Method for selecting an intratumoral turbulence device and products thereof

Abstract

The application discloses a method for selecting an intratumoral turbulence device and a product thereof. The method comprises: obtaining medical image data of a target object, wherein the target object at least comprises an aneurysm and a parent vessel; segmenting and three-dimensionally reconstructing the medical image data to obtain a segmentation model of the target object; and inputting the segmentation model into a selection model based on deep learning to perform model recommendation; wherein the selection model comprises: a geometric feature extraction module configured to extract geometric features of the target object from the segmentation model, the geometric features being explicit measurable parameter features; and a classifier configured to output a prediction result of each candidate model of the intratumoral turbulence device based on at least the geometric features. According to the method, end-to-end automatic prediction from image data to model recommendation is realized.

Description

Technical Field

[0001] This application generally relates to the field of image processing technology. More specifically, this application relates to a method, apparatus, electronic device, and computer-readable storage medium for selecting an intratumoral disturbance device. Background Technology

[0002] Aneurysms are common structural lesions of the cerebral blood vessels, and their treatment has gradually shifted from traditional open craniotomy to endovascular interventional surgery. Among these, the Woven EndoBridge (WEB) device has achieved good results in the treatment of bifurcation and wide-necked aneurysms due to its advantages in occluding the aneurysm neck and maintaining vascular patency. However, there are many models of WEB devices, and different models differ in size, shape, and mechanical properties. Choosing the appropriate model directly affects the embolization effect, surgical safety, and patient prognosis.

[0003] In current technologies, the selection of the model and size of the web-based device primarily relies on geometric parameters such as aneurysm height, width, and neck width measured manually on preoperative two-dimensional digital subtraction angiography (DSA), combined with the surgeon's clinical experience. While some existing software can obtain a three-dimensional model of the aneurysm through 3D reconstruction technology and provide corresponding measurement data, the final model selection is still manually performed by the physician, lacking automated intelligent decision support. Because aneurysms have complex and varied morphologies in three-dimensional space, two-dimensional measurements cannot fully reflect the true spatial morphology of the aneurysm, leading to significant differences in measurement results between different physicians, easily resulting in the selection of a device that is too large or too small. Deviations in model selection not only increase surgical time and radiation exposure but may also cause clinical risks such as incomplete embolization, increased postoperative recurrence rates, or device dislodgement.

[0004] In view of this, there is an urgent need to provide a solution for the selection of intratumoral disturbance devices, so as to realize the automation and accurate prediction of device models, reduce the reliance on doctors' subjective experience, and improve the consistency of selection and clinical prognosis. Summary of the Invention

[0005] In order to at least address one or more of the technical problems mentioned above, this application proposes methods, apparatuses, electronic devices, and computer-readable storage media schemes for selecting intratumoral turbulence devices in several aspects.

[0006] In a first aspect, this application provides a method for selecting an intratumoral flow disruptor, comprising: acquiring medical image data of a target object, wherein the target object includes at least an aneurysm and a tumor-bearing vessel; segmenting and three-dimensionally reconstructing the medical image data to obtain a segmentation model of the target object; and inputting the segmentation model into a deep learning-based selection model for model recommendation; wherein the selection model includes: a geometric feature extraction module for extracting geometric features of the target object from the segmentation model, wherein the geometric features are explicitly measurable parametric features; and a classifier for outputting prediction results for each candidate model of the intratumoral flow disruptor based at least on the geometric features.

[0007] In some embodiments, the selection model further includes: a three-dimensional morphology extraction module, used to obtain a three-dimensional voxel model of the target object based on the segmentation model; and to extract morphological features of the three-dimensional voxel model using a three-dimensional morphology extraction network, wherein the morphological features are implicit spatial structural features; a feature fusion module, used to fuse the geometric features and the morphological features to obtain fused features; and a classifier further used to classify the fused features and output prediction results for each candidate model of the intratumoral disturbance device.

[0008] In other embodiments, fusing the geometric features and the morphological features includes: encoding the geometric features through a geometric feature encoding layer to generate a geometric feature vector; encoding the three-dimensional voxel model through a morphological feature encoding layer to generate a morphological feature tensor; mapping the geometric feature vector and the morphological feature tensor to a unified dimensional space, and performing weighted fusion using an attention mechanism.

[0009] In some other embodiments, the three-dimensional morphology extraction network includes a three-dimensional convolutional neural network; the geometric feature encoding layer includes one or more fully connected networks; and the morphological feature encoding layer includes a three-dimensional convolutional neural network and a pooling layer.

[0010] In some embodiments, the geometric features include at least one of the following aneurysm geometric features: long diameter, width, height, transverse diameter, longest diameter, aspect ratio, neck diameter, ratio of transverse diameter to neck diameter, volume, neck area, non-sphericity index, aspect ratio, ratio of volume to neck area, and fluctuation index; and / or at least one of the following aneurysm-bearing vessel geometric features: proximal diameter of the aneurysm-bearing vessel, distal diameter of the aneurysm-bearing vessel, average diameter of the aneurysm-bearing vessel, and length of the aneurysm-bearing vessel; and / or at least one of the following aneurysm-bearing vessel correlation features: angle of incidence, aneurysm angle, and size ratio.

[0011] In other embodiments, before inputting the segmentation model into the deep learning-based selection model, the method further includes: constructing a training dataset, the training dataset including medical image data samples of historical cases and corresponding ground truth labels, the ground truth labels being the models of intratumoral perturbation devices that were successfully implanted during surgery and whose postoperative efficacy met preset conditions; training the selection model using the training dataset, and supervising the training of the selection model using a weighted cross-entropy loss function, wherein the weights of the weighted cross-entropy loss function are automatically adjusted according to the number of samples of different models in the training dataset.

[0012] In some other embodiments, the truth labels are constructed by extracting the actual implanted model during surgery from the surgical records and consumables management system of the historical cases; and retaining cases where postoperative imaging shows complete occlusion or embolism rate reaching a preset threshold, and using their actual implanted model as the truth label for positive samples.

[0013] In some embodiments, the training dataset further includes negative samples of cases with postoperative recurrence, prolapse, or incomplete embolization in the historical cases and their corresponding negative sample labels; the method further includes: jointly training the selection model based on the ground truth labels and the negative sample labels.

[0014] In other embodiments, the method further includes: after a new case completes surgery and the postoperative efficacy is confirmed, adding the medical imaging data and implant type of the new case to the training dataset; and using transfer learning or incremental learning strategies to update the parameters of the trained selection model using the newly added data in the training dataset.

[0015] In a second aspect, this application provides an apparatus for selecting an intratumoral aneurysm disturbance device, comprising: a data acquisition module configured to acquire medical image data of a target object, wherein the target object includes at least an aneurysm and a tumor-bearing vessel; a segmentation and reconstruction module configured to segment and reconstruct the medical image data in three dimensions to obtain a segmentation model of the target object; and a selection module configured to receive the segmentation model and output prediction results of each candidate model of the intratumoral aneurysm disturbance device for model recommendation; wherein the selection module includes: a geometric feature extraction module configured to extract geometric features of the target object from the segmentation model, wherein the geometric features are explicitly measurable parametric features; and a classifier configured to output prediction results of each candidate model of the intratumoral aneurysm disturbance device based at least on the geometric features.

[0016] In a third aspect, this application provides an electronic device comprising: a processor configured to execute program instructions; and a memory configured to store the program instructions, which, when loaded and executed by the processor, cause the processor to perform the method according to any one of the first aspects of this application.

[0017] In a fourth aspect, this application provides a computer-readable storage medium storing program instructions that, when loaded and executed by a processor, cause the processor to perform the method according to any one of the first aspects of this application.

[0018] Using the above-described scheme for selecting intraneural flow disturbance devices, this embodiment of the application constructs a deep learning selection model that includes a geometric feature extraction module and a classifier. The segmentation model of the target object after segmentation and reconstruction is directly input into the selection model. The geometric feature extraction module automatically extracts the explicit measurable geometric features of the aneurysm and the aneurysm-bearing vessel, and the classifier outputs the prediction results of each candidate model based at least on the geometric features. This achieves end-to-end automated prediction from image data to model recommendation, which can effectively reduce the reliance on manual measurement and doctors' subjective experience, improve the consistency and accuracy of model selection, and reduce the clinical risks caused by selection deviations. Attached Figure Description

[0019] 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:

[0020] Figure 1 An exemplary method flowchart for selecting an intratumoral disturbance device according to some embodiments of this application is shown; Figure 2 This diagram illustrates the results of segmenting and reconstructing medical image data. Figure 3 Exemplary flowcharts for selecting intratumoral turbulence devices according to other embodiments of this application are shown; Figure 4 A schematic block diagram of a system for selecting an intratumoral turbulence device according to an embodiment of this application is shown. Detailed Implementation

[0021] 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.

[0022] 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.

[0023] 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.

[0024] 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]."

[0025] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0026] Figure 1 A flowchart illustrating an exemplary method for selecting an intratumoral turbulence-disrupting device according to some embodiments of this application is shown. Figure 1As shown, method 100 may include: step S110, acquiring medical image data of the target object, wherein the target object includes at least an aneurysm and aneurysm-bearing vessel; step S120, segmenting and reconstructing the medical image data to obtain a segmentation model of the target object; and step S130, inputting the segmentation model into a deep learning-based selection model for model recommendation; wherein the selection model includes: a geometric feature extraction module, used to extract geometric features of the target object from the segmentation model, wherein the geometric features are explicitly measurable parametric features; and a classifier, used to output prediction results for each candidate model of the intraneural perturbation device based at least on the geometric features.

[0027] The aforementioned medical imaging data refers to medical images that clearly display vascular structures (such as cerebral vascular structures), specifically including enhanced vascular images such as computed tomography angiography (CTA), digital subtraction angiography (DSA), or magnetic resonance angiography (MRA). These images effectively present vascular morphology and the anatomical structure of aneurysms, providing a data foundation for subsequent segmentation and reconstruction. In actual clinical settings, this step can be automated by accessing the hospital's image archiving and communication system to acquire the patient's preoperative imaging data, or it can be completed through manual data import.

[0028] After acquiring medical image data, step S120 can be executed to segment and reconstruct the medical image data in three dimensions, thereby obtaining a segmentation model of the target object. This step aims to accurately segment and reconstruct the aneurysm and its carrier vessel from the original image. The carrier vessel refers to the parent vessel that supports the aneurysm and protrudes from its wall. Specifically, firstly, an image segmentation algorithm is used to process the medical image data to distinguish the vascular tissue from the background region. Segmentation methods can employ threshold segmentation, region growing, or deep learning-based semantic segmentation networks, such as U-Net or its variants, to obtain a more accurate vascular mask. After completing the vascular segmentation, since the aneurysm is essentially a local protrusion structure on the vessel wall, it is necessary to further separate the aneurysm region from the segmented vascular model to form an independent aneurysm mask. Subsequently, using a three-dimensional reconstruction algorithm, such as the Moving Cube algorithm, a three-dimensional surface model of the vessel and aneurysm is constructed based on the above mask, thus obtaining the segmentation model of the target object. The segmentation model is presented in the form of a three-dimensional mesh, which preserves the true spatial morphology of the aneurysm and its parent vessel, providing a structured input for subsequent feature extraction.

[0029] by Figure 2 As shown in the example, Figure 2 Figure (a) illustrates a type of enhanced vascular image (such as a DSA image), which, after vascular segmentation and 3D reconstruction, can yield images like... Figure 2 The vascular model shown in Figure (b) is then further... Figure 2 In the vascular model shown in Figure (b), the following formations are observed: Figure 2 The isolated aneurysm mask shown in Figure (c) (indicated by the red area in the figure). Figure 2 Figure (b) or (c) in the above can be used as a segmentation model in the embodiments of this application, and as input data for the selection model, wherein the target object may include the aneurysm region and its tumor-bearing vessel region.

[0030] After obtaining the segmentation model of the target object, step S130 is executed, in which the segmentation model is input into a pre-built deep learning-based selection model, which automatically completes the model analysis and recommendation.

[0031] Specifically, the selection model in this application embodiment is a deep learning network model that has been trained to predict model type, and it contains at least two core components: a geometric feature extraction module and a classifier.

[0032] The geometric feature extraction module is used to automatically extract the geometric features of the target object from the input segmentation model. Geometric features refer to explicit measurable parameters that can quantitatively describe the geometry of the aneurysm and the carrier vessel. These parameters correspond to the measurement indicators that clinicians rely on in the traditional selection process.

[0033] For example, in some embodiments, at least one of the following is considered: the long diameter, width, height, transverse diameter, longest diameter, aspect ratio, neck diameter, ratio of transverse diameter to neck diameter, volume, neck area, non-sphericity index, aspect ratio, ratio of volume to neck area, and fluctuation index of the aneurysm; and / or at least one of the following: the proximal diameter of the carrier vessel, the distal diameter of the carrier vessel, the average diameter of the carrier vessel, and the length of the carrier vessel; and / or at least one of the following: the angle of incidence, the aneurysm angle, and the size ratio.

[0034] The geometric features of an aneurysm directly reflect its shape and size. The geometric features of the carrier vessel reflect its shape and size. The angle of incidence, aneurysm angle, and size ratio are related characteristics of the aneurysm and the carrier vessel. The angle of incidence refers to the angle between the blood flow into the aneurysm and the axis of the carrier vessel, while the aneurysm angle refers to the angle between the long axis of the aneurysm and the axis of the carrier vessel. Both together reflect the relative spatial configuration of the aneurysm and the carrier vessel, significantly influencing the difficulty of device implantation and the selection of the appropriate device model.

[0035] Furthermore, the variability index reflects the irregularity of the aneurysm surface; a higher value indicates a more uneven aneurysm. The size ratio, the ratio of the aneurysm's maximum diameter to the average diameter of the parent vessel, is used to assess the size of the aneurysm relative to the vessel. The non-sphericity index reflects the degree to which the aneurysm deviates from an ideal spherical shape; a higher value indicates a more irregular shape. The aspect ratio, the ratio of the aneurysm height to the neck diameter, is often used to assess the risk of aneurysm rupture. The ratio of volume to neck area comprehensively reflects the relative relationship between the aneurysm body and the neck, and is an important reference indicator for evaluating the stability of internal flow control devices. These features quantify the geometric characteristics of the aneurysm and the parent vessel from different dimensions, providing rich and clinically significant geometric parameters for the selection model.

[0036] The aforementioned geometric features reflect both the geometric properties of the aneurysm and its carrier vessel, as well as the constraints imposed by the aneurysm and carrier vessel on device implantation. The geometric feature extraction module automatically calculates the aforementioned parameters by analyzing the three-dimensional geometric information in the segmentation model and organizes them into structured feature vectors, thereby transforming the complex three-dimensional morphology into a numerical representation that can be processed by the classifier.

[0037] The aforementioned classifier is used to predict the models of candidate intratumoral perturbation devices based at least on the geometric features of the target object. The classifier can be constructed using a softmax network, with the feature vectors output by the geometric feature extraction module as input and the predicted probability distributions for different models as output. During model training, the classifier undergoes supervised learning using a large amount of historical case data to learn the mapping relationship between geometric features and the optimal device model. In practical applications, the classifier automatically calculates the probability of each model based on the input features and outputs the model with the highest probability as the recommendation result. Alternatively, it can output multiple candidate models based on probability for doctors' reference.

[0038] Through the steps described above, the selection method provided in this application achieves end-to-end automated processing from imaging data to model recommendation. Compared with the traditional manual measurement and experience-based judgment mode, this method has the following significant effects: On the one hand, through automated geometric feature extraction, the subjective factors of measurement errors between different doctors are eliminated, improving the consistency and accuracy of feature extraction; on the other hand, the feature-model mapping relationship established based on a deep learning model can effectively utilize successful experiences from historical surgical data to provide data-driven intelligent support for clinical decision-making, thereby reducing clinical risks such as incomplete embolization, increased recurrence rate, or device dislodgement caused by selection deviations, and improving the precision and safety of interventional surgery.

[0039] In a preferred embodiment, the selection model provided in this application, based on geometric features, further introduces a three-dimensional morphological feature extraction and multimodal feature fusion mechanism, thereby more comprehensively capturing the complex spatial configuration of aneurysms and tumor-bearing vessels. The following, in conjunction with... Figure 3 The flowchart shown provides a detailed description of this preferred embodiment.

[0040] Figure 3 Exemplary flowcharts for selecting intratumoral flow disturbance devices according to other embodiments of this application are shown. Figure 3 As shown, method 300 may include: step S310, acquiring medical image data of the target object, wherein the target object includes at least an aneurysm and aneurysm-bearing vessel; step S320, segmenting and reconstructing the medical image data to obtain a segmentation model of the target object; and step S330, inputting the segmentation model into a deep learning-based selection model for model recommendation. Steps S310 and S320 are combined with the preceding text. Figure 1 Steps S110 and S120 are the same or similar, and will not be repeated here.

[0041] like Figure 3 As further shown in step S330, the selection model includes a geometric feature extraction module 331 and a classifier 334, and further includes: a three-dimensional morphology extraction module 332, used to obtain a three-dimensional voxel model of the target object based on the segmentation model; and to extract the morphological features of the three-dimensional voxel model using a three-dimensional morphology extraction network, wherein the morphological features are implicit spatial structural features; a feature fusion module 333, used to fuse the geometric features and morphological features to obtain fused features; and a classifier 334 further used to classify the fused features and output the prediction results of each candidate model of the in-tumor turbulence device.

[0042] As mentioned earlier, geometric features (such as aneurysm height, width, volume, and neck diameter) are explicit parameters commonly used by clinicians in traditional device selection, possessing clear physical meaning and interpretability. However, aneurysms often exhibit highly irregular geometric shapes in three-dimensional space, such as lobed, dumbbell-shaped, or irregular protrusions. These complex structures are difficult to fully characterize using only a limited number of geometric parameters. Relying solely on geometric features may result in the loss of morphological information crucial for device selection, such as local surface irregularities and neck shape distortions. Therefore, the selection model in this application adds a three-dimensional morphology extraction module and a feature fusion module in addition to the geometric feature extraction module, forming a dual-branch parallel processing architecture.

[0043] The 3D morphology extraction module 332 is used to obtain a 3D voxel model of the target object from the segmentation model and to extract the morphological features of the voxel model using a 3D morphology extraction network. The 3D voxel model is a regular 3D voxel grid representation converted from the segmented 3D mesh model of the aneurysm and its carrier vessel. Each voxel represents a spatial cube unit, and its value indicates whether the location belongs to the target tissue. Voxelization preserves the complete spatial structure of the target object, allowing subsequent convolution operations to be performed directly in 3D space.

[0044] Three-dimensional morphological extraction networks typically employ, for example, three-dimensional convolutional neural networks (3D CNNs). Similar to conventional convolutional networks processing two-dimensional images, 3D CNNs automatically learn local spatial features by sliding three-dimensional convolutional kernels across voxel space. They progressively expand the receptive field through multi-layer stacking, thereby extracting multi-level features ranging from low-level edges and corners to high-level semantic structures. Unlike geometric features, these morphological features automatically learned by 3D CNNs are not predefined parameters but rather abstract representations learned from a large number of cases in a data-driven manner; therefore, they are called implicit spatial structural features. These features can capture morphological details that are difficult to describe with geometric parameters, such as the depth of local depressions on the aneurysm surface, the irregular contour of the aneurysm neck, and the curvature changes at the junction of the aneurysm and the parent vessel.

[0045] The specific implementation of feature fusion can be varied. For example, in some other embodiments, fusing geometric features and morphological features may include: encoding geometric features through a geometric feature encoding layer to generate geometric feature vectors; encoding a three-dimensional voxel model through a morphological feature encoding layer to generate morphological feature tensors; mapping the geometric feature vectors and morphological feature tensors to a unified dimensional space; and using an attention mechanism for weighted fusion.

[0046] Mapping geometric feature vectors and morphological feature tensors to a unified dimensional space refers to transforming both from their original feature dimensions to feature representations of the same dimension through linear transformations (such as fully connected layers) or convolutional operations. This makes them mathematically comparable and composable, enabling effective feature fusion in the same space through weighting, concatenation, or attention mechanisms. For example, in some embodiments, the geometric feature encoding layer may include one or more fully connected networks; the morphological feature encoding layer may include a three-dimensional convolutional neural network and pooling layers.

[0047] After obtaining the geometric feature vector and the three-dimensional morphological feature tensor respectively, the feature fusion module is responsible for effectively integrating these two types of features to generate more informative fused features. The purpose of fusion is to balance the explicit physical meaning of geometric features with the deep spatial expressive power of morphological features, enabling the model to understand aneurysms from both quantitative empirical parameters and overall morphological patterns.

[0048] The attention mechanism automatically learns the importance of different feature dimensions to the final selection decision by calculating the relevance weights between queries, keys, and values, thereby strengthening the contribution of key dimensions and suppressing redundant or noisy information. The fused features retain key parameters commonly used in doctors' experience while incorporating complex morphological information that is difficult for the human eye to quantify directly, providing a more comprehensive basis for subsequent classification decisions.

[0049] After obtaining the fused features, the classifier no longer classifies based solely on geometric features, but processes the fused features. The input dimension of the classifier matches the dimension of the fused features, and the output layer still uses the Softmax function to calculate the predicted probability distribution of each candidate type. Because the fused features contain richer discriminative information, the classifier can make more stable and accurate type judgments under complex aneurysm morphologies, especially in cases with similar geometric parameters but significant morphological differences. The model can effectively distinguish the applicability of different types by using three-dimensional morphological features.

[0050] Through the above dual-branch fusion architecture, the selection model of this application achieves a comprehensive representation of aneurysm morphology: the geometric feature branch provides stable and interpretable prior knowledge, while the three-dimensional morphology branch supplements detailed and abstract spatial structural information. The fusion of the two significantly improves the model's generalization ability and selection accuracy for aneurysms of different morphologies, thereby further reducing the risk of selection bias caused by morphological complexity.

[0051] After introducing the prediction phase of the selection model, the training process of the model is further described below. The performance of the selection model is highly dependent on the quality of the training data and the accuracy of the labels. To this end, this application proposes a complete training data construction and model optimization scheme to ensure that the model can effectively learn the mapping relationship from aneurysm morphology to the optimal device model from historical experience.

[0052] In some embodiments, before inputting the segmentation model into the deep learning-based selection model for prediction, the aforementioned method 100 or method 300 may further include: constructing a training dataset, which includes medical image data samples of historical cases and corresponding ground truth labels, wherein the ground truth labels are the models of intratumoral perturbation devices that were successfully implanted during surgery and whose postoperative efficacy met preset conditions; training the selection model using the training dataset, and supervising the training of the selection model using a weighted cross-entropy loss function, wherein the weights of the weighted cross-entropy loss function are automatically adjusted according to the number of samples of different models in the training dataset.

[0053] The training dataset comprises two core elements: first, medical image data samples from historical cases (segmentation models obtained after segmentation and reconstruction); and second, ground truth labels corresponding to each sample. Ground truth labels refer to the models of intratumoral disruptive devices that have been successfully and effectively implanted in clinical practice. To obtain high-quality ground truth labels, the implantation model recorded in the surgical record cannot be simply used directly, as this model may not be the optimal choice. For example, it may be temporarily changed during surgery due to operational reasons, or the postoperative effect may be unsatisfactory. Therefore, this embodiment imposes strict screening conditions on the construction of ground truth labels.

[0054] Specifically, the ground truth label can be defined as the model of the intraneural disruptor that was successfully implanted intraoperatively and whose postoperative efficacy met the preset conditions. Successful intraoperative implantation refers to the model of the device that actually remained in the patient's body at the end of the surgery, distinguishing it from models that were attempted but removed during the surgery. Meeting the preset conditions for postoperative efficacy is verified through postoperative imaging follow-up. In one example, the preset conditions could be that angiography at 1 or 3 months postoperatively showed the aneurysm to be in a state of "complete occlusion" (e.g., using Raymond grade I in the Raymond classification), or that the embolization rate was not lower than a certain preset threshold (e.g., 90%). Only cases that simultaneously meet both conditions are labeled with the ground truth label for positive samples. In this way, cases of surgical failure or poor efficacy are excluded from the training dataset, ensuring that the model's learning decisions are based on clinically validated successful experiences.

[0055] In practice, the construction of truth labels can be achieved as follows: First, extract the actual implanted device model information from the surgical record system and consumable management system of historical cases; then, retrieve the postoperative imaging review reports of the same patient and screen out cases showing complete occlusion or embolization rates reaching a preset threshold; finally, use the implanted device model of the cases that meet the screening criteria as the truth label for that sample. Furthermore, for surgeries involving multiple intraoperative consumable changes, only the model that was successfully implanted is retained as the label; for cases with postoperative recurrence, device dislodgement, or incomplete embolization, they are marked as negative samples (see below) and not included in the truth label set of positive samples. The training dataset constructed in this way can truly reflect the best practices in clinical decision-making, avoiding subjective bias from relying solely on surgical records and ignoring postoperative outcomes.

[0056] During training, the model outputs the predicted probabilities of each candidate model, while the ground truth labels are discrete model categories. To ensure the model's predicted distribution approximates the true distribution, this embodiment uses a cross-entropy loss function as a supervision signal. However, in actual clinical data, the usage frequency of different models of intratumoral perturbation devices often varies significantly: for example, some common sizes have a large number of samples, while some special sizes have a small number of samples. This class imbalance problem causes the model to favor predicting models with a large sample size, neglecting the ability to identify a few models.

[0057] To address the aforementioned issues, this application employs a weighted cross-entropy loss function. This loss function, based on standard cross-entropy, introduces a weight factor for each category. The weights can be automatically adjusted based on the number of samples of different models in the training dataset. An exemplary weight calculation method is to make the weight proportional to the inverse of the number of samples; that is, the smaller the number of samples for a category, the greater its loss weight, thereby strengthening the model's focus on minority categories during training. Through this adaptive weighting mechanism, the model can treat all models equally during training, regardless of their sample size, thus significantly mitigating the performance bias caused by class imbalance and improving the model's accuracy in recognizing rare models.

[0058] In a further embodiment, this application not only uses positive samples (successful cases with satisfactory efficacy) for training, but also introduces negative samples to enhance the model's ability to identify incorrect selections. Specifically, in some embodiments, the training dataset may also include cases of postoperative recurrence, device dislodgement, or incomplete embolization from historical cases. These cases are labeled as negative samples, and a corresponding negative sample label is assigned to each negative sample. The negative sample label does not point to a specific correct model, but rather identifies the implantation model corresponding to that sample as not recommended or incorrectly selected.

[0059] During model training, embodiments of this application can jointly train the model selection process based on the ground truth labels of positive samples and the negative sample labels of negative samples. This joint training can be achieved by directly incorporating negative samples into the classification training, embedding them as incorrect categories, and using cross-entropy loss to reduce the model's prediction probability for that category. The introduction of negative samples allows the model to learn not only which categories are suitable but also which are unsuitable, thus enabling it to proactively avoid clinically proven ineffective categories when encountering cases with morphological similarities to negative samples.

[0060] Through the aforementioned training strategies, including ground truth labels based on postoperative efficacy screening, weighted cross-entropy loss to handle class imbalance, and joint training with positive and negative samples, the selection model constructed in this application embodiment can efficiently learn from limited clinical data and gradually approximate the decision-making logic of clinical experts. This training scheme significantly improves the model's generalization ability and robustness in real clinical environments, providing a solid technical foundation for subsequent automated model prediction and recommendation.

[0061] Furthermore, after the model is deployed and applied, new surgical cases are constantly generated in clinical practice. These new cases include both successful and satisfactory positive samples, as well as negative samples with poor results, all of which contain valuable selection experience. In order to enable the selection model to continuously adapt to changes in clinical data, absorb new successful experiences, and avoid new failures, this application embodiment further provides a mechanism for dynamic model updating.

[0062] In other embodiments, method 100 or method 300 may further include: after a new case completes surgery and the postoperative efficacy is confirmed, adding the medical imaging data and implant model of the new case to the training dataset; and using transfer learning or incremental learning strategies to update the parameters of the trained selection model using the newly added data in the training dataset.

[0063] Confirming postoperative efficacy involves assessing the degree of aneurysm occlusion through postoperative imaging follow-up (e.g., angiography at 1 or 3 months post-surgery) to determine whether the surgery achieved complete occlusion or met a preset threshold for embolization rate. Only cases with validated efficacy are included in the dataset, ensuring that new data also possesses clinical credibility. For successful cases with satisfactory efficacy, the implantation model serves as the ground truth label for positive samples; for cases with poor efficacy, it serves as a negative sample. Through continuous accumulation, the size and quality of the training dataset continuously improve over time.

[0064] After acquiring new data, this embodiment does not retrain the model from scratch, as retraining would consume significant computational resources and time, and old data might not be fully retained due to storage limitations. Therefore, this embodiment employs transfer learning or incremental learning strategies, using newly added data from the training dataset to update the parameters of the already trained selection model.

[0065] Transfer learning refers to using the knowledge (i.e., model parameters) learned by a model on an existing dataset as initial values, and then continuing training on new data with limited parameters. In practice, some of the lower-level network layers of the model (these layers typically learn general morphological features) can be fixed, while only the higher-level classification layers or the last few layers of the network are fine-tuned. Because the amount of new data is relatively small, the fine-tuning process can converge quickly, while avoiding catastrophic forgetting of existing knowledge.

[0066] Incremental learning is a more flexible continuous learning strategy that allows the model to gradually adapt to new data without revisiting all historical data. Incremental learning can be implemented in several ways, such as using Elastic Weight Consolidation (EWC) to add a regularization term to the loss function to protect parameters important to older tasks; or using Memory Replay to simultaneously train new data with random sampling of some older data, thus maintaining the model's memory of previous knowledge. The advantage of incremental learning is that it does not require storing all historical data, making it suitable for long-term applications where data accumulates continuously.

[0067] Through the aforementioned model update mechanism, the selection model can achieve the following beneficial effects: First, the model can continuously absorb new successful experiences and lessons learned in clinical practice, constantly optimizing its own parameters to adapt to the characteristics of patient groups, surgical habits, and the introduction of new device models in different medical institutions; Second, the transfer learning and incremental learning strategies avoid the high computational costs and storage overhead of retraining from scratch, making the model update process efficient and lightweight, and easy to execute regularly or on demand in the clinical environment; Third, the dynamic model update mechanism enables the technical solution of this application embodiment to have the ability to self-evolve, and the selection accuracy can be continuously improved with the accumulation of data, without gradually becoming ineffective due to changes in clinical practice.

[0068] In summary, this application not only provides a static selection model that is trained once based on historical data, but also constructs a dynamic learning system that can be continuously optimized along with clinical practice, which has significant practical application value.

[0069] In a second aspect, this application also provides an apparatus for selecting an intraneural perturbation device, comprising: a data acquisition module configured to acquire medical image data of a target object, wherein the target object includes at least an aneurysm and a tumor-bearing vessel; a segmentation and reconstruction module configured to segment and reconstruct the medical image data in three dimensions to obtain a segmentation model of the target object; and a selection module configured to receive the segmentation model and output prediction results of each candidate model of the intraneural perturbation device for model recommendation; wherein the selection module includes: a geometric feature extraction module configured to extract geometric features of the target object from the segmentation model, wherein the geometric features are explicitly measurable parametric features; and a classifier configured to output prediction results of each candidate model of the intraneural perturbation device based at least on the geometric features.

[0070] The description of the apparatus in this application embodiment can be combined with the foregoing text. Figures 1-3 The methods described in any of them are the same or similar, and will not be repeated here.

[0071] This application also provides an electronic device, including: a processor configured to execute program instructions; and a memory configured to store the program instructions, which, when loaded and executed by the processor, cause the processor to perform the functions described above according to this application. Figures 1-3 Any of the methods described herein. The following will combine... Figure 4 The system shown is described.

[0072] Figure 4 A schematic block diagram of a system for selecting an intratumoral flow disruptor according to an embodiment of this application is shown. The system 400 may include an electronic device 401 according to an embodiment of this application, as well as its peripheral devices and an external network, wherein the electronic device 401 is used to perform the operation of selecting an intratumoral flow disruptor to achieve the aforementioned combination. Figures 1-3 The technical solutions of any of the embodiments described in this application.

[0073] like Figure 4 As shown, electronic device 401 may include CPU 4011, which may be a general-purpose CPU, a dedicated CPU, or other information processing and program execution unit. Furthermore, electronic device 401 may also include a large-capacity memory 4012 and a read-only memory (ROM) 4013. The large-capacity memory 4012 may be configured to store various types of data, including medical images, vascular models, historical cases, model parameters, and other data, as well as various programs required for running methods for selecting intratumoral perturbation devices. The ROM 4013 may be configured to store data required for power-on self-test of electronic device 401, initialization of various functional modules in the system, drivers for basic input / output of the system, and data required to boot the operating system.

[0074] Furthermore, electronic device 401 also includes other hardware platforms or components, such as the TPU 4014, GPU 4015, FPGA 4016, and MLU 4017 shown. It is understood that although various hardware platforms or components are shown in electronic device 401, they are merely exemplary and not restrictive, and those skilled in the art can add or remove corresponding hardware as needed. For example, electronic device 401 may include only a CPU as a known hardware platform and another hardware platform as the hardware platform of this application.

[0075] The electronic device 401 of this application also includes a communication interface 4018, through which it can connect to a local area network / wireless local area network (LAN / WLAN) 405, and further connect to a local server 406 or the Internet 407 via the LAN / WLAN. Alternatively or additionally, the electronic device 401 of this application can also directly connect to the Internet or a cellular network via the communication interface 4018 based on wireless communication technology, such as third-generation ("3G"), fourth-generation ("4G"), or fifth-generation ("5G") wireless communication technology. In some application scenarios, the electronic device 401 of this application can also access a server 308 on an external network and a possible database 409 as needed to obtain data such as various known intratumoral perturbation device models, historical cases, and geometric characteristic parameters, and can remotely store various parameters or intermediate data.

[0076] Peripherals of electronic device 401 may include a display device 402, an input device 403, and a data transmission interface 404. In one embodiment, the display device 402 may include, for example, one or more speakers and / or one or more visual displays, configured to provide voice prompts and / or display images and videos of the prediction process or recommendation results of the electronic device. The input device 403 may include, for example, a keyboard, mouse, microphone, or other input buttons or controls, configured to receive input of medical image data or segmentation models or user instructions. The data transmission interface 404 may include, for example, a serial interface, parallel interface, or Universal Serial Bus interface (“USB”), Small Computer System Interface (“SCSI”), Serial ATA, FireWire (“FireWire”), PCI Express, and High Definition Multimedia Interface (“HDMI”), configured for data transmission and interaction with other devices or systems. According to the present application, the data transmission interface 404 can receive medical image data, etc., and transmit various types of data and results to electronic device 401.

[0077] The CPU 4011, mass storage 4012, read-only memory ROM 4013, TPU 4014, GPU 4015, FPGA 4016, MLU 4017, and communication interface 4018 of the electronic device 401 of this application can be interconnected via bus 4019, and can interact with peripheral devices through this bus. In one embodiment, the CPU 4011 can control other hardware components in the electronic device 401 and its peripheral devices through this bus 4019.

[0078] In operation, the processor CPU 4011 of the electronic device 401 of this application can receive medical images, segmentation models, etc., through the input device 403 or the data transmission interface 404, and retrieve computer program instructions or code stored in the memory 4012 to select an intratumoral turbulence device from the received segmentation model, so as to obtain a recommended result for the selection of the intratumoral turbulence device. After the CPU 4011 determines the recommended result by executing the program instructions, it can display the result on the display device 402 or output the result through voice prompts. In addition, the electronic device 401 can also upload the recommended result to a network, such as a remote database 409, through the communication interface 4018.

[0079] It should also be understood that any module, unit, component, server, computer, terminal, or device of the executable instructions in this application may include or otherwise access computer-readable media, such as storage media, computer storage media, or data storage devices (removable) and / or non-removable) such as disks, optical discs, or magnetic tapes. Computer storage media may include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, program modules, or other data.

[0080] Based on the foregoing, this application also provides a computer-readable storage medium storing computer-readable instructions thereon, which, when executed by one or more processors, implement the above-described combination of instructions. Figures 1-3 The method described in any of the embodiments.

[0081] Computer-readable storage media can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions that can be stored or otherwise maintained by such a computer-readable medium.

[0082] 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 selecting an intratumoral flow disturbance device, characterized in that, include: Acquire medical imaging data of a target object, wherein the target object includes at least an aneurysm and a tumor-bearing vessel; The medical image data is segmented and reconstructed in three dimensions to obtain a segmentation model of the target object; as well as The segmentation model is input into a deep learning-based selection model for model recommendation; The selection model mentioned above includes: A geometric feature extraction module is used to extract the geometric features of the target object from the segmentation model, wherein the geometric features are explicitly measurable parametric features; A classifier is used to output prediction results for each candidate model of the in-tumor turbulence device, based at least on the geometric features.

2. The method according to claim 1, characterized in that, The selection model further includes: A 3D morphology extraction module is used to obtain a 3D voxel model of the target object based on the segmentation model; and to extract the morphological features of the 3D voxel model using a 3D morphology extraction network, wherein the morphological features are implicit spatial structural features. The feature fusion module is used to fuse the geometric features and the morphological features to obtain fused features; The classifier is further used to classify the fused features and output the prediction results for each candidate model of the intratumoral disturbance device.

3. The method according to claim 2, characterized in that, The fusion of the geometric features and the morphological features includes: The geometric features are encoded through a geometric feature encoding layer to generate a geometric feature vector; The 3D voxel model is encoded through a morphological feature encoding layer to generate a morphological feature tensor; The geometric feature vectors and the morphological feature tensors are mapped to a unified dimensional space, and weighted fusion is performed using an attention mechanism.

4. The method according to claim 3, characterized in that, The three-dimensional morphology extraction network includes a three-dimensional convolutional neural network; The geometric feature encoding layer includes one or more fully connected networks; The morphological feature encoding layer includes a three-dimensional convolutional neural network and a pooling layer.

5. The method according to claim 1, characterized in that, The geometric features include: At least one of the following geometric characteristics of an aneurysm: long diameter, width, height, transverse diameter, longest diameter, aspect ratio, neck diameter, ratio of transverse diameter to neck diameter, volume, neck area, non-sphericity index, aspect ratio, volume-to-neck area ratio, and fluctuation index; and / or At least one of the following geometric characteristics of the tumor-bearing vessel: proximal diameter, distal diameter, average diameter, and length; and / or At least one of the following characteristics is considered in relation to the carrier vessel of the aneurysm: angle of incidence, aneurysm angle, and size ratio.

6. The method according to any one of claims 1-5, characterized in that, Before inputting the segmentation model into the deep learning-based selection model, the method further includes: A training dataset is constructed, which includes medical image data samples of historical cases and corresponding ground truth labels. The ground truth labels are the models of intratumoral perturbation devices that were successfully implanted during the operation and whose postoperative efficacy met the preset conditions. The selection model is trained using the training dataset, and the training of the selection model is supervised by a weighted cross-entropy loss function, wherein the weights of the weighted cross-entropy loss function are automatically adjusted according to the number of samples of different models in the training dataset.

7. The method according to claim 6, characterized in that, The truth labels are constructed in the following way: Extract the actual implanted model during surgery from the surgical records and consumables management system of the aforementioned historical cases; and Cases where postoperative imaging showed complete occlusion or an embolism rate reaching a preset threshold were retained, and their actual implanted model was used as the true value label for positive samples.

8. The method according to claim 7, characterized in that, The training dataset also includes negative samples of cases with postoperative recurrence, prolapse, or incomplete embolization from the historical cases, along with their corresponding negative sample labels; the method further includes: The selection model is jointly trained based on the true value label and the negative sample label.

9. The method according to any one of claims 6-8, characterized in that, Also includes: Once a new patient completes surgery and the postoperative efficacy is confirmed, the medical imaging data and implant type of the new patient will be added to the training dataset. The selection model is updated with parameters using newly added data from the training dataset by employing transfer learning or incremental learning strategies.

10. A device for selecting intratumoral turbulence devices, characterized in that, include: The data acquisition module is configured to acquire medical image data of a target object, wherein the target object includes at least an aneurysm and a tumor-bearing vessel; The segmentation and reconstruction module is configured to segment and reconstruct the medical image data in three dimensions to obtain a segmentation model of the target object. as well as The selection module is configured to receive the segmentation model and output the prediction results of each candidate model of the intratumoral disturbance device for model recommendation. The selection module includes: A geometric feature extraction module is configured to extract geometric features of the target object from the segmentation model, wherein the geometric features are explicitly measurable parametric features; A classifier, configured to output prediction results for each candidate model of the intratumoral perturbation device based at least on the geometric features.

11. An electronic device, characterized in that, include: A processor, configured to execute program instructions; as well as A memory configured to store the program instructions, which, when loaded and executed by the processor, cause the processor to perform the method according to any one of claims 1-9.

12. A computer-readable storage medium storing program instructions, characterized in that, When the program instructions are loaded and executed by the processor, the processor performs the method according to any one of claims 1-9.

Images