Intracranial aneurysm detection method and device based on bottom double-branch network and confidence coefficient calibration

Abstract

The invention discloses an intracranial aneurysm detection method and device based on a bottom double-branch network and confidence coefficient calibration, and the method comprises the steps: carrying out the low-dimensional feature extraction of an image in the bottom double-branch network, and obtaining a feature map, dividing the feature map into a first sub-feature map and a second sub-feature map, and performing feature extraction and fusion through different convolution operation branches to obtain a final feature map; performing prediction and matching of a real bounding box through an optimal transmission algorithm, and training a bottom double-branch network according to a matching result; in the bottom double-branch network training process, confidence coefficient calibration is carried out at the output end of the model by constructing a loss function based on confidence coefficient scores, and finally the trained bottom double-branch network is obtained to serve as an intracranial aneurysm detection model; and inputting the new to-be-detected picture into the intracranial aneurysm detection model, and detecting to obtain a high-confidence-coefficient intracranial aneurysm prediction bounding box so as to accurately position the intracranial aneurysm. According to the invention, the accuracy and reliability of intracranial aneurysm detection can be obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and particularly relates to a method and device for intracranial aneurysm detection based on a bottom dual-branch network and confidence calibration. Background Art

[0002] Intracranial aneurysms (IAs) are abnormal bulges in the cerebrovascular wall, and their rupture may lead to fatal subarachnoid hemorrhage. Currently, the clinical diagnosis of IAs mainly relies on the manual interpretation of magnetic resonance angiography (MRA), which is not only time-consuming and laborious but also highly subjective. Therefore, automatic detection methods based on deep learning have gradually become a research hotspot. However, the existing technologies still face significant challenges in detecting small-sized aneurysms, reducing the false positive rate (FPR), and improving the localization accuracy.

[0003] Traditional automatic recognition research on IAs mostly focuses on image segmentation techniques. For example, Bizjak et al. proposed a method for segmenting MRA and CTA images based on deep geometric learning, which achieved precise division of the aneurysm contour through three-dimensional geometric feature extraction. Sichtermann et al. adopted the DeepMedic convolutional neural network model and achieved sensitivity comparable to that of professional doctors in 3D TOF-MRA data. The VA-Unet model developed by You et al. optimized the segmentation performance of aneurysms of different sizes through residual connections. However, these segmentation methods have the following problems: (1) Insufficient sensitivity to small targets: The high voxel resolution of MRA images and the small size of aneurysms (usually only occupying a very small area of the image) result in extremely imbalanced data distribution. Although the segmentation model can capture local texture features, its sensitivity to tiny aneurysms significantly decreases. Especially in cases with complex morphology or blurred boundaries, false negatives (FN) or breaks are likely to occur in the segmentation results; (2) High computational redundancy and false positive rate: The segmentation task requires pixel-by-pixel classification, and most regions in high-resolution MRA images are normal blood vessels or backgrounds, which causes the model to be easily interfered by noise during redundant calculations. For example, the method of combining 3D U-Net and residual connections proposed by Zhu et al., although performing well on large datasets, its FPs / case (the number of false positives per case) is still as high as 1.4783 on the ADAM dataset, far exceeding the actual clinical requirements; (3) Dependence on high-quality labeled data: The training of the segmentation model requires a large amount of accurately labeled vascular and aneurysm contour data, and the labeling cost of medical images is high and is easily affected by the subjectivity of experts. For example, the manual feature extraction method based on multi-range filters and local variance proposed by Chung et al. reduces data dependence, but its feature expression ability is limited and it is difficult to adapt to complex cases.

[0004] To overcome the limitations of the above segmentation methods, some studies have turned to object detection-based frameworks. For example, some studies have proposed using the 3D-YOLO architecture to achieve efficient inference through single-stage detection. However, its application in three-dimensional medical images faces the following challenges: (1) Weak multi-scale feature fusion ability: 3D-YOLO lacks the support of a feature pyramid network and is difficult to effectively fuse semantic information at different levels. Its AP (IoU-0.5) on the ADAM dataset is only 0.3011. Moreover, due to the lack of an encoder-decoder structure, the model has insufficient localization accuracy for small-sized objects, resulting in a sharp drop in performance at high IoU thresholds; (2) Inefficient post-processing algorithm: Traditional non-maximum suppression (NMS) selects the detection box with the highest confidence through a greedy strategy but cannot globally optimize the matching relationship between the detection box and the ground truth box. For example, although nnDetection improves sensitivity through multi-scale detection heads, its FPs / case at IoU-0.1 on the ADAM dataset is still 0.2142, indicating that redundant detection boxes cannot be effectively removed; (3) Insufficient confidence reliability: There is a deviation between the confidence output by existing detection models and the true prediction accuracy, and the problem of high-confidence false positives (IHC) is prominent. Taking VoxelNet as an example, although it performs excellently in the field of autonomous driving, when directly applied to medical images, the lack of a confidence calibration mechanism increases the risk of clinical misdiagnosis.

[0005] In addition to the above methods, some studies have tried to combine multi-modal data or new attention mechanisms to improve performance, but the core problem of accurately locating the polarity of intracranial aneurysms has not been fundamentally solved, manifested as: (1) Multi-modal fusion: Chengjie et al. proposed a multi-modal convolutional neural network based on MRI and CTA to enhance detection robustness through complementary information. However, the registration of multi-modal data is difficult, and the computational complexity of the model surges, making it difficult to be deployed in clinical real-time scenarios; (2) Attention mechanism: Nan et al. introduced a boundary attention module in U-Net to focus on feature extraction in the vascular bifurcation area. However, the attention mechanism is sensitive to noise and is prone to false activation in low signal-to-noise ratio MRA images, which instead exacerbates the false positive problem; (3) Point cloud and voxel fusion: Hui et al. proposed the Shape-Adaptive Set Abstraction Network (SASAN), which adapts to lesions of different sizes through dynamic spherical radius adjustment. However, it relies on the three-dimensional representation of point cloud data, while medical images are mostly regular voxel grids, and serious information loss occurs during the conversion process, limiting its application potential in MRA.

[0006] In summary, the existing technologies have common defects in the following aspects: (1) The contradiction between small target detection and feature diversity: In high-resolution MRA images, the tiny size of aneurysms and complex backgrounds make it difficult for the model to balance the receptive field and detail retention; (2) The detection box matching strategy is sub-optimal: Traditional NMS only relies on local confidence ranking and lacks a global optimal matching perspective; (3) The confidence level is disconnected from the true accuracy: The confidence level output by the model fails to accurately reflect the prediction reliability, and high-confidence false positives seriously interfere with clinical decisions. In the future, improving these common defects will be the key to promoting the development of intracranial aneurysm automatic detection technology. Summary of the Invention

[0007] In view of the above, the object of the present invention is to provide a method and device for intracranial aneurysm detection based on a bottom dual-branch network and confidence calibration, which enhances multi-scale feature diversity through the bottom dual-branch network, introduces the optimal transport theory to optimize the detection box matching, and further combines the confidence calibration mechanism to significantly improve the accuracy and reliability of intracranial aneurysm detection. The present invention particularly focuses on enhancing the feature extraction ability for tiny objects, and by optimizing the bounding box matching and adjusting the output confidence score, effectively reduces false positive results, can significantly reduce the number of false positives per patient without sacrificing sensitivity, thereby improving the overall detection performance, and is expected to provide a more reliable automatic detection tool for clinicians, thus promoting early diagnosis and improving the prognosis of patients, which not only helps to improve medical efficiency, but also significantly reduces the additional risks and costs caused by misdiagnosis.

[0008] To achieve the above object of the invention, the technical solutions provided by the present invention are as follows:

[0009] In a first aspect, an intracranial aneurysm detection method based on a bottom dual-branch network and confidence calibration provided by an embodiment of the present invention includes the following steps:

[0010] Input intracranial aneurysm magnetic resonance angiography images into the bottom dual-branch network to extract low-dimensional features to obtain a feature map, and divide the feature map into a first sub-feature map and a second sub-feature map, and then perform feature extraction and fusion through different convolutional operation branches to obtain a final feature map;

[0011] Based on the predicted bounding boxes of intracranial aneurysms obtained from the final feature map, match the predicted bounding boxes and the true bounding boxes through the optimal transport algorithm to generate a final matching result, and train the bottom dual-branch network according to the matching result;

[0012] During the training process of the bottom dual-branch network, perform confidence calibration at the model output end by constructing a loss function based on the confidence score, and finally obtain the trained bottom dual-branch network as an intracranial aneurysm detection model;

[0013] Input the new image to be detected into the intracranial aneurysm detection model to detect the intracranial aneurysm prediction bounding box with high confidence to accurately locate the intracranial aneurysm.

[0014] Preferably, after dividing the feature map into a first sub-feature map and a second sub-feature map, feature extraction and fusion are performed through different convolutional operation branches to obtain the final feature map, including:

[0015] Divide the feature map by channels to obtain a first sub-feature map and a second sub-feature map. The first convolutional operation branch in the bottom double-branch network is used to copy and retain the original features of the first sub-feature map during each feature extraction. The second convolutional operation branch in the bottom double-branch network is used to perform multiple high-dimensional feature extractions on the second sub-feature map. During each high-dimensional feature extraction, the original features of the first sub-feature map are fused with the features of the previous and all previous high-dimensional feature extractions, and the finally obtained high-dimensional features are fused with the original features of the first sub-feature map to generate the final feature map.

[0016] Preferably, the optimal transport algorithm is used to match the predicted bounding box and the ground truth bounding box to generate the final matching result, including:

[0017] Calculate the transport cost matrix based on the predicted bounding box and the ground truth bounding box, and obtain the optimal transport cost matrix by minimizing the total transport cost, thereby generating the final matching result of the predicted bounding box and the ground truth bounding box, expressed as:

[0018]

[0019] Among them, represents the transport cost matrix, P i,j represents the matching degree between the i-th predicted bounding box d i and the j-th ground truth bounding box g j , N p represents the total number of predicted bounding boxes, N g represents the total number of ground truth bounding boxes, represents the cost function, M represents the initial matrix, u(α,β) represents the marginal condition satisfied by M, α i represents the probability mass assigned to the i-th predicted bounding box, β j represents the probability mass assigned to the j-th ground truth bounding box.

[0020] Preferably, a regularization term is introduced into the transport cost matrix to reduce the overfitting risk and make the parameters smoother during the training process of the bottom double-branch network, expressed as:

[0021]

[0022] Among them, γ represents the weight parameter, and H(M) represents the regularization term.

[0023] Preferably, the confidence calibration is performed by constructing a loss function based on the confidence score at the model output end, including:

[0024] Introduce the confidence score as a regularization term to construct the loss function L cc , expressed as:

[0025]

[0026] where CR represents the objective function based on the confidence score, and t AHC represents the confidence score of high-confidence accuracy, and t ILC represents the confidence score of low-confidence inaccuracy, and t ALC represents the confidence score of low-confidence accuracy, and t IHC represents the confidence score of high-confidence inaccuracy.

[0027] Preferably, in the post-processing stage of the bottom dual-branch network, the prediction results are sorted according to the calibrated confidence score, and the prediction bounding box with the highest confidence is selected as the final output.

[0028] In a second aspect, to achieve the above-mentioned invention purpose, an intracranial aneurysm detection device based on a bottom dual-branch network and confidence calibration is further provided in an embodiment of the present invention, which is implemented by using the above-mentioned intracranial aneurysm detection method based on a bottom dual-branch network and confidence calibration, including: a dual-branch feature extraction module, an optimal transport matching module, a confidence calibration module, and a model detection module;

[0029] The dual-branch feature extraction module is used to input intracranial aneurysm magnetic resonance angiography pictures into the bottom dual-branch network, perform low-dimensional feature extraction to obtain a feature map, and divide the feature map into a first sub-feature map and a second sub-feature map, and then perform feature extraction and fusion through different convolutional operation branches to obtain a final feature map;

[0030] The optimal transport matching module is used to match the predicted bounding box of the intracranial aneurysm obtained based on the final feature map with the ground truth bounding box through the optimal transport algorithm, generate a final matching result, and train the bottom dual-branch network according to the matching result;

[0031] The confidence calibration module is used to perform confidence calibration by constructing a loss function based on the confidence score at the model output end during the training process of the bottom dual-branch network, and finally obtain the trained bottom dual-branch network as an intracranial aneurysm detection model;

[0032] The model detection module is used to input new pictures to be detected into the intracranial aneurysm detection model to detect high-confidence predicted bounding boxes of intracranial aneurysms to accurately locate intracranial aneurysms.

[0033] In a third aspect, to achieve the above object of the invention, an embodiment of the present invention further provides an electronic device, including a memory and one or more processors. The memory is used to store a computer program, and the processor is used to implement the above-mentioned intracranial aneurysm detection method based on the bottom dual-branch network and confidence calibration when executing the computer program.

[0034] In a fourth aspect, to achieve the above object of the invention, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a computer, the above-mentioned intracranial aneurysm detection method based on the bottom dual-branch network and confidence calibration is implemented.

[0035] In a fifth aspect, to achieve the above object of the invention, an embodiment of the present invention further provides a computer product, which includes a computer program. When the computer program is executed by a processor, the above-mentioned intracranial aneurysm detection method based on the bottom dual-branch network and confidence calibration is implemented.

[0036] Compared with the prior art, the beneficial effects of the present invention at least include:

[0037] (1) The present invention extracts features through the bottom dual-branch network, fully captures the detailed features of small target areas, can improve the richness and accuracy of feature representation, achieve higher localization accuracy under a high IoU threshold, and combines the optimal transport theory to match the predicted bounding box and the ground truth bounding box, further optimizing the box matching process in object detection, effectively improving the detection accuracy and efficiency.

[0038] (2) On the basis of the bottom dual-branch feature extraction and optimal transport, the present invention further performs confidence calibration at the output end to improve the reliability and accuracy of the prediction results, constructs an additional regularization term to ensure the stability of high-confidence predictions, reduces the false positive rate while maintaining high sensitivity, can enhance the generalization ability of the model, reduce the risk of overfitting, and further optimizes the detection results by combining effective post-processing strategies, improving the performance and robustness of the overall system. Description of the Drawings

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0040] Figure 1It is a schematic flowchart of the intracranial aneurysm detection method based on the bottom double-branch network and confidence calibration provided by an embodiment of the present invention;

[0041] Figure 2 It is a schematic framework diagram of the intracranial aneurysm detection method based on the bottom double-branch network and confidence calibration provided by an embodiment of the present invention;

[0042] Figure 3 It is a schematic diagram of feature extraction by the bottom double-branch network provided by an embodiment of the present invention;

[0043] Figure 4 It is a schematic structural diagram of the intracranial aneurysm detection device based on the bottom double-branch network and confidence calibration provided by an embodiment of the present invention. Detailed implementation manners

[0044] To make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation manners described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0045] The inventive concept of the present invention is as follows: Aiming at the problems that the existing intracranial aneurysm detection methods generally have a high false positive rate (FPs) and low localization accuracy when dealing with small and complex lesions, especially the poor performance at high IoU thresholds, which affects the diagnostic accuracy, and may also lead to unnecessary further examinations or treatments, increasing the medical cost and patient burden. Embodiments of the present invention provide an intracranial aneurysm detection method and device based on a bottom double-branch network and confidence calibration, and design an innovative 3D voxel detection framework: Bottom Double Branch Path with Confidence Calibration (BCOC). Specifically, feature extraction is enhanced through the bottom double-branch path network (BDBP), the bounding box matching is optimized through the Optimal Transport (OT) theory, and the output confidence score is adjusted through Confidence Calibration (CC), thereby significantly improving the accuracy and reliability of intracranial aneurysm detection.

[0046] Figure 1 It is a schematic flowchart of the intracranial aneurysm detection method based on the bottom double-branch network and confidence calibration provided by an embodiment of the present invention. As Figure 1 shown, the embodiment provides an intracranial aneurysm detection method based on a bottom double-branch network and confidence calibration, including the following steps:

[0047] S1. Input the magnetic resonance angiography (MRA) images of intracranial aneurysms into the bottom dual-branch network to extract low-dimensional features and obtain a feature map. After dividing the feature map into a first sub-feature map and a second sub-feature map, perform feature extraction and fusion through different convolutional operation branches to obtain the final feature map.

[0048] In the embodiment, a bottom dual-branch path network (BDBP) based on the Feature Pyramid Network (FPN) architecture, combined with Optimal Transport (OT) and Confidence Calibration (CC), is provided for the detection of intracranial aneurysms (IAs) in 3D MRA images. As Figure 2 shown, the model mainly presents a multi-scale feature pyramid architecture, including a bottom path, a pyramid path, and a detection head. A BDBP connection is embedded in the bottom path, dividing the entire network into two branches: one branch transmits features, and the other is responsible for feature enhancement. At the same time, OT optimization is used in the post-processing stage to solve the matching problem between the predicted bounding boxes and the ground truth bounding boxes, enabling them to be better aligned. Finally, the output probability distribution is adjusted through confidence calibration.

[0049] The design of the bottom dual-branch network (BDBP) takes into account the small size characteristics of IAs. To enhance feature diversity and improve gradient propagation, and to enhance the diversity of feature maps at different levels, the feature map is divided into two independent branches, and convolutional operations are performed separately to generate diverse feature representations, while reducing computational and memory loads. This method introduces additional branches during the feature extraction process to increase gradient paths and minimizes the learning of redundant gradient information by splitting the gradient flow.

[0050] Specifically, BDBP includes the following parts:

[0051] (1) Feature map segmentation: As Figure 3 shown, in the bottom dual-branch network, the feature map X0 is divided by channels to obtain a first sub-feature map X0 ′ and a second sub-feature map X0″.

[0052] X0 ′ = X0[:, :, :, :, channels / / 2:]

[0053] X0″ = X0[:, :, :, :, :channels / / 2]

[0054] where, : represents other channel features, channels / / 2: represents the channel features from the end to the head separated by a certain channel, and :channels / / 2 represents the channel features from the head to the end separated by a certain channel.

[0055] (2) Convolutional operation: Apply different convolutional kernels to each sub-feature map, and generate a new feature representation X through k times of dense convolutionlocal_dense .

[0056] X local_dense = Concatenate([X0″, Convolution3D(X0″), …, Convolution3D k (X0″)])

[0057] Among them, Concatenate(·) represents the concatenation operation, and Convolution3D(·) represents the three-dimensional convolution operation. Specifically, as Figure 3 shown, in the bottom double-branch network, the first convolution operation branch is used to copy and retain the original features of the first sub-feature map X0 during each feature extraction ′ . The second convolution operation branch in the bottom double-branch network is used to perform multiple high-dimensional feature extractions on the second sub-feature map X0″ through a local dense block (including k dense layers). During each high-dimensional feature extraction, the original features of the first sub-feature map are concatenated and fused with the features of the previous and all previous high-dimensional feature extractions to obtain X local_dense .

[0058] (3) Feature fusion: The finally obtained high-dimensional feature X local_dense is fused with the original features X0 of the first sub-feature map ′ through a partial transition layer to generate the final feature map X out .

[0059] X out = Transition(Concatenate([X0 ′ , X local_dense ))

[0060] Among them, Concatenate(·) represents the concatenation operation, and Transition(·) includes the conversion operations of dimensionality reduction and pooling.

[0061] S2. The predicted bounding box of the intracranial aneurysm obtained based on the final feature map is matched with the ground truth bounding box through the optimal transport algorithm to generate the final matching result, and the bottom double-branch network is trained according to the matching result.

[0062] In the embodiment, for the predicted bounding box of the intracranial aneurysm recognized based on the final feature map, the optimal transport (OT) algorithm is introduced to optimize the matching between the predicted bounding box and the ground truth bounding box, thereby improving the localization accuracy. Given two distributions α and β, where α is associated with the predicted object d i (i = 1, 2, …, N p ), and β corresponds to the ground truth object g i (i = 1, 2, …, N g),OT considers the prediction object d i and the real object g i The one-to-one matching cost function between Its goal is to find the optimal matching matrix that minimizes, specifically including the following steps.

[0063] (1) Transmission cost matrix calculation: Calculate the transmission cost matrix according to the predicted bounding box and the real bounding box (the label in the ADAM dataset), expressed as:

[0064]

[0065] Among them, represents the transmission cost matrix, P i,j represents the i-th predicted bounding box d i and the matching degree between the j-th real bounding box g j N p represents the total number of predicted bounding boxes, N g represents the total number of real bounding boxes, represents the cost function, M represents the initial matrix, u(α,β) represents the marginal condition satisfied by M, α i represents the probability mass assigned to the i-th predicted bounding box, β j represents the probability mass assigned to the j-th real bounding box.

[0066] (2) Solving the optimal transport problem: Introduce a regularization term in the transmission cost matrix to reduce the overfitting risk and make the parameters smoother during the training process of the bottom double-branch network, expressed as:

[0067]

[0068] Among them, γ represents the weight parameter, and H(M) represents the regularization term.

[0069] (3) Bounding box matching: Obtain the optimal transmission cost matrix by minimizing the total transmission cost, match the predicted bounding box with the real bounding box, and thus generate the final matching result.

[0070] S3. During the training process of the bottom double-branch network, confidence calibration is performed by constructing a loss function based on the confidence score at the model output end, and finally the trained bottom double-branch network is obtained as the intracranial aneurysm detection model.

[0071] In the embodiments, to improve the accuracy of IAs detection, a confidence calibration (CC) strategy is integrated into the model. By adjusting the confidence scores to better reflect the actual prediction accuracy, the correct detections have higher scores and the incorrect ones have lower scores, thereby reducing the false positive rate. The CC process first divides the confidence and precision spaces into four categories: accurately high confidence (AHC), accurately low confidence (ALC), inaccurately high confidence (IHC), and inaccurately low confidence (ILC), and defines an objective function CR to maximize the confidence scores of accurate detections while minimizing the scores of inaccurate detections. The constructed loss function L cc is expressed as:

[0072]

[0073] where CR represents the objective function based on the confidence scores, and t AHC represents the confidence score of accurately high confidence, and t ILC represents the confidence score of inaccurately low confidence, and t ALC represents the confidence score of accurately low confidence, and t IHC represents the confidence score of inaccurately high confidence.

[0074] In the post - processing stage of the bottom double - branch network, the prediction results are sorted according to the calibrated confidence scores, and the prediction bounding box with the highest confidence is selected as the final output. The bottom double - branch network is trained through a model training strategy that combines OT and CC to further improve the prediction accuracy, and finally the trained bottom double - branch network is obtained as the intracranial aneurysm detection model (BCOC model).

[0075] S4. Input the new image to be detected into the intracranial aneurysm detection model to obtain a high - confidence intracranial aneurysm prediction bounding box for accurately locating the intracranial aneurysm.

[0076] In the embodiments, for the new MRA image of the intracranial aneurysm to be detected, it is detected through the trained BCOC model in the embodiments, and a high - confidence intracranial aneurysm prediction bounding box is predicted to accurately locate the intracranial aneurysm, realizing high - precision detection of the intracranial aneurysm.

[0077] In summary, a method for detecting intracranial aneurysms based on a bottom double - branch network and confidence calibration provided by the embodiments of the present invention significantly improves the accuracy and reliability of intracranial aneurysm detection, and experimentally verifies the overall performance of the model. The specific effects are as follows:

[0078] First, in terms of technical performance, the experimental results of the BCOC model on two public datasets, ADAM and LAU, show that the model outperforms existing methods at different IoU thresholds. For example, at the IoU-0.1 threshold, the average precision (AP) of the BCOC model on the ADAM dataset reaches 0.8186, the sensitivity is 93.91%, and the number of false positives per case (FPs / case) is only 0.1332. At the more stringent IoU-0.5 threshold, the AP can still reach 0.3799, the sensitivity is 53.33%, and the FPs / case is 0.1423. In contrast, the performance of other advanced models such as nnDetection and 3D-YOLO on the same dataset is as follows: for nnDetection, the AP is 0.7903 at IoU-0.1, the sensitivity is 89.50%, and the FPs / case is 0.2142; at IoU-0.5, the AP is 0.3416, the sensitivity is 48.41%, and the FPs / case is 0.2415. This shows that the BCOC model not only maintains a low false positive rate at high sensitivity but also performs well under strict localization requirements.

[0079] Second, in terms of social and economic benefits, the high accuracy of the BCOC model can significantly reduce the risks of misdiagnosis and missed diagnosis, and lower the medical costs brought by unnecessary further examinations or treatments. By detecting intracranial aneurysms early, timely intervention measures can be taken, thus avoiding potential life-threatening situations and improving the quality of life of patients. In addition, reducing false positive results can also relieve the psychological burden of patients and their families and optimize the allocation of medical resources, enabling more resources to be used for cases that truly need them.

[0080] Finally, from the perspective of technological development, the design concept and technical means of the BCOC model provide new ideas and tools for the field of medical image analysis. Its innovative feature extraction, bounding box matching, and confidence calibration strategies are not only applicable to the detection of intracranial aneurysms but can also be extended to other recognition tasks of small and complex medical objects. Therefore, the present invention not only solves the main problems in current intracranial aneurysm detection but also lays a solid foundation for future technological progress. These scientific analyses and experimental results fully prove the effectiveness and superiority of the BCOC model.

[0081] Based on the same inventive concept, as Figure 4 shown, the embodiment of the present invention also provides an intracranial aneurysm detection device 400 based on a bottom dual-branch network and confidence calibration, including: a dual-branch feature extraction module 410, an optimal transport matching module 420, a confidence calibration module 430, and a model detection module 440.

[0082] The dual-branch feature extraction module 410 is used to input the intracranial aneurysm magnetic resonance angiography images into the bottom dual-branch network, perform low-dimensional feature extraction to obtain a feature map, and divide the feature map into a first sub-feature map and a second sub-feature map, and then perform feature extraction and fusion through different convolutional operation branches to obtain a final feature map.

[0083] The optimal transport matching module 420 is used to obtain the predicted bounding box of the intracranial aneurysm based on the final feature map, match the predicted bounding box and the ground truth bounding box through the optimal transport algorithm, generate the final matching result, and train the bottom dual-branch network according to the matching result.

[0084] The confidence calibration module 430 is used to perform confidence calibration by constructing a loss function based on the confidence score at the output end of the model during the training process of the bottom dual-branch network, and finally obtain the trained bottom dual-branch network as the intracranial aneurysm detection model.

[0085] The model detection module 440 is used to input the new image to be detected into the intracranial aneurysm detection model to detect a predicted bounding box of the intracranial aneurysm with high confidence to accurately locate the intracranial aneurysm.

[0086] Based on the same inventive concept, an embodiment of the present invention further provides an electronic device, including a memory and one or more processors. The memory is used to store a computer program, and the processor is used to implement the above-mentioned intracranial aneurysm detection method based on the bottom dual-branch network and confidence calibration when executing the computer program.

[0087] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a computer, the above-mentioned intracranial aneurysm detection method based on the bottom dual-branch network and confidence calibration is implemented.

[0088] Based on the same inventive concept, an embodiment of the present invention further provides a computer product, which includes a computer program. When the computer program is executed by a processor, the above-mentioned intracranial aneurysm detection method based on the bottom dual-branch network and confidence calibration is implemented.

[0089] It should be noted that the above-mentioned intracranial aneurysm detection device, electronic device, computer-readable storage medium, and computer product based on the bottom dual-branch network and confidence calibration all belong to the same inventive concept as the intracranial aneurysm detection method based on the bottom dual-branch network and confidence calibration. The specific implementation process can be seen in the embodiments of the intracranial aneurysm detection method based on the bottom dual-branch network and confidence calibration, which will not be elaborated here.

[0090] The specific embodiments described above have elaborated in detail on the technical solutions and beneficial effects of the present invention. It should be understood that the above description is only the most preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, supplements, equivalent replacements, etc. made within the principle scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting intracranial aneurysms based on a bottom double-branch network and confidence calibration, characterized in that: The following steps are involved: The intracranial aneurysm magnetic resonance angiography image is input into the bottom double-branch network to extract low-dimensional features to obtain a feature map, and the feature map is divided into a first sub-feature map and a second sub-feature map, and then feature extraction and fusion are performed through different convolution operation branches to obtain the final feature map; Based on the intracranial aneurysm prediction bounding box obtained from the final feature map, the predicted bounding box and the real bounding box are matched through the optimal transmission algorithm to generate the final matching result, and the bottom double-branch network is trained according to the matching result; During the bottom double-branch network training process, confidence calibration is performed at the model output by constructing a loss function based on the confidence score, and finally the trained bottom double-branch network is obtained as the intracranial aneurysm detection model; The new image to be detected is input into the intracranial aneurysm detection model to obtain a high-confidence intracranial aneurysm prediction bounding box to accurately locate the intracranial aneurysm.

2. The intracranial aneurysm detection method based on bottom double-branch network and confidence calibration according to claim 1, characterized in that: The step of dividing the feature map into a first sub-feature map and a second sub-feature map and then extracting and fusing features through different convolution operation branches to obtain a final feature map includes: The feature map is divided according to the channel to obtain the first sub-feature map and the second sub-feature map. The first convolution operation branch in the bottom double-branch network is used to copy and retain the original features of the first sub-feature map during each feature extraction. The second convolution operation branch in the bottom double-branch network is used to perform multiple high-dimensional feature extractions on the second sub-feature map. Each time the high-dimensional feature is extracted, the original features of the first sub-feature map are fused with the features of the previous and all previous high-dimensional feature extractions, and the final high-dimensional features are fused with the original features of the first sub-feature map to generate the final feature map.

3. The intracranial aneurysm detection method based on bottom double-branch network and confidence calibration according to claim 1, characterized in that: The matching of the predicted bounding box and the real bounding box by the optimal transmission algorithm to generate the final matching result includes: The transmission cost matrix is ​​calculated based on the predicted bounding box and the true bounding box. The optimal transmission cost matrix is ​​obtained by minimizing the total transmission cost, thereby generating the final matching result of the predicted bounding box and the true bounding box, which is expressed as: in, represents the transmission cost matrix, P i,j represents the i-th predicted bounding box d i and the jth ground-truth bounding box g j The matching degree between p Represents the total number of predicted bounding boxes, N g represents the total number of ground-truth bounding boxes, represents the cost function, M represents the initial matrix, u(α,β) represents the marginal condition satisfied by M, α i represents the probability mass assigned to the i-th predicted bounding box, β j represents the probability mass assigned to the jth ground-truth bounding box.

4. The method for detecting intracranial aneurysms based on bottom double-branch network and confidence calibration according to claim 3, characterized in that: A regularization term is introduced into the transmission cost matrix to reduce the risk of overfitting and make the parameters smoother during the training of the bottom two-branch network, which is expressed as: Among them, γ represents the weight parameter and H(M) represents the regularization term.

5. The method for detecting intracranial aneurysms based on bottom double-branch network and confidence calibration according to claim 1, characterized in that: The confidence calibration is performed at the model output end by constructing a loss function based on the confidence score, including: The confidence score is introduced as a regularization term to construct the loss function L cc , expressed as: Where CR represents the objective function based on the confidence score, t AHC Indicates the confidence score with high confidence, t ILC The confidence score indicating low confidence is inaccurate, t ALC Indicates the confidence score of low confidence accuracy, t IHC Confidence scores indicating high confidence that the score is inaccurate.

6. The method for detecting intracranial aneurysms based on bottom double-branch network and confidence calibration according to claim 1, characterized in that: In the post-processing stage of the bottom two-branch network, the prediction results are ranked according to the calibrated confidence scores, and the predicted bounding box with the highest confidence is selected as the final output.

7. An intracranial aneurysm detection device based on a bottom double-branch network and confidence calibration, implemented by using the intracranial aneurysm detection method based on a bottom double-branch network and confidence calibration according to any one of claims 1 to 6, characterized in that: include: Dual-branch feature extraction module, optimal transmission matching module, confidence calibration module and model detection module; The dual-branch feature extraction module is used to input the intracranial aneurysm magnetic resonance angiography image into the bottom dual-branch network, perform low-dimensional feature extraction to obtain a feature map, and divide the feature map into a first sub-feature map and a second sub-feature map, and then perform feature extraction and fusion through different convolution operation branches to obtain a final feature map; The optimal transmission matching module is used to obtain the intracranial aneurysm prediction bounding box based on the final feature map, match the predicted bounding box with the real bounding box through the optimal transmission algorithm, generate the final matching result, and train the bottom double-branch network according to the matching result; The confidence calibration module is used to perform confidence calibration at the output end of the model by constructing a loss function based on the confidence score during the bottom double-branch network training process, and finally obtain the trained bottom double-branch network as an intracranial aneurysm detection model; The model detection module is used to input a new image to be detected into the intracranial aneurysm detection model to detect and obtain a high-confidence intracranial aneurysm prediction boundary box to accurately locate the intracranial aneurysm.

8. An electronic device comprising a memory and one or more processors, wherein the memory is used to store a computer program, characterized in that: The processor is used to implement the intracranial aneurysm detection method based on bottom double-branch network and confidence calibration as described in any one of claims 1-6 when executing the computer program.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a computer, the intracranial aneurysm detection method based on a bottom double-branch network and confidence calibration as described in any one of claims 1 to 6 is implemented.

10. A computer product comprising a computer program, characterized in that When the computer program is executed by a processor, the intracranial aneurysm detection method based on a bottom double-branch network and confidence calibration as described in any one of claims 1 to 6 is implemented.

Images