Patent Application
20250078263
The present disclosure relates to the field of artificial intelligence, specifically to a method, system, and computer program product for the diagnosis of intracranial aneurysms.
Intracranial aneurysm is a dangerous condition in the brain. If it ruptures and starts bleeding, it can lead to a high rate of death and severe disability. Thanks to advances in medical technology, it is now possible to detect unruptured intracranial aneurysms early using non-invasive imaging techniques such as magnetic resonance angiography (MRA) or computed tomography angiography (CTA).
Early diagnosis is crucial for effective clinical management and can help prevent rupture through interventional procedures. However, diagnosing intracranial aneurysms can be challenging due to the small size of the lesions, which can be as small as 1.5 mm, and the large number of CTA and MRA images, which can reach hundreds in one set. Additionally, the complex structure of the intracranial arteries makes manual interpretation time-consuming and labor-intensive. The accuracy of manual interpretation depends heavily on the qualifications of the radiologist, which increases the risk of misdiagnosis and discrepancies in interpretations. It often takes several years of training as a radiologist to develop a consistent reading, and even then, small intracranial aneurysms may still be missed.
To solve the technical problem of difficult interpretation of intracranial aneurysms, this disclosure provides a method, system, and computer-readable recording medium for determining intracranial aneurysms. By optimizing the intracranial aneurysms recognition process with an AI-based auxiliary diagnostic model for intracranial aneurysms, the accuracy of intracranial aneurysms recognition is significantly improved. Specifically, the technical solutions of this application are as follows:
In the first aspect, the present application discloses an embodiment of a method for determining intracranial aneurysms. The method includes the following steps:
In some embodiments, Step (b), which enhances a intracranial artery in the first image and performs threshold segmentation on an enhanced first image to obtain a second image containing only the intracranial artery, further includes the following steps:
In some embodiments, Step (c), which uses a first-stage detection model to predict a intracranial aneurysms in the second image to screen out a third image containing a suspected intracranial aneurysm region, further includes the following steps:
In some embodiments, the first-stage detection model is a 3D CNN model.
In some embodiments, Step (d), which using a second-stage segmentation model to segment the intracranial aneurysm region in the third image, obtaining a fourth image containing the intracranial aneurysm region, further includes the following steps:
In some embodiments, the second-stage segmentation model is a 3D U-Net model.
In some embodiments, prior to using the first-stage detection model to predict the intracranial aneurysms in the second image, the method further includes the following steps:
In some embodiments, prior to training the first-stage detection model and the second-stage segmentation model using a semi-supervised learning method with consistency execution strategy, the method further includes the following steps:
In another aspect, the present application also discloses a system for determining intracranial aneurysms used for performing the method according to any of the above features. The system includes:
In other aspect, the present application further discloses another embodiment of a method for determining intracranial aneurysms. The method includes the following steps:
In some embodiments, Step (b), generating a mask, and performing intracranial artery segmentation on the first image to obtain a intracranial artery region, further includes the following steps:
In some embodiments, the grouping result comprises the top 15 to 30 intracranial aneurysm pixel groups by total number, and groups with more than 100 to 200 pixels.
In some embodiments, Step (c), using a segmentation model to perform intracranial aneurysm segmentation on the intracranial artery region in the first image to generate a segmentation result, further includes the following steps:
In some embodiments, the segmentation model preferably is a 3D U-Net model.
In some embodiments, Step (d), using a second-stage segmentation model to segment the intracranial aneurysm region in the third image, obtaining a fourth image containing the intracranial aneurysm region, further includes the following steps:
In some embodiments, the intracranial aneurysm mask map is used to display the boundary, location, size, and shape of the intracranial aneurysm.
In some embodiments, the classification model preferably is a 3D CNN model.
In some embodiments, prior to using the segmentation model to perform intracranial aneurysm segmentation on the intracranial artery region in the first image to generate a segmentation result, further includes the following steps:
In some embodiments, the training of the segmentation model and the classification model using the semi-supervised learning method with consistency execution strategy, further includes the following steps:
In others aspect, the present application also discloses another system for determining intracranial aneurysms used for performing the method according to any of the above features. The system includes:
In other aspect, the present application further discloses a computer-readable recording medium for determining intracranial aneurysms. The computer-readable recording medium includes computer program or an instruction, wherein the computer program or instruction, when executed by a processor, implement any above of the method.
Compared to the prior art, the present application has at least the following beneficial effects or unexpected results:
1. Establishing a two-stage deep learning model for intracranial aneurysm detection, which retains high sensitivity in the first-stage detection model while quickly screening high-risk image slices. The second-stage segmentation model performs pixel-level segmentation, achieving high sensitivity, high accuracy, and high-quality intracranial aneurysm segmentation.
2. The pre-processed images input into the intracranial aneurysm detection model are segmented along the vessels based on the 3D vascular segmentation modeling results of the Modified Hessian Matrix algorithm to detect community block images. This method avoids misidentifications outside the vessels, significantly reducing computation time during the prediction of the intracranial aneurysm detection model. Additionally, during the training of the intracranial aneurysm detection model, the data only includes vascular parts, excluding other head tissues, making the data distribution more concentrated and allowing the model to better learn the features of the vessels and intracranial aneurysms.
3. Adopting a semi-supervised learning training strategy for further training of the two-stage deep learning model, incorporating the Consistency-Enforcing training strategy and dataset expansion in the training, with the goal of using the unlabeled data to regularize the network, and requiring that the output predictions should not change significantly with respect to the input image applying the actual disturbances to the outputs, so that the outputs have consistency.
4. Providing a comprehensive intracranial aneurysm reading system that integrates vascular segmentation imaging technology and deep learning models to improve detection sensitivity, perform precise quantitative analysis of intracranial aneurysm size and shape, and improve diagnostic accuracy to detect even small lesions, which helps detect intracranial aneurysms at an early stage for timely medical intervention. It can also predict the risk of rupture of the intracranial aneurysm for proactive medical management.
One or more specific embodiments are shown in the accompanying drawings by way of example and not by way of limitation. The accompanying drawings are not to scale unless otherwise disclosed. The present disclosure should be understood by those of ordinary knowledge in the art in light of the following detailed description of preferred specific embodiments and with reference to the accompanying drawings.
The above features, technical characteristics, advantages, and their realizations of the present disclosure will be further described below in a clear and easy-to-understand manner in conjunction with the illustrations of the preferred specific embodiments.
The following descriptions in the present disclosure are primarily intended to illustrate the content of the disclosure and not to limit its scope. Therefore, specific details such as system architecture and techniques are provided to thoroughly explain relevant specific embodiments of this application. However, one person having ordinary skill in the art to which this invention pertains should be able to implement the essence of this disclosure without these specific details. In certain instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details that might obscure the description of the disclosure.
It should be understood that when terms like “comprising,” “including” or “includes” are used in the specification and claims of this disclosure, they indicate the presence of 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.
For clarity in the drawings, only parts related to the disclosure are shown schematically, and they do not represent the actual structure as a product. Furthermore, for simplicity and ease of understanding, in some figures, parts with the same structure or function are schematically shown, or only one of them is labeled. Herein, “one” does not only mean “only one” but can also mean “more than one.”
Furthermore, it should be understood that the term “and/or” used in the specification and claims of this application means any combination and all possible combinations of one or more of the associated listed items and includes such combinations.
In the present disclosure, it needs to be explained that unless otherwise explicitly specified and limited, the terms “install,” “connect,” and “link” should be broadly understood. For example, they can be fixed connections, removable connections, or integrated connections; they can be mechanical connections or electrical connections; they can be direct connections or indirect connections through intermediaries, and they can be internal communications between two components. For those skilled in the art, the specific meanings of these terms in this application can be understood according to specific circumstances.
In specific embodiments, the terminal devices described in the specific embodiments disclosed in this application include but are not limited to portable devices such as mobile phones, notebook computers, educational tablets, or other tablets with touch-sensitive surfaces (e.g., touch screen displays and/or touchpads). It should also be understood that in certain specific embodiments, the terminal device is not a portable communication device but a desktop computer with a touch-sensitive surface (e.g., touch screen display and/or touchpad).
Additionally, in the description of the present application, terms like “first” and “second” are only used to distinguish descriptions and should not be understood as indicating or implying relative importance.
To more clearly describe the technical solutions in the specific embodiments or the existing technology in this disclosure, specific implementations of this application will be described below with reference to the accompanying drawings. It is obvious that the accompanying drawings described below are merely some specific embodiments of this application. For those persons having ordinary skill in the art to which this invention pertains, other drawings can be obtained and other possible specific implementations can be derived without inventive labor based on these drawings.
A intracranial aneurysm refers to an abnormal localized dilation of an artery wall caused by the expansion of the artery's lumen. Intracranial aneurysms most commonly rupture in the space between the brain and the thin tissues covering the brain. The formation and growth are due to the pressure exerted by blood flowing through the vessels on the weakened areas of the vessel wall. Intracranial aneurysms are common. However, most intracranial aneurysms are not severe, especially smaller ones. Most intracranial aneurysms do not rupture. They typically do not cause symptoms or health problems. Intracranial aneurysms are often detected when diagnosing other medical conditions. Once an aneurysm ruptures and bleeds, it can seriously cause consciousness disorders or even coma in the patient.
Therefore, early detection and accurate assessment of intracranial aneurysms are crucial for effective medical intervention and reducing the risk of rupture. Current diagnostic methods can early detect unruptured intracranial aneurysms through non-invasive imaging techniques such as Magnetic Resonance Angiography (MRA) or Computed Tomography Angiography (CTA). Early diagnosis aids early clinical handling and potential prevention of rupture through interventional surgery.
However, there are still unmet clinical needs in the current clinical diagnosis of intracranial aneurysms: intracranial aneurysms are relatively small, with diameters as small as 2 millimeters, and CTA and MRA imaging examinations involve substantial quantities, reaching about 200 to 300 images per set, and the intracranial artery structure is complex. Hence, diagnosing intracranial aneurysms is time-consuming and laborious, often requiring years of training for radiology specialists to stabilize readings, posing two challenges. Firstly, there is a relatively high degree of physician discrepancy in clinical CTA and MRA intracranial aneurysm detection, as lesion interpretation may vary with professional level, with studies in prominent international journals such as Stroke indicating that diagnostic rates for intracranial aneurysms grow with specialization and experience. However, limited medical workforce means not every brain imaging examination can be interpreted by a neuroradiologist, posing hidden issues in clinical diagnosis. Moreover, the diagnostic sensitivity for intracranial aneurysms is severely influenced by the lesion's size, with smaller aneurysms more likely to be missed. Considering the aforementioned clinical difficulties and unmet needs, establishing high-performance diagnostic tools to assist detection, improve efficiency, and reduce inconsistent physician judgments is necessary, aiding in improving patient clinical care.
Please refer to
S100, Step (a): Acquire the first image, which is a pre-processed three-dimensional image of the intracranial aneurysm. Specifically, current non-invasive imaging techniques like CTA have limitations, including the need for contrast agents and radiation exposure. In contrast, MRA does not require contrast agents and has no radiation, making it more suitable for screening unruptured intracranial aneurysms. Thus, this implementation uses MR three-dimensional imaging as an example. Analysis of CTA images. CTA has a shorter detection time than MRI and is often used in emergencies. However, it is not available to all patients because of the need to inject a contrast agent into the patient. However, since the patient must be injected with contrast material first, not all patients can use it. In addition, because of the contrast material injection, the threshold for AI model construction is lower, depending on the high contrast of the images. Of course, based on the same technical concept as this application, in other implementations, the diagnosis of intracranial aneurysms using images obtained by other existing techniques, such as CT or CBCT, should also be within the scope of protection of this application.
S200, Step (b): Enhance the intracranial artery in the first image and perform threshold segmentation on the enhanced first image to obtain a second image containing only the intracranial artery. Specifically, brain vessel 3D modeling using the Hessian matrix algorithm identifies the brain vessels' locations, segments out community block images along the vessel location for input into the aneurysm detection model, avoiding false positives outside the vessel area and significantly reducing sliding window computation time.
S300, Step (c): Use a first-stage detection model to predict intracranial aneurysms in the second image and further screen out a third image containing suspected intracranial aneurysm regions.
S400, Step (d): Use a second-stage segmentation model to segment the intracranial aneurysm regions in the third image, and further obtaining a fourth image containing the intracranial aneurysm regions. Specifically, in intracranial aneurysm detection, a single-stage AI detection model may produce many false positives due to high sensitivity requirements and difficulty balancing quality parameters. Therefore, this application designs a two-stage architecture where the first-stage detection model quickly screens high-risk image slices with high sensitivity, and the second-stage segmentation model produces pixel-level segmentation, reducing false positives in small positive sample areas, achieving high-sensitivity, high-accuracy, and high-quality intracranial aneurysm segmentation.
Another embodiment of a method for determining intracranial aneurysms disclosed in the present application is based on the embodiments of the method described above. In Step (b) (i.e., Step S200): enhance the intracranial artery in the first image and perform threshold segmentation on the enhanced first image to obtain a second image containing only the intracranial artery, specifically including the following steps:
S210, Step (b1): Use a second-order Hessian matrix to perform feature analysis on the first image, obtaining the characteristic value of the second-order intensity variations for each pixel point in the first image.
S220, Step (b2): Determine that a target pixel point belongs to an intracranial artery if the characteristic value of the target pixel point meets a threshold condition, and further merge all pixel points belonging to the intracranial artery to form an intracranial artery path.
S230, Step (b3): Segment and retain the image content on the intracranial artery path, remove the image content outside the intracranial artery path, and obtain the second image containing only the intracranial artery.
In other specific embodiment of the present disclosure, Step (b) (i.e., Step S200) is specifically explained as follows: by performing feature analysis on the Hessian matrix, the image features are obtained, capturing the local second-order intensity variations near each pixel point. The specific formula is as follows:
Wherein ∇ is the gradient operator; I is the target pixel; Ixx is the second-order intensity variation of the target pixel in the xx plane; Ixy is the second-order intensity variation of the target pixel in the xy plane; Ixz is the second-order intensity variation of the target pixel in the xz plane; lyx is the second-order intensity variation of the target pixel in the yx plane; Iyy is the second-order intensity variation of the target pixel in the yy plane; lyz is the second-order intensity variation of the target pixel in the yz plane; Izx is the second-order intensity variation of the target pixel in the zx plane; Izy is the second-order intensity variation of the target pixel in the zy plane; Izz is the second-order intensity variation of the target pixel in the zz plane.
Building the Hessian matrix of 3D medical images, we can calculate the intensity variation feature values in the X, Y, Z three directions: λ1, λ2, and λ3. Among them, λ1 represents the intensity variation along the vessel direction, λ2 and λ3 represent the intensity variations perpendicular to the vessel direction. According to the vessel structure, in TOF-MRA images, the gray value of the vessels is high, contrasting with the relatively darker background. The intensity variation along the main vessel direction is significantly smaller than the variation perpendicular to the vessel direction. This can be used as a consistency criterion to distinguish vessel structures. Based on this, the feature values well describe the vessel structures. When the values of λ2 and λ3 of a pixel point are large, and the value of λ1 is small, it is likely to belong to a vessel. The specific judgment condition is shown in the following formula:
Wherein λ1 is the intensity variation feature value of the target pixel in the X direction; λ2 is the intensity variation feature value of the target pixel in the Y direction; λ3 is the intensity variation feature value of the target pixel in the Z direction.
Thus, the intensity variations of each pixel point can be combined to form a vessel path. Refer to
If the whole brain image is input into the intracranial aneurysm detection model, it may falsely identify bright spots outside the vessel as suspected aneurysms, generating too many false positives, making it difficult to assist doctors. Therefore, in this implementation, the pre-processed images input into the aneurysm detection model are segmented along the vessel path based on the 3D vessel segmentation modeling results using the Hessian matrix algorithm. Refer to
In another embodiment of a method for determining intracranial aneurysms disclosed in the present application, based on any embodiments of the aforementioned method. In Step (c) (i.e., Step S300): Use a first-stage detection model to predict a intracranial aneurysms in the second image, and screen out a third image containing a suspected intracranial aneurysm region, specifically including the following steps:
S310, Step (c1): Use a sliding window segmentation method to divide the second image along the intracranial artery path to obtain a plurality of sub-images.
S320, Step (c2): Send the plurality of sub-images in batches to the first-stage detection model to predict a intracranial aneurysms location.
S330, Step (c3): Process a prediction result and screen out the third image containing suspected intracranial aneurysm region.
In the present embodiment, the first-stage detection model uses a 3D CNN model as an example. Using other existing models, such as VGG, ResNet, YOLO, U-Net, etc., as the first-stage detection model should also fall within the scope of the present disclosure.
Refer to
In another embodiment of the method for determining intracranial aneurysms disclosed in the present disclosure, Step (d) (i.e., step S400): using the second-stage segmentation model to segment the intracranial aneurysm regions in the third image, obtaining the fourth image containing the intracranial aneurysm regions, includes the following steps:
In the present embodiment, the second-stage model is a 3D U-Net model. Using other existing models as the first-stage detection model should also fall within the scope of this application.
Specifically, the second-stage aneurysm segmentation model refines the morphology of the aneurysm predicted by the first-stage aneurysm detection model. Through 3D U-Net, it precisely segments the lesion's position, size, and shape. Refer to
In practical applications, this application's AI-assisted diagnostic model for intracranial aneurysm is divided into a two-stage model. In the model test set (including 90 cases of intracranial aneurysms), the first-stage aneurysm detection model's (Aneurysm Detection Model) performance in screening high-risk intracranial aneurysms is excellent (accuracy rate 95.2%, specificity 96.0%). Additionally, the overall detection sensitivity of the second-stage aneurysm segmentation model (Aneurysm Segmentation Model) is 82.2%. Notably, the model's performance is unaffected by the aneurysm size, maintaining 80% sensitivity even for aneurysms less than or equal to 3 millimeters, equivalent to the aforementioned neuroradiologists' level.
Comparing (1) the AI model alone; (2) physicians without AI assistance; and (3) physicians using our AI platform. The test results show that (1) the AI model alone detects intracranial aneurysms with 84.7% sensitivity (312/367); (2) physicians without AI assistance have about 84.2% sensitivity (310/367); (3) physicians using AI achieve 96.7% sensitivity (355/367), indicating the AI model's performance is comparable to physicians' interpretations, consistent with the aforementioned internal test set results. More notably, we found that the AI model and physicians' interpretations are complementary: the lesions missed by the AI model and those missed by physicians often do not overlap. Physicians are more likely to make mistakes when the patient has multiple intracranial aneurysms. When there is more than one aneurysm, physicians are likely to miss the second or third aneurysm, even if these lesions are near the already diagnosed lesion. This may relate to human inertia and fatigue when performing high-intensity repetitive tasks, leading to relaxation after finding the first lesion and missing others, whereas AI does not have this issue. Conversely, the AI model is more likely to miss fewer common lesions in the training set (rare locations and types), even if the lesion size is not necessarily very small, while physicians are less likely to miss such lesions. This complementarity between the AI model and physicians' interpretations allows physicians using AI-assisted interpretations to achieve up to 96.7% diagnostic sensitivity, meaning physicians using AI can reduce the misdiagnosis rate by 12.5% compared to those not using AI, with significant benefits. This further enables early diagnosis, regular tracking, and proper management to prevent aneurysm rupture and hemorrhagic stroke complications.
In another embodiment of a method for determining intracranial aneurysms disclosed in the present application, based on any embodiments of the aforementioned method, before using the first-stage detection model to predict intracranial aneurysms in the second image, the following steps are included:
Step S010: Construct the first-stage detection model and the second-stage segmentation model.
Step S020: Train the first-stage detection model and the second-stage segmentation model using a semi-supervised learning method with consistency execution strategy.
In this disclosure, Step S020: training the first-stage detection model and the second-stage segmentation model using a semi-supervised learning method with consistency execution strategy further includes the following steps:
Step S021: Construct the teacher-student models for both the first-stage detection model and the second-stage segmentation model.
Step S022: Use labeled image data as a training set to train the teacher model, obtaining the teacher labeling loss.
Step S023: Use the teacher model to predict unlabeled image data, obtaining the prediction results. Add labels to the unlabeled image data, obtaining the confidence of the prediction results.
Step S024: Add the high-confidence prediction results as pseudo-labels to the training set. Use the training set to train the student model, obtaining the student unlabeled loss.
Step S025: Combine the teacher labeling loss and the student unlabeled loss to obtain the total loss function. Update the student's model weights using the total loss function and replace the student model with the new teacher model. Continue training until the model converges.
Specifically, to further optimize the intracranial aneurysm segmentation model, this disclosure additionally collects a large amount of clinical unlabeled TOF-MRA images as the training set, using a semi-supervised learning strategy to further train the model.
In other embodiment of the present disclosure, further improvements are made to the training method. Before model training, the labeled data undergoes data augmentation operations, including random flipping, random rotation, random zooming, random translation, and random shearing. For small amounts of labeled training data, significant improvements can be achieved, and even for large amounts of data, improvements can be made.
After performing data augmentation on the training set data, the teacher model is input to obtain prediction labeled, calculating the teacher labeled loss, without updating the model weights. The unlabeled data first undergoes data augmentation operations such as random flipping, then is input into the teacher model to obtain prediction labeled. To convert prediction labeled to pseudo label, the present disclosure incorporates the Consistency-Enforcing training strategy, aiming to use unlabeled data to regularize the network, ensuring the output predictions remain consistent despite applying actual perturbations to the input images.
In other embodiments of the present disclosure, the semi-supervised learning training process is divided into three main parts, as shown in
Specifically, in this implementation, the actual practice is to first use a specific threshold to convert soft labels to hard labels, and apply actual perturbations to the images and prediction labeled. These actual perturbations are achieved through data augmentation, finally generating image (X unlabeled) and pseudo label (Y unlabeled). Input X unlabeled into the student model to obtain P unlabeled, and calculate the student unlabeled loss with Y unlabeled, without updating the model weights. Finally, combine the teacher labeled loss and the student unlabeled loss into the total loss, update the student's model weights, and replace the new teacher model, continuing training until the model converges.
In practical applications, the semi-supervised learning method adopted in the present disclosure improves the performance for various amounts of labeled data. As shown in
In other embodiments of the present disclosure, the method for determining intracranial aneurysms also includes using a risk prediction module to predict and quantify the rupture risk of lesions. Specifically, the segmented intracranial aneurysm images from the two-stage deep learning model are input into the risk prediction module, which integrates clinical data, lesion characteristics, and related patient information to predict the likelihood of intracranial aneurysm rupture, helping medical staff identify high-risk cases needing immediate attention, thereby reducing the risk of hemorrhagic stroke-related patient death and disability.
Based on the same technical concept, this disclosure also reveals a system for determining intracranial aneurysms, which can implement any method for determining intracranial aneurysms. Specifically, as shown in
In another embodiment of the system for determining intracranial aneurysms disclosed in the present application, based on the embodiments of the aforementioned system, the vessel segmentation module further includes the follows. A feature analysis submodule used for performing feature analysis on the first image using the second-order Hessian matrix and obtaining the second-order intensity variations for each pixel point in the first image. A feature value calculation submodule used for constructing the third-order Hessian matrix, analyzing the second-order intensity variations for each pixel point, and obtaining the characteristic value of the third-order intensity variations for each pixel point. When the characteristic value of the target pixel point meets a certain threshold condition, it is determined to belong to a vessel. All pixel points belonging to the vessel are merged to form a vessel path. A path segmentation submodule used for segmenting and retaining the image content on the vessel path, removing the image content outside the vessel path, and obtaining the second image containing only the intracranial artery.
The intracranial aneurysm region screening module is specifically used to perform the following steps:
The intracranial aneurysm region segmentation module is specifically used to perform the following steps, wherein the first-stage detection model is a 3D CNN model:
Another embodiments of the system for determining intracranial aneurysms provided in the present application, based on any one of the aforementioned embodiments, the model construction module is also used to train the first-stage detection model and the second-stage segmentation model using a semi-supervised learning method with a consistency execution strategy.
Specifically, the model construction module is used to perform the following steps:
Please refer to
Another preferable embodiment of the method for determining intracranial aneurysms disclosed herein is based on the specific embodiment of the method for determining intracranial aneurysms mentioned above. Step (b) (i.e., Step S2000): Generating a intracranial artery region mask, and performing intracranial artery segmentation on the first image to obtain a intracranial artery region, includes the following steps:
Another specific embodiment of the method for determining intracranial aneurysms disclosed herein is based on the specific embodiment of the method for determining intracranial aneurysms mentioned above. Step (c) (i.e., Step S3000): Using a segmentation model, performing intracranial aneurysm segmentation on the intracranial artery region in the first image to generate a segmentation result. The specific details and description of the segmentation model in this step can refer to the content of the segmentation model in other specific embodiments previously mentioned. Therefore, another preferred embodiment of the method for determining intracranial aneurysms includes the following steps:
In this preferable embodiment, the preferred segmentation model is a 3D U-Net model. However, other similar models used as segmentation models should also be within the protection scope of the present application.
Another specific embodiment of the method for determining intracranial aneurysms disclosed herein, Step (d) (i.e., Step S4000): Using a classification model to classify the segmentation result into intracranial aneurysm and non-intracranial aneurysm, and screening out the true positive result in the classification result, includes the following steps:
In this preferable embodiment, the preferred second-stage model (i.e., the classification model) is a 3D CNN model. However, other similar models should also be within the protection scope of the present application.
A preferred embodiment of the method for determining intracranial aneurysms disclosed herein has specific preferred effects including: significantly reducing processing time (reducing the analysis time for one sample from 7 minutes to about 30 seconds), greatly improving the segmentation effect of intracranial artery, significantly improving the detection effect of intracranial aneurysms, and can be applied to 3D intracranial artery presentation to assist clinical diagnosis. Furthermore, the applicable fields include: it can be applied to intracranial artery skeleton or intracranial artery diameter analysis, intracranial artery stenosis analysis, intracranial artery curvature and hemodynamics analysis, and can be integrated into a complete intracranial artery analysis platform in the future.
Another specific embodiment of the intracranial aneurysm interpretation system provided by the present application, based on any one of the embodiments mentioned above, further comprises the following steps before using the segmentation model to predict intracranial aneurysms in the first image:
Another specific embodiment of the intracranial aneurysm interpretation system provided by the present application, based on any one of the embodiments mentioned above, wherein the training of the segmentation model and the classification model using the semi-supervised learning method with a consistency execution strategy includes the following steps:
Please refer to Table 1. Table 1 shows a preferred embodiment of the method for determining intracranial aneurysms disclosed herein, using 800 known cases (including 112 intracranial aneurysms) for training to optimize the model. First, the terms in the table are explained: Sensitivity represents the ability to correctly identify positive cases, with high sensitivity indicating a low false-negative rate; Specificity represents the ability to correctly identify negative cases, with high specificity indicating a low false-positive rate. Therefore, a positive test result is more certain to indicate the presence of the lesion; FPS (Average of False Positive Per Scan) represents the average number of false positives per scan. The first detection method represents the detection method mentioned at the beginning of the disclosure, which includes the first-stage detection model and the second-stage segmentation model; The second detection method represents the detection method mentioned later in the disclosure, which includes using the segmentation model first and then the classification model; Thr. 1 (Threshold 1) represents threshold 1 or FPS>1; Thr. 2 (Threshold 2) represents threshold 2 or FPS<1. As shown in Table 1, all detections are performed using the method for determining intracranial aneurysms disclosed herein (including both specific embodiments).
Furthermore, when using the preferred embodiment of the method for determining intracranial aneurysms disclosed herein (i.e., the second detection method in Table 1), the model sensitivity is significantly improved. In the same validation set as the first detection method, the sensitivity for intracranial aneurysms of size≥3 mm reaches 97.2% (i.e., [(96.7%*69+100%*13)/(69+13)]*100), and FPS is controlled below 1.45.
Please refer to Table 2. Table 2 shows the test results on a test set of 230 cases obtained from Taipei University Hospital (including a total of 311 intracranial aneurysms). As shown in Table 2, using the preferred embodiment of the method for determining intracranial aneurysms disclosed herein, the model sensitivity is significantly improved, with a sensitivity for intracranial aneurysms of size≥3 mm reaching 97.3% (i.e., [(97.0%*202+100%*25)/(202+25)]*100), and FPS is controlled below 1.69.
Notably, existing techniques (or prior techniques) or products on the market can only effectively detect intracranial aneurysms of size≥4 mm. However, this technology can effectively detect intracranial aneurysms of size≥1 mm using MRA, with even greater efficacy for detecting intracranial aneurysms of size≥3 mm. Therefore, it has an unexpected result compared to existing technologies.
Based on the same technical concept, the present disclosure also discloses a computer program, including computer programs/instructions characterized in that the computer programs/instructions, when executed by a processor, implement the steps of any one of the methods described in the embodiments.
The method, system, and computer program product for determining intracranial aneurysms disclosed in the present application have the same technical features, and the technical details of the embodiments of the three can be applied to each other. To reduce repetition, these are not elaborated here.
One person having ordinary skill in the art to which this invention pertains can understand that for convenience and brevity of description, only the above program modules' division is exemplified. In practical applications, the functions can be assigned to different program modules as needed, completing all or part of the functions described above. The program modules in the embodiments can be integrated into a processing unit, or they can physically exist separately. Multiple units or components can be integrated into one processing unit, which can be realized in hardware form or software program units. Additionally, the specific names of each program module are only for mutual distinction and do not limit the scope of protection of the present disclosure.
In the above embodiments, each embodiment's description has its focus, and the parts not described in detail in a particular embodiment can refer to other embodiments' relevant descriptions.
One person having ordinary skill in the art to which this invention pertains can be aware that the units and algorithm steps described in each example disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The execution of these functions in hardware or software depends on the specific application and design constraints. A professional technician can choose different methods to implement the described functions for each specific application, but such embodiments should not be considered beyond the scope of the present disclosure.
In the disclosed embodiments, it should be understood that the disclosed devices and methods can be realized in other ways. For example, the above-described device embodiments are merely exemplary. For example, the division of modules or units is only for logical function division. There can be other division methods in practical implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device, or unit, which can be electrical, mechanical, or other forms.
The units described as separate components can be or not be physically separate. The components displayed as units can be or not be physical units and can be located in one place or distributed across multiple network units. The purpose of implementing the solution of this embodiment can be achieved by selecting part or all of the units according to actual needs.
Additionally, in the various embodiments of the present disclosure, each functional unit may be integrated into one processing unit or may be physically separate units. The integrated unit may be implemented in hardware form or software functional units.
Although specific embodiments of the present disclosure have been described, a person having ordinary skill In the art once knowing the basic inventive concept can make various changes and modifications to these embodiments. Therefore, the appended claims should be construed to include the preferred embodiments and all changes and modifications falling within the scope of the present disclosure.
It is evident that those person having ordinary skill In the art can make various modifications and variations to this application without departing from the spirit and scope of the present application. Therefore, if such modifications and variations fall within the scope of the claims and their equivalent technologies, the present application also intends to include these modifications and variations.
This application claims priority of Provisional Application No. 63/580,389, filed on Sep. 4, 2023, the content of which is incorporated herein in its entirety by reference.
| Number | Date | Country | |
|---|---|---|---|
| 63580389 | Sep 2023 | US |
