Robust intracranial aneurysm segmentation method for unmarked CTA scanning

Abstract

The invention provides a robust intracranial aneurysm segmentation method for unlabeled CTA scanning, and the method comprises the following steps: designing a triple learning framework comprising pseudo-supervised learning, comparative learning and confidence learning, and learning a robust intracranial aneurysm segmentation model from brain CTA scanning with unreliable labels. The brain CTA scan is pre-processed to remove redundant information and generate various enhanced views of the initial input. And the preprocessed CTA image is sent to a pseudo-supervised learning module, and an enhanced average teacher model is adopted to carry out pseudo-supervised learning on CTA scanning with unreliable labels. A conversion network is added to an enhanced teacher network and is used for voxel-level contrast learning. According to the method and the diagnosis system, the accuracy of intracranial aneurysm segmentation is improved, the diagnosis time of a doctor can be effectively shortened, and the efficiency is improved. Meanwhile, missed diagnosis or misdiagnosis caused by different diagnosis levels or fatigue of doctors is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional image processing, and in particular to a triple learning framework and system for segmenting intracranial aneurysms in unreliable labeled CTA scans. Background Art

[0002] Intracranial aneurysms are tumor-like protrusions caused by local congenital defects or damage to the intracranial arterial wall, causing abnormal enlargement of the local lumen. According to statistics, 80% of non-traumatic subarachnoid hemorrhages are caused by rupture of intracranial aneurysms, leading to poor prognosis or even death, especially in the case of secondary bleeding. Digital subtraction angiography, computed tomography angiography, and magnetic resonance angiography are commonly used methods for detecting intracranial aneurysms. Although digital subtraction angiography is considered the gold standard, it is hindered by time-consuming, invasive, and high technical requirements. On the other hand, the low vascular spatial resolution of magnetic resonance angiography hinders the segmentation of intracranial aneurysms. In contrast, CTA has the advantages of non-invasiveness, time-saving, and high image spatial resolution, making it more suitable for intracranial aneurysm segmentation research.

[0003] Deep learning has shown great promise in the field of medical image segmentation, often reaching or even exceeding the diagnostic accuracy of human experts, especially in areas such as brain tumor and lung nodule segmentation. With the continuous advancement of CTA and other advanced medical imaging technologies, the number of unlabeled scans related to intracranial aneurysms has increased significantly. However, the process of obtaining a large amount of reliable labeled data is not only time-consuming but also laborious, which poses a huge challenge to the study of deep learning in intracranial aneurysm segmentation.

[0004] Currently, there are two main paradigms used to address these challenges. The first paradigm is called semi-supervised segmentation, which leverages unlabeled data to improve segmentation performance through pseudo-supervised learning. This approach involves using the predictions of the segmentation network on unlabeled images as pseudo-labels during training. Pseudo-supervision can be divided into offline and online training techniques. Offline training generates pseudo-labels for unlabeled data to expand the training set, while online training introduces perturbations to the model or input data and uses consistency learning to ensure the similarity between the perturbed inputs and the predictions of the pseudo-labels. The second paradigm focuses on automating the labeling process to quickly generate unreliable labels, which may arise due to the lack of automated segmentation tools. Subsequently, experts are responsible for re-labeling or performing quality assessments on subsets of these data, called reliable labels. However, unreliable labels introduce uncertainty into the training of deep neural networks.

[0005] Despite significant progress in studying these two methods, there are still two major challenges to be addressed in the intracranial aneurysm segmentation task. First, pseudo-supervision employs a segmentation model to generate pseudo labels for unlabeled data. However, the model is often susceptible to label noise, which may lead to instability and reduce the effectiveness of pseudo-supervision. Second, since the region of interest is relatively small relative to the entire image, the voxel averaging method commonly used in pseudo-supervision may destroy the information value of the region. In addition, most pseudo-supervision techniques only provide guidance in the label space and lack explicit support in the feature space, which limits the improvement of class separability. In addition, directly merging labeled data of different qualities may confuse the network training process, offset the advantages of reliable labels, and lead to performance degradation. Therefore, a novel and reliable method to exploit the discriminative information contained in unreliable labeled data is the key to solving the above problems. Summary of the invention

[0006] The main purpose of the present invention is to provide a robust intracranial aneurysm segmentation method for unlabeled CTA scanning, which method and system improve the accuracy and robustness of intracranial aneurysm segmentation in real scenarios.

[0007] To achieve the above objectives, the present invention provides a robust intracranial aneurysm segmentation method for unlabeled CTA scans, which includes collecting low-precision labeled brain CTA scans to be segmented and high-precision labeled brain CTA scans containing intracranial aneurysms;

[0008] An intracranial aneurysm segmentation network model based on three learning paradigms was constructed, and intracranial aneurysm features were extracted.

[0009] The intracranial aneurysm segmentation network model is trained according to the intracranial aneurysm data to obtain a trained model weight.

[0010] The trained model weights are applied, and a model test is performed based on the intracranial aneurysm data to generate a result of intracranial aneurysm segmentation.

[0011] As a preferred solution of the robust intracranial aneurysm segmentation method based on unlabeled CTA scans of the present invention, the collecting of low-precision labeled brain CTA scans to be segmented and high-precision labeled brain CTA scans containing intracranial aneurysms includes:

[0012] The collected CTA scans are converted into different formats and the data is annotated;

[0013] The preprocessed intracranial aneurysm data is used as the training and testing data set for the segmentation network.

[0014] As a preferred solution of the robust intracranial aneurysm segmentation method based on unlabeled CTA scanning described in the present invention, the intracranial aneurysm segmentation network model includes a preprocessing module, a pseudo-supervised learning module, a comparative learning module, and a multi-scale confident learning module.

[0015] The pre-processing module comprises:

[0016] Applying intensity normalization, bilinear interpolation, and random gridding to the original view distorts the original input data to generate an enhanced view;

[0017] The intracranial aneurysm area is cropped into more accurate three-dimensional image block data to reduce the interference of redundant information on model training.

[0018] The pseudo-supervised learning module includes:

[0019] The preprocessed intracranial aneurysm data is taken as input;

[0020] The enhanced mean teacher model is used to process intracranial aneurysm data, extract the characteristics of intracranial aneurysms, and use the soft voting mechanism to reduce noise interference and improve the effectiveness of model training;

[0021] A voxel selection strategy is adopted instead of the traditional voxel averaging method, and the uncertainty region is integrated into the consistency loss function to pay more attention to these information-rich but unstable regions.

[0022] The contrastive learning module includes:

[0023] In the decoding block of the enhanced mean teacher network, a conversion network is used to extract high-level semantic features;

[0024] Construct positive and negative training pairs using the same position in different views in the high-level semantic feature space;

[0025] The multi-scale confidence learning module includes:

[0026] The predicted pseudo labels in the pseudo-supervised learning module and the low-precision intracranial aneurysm CTA labels are used as input;

[0027] Use multi-scale confidence strategies to effectively perceive a certain range of structures rather than just the information of a single voxel;

[0028] By calculating the confidence and threshold, the corrected low-precision labels are iteratively updated to provide additional training information for training.

[0029] As a preferred solution of the robust intracranial aneurysm segmentation method based on unlabeled CTA scanning according to the present invention, the specific process of the voxel selection strategy includes:

[0030] Receive the predicted pseudo labels of the student model and the teacher model as input;

[0031] By calculating the average threshold and classification probability of each class, the potential voxel entities within the uncertainty region can be mapped;

[0032] Since intracranial aneurysms are relatively small in CTA scans, latent voxel entities are integrated into the model training process to focus more on these informative but unstable regions.

[0033] As a preferred solution of the robust intracranial aneurysm segmentation method based on unlabeled CTA scanning according to the present invention, the specific design of the conversion network includes:

[0034] It consists of two 3D convolution blocks and a PRelu activation function;

[0035] Project the high-dimensional hidden representation of the decoding part of the enhanced teacher model into a low-dimensional space;

[0036] As a preferred solution of the robust intracranial aneurysm segmentation method based on unlabeled CTA scanning described in the present invention, the specific design of the multi-scale confidence strategy includes:

[0037] The predicted pseudo labels in the pseudo-supervised learning module and the low-precision intracranial aneurysm CTA labels are used as input;

[0038] Use 1×1, 2×2 and 3×3 multi-scale sliding windows to learn pseudo labels, thereby obtaining a joint confidence matrix of three scales;

[0039] After adding the three scales separately, we get the multi-scale joint confidence matrix;

[0040] Correcting the multi-scale joint confidence matrix to obtain a correction matrix;

[0041] Further normalization is performed to generate a multi-scale label matrix;

[0042] The uncertainty of each voxel is estimated using the predicted entropy to obtain a potential target map.

[0043] As a preferred solution of the robust intracranial aneurysm segmentation method based on unlabeled CTA scanning according to the present invention, the training of the intracranial aneurysm segmentation network model includes:

[0044] Use pseudo-supervised learning modules to enhance the model’s training effect on small target intracranial aneurysms;

[0045] Improve feature expression capabilities and capture feature space through contrastive learning modules;

[0046] Utilizes a multi-scale confidence learning module to focus on broader local structures rather than individual voxels, thus improving correction efficiency;

[0047] By effectively combining the three learning paradigms, the model can be robustly trained on a mixture of low-precision labeled brain CTA scans and high-precision labeled brain CTA scans containing intracranial aneurysms.

[0048] As a preferred solution of the robust intracranial aneurysm segmentation method based on unlabeled CTA scanning according to the present invention, the generated intracranial aneurysm segmentation result includes:

[0049] Use the trained model weights to perform intracranial aneurysm data segmentation test;

[0050] Output and verify the accuracy of segmentation results and model performance;

[0051] The accuracy of the entire segmentation was calculated by the ratio of correctly predicted voxels to the total voxels, and the segmentation intersection-union ratio was calculated by comparing the ratio of the number of correctly predicted voxels to the actual number of voxels of intracranial aneurysms and the union of the predicted results.

[0052] This paper develops a triple learning framework to learn a robust intracranial aneurysm segmentation model from brain CTA scans with unreliable labels. TLF mainly consists of three learning paradigms: pseudo-supervised learning, contrastive learning, and confident learning. These three learning methods work together to iteratively extract semantic information of intracranial aneurysms from unreliable labeled brain CTA scans. Experiments show that the model has good adaptability and performance, and can effectively improve the accuracy of intracranial aneurysm segmentation.

[0053] Compared with the prior art, the present invention has the following beneficial effects: In the intracranial aneurysm segmentation network, high-quality intracranial aneurysm segmentation is achieved from unreliable labeled brain CTA scans by collaboratively using three learning methods. In the enhanced mean teacher model, the encoder-decoder network is used as the backbone network of the student network and the teacher network, and the soft voting method is used for decision-making to improve the stability and effectiveness of the model. The voxel selection strategy is used to integrate the uncertain region into the consistency loss function to pay more attention to these information-rich but unstable regions. Secondly, additional enhanced views are added on the basis of the original view, and stricter consistency regularization is used, which helps to further stabilize the training process of the proposed method. At the same time, positive and negative training pairs are constructed in the high-level semantic feature space to realize contrastive learning, and the top K samples are selected as negative samples according to the prediction confidence, which reduces the false negative rate in the sampling process. Through multi-scale confidence learning, confidence learning is performed from a larger range of local structures rather than a single voxel, which improves the effectiveness of unreliable label correction. In summary, the robust intracranial aneurysm segmentation method for unlabeled CTA scans proposed in the present invention improves the segmentation accuracy of intracranial aneurysms and achieves a high accuracy on the intracranial aneurysm dataset, which can play a certain auxiliary role in the doctor's diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0055] Figure 1 An overall flow chart of a robust intracranial aneurysm segmentation method for unlabeled CTA scans provided by one embodiment of the present invention.

[0056] Figure 2 A schematic diagram of the structure of a pseudo-supervised learning module in a robust intracranial aneurysm segmentation method for unlabeled CTA scans provided by an embodiment of the present invention.

[0057] Figure 3 A flowchart of how a multi-scale confidence learning module identifies potential noise in a label in a robust intracranial aneurysm segmentation method for unlabeled CTA scans provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the above-mentioned purposes, features and advantages of the present invention more understandable, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0059] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0060] Figure 1 The overall flow chart of a robust intracranial aneurysm segmentation method for unlabeled CTA scans is shown, including:

[0061] Step S1: collecting low-precision labeled brain CTA scans to be segmented and high-precision labeled brain CTA scans containing intracranial aneurysms;

[0062] The collected CTA scans are converted into different formats and the data is annotated;

[0063] The preprocessed intracranial aneurysm data is used as the training and testing data set for the segmentation network.

[0064] Step S2: construct a robust intracranial aneurysm segmentation network model for unlabeled CTA scans and extract intracranial aneurysm features;

[0065] Furthermore, the intracranial aneurysm segmentation network model includes a preprocessing module, a pseudo-supervised learning module, a contrastive learning module, and a multi-scale confidence learning module.

[0066] Furthermore, the preprocessing module includes,

[0067] Applying intensity normalization, bilinear interpolation, and random gridding to the original view distorts the original input data to generate an enhanced view;

[0068] The original image input is cropped from 512×512×D (D is the depth of the image) to 128×128×128 according to the area where the intracranial aneurysm is located to reduce the interference of redundant information.

[0069] Noise is added to some annotations to simulate the scenario where a large amount of high-precision annotation data cannot be obtained and only a large amount of low-precision annotations and a small amount of high-precision annotations can be used for mixed training.

[0070] Furthermore, Figure 2As shown, the pseudo-supervised learning module includes,

[0071] The preprocessed intracranial aneurysm data is used as the input of the network, and the size of the input data is 128×128×128.

[0072] Use the enhanced mean teacher model to perform feature extraction on the input;

[0073] A voxel selection strategy is adopted to enhance the effectiveness of informative regions in the annotated data, thereby improving the network's ability to learn features.

[0074] Furthermore, the comparative learning module includes,

[0075] In the decoding block of the enhanced mean teacher network, a conversion network is used to extract high-level semantic features;

[0076] Construct positive and negative training pairs using the same position in different views in the high-level semantic feature space;

[0077] Furthermore, the multi-scale confidence learning module includes,

[0078] The predicted pseudo labels in the pseudo-supervised learning module and the low-precision intracranial aneurysm CTA labels are used as input;

[0079] Use multi-scale confidence strategies to effectively perceive a certain range of structures rather than just the information of a single voxel;

[0080] By calculating the confidence and threshold, the corrected low-precision labels are iteratively updated to provide additional training information for training.

[0081] Step S3: using the collected intracranial aneurysm data as input, training the intracranial aneurysm segmentation network model to obtain the trained model weights;

[0082] Furthermore, the strategies used in model training include voxel selection strategy and multi-scale confidence learning strategy;

[0083] Specifically, voxel selection strategies include,

[0084] Receive the predicted pseudo labels of the student model and the teacher model as input;

[0085] By calculating the average threshold and classification probability of each class, the potential voxel entities within the uncertainty region can be mapped;

[0086] Since intracranial aneurysms are relatively small in CTA scans, latent voxel entities are integrated into the model training process to focus more on these informative but unstable regions.

[0087] Specifically, Figure 3 As shown, the multi-scale confidence learning strategy includes,

[0088] The predicted pseudo labels in the pseudo-supervised learning module and the low-precision intracranial aneurysm CTA labels are used as input;

[0089] Use 1×1, 2×2 and 3×3 multi-scale sliding windows to learn pseudo labels, thereby obtaining a joint confidence matrix of three scales;

[0090] After adding the three scales separately, we get the multi-scale joint confidence matrix;

[0091] Correcting the multi-scale joint confidence matrix to obtain a correction matrix;

[0092] Further normalization is performed to generate a multi-scale label matrix;

[0093] The uncertainty of each voxel is estimated using the predicted entropy to obtain a potential target map.

[0094] Furthermore, the intracranial aneurysm segmentation network model is trained including:

[0095] Use pseudo-supervised learning modules to enhance the model’s training effect on small target intracranial aneurysms;

[0096] Improve feature expression capabilities and capture feature space through contrastive learning modules;

[0097] Utilizes a multi-scale confidence learning module to focus on broader local structures rather than individual voxels, thus improving correction efficiency;

[0098] By effectively combining the three learning paradigms, the model can be robustly trained on a mixture of low-precision labeled brain CTA scans and high-precision labeled brain CTA scans containing intracranial aneurysms.

[0099] Step S4: Apply the trained model weights and test the intracranial aneurysm data to generate intracranial aneurysm segmentation results.

[0100] Furthermore, the generated intracranial aneurysm segmentation results include:

[0101] Use the trained model weights to perform intracranial aneurysm data segmentation test;

[0102] Output and verify the accuracy of segmentation results and model performance;

[0103] The accuracy of the entire segmentation was calculated by the ratio of correctly predicted voxels to the total voxels, and the segmentation intersection-union ratio was calculated by comparing the ratio of the number of correctly predicted voxels to the actual number of voxels of intracranial aneurysms and the union of the predicted results.

[0104] The present invention proposes a robust intracranial aneurysm segmentation method for unlabeled CTA scans, which includes the following steps: a triple learning framework is designed to learn a robust intracranial aneurysm segmentation model from brain CTA scans with unreliable labels. The framework mainly consists of three learning paradigms: pseudo-supervised learning, contrastive learning, and confidence learning. The three learning methods work together to iteratively extract semantic information of intracranial aneurysms from unreliable labeled brain CTA scans. First, the brain CTA scans are preprocessed to remove redundant information and various enhanced views of the initial input are generated. Secondly, the preprocessed CTA images are fed into the pseudo-supervised learning module, and an enhanced average teacher model is used to perform pseudo-supervised learning on the CTA scans with unreliable labels. In addition, a conversion network is attached to the enhanced teacher network for voxel-level contrastive learning. The contrastive learning module further improves the supervision of the feature space by reducing the differences between voxels of the same type and increasing the dissimilarity between voxels of different classes. At the same time, the multi-scale confidence learning module iteratively updates the labels of unreliable labeled data, thereby minimizing the impact of noisy labels on training. The method and diagnostic system of the present invention improve the accuracy of intracranial aneurysm segmentation, can effectively reduce the time of doctors' diagnosis, improve efficiency, and at the same time reduce missed diagnosis or misdiagnosis caused by different diagnostic levels or fatigue of doctors.

[0105] The above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A robust intracranial aneurysm segmentation method for unlabeled CTA scans, characterized in that: The steps include: Step S1: collecting low-precision labeled brain CTA scans to be segmented and high-precision labeled brain CTA scans containing intracranial aneurysms; Step S2: constructing an intracranial aneurysm segmentation network model based on three learning paradigms and extracting intracranial aneurysm features; Step S3: training the intracranial aneurysm segmentation network model according to the intracranial aneurysm data to obtain the trained model weights; Step S4: applying the trained model weights and performing model testing based on the intracranial aneurysm data to generate intracranial aneurysm segmentation results.

2. The robust intracranial aneurysm segmentation method for unlabeled CTA scanning according to claim 1, characterized in that: The step S1 of collecting the low-precision labeled brain CTA scan to be segmented and the high-precision labeled brain CTA scan containing the intracranial aneurysm specifically includes: The collected CTA scans are converted into different formats and the data is annotated; The preprocessed intracranial aneurysm data is used as the training and testing data set for the segmentation network.

3. The robust intracranial aneurysm segmentation method for unlabeled CTA scanning according to claim 1, characterized in that: The intracranial aneurysm segmentation network model described in step S2 includes a preprocessing module, a pseudo-supervised learning module, a contrastive learning module, and a multi-scale confidence learning module; The pre-processing module comprises: Applying intensity normalization, bilinear interpolation, and random gridding to the original view distorts the original input data to generate an enhanced view; Crop the data into three-dimensional image blocks according to the intracranial aneurysm area; The pseudo-supervised learning module includes: The preprocessed intracranial aneurysm data is taken as input; The enhanced mean teacher model is used to process intracranial aneurysm data, extract the features of intracranial aneurysms, and use the soft voting mechanism to reduce the interference of noise; A voxel selection strategy is adopted to integrate the uncertainty region into the consistency loss function; The contrastive learning module includes: In the decoding block of the enhanced mean teacher network, a conversion network is used to extract high-level semantic features; Construct positive and negative training pairs using the same position in different views in the high-level semantic feature space; The multi-scale confidence learning module includes: The predicted pseudo labels in the pseudo-supervised learning module and the low-precision intracranial aneurysm CTA labels are used as input; Use multi-scale confidence strategies to effectively perceive a certain range of structures rather than just the information of a single voxel; By calculating the confidence and threshold, the corrected low-precision labels are iteratively updated to provide additional training information for training.

4. The robust intracranial aneurysm segmentation method for unlabeled CTA scanning according to claim 3, characterized in that: The specific process of voxel selection strategy includes: Receive the predicted pseudo labels of the student model and the teacher model as input; By calculating the average threshold and classification probability of each class, the potential voxel entities within the uncertainty region can be mapped; Since intracranial aneurysms are relatively small in CTA scans, latent voxel entities are integrated into the model training process to focus more on these informative but unstable regions.

5. The robust intracranial aneurysm segmentation method for unlabeled CTA scanning according to claim 3, characterized in that: The specific design of the conversion network includes: It consists of two 3D convolution blocks and a PRelu activation function; The high-dimensional hidden representation of the decoding part of the augmented teacher model is projected into a low-dimensional space.

6. The robust intracranial aneurysm segmentation method for unlabeled CTA scanning according to claim 3, characterized in that: The specific design of the multi-scale confidence strategy includes: The predicted pseudo labels in the pseudo-supervised learning module and the low-precision intracranial aneurysm CTA labels are used as input; Use 1×1, 2×2 and 3×3 multi-scale sliding windows to learn pseudo labels, thereby obtaining a joint confidence matrix of three scales; After adding the three scales separately, we get the multi-scale joint confidence matrix; Correcting the multi-scale joint confidence matrix to obtain a correction matrix; Further normalization is performed to generate a multi-scale label matrix; The uncertainty of each voxel is estimated using the predicted entropy to obtain a potential target map.

7. The robust intracranial aneurysm segmentation method for unlabeled CTA scanning according to claim 4, 5 or 6, characterized in that: The training of the intracranial aneurysm segmentation network model includes: Use pseudo-supervised learning modules to enhance the model’s training effect on small target intracranial aneurysms; Improve feature expression capabilities and capture feature space through contrastive learning modules; Utilizes a multi-scale confidence learning module to focus on broader local structures rather than individual voxels, thus improving correction efficiency; By effectively combining the three learning paradigms, the model can be robustly trained on a mixture of low-precision labeled brain CTA scans and high-precision labeled brain CTA scans containing intracranial aneurysms.

8. The robust intracranial aneurysm segmentation method for unlabeled CTA scanning according to claim 1, characterized in that: The generated intracranial aneurysm segmentation results include: Use the trained model weights to perform intracranial aneurysm data segmentation test; Output and verify the accuracy of segmentation results and model performance; The accuracy of the entire segmentation was calculated by the ratio of correctly predicted voxels to the total voxels, and the segmentation intersection-union ratio was calculated by comparing the ratio of the number of correctly predicted voxels to the actual number of voxels of intracranial aneurysms and the union of the predicted results.

Images