A semi-supervised medical image object detection method based on detr

Abstract

The application discloses a semi-supervised medical image target detection method based on DETR, first, through the deblurring processing of the image, the inaccurate blur kernel and the additive Gaussian noise are used to model the image degradation, and a variety of prior knowledge and optimization algorithms are combined to restore the clear image. Secondly, a pair of teacher and student models with the same network architecture are adopted, and the strong and weak enhanced versions of the unlabeled image are used for training, and the training of the student model is guided by the pseudo label of the teacher model. At the same time, the cost-based pseudo label mining module dynamically mines the reliable pseudo bounding box in the unlabeled data, which is used for cross-view query consistency calculation. Finally, the supervised loss, unsupervised loss and cross-view consistency loss are combined with the corresponding weights to guide the training of the model. The method is simple and efficient, and can adapt to complex and changeable medical image structures, and provides accurate target detection information for clinical diagnosis.

Description

Technical Field

[0001] This invention relates to a semi-supervised medical image target detection method based on DETR, belonging to the field of medical image processing. Background Technology

[0002] Medical image analysis plays a crucial role in clinical diagnosis and treatment. It involves interpreting medical images to identify and locate anatomical structures, pathological changes, or other regions of interest. Object detection is a core task in medical image analysis, requiring algorithms to accurately identify and locate structures of interest, such as tumors, organs, or other abnormalities, within images. This helps physicians make more accurate diagnoses, monitor diseases, and evaluate treatment effectiveness.

[0003] However, annotating medical images typically requires specialized knowledge and experienced radiologists, making the process time-consuming and costly. Furthermore, the diversity and complexity of medical images further increase the difficulty of annotation. Therefore, how to effectively utilize limited annotated data to train high-performance object detection models has become an important research topic.

[0004] Traditional object detection methods typically rely on large amounts of labeled data for training, which is difficult to achieve in the field of medical imaging. To overcome this challenge, semi-supervised learning, as a technique that utilizes unlabeled data to improve model performance, has attracted widespread attention. By learning useful information from unlabeled data and combining it with limited labeled data, semi-supervised learning can improve the model's generalization ability without requiring a large amount of labeling.

[0005] DETR (Detection Transformer) is an end-to-end object detection model based on Transformer. It transforms the object detection problem into an ensemble prediction problem, avoiding traditional anchor point generation and NMS post-processing steps. DETR has demonstrated strong performance in object detection of natural images, but when directly applied to object detection of medical images, it still faces challenges such as insufficient labeled data and weak model generalization ability.

[0006] Therefore, this patent proposes a semi-supervised medical image target detection method based on DETR. It aims to fully utilize unlabeled medical image data and improve the performance and generalization ability of the target detection model by employing innovative techniques such as deblurring, teacher-student models, stage-based hybrid matching, cross-view query consistency, and cost-based pseudo-label mining. This method not only reduces reliance on large amounts of labeled data but also improves the accuracy and efficiency of the model in practical clinical applications. Summary of the Invention

[0007] The purpose of this invention is to propose a semi-supervised medical image target detection method based on DETR. Through the design of a series of innovative technical modules and strategies, this method aims to address the problems of insufficient labeled data and weak model generalization ability in medical image target detection. The method includes key technologies such as deblurring, teacher-student model, stage-based hybrid matching, cross-view query consistency, and cost-based pseudo-label mining. Through these designs, an efficient and accurate semi-supervised medical image target detection method is achieved. This method can improve model performance by utilizing unlabeled data with limited labeled data, reducing reliance on large amounts of labeled data and improving the accuracy and efficiency of the model in practical clinical applications.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] A semi-supervised medical image target detection method based on DETR includes the following steps:

[0010] Step 1: Deblur

[0011] Image degradation is modeled using an inaccurate blur kernel and kernel error, along with additive Gaussian noise. The solution space is constrained by combining depth residual priors, total variation priors, and sparsity priors, modeling residuals, sharp images, and artifacts separately. The minimization problem is decomposed into two subproblems, solved using the gradient-based ADAM algorithm and proximal gradient descent. The input is a blurred image and parameters; the output is the restored image and residuals.

[0012] Step 2: Training Data

[0013] For the recovered image x, a labeled image set can be used during training. and unlabeled image set N s and N u These represent the number of labeled and unlabeled images, respectively. For the labeled image x... s , annotation y s Includes the coordinates and object categories of all bounding boxes.

[0014] Step 3: Teacher-Student Model

[0015] While following the popular teacher-student paradigm in SSOD, semi-supervised DETR employs a pair of teacher and student models with the same network architecture. Specifically, in each training iteration, weakly and strongly augmented unlabeled images are fed into the teacher and student models, respectively. The pseudo-labels generated by the teacher are then assigned confidence scores greater than τ. s This serves as supervision for training students. The student's parameters are updated via backpropagation, while the teacher model's parameters are the student's exponential moving average (EMA).

[0016] Step 4: Phased Mixed Matching

[0017] The DETR-based framework relies on a one-to-one mapping for end-to-end object detection. For DETR-based multi-task learning systems, the Hungarian algorithm can be used to achieve an optimal one-to-one match between student predictions and teacher-generated pseudo-labels.

[0018] However, in the early stages of SSOD training, teacher-generated pseudo-labels are often inaccurate and unreliable, which poses a high risk of sparse, low-quality suggestions under a one-to-one assignment strategy. To achieve efficient semi-supervised learning by leveraging multiple positive queries, a one-to-many assignment is proposed instead of a one-to-one assignment. Higher-order combinations are used to classify the score s and the IoU value u as matching cost metrics. In the one-to-one assignment, the M proposals with the largest m values ​​are selected as positive samples, while the remaining proposals are treated as negative samples.

[0019] In the early stages of semi-supervised training, the model is paired with a one-to-one assignment for T1 iterations. During this stage, the classification loss and regression loss are modified. In the second stage, the pairing is switched back to one-to-one during training.

[0020] Step 5: Cross-view query consistency

[0021] In non-DETR-based SSOD frameworks, consistency regularization can often be conveniently used to minimize the teacher model f. θ And student model f′ θ This paper proposes a cross-view query consistency module that enables the DETR-based architecture to learn object query semantically invariant features across different augmented views. For each unlabeled image, on the pseudo-boundary box set b, we apply several MLPs to the RoI features extracted via RoIAlign. Then, c t and c s The original object query, treated as a cross-view query embedding and appended to another view, serves as input to the decoder. Guided by the semantics of the input cross-view query embedding, the correspondence of decoded features can be naturally guaranteed.

[0022] Step 6: Cost-based pseudo-label mining

[0023] To uncover more pseudo-bounding boxes with meaningful semantic information for consistent learning across views, a cost-based pseudo-label mining module is proposed. This module can dynamically mine reliable pseudo-bounding boxes from unlabeled data. Specifically, an additional binary matching is performed between the initially filtered pseudo-bounding boxes and the predicted proposals, and the matching cost is used to describe the reliability of the pseudo-bounding boxes. Subsequently, in each training batch, the initial pseudo-bounding boxes are clustered into two states by fitting a Gaussian mixture model. Furthermore, a threshold is set for the cost of the reliable pseudo-bounding box cluster centers, and all pseudo-bounding boxes below the threshold are collected for cross-view query consistency calculation.

[0024] Step 7: Loss Function

[0025] The final loss L is shown below:

[0026]

[0027] in and These are supervised loss and unsupervised loss, including classification loss and regression loss. L c Indicates cross-view Figure 1 Caused sexual harm. u and w c These are the unsupervised loss weights and consistency loss weights. t is the current training iteration number, and T1 is the training time of the first phase of the SHM module.

[0028] Through the above steps, end-to-end semi-supervised medical image target detection is achieved.

[0029] This invention innovatively integrates multiple techniques, including deblurring, teacher-student models, staged hybrid matching, cross-view query consistency, and cost-based pseudo-label mining. First, images are deblurred using an inaccurate blur kernel and additive Gaussian noise to model image degradation, combining various prior knowledge and optimization algorithms to recover a clear image. Second, a pair of teacher and student models with identical network architectures are used for training with strongly and weakly enhanced versions of unlabeled images. The pseudo-labels from the teacher model guide the training of the student model. Furthermore, the method uses a one-to-many assignment strategy in the early stages of training to avoid the risk of low-quality suggestions, subsequently switching back to a one-to-one pairing strategy. A cross-view query consistency module enables the model to learn semantically invariant features of object queries across different enhanced views. Simultaneously, a cost-based pseudo-label mining module dynamically mines reliable pseudo-boundary boxes from unlabeled data for cross-view query consistency calculation. Finally, supervised loss, unsupervised loss, and cross-view... Figure 1Consistency loss and corresponding weights are used to guide model training. These features together constitute the core of the DETR-based semi-supervised medical image object detection method, enabling it to improve model performance by utilizing unlabeled data with limited labeled data.

[0030] The beneficial effects of this invention are as follows: This invention improves annotation efficiency, enhances model generalization ability, increases detection accuracy, reduces false positive rate, exhibits strong adaptability, provides real-time feedback and continuous learning, is easy to deploy and use, is cost-effective, promotes medical research, and improves the quality of medical care. By utilizing a large amount of unlabeled medical image data and reducing reliance on expensive and time-consuming manual annotation, the method of this invention significantly improves annotation efficiency. The design of the teacher-student model and cross-view query consistency module enables the model to learn invariant semantic features from different perspectives and enhancement strategies, enhancing the model's generalization ability to unknown data. The application of deblurring and cost-based pseudo-label mining techniques improves image quality and the reliability of pseudo-labels, thereby improving the accuracy of target detection. Dynamically mining reliable pseudo-boundary boxes, combined with cross-view query consistency, effectively reduces false positive detection results and improves detection reliability. Multi-scale and multi-angle training strategies enable the model to adapt to medical images of different scales and perspectives, improving the model's adaptability. The real-time feedback mechanism allows professional physicians to evaluate and correct the model's detection results, enabling the model to continuously learn and improve from practical applications. Model compression and acceleration techniques enable the deployment and use of models with limited computing resources, lowering the barrier to practical application. By reducing reliance on large amounts of labeled data and improving model training efficiency, the method of this invention offers significant cost-effectiveness advantages. By providing an efficient and accurate tool for medical image target detection, the method of this invention helps advance medical research, particularly in disease diagnosis, treatment monitoring, and biomedical research. In practical clinical applications, the method of this invention can assist doctors in making faster and more accurate diagnoses, thereby improving the quality of medical services and patient satisfaction. Attached Figure Description

[0031] Figure 1 This is a flowchart of the semi-supervised medical image target detection method based on DETR in this invention.

[0032] Figure 2 This is a schematic diagram of the deblurring process in this invention. Detailed Implementation

[0033] The invention will now be further described with reference to the accompanying drawings.

[0034] like Figure 1 , 2 As shown, a semi-supervised medical image target detection method based on DETR includes the following steps:

[0035] Step 1: Deblur

[0036] Image degradation is modeled using an inaccurate blur kernel and kernel error, along with additive Gaussian noise. The solution space is constrained by combining depth residual priors, total variation priors, and sparsity priors, modeling residuals, sharp images, and artifacts separately. The minimization problem is decomposed into two subproblems, solved using the gradient-based ADAM algorithm and proximal gradient descent. The input is a blurred image and parameters; the output is the restored image and residuals.

[0037] Step 1.1: For the deblurring problem, considering kernel error and potential artifacts, the degradation process is expressed as:

[0038]

[0039] Where y and x represent the blurred image and the sharp image, respectively. Δk represents the inaccurate fuzzy kernel and kernel error, and n represents additive Gaussian noise. It is a convolution operator. The residual is caused by kernel error, and h represents the artifact.

[0040] Step 1.2: Deriving x, h, and r from the blurred image y is a typical ill-posed problem. To constrain the solution space, prior information constraints need to be applied to x, h, and r. The DRP with residual term r containing sparse prior terms is used to fuse the total variational term with the depth image prior of the sharp image x, as well as the sparse prior term in the discrete cosine transform domain, to handle the artifact h.

[0041] To model the residual *r*, a dataset-free DRP guided by sparse priors in the spatial domain is proposed, using a customized untrained network U-Net representation to capture the residual *r*. A total variation prior is used to model the sharp image *x*. A sparse prior is applied to *h* in the DCT domain to model the distortion term *h*. The corresponding optimization problem can be represented as follows:

[0042]

[0043] Where, y∈R n1·n2 and x∈R n1·n2 These represent blurred and sharp images, respectively. Indicates an inaccurate fuzzy kernel; I θ (z x )∈R n1·n2 and It is made by an untrained neural network I θ (·)and Estimates of the sharp image x and residual r are generated respectively; θ and ζ collect the corresponding network parameters; z x~N(0,σI) and z r ~N(0,σI) is the random input to the neural network; ||·|| TV It is total variation regularization. Given an inaccurate kernel and a blurred image, unsupervised inference can be performed to obtain a sharp image, residuals, and artifacts.

[0044] Step 1.3: To solve this modeling problem, an alternating minimization algorithm is used. Since the variables are coupled together, the minimization problem is decomposed into two more manageable subproblems using the following alternating minimization scheme:

[0045]

[0046] Here, the superscript "i" indicates the iteration number, and L(θ,ζ,υ) is the objective function. Detailed solutions to these two subproblems are as follows: 1) Subproblem (θ,ζ): The (θ,ζ) subproblem is solved using the gradient-based ADAM algorithm. The gradients with respect to θ and ζ can be computed using the standard backpropagation algorithm; here, we jointly update the weights θ and ζ during iteration.

[0047] 2) Subproblem v: Assume So the V-shaped problem is... This can be solved precisely using the proximal gradient descent method.

[0048] Step 1.4: The entire process inputs a blurred image y, with an inaccurate kernel. Parameter λ s (s=1,2,3), with T iterations, the recovered image x and residual r were finally obtained.

[0049] Step 2: Training Data

[0050] For the recovered image x, a labeled image set can be used during training. and unlabeled image set N s and N u These represent the number of labeled and unlabeled images, respectively. For the labeled image x... s , annotation y s Includes the coordinates and object categories of all bounding boxes.

[0051] Step 3: Teacher-Student Model

[0052] While following the popular teacher-student paradigm in SSOD, semi-supervised DETR employs a pair of teacher and student models with the same network architecture. Specifically, in each training iteration, weakly and strongly augmented unlabeled images are fed into the teacher and student models, respectively. The pseudo-labels generated by the teacher are then assigned confidence scores greater than τ. sThis serves as supervision for training students. The student's parameters are updated via backpropagation, while the teacher model's parameters are the student's exponential moving average (EMA).

[0053] Step 4: Phased Mixed Matching

[0054] The DETR-based framework relies on a one-to-one mapping for end-to-end object detection. For DETR-based multi-task learning systems, the Hungarian algorithm can be used to achieve an optimal one-to-one match between student predictions and teacher-generated pseudo-labels:

[0055]

[0056] Where ξ N It is a set of permutations of N elements. It's a fake tag. The matching cost between the student's predicted outcome σ(i).

[0057] However, in the early stages of SSOD training, teacher-generated pseudo-labels are often inaccurate and unreliable, which poses a high risk of sparse, low-quality recommendations under a one-to-one assignment strategy. To achieve efficient semi-supervised learning by leveraging multiple positive queries, a one-to-many assignment is proposed instead of a one-to-one assignment:

[0058]

[0059] in It is a combination of M and N, representing each pseudobox. A subset of M proposals was assigned. Higher-order combinations were used to classify the score s and the IoU value u as a measure of matching cost.

[0060] m = s α ·u β

[0061] Here, α and β control the influence of classification scores and IoU during the assignment. In a one-to-one assignment, the M proposals with the largest m values ​​are selected as positive samples, while the remaining proposals are considered negative samples.

[0062] In the early stages of semi-supervised training, the model is iterated T1 times using a one-to-one assignment, during which the classification loss and regression loss are modified:

[0063]

[0064] By assigning multiple positive samples to each pseudo-label, potentially high-quality positive samples also have the opportunity to be optimized, which greatly improves the convergence speed and correspondingly obtains better-quality pseudo-labels. However, multiple positive samples for each pseudo-label lead to duplicate predictions. To mitigate this problem, the training phase switches back to one-to-one pairing in the second stage. By doing so, high-quality pseudo-labels are enjoyed after the first stage of training, and duplicate predictions are gradually reduced to achieve a one-to-one assignment training detector without NMS in the second stage. The loss function for this stage is:

[0065]

[0066] Step 5: Cross-view query consistency

[0067] In non-DETR-based SSOD frameworks, consistency regularization can often be conveniently used to minimize the teacher model f. θ And student model f′ θ The difference between outputs given the same input x but with different random augmentations:

[0068]

[0069] However, for DETR-based architectures, performing consistency regularization becomes impossible because there is no explicit (or deterministic) relationship between the input object query and its output prediction. To address this issue, a cross-view query consistency module is proposed, enabling DETR-based architectures to learn semantically invariant features of object queries across different augmented views. For each unlabeled image, on the pseudo-boundary box set b, we apply several MLPs to the RoI features extracted via RoIAlign:

[0070] c t =MLP(RoIAlign(F t ,b))

[0071] c s =MLP(RoIAlign(F s ,b))

[0072] Where F t and F s These represent the core characteristics of teachers and students, respectively. Then, c t and c s The original object query, treated as a cross-view query embedding and appended to another view, serves as input to the decoder.

[0073]

[0074] Where q. and E. represent the original object query and the encoded image features, respectively. Let 'o' represent the decoded features of the cross-view query and the original object query. Note that the subscripts 't' and 's' represent teacher and student, respectively. Guided by the semantics of the input cross-view query embedding, the correspondence of the decoded features can be naturally guaranteed, and we apply the following consistency loss:

[0075]

[0076] Step 6: Cost-based pseudo-label mining

[0077] To uncover more pseudo-bounding boxes with meaningful semantic information for consistent learning across view queries, a cost-based pseudo-label mining module is proposed. This module can dynamically mine reliable pseudo-bounding boxes from unlabeled data. Specifically, an additional binary matching is performed between the initially filtered pseudo-bounding boxes and the predicted proposals, and the matching cost is used to describe the reliability of the pseudo-bounding boxes.

[0078]

[0079] Where p i b i Let represent the classification and regression results for the i-th prediction proposal, and This represents the category label and bounding box coordinates of the j-th pseudo-label. Subsequently, in each training batch, the initial pseudo-boundings are clustered into two states by fitting a Gaussian mixture model. The matching cost is well aligned with the quality of the pseudo-boundings. Furthermore, a threshold is set for the cost of reliable pseudo-bounding cluster centers, and all pseudo-boundings below the threshold are collected for cross-view query consistency calculations.

[0080] Step 7: Loss Function

[0081] The final loss L is shown below:

[0082]

[0083] in and These are supervised loss and unsupervised loss, including classification loss and regression loss. L c Indicates cross-view Figure 1 Caused sexual harm. u and w c These are the unsupervised loss weights and consistency loss weights. t is the current training iteration number, and T1 is the training time of the first phase of the SHM module.

[0084] Through the above steps, end-to-end semi-supervised medical image target detection is achieved.

[0085] This invention preprocesses medical images before deblurring, including normalization, cropping, and resizing, to ensure consistent image quality and standardized model input. It improves the blur kernel estimation method by using deep learning techniques to more accurately estimate the blur kernel, resulting in clearer images during deblurring. The invention optimizes the teacher model by using larger model capacity, more complex network structures, or more efficient training strategies to improve the accuracy and reliability of pseudo-label generation. It also adjusts the architecture and parameters of the student model to better suit the characteristics of medical images, for example, by using more convolutional layers to capture local features or introducing attention mechanisms to enhance the model's focus on important regions. This invention introduces a cross-view query consistency module, enabling the model to learn semantically invariant features of object queries across different augmented views. This helps improve the model's generalization ability across different views, thereby improving the accuracy and reliability of object detection. Finally, this invention proposes a cost-based pseudo-label mining module to dynamically mine reliable pseudo-boundary boxes in unlabeled data for cross-view query consistency calculation. This method can extract more useful information from unlabeled data, thereby improving model performance. This invention designs a method that combines supervised loss, unsupervised loss, and cross-view loss. Figure 1 The model training is guided by a loss function based on consistency loss and its corresponding weights. This approach can effectively balance the relationship between different losses, thereby improving model performance.

[0086] In addition to the above embodiments, the present invention may have other implementation methods. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A semi-supervised medical image target detection method based on DETR, characterized in that, Includes the following steps: Step 1: Deblurring Inaccurate blur kernels and kernel errors, additive Gaussian noise are used to model the image degradation process; The solution space is constrained by combining deep residual priors, total variation priors, and sparse priors, and residuals, sharp images, and artifacts are modeled respectively. The problem is decomposed into two subproblems and solved using the gradient-based ADAM algorithm and proximal gradient descent; the input is a blurred image and parameters, and the output is the restored image and residual. Step 2: Supervised Pre-training For the recovered image x, a labeled image set is used during training. and unlabeled image set N s and N u These represent the number of labeled and unlabeled images, respectively; for a labeled image x s , annotation y s Includes the coordinates and object categories of all bounding boxes; Step 3: Pseudo-tag generation While following the popular teacher-student paradigm in SSOD, semi-supervised DETR employs a pair of teacher and student models with the same network architecture; specifically, in each training iteration, weakly and strongly augmented unlabeled images are fed into the teacher and student models, respectively; then, the pseudo-labels generated by the teacher are assigned with confidence scores greater than τ. s As supervision for training students; the parameters of the students are updated through backpropagation, while the parameters of the teacher model are the exponential moving averages of the students; Step 4: Two-stage hybrid matching strategy The DETR-based framework relies on a one-to-one mapping for end-to-end object detection; for DETR-based multi-task learning systems, the Hungarian algorithm is used to perform optimal one-to-one matching between student predictions and teacher-generated pseudo-labels. To achieve efficient semi-supervised learning by utilizing multiple positive queries, a one-to-many allocation is proposed instead of a one-to-one allocation. Higher-order combinations are used to classify scores s and IoU values ​​u as matching cost measures; In a one-to-many assignment, the M proposals with the largest m value are selected as positive samples, while the remaining proposals are considered as negative samples. In the early stages of semi-supervised training, a one-to-many assignment is used to iterate the model for T1 times. During this stage, the classification loss and regression loss are modified. In the second stage, the assignment is switched back to one-to-one during training. Step 5: Cross-view query consistency A cross-view query consistency module is proposed, enabling the DETR-based architecture to learn object query semantically invariant features across different augmented views. For each unlabeled image, several MLPs are applied to the RoI features extracted via RoIAlign on the pseudo-boundary box set b; then, c t and c s It is treated as a cross-view query embedding and attached to the original object query in another view as input to the decoder; ； Where F t and F s These respectively represent the core characteristics of teachers and students; Guided by the semantics of the input cross-view query embedding, the correspondence of decoded features is guaranteed; Step 6: Reliable Pseudo-Label Mining To uncover more pseudo-bounding boxes with meaningful semantic information for consistent learning across view queries, a cost-based pseudo-label mining module is proposed. This module dynamically mines reliable pseudo-bounding boxes from unlabeled data. Specifically, an additional binary matching is performed between the initially filtered pseudo-bounding boxes and the predicted proposals, and the matching cost is used to describe the reliability of the pseudo-bounding boxes. Subsequently, in each training batch, the initial pseudo-boxes are clustered into two states by fitting a Gaussian mixture model. Furthermore, the cost value of the reliable pseudo-box cluster center is set as a threshold, and all pseudo-boxes below the threshold are collected for cross-view query consistency calculation. Step 7: Train the model based on the loss function to achieve semi-supervised object detection. The final loss L is as follows: ； in and These are supervised loss and unsupervised loss, including classification loss and regression loss; L c Indicates cross-view consistency loss; ω u and ω c These are unsupervised loss weights and consistency loss weights; t is the current training iteration number, T1 is the iteration number of the first stage of training in the two-stage hybrid matching; o2m refers to one-to-many assignment, o2o refers to one-to-one assignment; Through the above steps, end-to-end semi-supervised medical image target detection is achieved.

2. The semi-supervised medical image target detection method based on DETR as described in claim 1, characterized in that, The DETR-based framework in step 4 relies on a one-to-one mapping of end-to-end object detection; for a DETR-based multi-task learning system, the Hungarian algorithm is used to perform optimal one-to-one matching between student predictions and teacher-generated pseudo-labels: ； Where ξ N It is a set of permutations of N elements. It's a fake tag. and student prediction results Matching costs between them; To achieve efficient semi-supervised learning by leveraging multiple positive queries, a one-to-many allocation is proposed instead of a one-to-one allocation: ； in It is a combination of M and N, representing each pseudobox. A subset of M proposals was allocated; Use higher-order combinations to classify scores s and IoU values ​​u as matching cost metrics: ； In this context, α and β control the influence of classification scores and IoU during the allocation process; in one-to-many allocations, the M proposals with the largest m values ​​are selected as positive samples, while the remaining proposals are considered as negative samples. In the early stages of semi-supervised training, a one-to-many assignment is used to iterate the model for T1 times, during which the classification loss and regression loss are modified: , ; ； By assigning multiple positive samples to each pseudo-label, potentially high-quality positive samples also have the opportunity to be optimized, which greatly improves the convergence speed and correspondingly obtains better-quality pseudo-labels; however, multiple positive samples for each pseudo-label lead to duplicate predictions; to alleviate this problem, the training phase switches back to one-to-one assignment in the second stage; by doing so, high-quality pseudo-labels are enjoyed after the first stage of training, and duplicate predictions are gradually reduced, so as to achieve a one-to-one assignment training detector without NMS in the second stage; the loss function for this stage is: 。 3. The semi-supervised medical image target detection method based on DETR as described in claim 2, characterized in that, Step 5 proposes a cross-view query consistency module, enabling the DETR-based architecture to learn object query semantically invariant features across different augmented views; for each unlabeled image, several MLPs are applied to the RoI features extracted via RoIAlign on the pseudo-boundary box set b: ； Then, c t and c s It is treated as a cross-view query embedding and attached to the original object query in another view as input to the decoder; Guided by the semantics of the input cross-view query embedding, the correspondence of decoded features is guaranteed, and the following consistency loss is applied: 。

Images