] interval; then, based on the normalized gradient magnitude, determine the perturbation probability at each location of the high-frequency component using the following formula. : (1) in, The normalized gradient magnitude. This represents the global perturbation intensity coefficient. For example, The value can be 0.3, and those skilled in the art can adjust this coefficient according to the actual training effect.

[0040] In step S104, the CTA image is input into the pre-trained teacher detection model to obtain the corresponding cerebral aneurysm detection pseudo-label and intermediate layer teacher features.

[0041] Specifically, the teacher detection model is a high-capacity object detection network pre-trained on labeled head CTA images. For example, the nnDetection network or the Faster R-CNN network based on the ResNet-50 backbone network can be used. The pre-trained teacher detection model has stable cerebral aneurysm detection capabilities on CTA images. Its output detection pseudo-labels include the bounding box coordinates of the aneurysm and the class confidence score. The intermediate layer teacher features are the feature maps output by the last convolutional layer of the teacher detection model's backbone network, used to provide anchor points for feature learning in the student model.

[0042] It is understood that the execution order of step S104 and the following step S105 is not limited; they can be executed in parallel or in a preset order.

[0043] In step S105, all input views are input into the student detection model to be trained to obtain the cerebral aneurysm detection results and intermediate layer student features corresponding to each group of input views.

[0044] Specifically, the student detection model is a lightweight single-stage object detection network. For example, lightweight RetinaNet, YOLOX-S, or CenterNet networks can be used, with the backbone network employing lightweight convolutional neural networks like MobileNetV2 or ShuffleNet to lower the model deployment threshold and adapt to the computational capabilities of primary healthcare equipment. The intermediate-layer student features are the feature maps output from the last convolutional layer of the student detection model's backbone network. Their channel count and size match those of the intermediate-layer teacher features to ensure the effectiveness of subsequent feature alignment and loss calculation.

[0045] Finally, in step S106, a joint loss function is constructed based on the pseudo-labels of cerebral aneurysm detection, the characteristics of intermediate-layer teachers, the results of cerebral aneurysm detection, and the characteristics of intermediate-layer students. The student detection model is then trained using the joint loss function.

[0046] In this embodiment, the joint loss function may include detection loss, contrastive distillation loss, and frequency domain consistency loss. Therefore, the expression for the joint loss function is: (2) in, For the joint loss function, To detect the loss, To compare distillation losses, For frequency domain consistency loss, , and This is the preset loss balance coefficient.

[0047] For example, the loss balance coefficient can be set to =1, =0.1, =0.5, and those skilled in the art can adjust the coefficient according to the actual training effect.

[0048] Among them, detection loss The consistency between the cerebral aneurysm detection results output by the student detection model and the pseudo-labels for cerebral aneurysm detection is constrained, and the expression for the detection loss is: (3) in, The number of perturbation views, This represents the detection result of the cerebral aneurysm corresponding to the k-th perturbation view. False label for cerebral aneurysm detection For classifying losses, The bounding box regression loss is used for classification loss (focus loss or cross-entropy loss) and bounding box regression loss (smoothing L1 loss or GIoU loss).

[0049] Comparison of distillation losses The InfoNCE loss is used to constrain the feature consistency between the intermediate layer student features and the intermediate layer teacher features. The expression for the contrastive distillation loss is as follows: (4) in, The number of perturbation views, For temperature coefficient, For the intermediate layer student features corresponding to the k-th perturbation view, Characteristics of middle-level teachers For cosine similarity, (5) pos represents the set of positive samples, corresponding to the features of teachers in the intermediate layer. negative sample set This includes intermediate-layer student features and intermediate-layer teacher features corresponding to other images within the current batch. Specifically, it includes all intermediate-layer student features and intermediate-layer teacher features corresponding to other images within the current training batch, excluding the currently processed NCCT image.

[0050] Frequency domain consistency loss The expression for the frequency domain consistency loss is used to constrain the feature consistency of intermediate layer student features corresponding to different input views in the frequency domain: (6) in, The number of perturbation views, Let be the high-frequency component of the intermediate layer student feature corresponding to the i-th perturbation view. The high-frequency component is the intermediate layer student feature corresponding to the j-th perturbation view. The high-frequency component is extracted by performing a frequency domain transformation on the intermediate layer student feature. The frequency domain transformation is consistent with the wavelet transform in step S102 to ensure that the extraction rules of the high-frequency component are consistent.

[0051] The above text combined Figure 1This paper describes a method for training a student detection model. In this embodiment, the high-frequency components carrying vessel edges and texture details are obtained by frequency domain transformation decomposition of NCCT images, specifically adapting to the image characteristics of weak vessel edge information in NCCT images. By applying spectral energy perturbation to the high-frequency components to generate multiple sets of input views, the student model can be guided to learn robust vessel edge features under different perturbation conditions, enhancing the model's ability to perceive aneurysm morphology and edge information. Model training is completed by constructing a joint loss function by combining the detection pseudo-labels output by the teacher model and intermediate layer features, achieving cross-modal knowledge transfer from the teacher model trained on CTA images to the student model trained on NCCT images. This alleviates the detection performance degradation problem that occurs when existing detection models are transferred from CTA images to NCCT images, providing a well-adapted model foundation for cerebral aneurysm detection in contrast-free scenarios.

[0052] Corresponding to the above model training method, this disclosure also provides a method for detecting cerebral aneurysms. Figure 2 An exemplary flowchart of a method 200 for detecting cerebral aneurysms according to an embodiment of this application is shown. It is understood that the entity executing method 200 is the same as the entity executing the training method 100 described above.

[0053] like Figure 2 As shown, in step S201, method 200 acquires the non-contrast computed tomography (NCCT) image to be tested. The NCCT image to be tested is a non-contrast CT three-dimensional volumetric image of the subject's head. After acquiring the NCCT image to be tested, preprocessing operations consistent with those in the training phase can be performed on it, including grayscale normalization, skull removal, and region of interest cropping, to ensure the consistency of the distribution between the input data and the model training data.

[0054] In step S202, the NCCT image to be tested is input into the pre-trained student detection model. This student detection model is trained using the method for training student detection models described in the preceding embodiments.

[0055] Finally, in step S203, the detection results of cerebral aneurysms in the NCCT image to be tested are output by the student detection model. The detection results are multiple sets of aneurysm prediction boxes output by the student detection model, and each set of prediction boxes includes the bounding box coordinates and confidence score.

[0056] Furthermore, the detection results output by the student detection model are post-processed, specifically by using the non-maximum suppression (NMS) algorithm to remove redundant prediction boxes to obtain the final detection results. The confidence threshold and intersection-over-union (IoU) threshold for NMS can be set according to clinical detection needs. For example, the confidence threshold can be set to 0.5 and the IoU threshold can be set to 0.3.

[0057] In practice, the final detection result can be a candidate list of cerebral aneurysms, where each aneurysm candidate contains the following information: Bounding Box: Represents the spatial location of an aneurysm in three dimensions. It can be represented in two ways: one is the coordinates (x, y, z) of the center point of the minimum bounding cube of the aneurysm and the corresponding dimensions (dx, dy, dz); the other is the coordinates (x1, y1, z1, x2, y2, z2) of the diagonal points of the minimum bounding cube of the aneurysm. Confidence: The probability that the model predicts the existence of an aneurysm within the bounding box; the value ranges from 0 to 1. Optional center point coordinates: can be calculated directly from the center position of the corresponding bounding box.

[0058] Ultimately, a standardized test report can be generated based on the test results, outputting the spatial location, size, and confidence score of all detected aneurysms in a list format. The test results can also be overlaid on the three-dimensional vascular reconstruction image of the NCCT image to be tested for visualization, which can be viewed and used as a reference for clinicians.

[0059] The above text combined Figure 2 This application describes a method for detecting cerebral aneurysms. In this embodiment, a student detection model trained using the aforementioned method is employed. This model requires only non-contrast-enhanced NCCT images to detect cerebral aneurysms, eliminating the need for iodine contrast agents and avoiding the risks of allergic reactions and nephrotoxicity associated with contrast agents. It also reduces the radiation dose and economic cost required for detection. The trained student model's excellent perception of vessel edges and aneurysm features in NCCT images improves the detection rate of cerebral aneurysms in contrast-free scenarios, reducing the reliance on physician clinical experience in traditional manual image interpretation. This method is suitable for large-scale population screening and cerebral aneurysm detection needs in individuals contraindicated to contrast agents.

[0060] Next, combine Figure 3 An electronic device 300 provided in an embodiment of this application will be described by way of example. Figure 3 As shown, the electronic device 300 in this application embodiment may include a processor 301, a memory 302, and a communication bus 303.

[0061] In specific embodiments, the processor 301 described above can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), CPU, controller, microcontroller, and microprocessor. It is understood that for different devices, the electronic device used to implement the above processor function can also be other types, and this embodiment does not specifically limit it.

[0062] In this embodiment, the communication bus 303 is used to establish communication between the processor 301 and the memory 302; the memory 302 stores program instructions for training a student detection model or for detecting cerebral aneurysms; when the processor 301 executes the program instructions stored in the memory 302, it implements the combination of this application. Figure 1 The described method for training a student detection model, or in combination with... Figure 2 The method described is for detecting cerebral aneurysms.

[0063] The above combination Figure 3 This document describes electronic devices that can be used to train student detection models or to detect cerebral aneurysms, which can be used to execute this application. It should be understood that the device structures or architectures described herein are merely exemplary, and the implementation methods and entities of this application are not limited thereto, but can be modified without departing from the spirit of this application. It is understood that the descriptions of the various embodiments in this disclosure emphasize the differences between the various embodiments, while their similarities or corresponding parts can be referred to mutually. For the sake of brevity, this disclosure will not elaborate on each one.

[0064] Based on the foregoing description in conjunction with the accompanying drawings, those skilled in the art will understand that the embodiments of this application can also be implemented by software programs. Therefore, this application also provides a computer-readable storage medium. This computer-readable storage medium stores program instructions for training a student detection model, or program instructions for detecting cerebral aneurysms, which can be used to implement the embodiments of this application. Figure 1 The described method for training a student detection model, or in combination with... Figure 2The method described is for detecting cerebral aneurysms.

[0065] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0066] While numerous embodiments of this application have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will arise for those skilled in the art without departing from the spirit and intent of this application. It should be understood that various alternatives to the embodiments of this application described herein may be employed in the practice of this application. The appended claims are intended to define the scope of protection of this application and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for training a student detection model, characterized in that, The student detection model is used for cerebral aneurysm detection in plain computed tomography (NCCT) images, and the training method includes the following operations: Obtain a training dataset, which includes multiple sets of paired NCCT images and corresponding computed tomography angiography (CTA) images, wherein the paired NCCT images and CTA images are spatially registered images of the same subject; The NCCT image is subjected to frequency domain transformation to decompose it into high-frequency components containing information on blood vessel edges and texture details; Apply spectral energy perturbation to the high-frequency components to generate at least one perturbation view, and combine it with the NCCT image as an unperturbed view to obtain multiple sets of input views; The CTA image is input into the pre-trained teacher detection model to obtain the corresponding cerebral aneurysm detection pseudo-label and intermediate layer teacher features; Input all input views into the student detection model to be trained to obtain the cerebral aneurysm detection results and intermediate layer student features corresponding to each set of input views; Based on the pseudo-labels for cerebral aneurysm detection, intermediate-layer teacher features, cerebral aneurysm detection results, and intermediate-layer student features, a joint loss function is constructed, and the student detection model is trained using the joint loss function.

2. The method according to claim 1, characterized in that, The spectral energy perturbation includes at least one of random masking perturbation, random enhancement perturbation, and random phase shift perturbation; wherein, the random masking perturbation is to randomly select a preset proportion of coefficients in the high-frequency components and set them to zero; the random enhancement perturbation is to multiply the coefficients of the high-frequency components by a random scaling factor within a preset range; and the random phase shift perturbation is to randomly shift the phase of the high-frequency components.

3. The method according to claim 2, characterized in that, The perturbation probability and intensity of the spectral energy perturbation are adaptively adjusted based on the local gradient magnitude of the high-frequency component; and applying the spectral energy perturbation to the high-frequency component includes the following operations: For each position of the high-frequency component, generate a random number r located in the interval [0,1]. Determine whether the random number r at this location is less than its perturbation probability. ; If the random number r at that position is less than its perturbation probability If a spectral energy perturbation is applied to that location, the larger the gradient magnitude, the smaller the perturbation intensity. If the random number r at that position is greater than or equal to its perturbation probability Then no spectral energy perturbation is applied to that position.

4. The method according to claim 3, characterized in that, Determine the perturbation probability at each location of the high-frequency component. Includes the following operations: Calculate the gradient magnitude at each position of the high-frequency component and normalize it to the [0,1] interval; Based on the normalized gradient magnitude, the perturbation probability at each location of the high-frequency component is determined using the following formula. : in, The normalized gradient magnitude. This represents the global disturbance intensity coefficient.

5. The method according to claim 1, characterized in that, The joint loss function includes detection loss, contrastive distillation loss, and frequency domain consistency loss; the expression for the joint loss function is: in, For the joint loss function, To detect the loss, To compare distillation losses, For frequency domain consistency loss, , and This is the preset loss balance coefficient.

6. The method according to claim 5, characterized in that, The detection loss is used to constrain the consistency between the cerebral aneurysm detection results output by the student detection model and the cerebral aneurysm detection pseudo-labels; the expression for the detection loss is: in, The number of perturbation views, This represents the detection result of the cerebral aneurysm corresponding to the k-th perturbation view. False label for cerebral aneurysm detection For classifying losses, The bounding box regression loss is defined as either focus loss or cross-entropy loss, while the classification loss is defined as either smoothing L1 loss or GIoU loss.

7. The method according to claim 5, characterized in that, The contrastive distillation loss employs InfoNCE loss to constrain the feature consistency between the intermediate layer student features and the intermediate layer teacher features; the expression for the contrastive distillation loss is: in, The number of perturbation views, For temperature coefficient, For the intermediate layer student features corresponding to the k-th perturbation view, Characteristics of middle-level teachers For cosine similarity, pos represents the set of positive samples, corresponding to the features of teachers in the intermediate layer. negative sample set This includes intermediate-layer student features and intermediate-layer teacher features corresponding to other images within the current batch.

8. The method according to claim 5, characterized in that, The frequency domain consistency loss is used to constrain the feature consistency of intermediate layer student features corresponding to different input views in the frequency domain; the expression for the frequency domain consistency loss is: in, The number of perturbation views, Let be the high-frequency component of the intermediate layer student feature corresponding to the i-th perturbation view. The high-frequency component is the intermediate layer student feature corresponding to the j-th perturbation view, which is extracted by performing a frequency domain transformation on the intermediate layer student feature.

9. A method for detecting cerebral aneurysms, characterized in that, include: Acquire plain computed tomography (NCCT) images of the subjects to be tested, wherein the NCCT images are CT images without contrast enhancement; The NCCT image to be tested is input into a pre-trained student detection model; wherein the student detection model is trained using the method for training a student detection model as described in any one of claims 1 to 8; The student detection model outputs the detection results of cerebral aneurysms in the NCCT images to be tested.

10. An electronic device, characterized in that, Including processors; and A memory storing program instructions for training a student detection model or for detecting cerebral aneurysms, wherein when the program instructions are executed by the processor, the method according to any one of claims 1 to 8 is implemented, or the method according to claim 9 is implemented.

Images