Graffiti supervision medical image segmentation method based on multi-scale contrast learning
Through the dual-branch network structure and hard pseudo-label mechanism, the problem of failure to effectively learn global shape prior knowledge and noise label interference in the prior art is solved, and the accuracy and stability of the medical image segmentation model are improved.
Patent Information
- Application Number
- CN202410938245.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-01
AI Technical Summary
Existing graffiti-supervised medical image segmentation models have difficulty learning global shape prior knowledge without fully annotated unpaired masks, and noise labels can lead to unstable segmentation performance.
Adopt a dual-branch network structure, a shared encoder, the main branch uses a CNN decoder, and the auxiliary branch uses an ASPP architecture, introducing a hard pseudo-label mechanism based on thresholds, allowing the supervised signal to propagate to the entire image through the dynamic combination of boundary loss and regional loss. Enhance the accuracy and generalization ability of segmentation models.
It realizes implicitly learning global shape prior knowledge without complete annotation, improves the accuracy and robustness of the segmentation model, reduces interference from noise labels, and enhances the effect of contrast learning.
Smart Images

Figure CN118887247A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image segmentation, and in particular to a graffiti supervised medical image segmentation method based on multi-scale contrast learning. Background Art
[0002] Medical image segmentation is a basic and critical task in computer-aided diagnosis, which aims to accurately identify target organs, tissues or lesion areas at the pixel level. Compared with images in natural scenes, medical images have more complex textures. Due to differences in imaging technology and equipment limitations, medical images are usually accompanied by noise and artifacts of varying degrees. In addition, most human organs are soft tissues with similar features and fuzzy boundaries, which increases the difficulty of discrimination. In recent years, deep learning has achieved remarkable results in automatic medical image segmentation, but most methods are fully supervised and require learning through a large number of labeled training samples. However, the manual annotation process of creating high-quality medical image datasets is time-consuming and expensive, relying on medical experts with cognition and experience, and the quality of annotation depends largely on the clinical experience and patience of the experts. To alleviate this problem, existing research is helping doctors gain more insights with less effort by reducing the need for complete and accurate labels, such as semi-supervised learning and weakly supervised learning. Semi-supervised learning uses unlabeled training data to assist model training, but requires that some of the labeled images in the dataset must be accurate. Weakly supervised learning proposes weak annotations that are easier to obtain (such as image-level annotations, sparse annotations, and noisy annotations), among which graffiti is the most convenient special weak annotation. Therefore, using graffiti-annotated datasets for model training is very attractive in medical image segmentation.
[0003] Existing graffiti supervised medical image models can be roughly divided into two categories. The first type of models introduces unpaired segmentation masks to train additional discriminative networks or autoencoder networks. However, the unpaired segmentation masks additionally required by such models are usually fully annotated. The second type of models expands graffiti annotations by generating pseudo labels or utilizing data augmentation strategies. However, noisy labels may be generated during the training process of such models. Since graffiti annotations are a rough representation of the region of interest and lack sufficient supervision information, noisy labels may mislead the training of the model, making its segmentation performance worse and unstable.
[0004] The existing solutions fail to address the two problems of learning global shape prior knowledge without fully annotated unpaired masks and generating noisy segmented regions caused by noisy labels. Summary of the invention
[0005] The purpose of the present invention is to provide a graffiti supervised medical image segmentation method based on multi-scale contrastive learning, aiming to solve the problem raised in the third part of the background technology.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] The present invention proposes a dual-branch network structure. The network shares an encoder, and the main branch uses a CNN decoder to generate prediction labels for target category slices. Since the receptive field of the deep features extracted by the encoder is limited, the auxiliary branch uses a decoder of the ASPP architecture. The present invention introduces a threshold-based hard pseudo-label mechanism to supervise unlabeled pixels so that the two branches can learn from each other. In this way, the local feature details obtained by the main branch can be integrated with the global representation of the auxiliary branch at multiple scales, and global shape prior knowledge can be implicitly learned without fully labeled unpaired masks.
[0008] In addition, in order to balance the local and multi-scale global information provided by the CNN branch and the ASPP branch, the independence of the two branches is maintained. The present invention uses a mechanism that dynamically combines boundary loss with regional loss to constrain pseudo-label supervision and allows the supervision signal to propagate from limited annotated pixels to all unlabeled pixels, thereby amplifying to the entire image, thereby improving the accuracy and generalization ability of the segmentation model.
[0009] The present invention introduces an attention gating mechanism in the main branch and designs a process specifically for processing non-target slices in the auxiliary branch. By constructing positive and negative samples and applying contrastive learning, the model can more accurately distinguish the target category foreground and non-target category background. Most existing segmentation methods rely on images with target classes, while the present invention uses slices that do not contain target categories for auxiliary supervision in the graffiti supervision scenario, constructs an end-to-end contrastive learning segmentation framework, and then guides image block-level contrastive learning through hard pseudo labels, reducing the interference of noise labels and helping the network learn more discriminative features.
[0010] A graffiti supervised medical image segmentation method based on multi-scale contrastive learning, comprising the following steps:
[0011] S1. Obtain a medical image to be segmented and annotate the medical image at the graffiti level;
[0012] S2, preprocessing the medical image to be segmented and the corresponding annotated image, and dividing the preprocessed data set into a training set, a validation set, and a test set according to the ratio;
[0013] S3. Use the divided training set and validation set to train the graffiti supervised medical image segmentation model, and use the divided test set to test the model. During the training, a dual-branch network and a hybrid supervised learning mechanism are used for model training.
[0014] S4. Use the tested graffiti supervised medical image segmentation model to perform image segmentation on the medical image.
[0015] Furthermore, the graffiti supervised medical image segmentation model includes:
[0016] A shared parameter encoder to extract global features of medical images in the training set;
[0017] The dual-branch network decoder, which includes a CNN decoder and an ASPP decoder, is used to map the medical image from graffiti annotations to the target object segmentation according to the global features to obtain the segmentation result.
[0018] Furthermore, the graffiti supervised medical image segmentation model is trained as follows:
[0019] Divide the images in the training set, validation set, and test set into target category slices and non-target slices;
[0020] According to different medical image training samples in a batch selected from the training set, a shared parameter encoder is used to extract global features, and the extracted global features are target category features and non-target features;
[0021] According to the extracted target category features and non-target features, a dual-branch network decoder consisting of a CNN decoder and an ASPP decoder is used to predict and segment the medical image. Two prediction results are output by the dual-branch network decoder respectively. The graffiti label of the target category slice is used to calculate the graffiti supervision loss. Then, the two prediction results are fused to generate a hard pseudo label and the corresponding uncertainty map. The hard pseudo label supervision loss is calculated based on the generated hard pseudo label and the prediction results output by the two branches.
[0022] The prediction head of the ASPP auxiliary branch samples the image blocks, and the predicted segmentation mask obtained by the auxiliary branch guides the target category slice. The image blocks containing the target category are taken as positive samples, and mask average pooling MAP is performed to obtain positive sample projections. The image blocks in the background area and all image blocks in the non-target slices are taken as negative samples. The negative sample projections are obtained through nonlinear projection operation MLP, and the positive sample projections and negative sample projections are compared and learned. Loss;
[0023] The trained graffiti supervised medical image segmentation model is verified using the validation set to complete the training of the graffiti supervised medical image segmentation model.
[0024] Furthermore, the graffiti supervision loss is calculated according to the following formula:
[0025]
[0026]
[0027] in, is the graffiti supervision loss, s represents the one-hot graffiti annotation of the input image, is the partial cross entropy loss, where C is the category set of the input image x, Ω s is the set of marked pixels for the graffiti label s, and They represent the probability that the i-th pixel in the graffiti annotation belongs to the c-th class and the probability that the i-th pixel in the prediction result belongs to the c-th class respectively.
[0028] Furthermore, hard fake labels are generated using a threshold-based method, according to the following formula:
[0029] Y t =argmax[α×(y CNN >t)×y CNN +(1-α)×(y ASPP >t)×y ASPP ]
[0030] where Y t represents a hard false label, α is randomly generated in each iteration, the value range of α is [0,1], t is the threshold, the value of the threshold is 0.5, y CNN is the prediction result output by the CNN decoder of the CNN branch, y ASPP It is the prediction result output by the ASPP decoder of the ASPP branch.
[0031] Furthermore, the hard pseudo-label supervision loss is calculated according to the following formula:
[0032]
[0033]
[0034] in, is the Dice loss function, is the hard pseudo-label supervision loss, β is the boundary loss weight, and the value range of β is [0,1]. Its initial value is greater than 0 but relatively small. As the training progresses, β becomes larger and larger.
[0035]
[0036] in represents the boundary loss, and They are the boundary representations of the prediction results and hard pseudo labels respectively. Dist representation adopts the form of distance measurement in the regional boundary space.
[0037] Furthermore, when sampling the image block, the features obtained by the target category slices, non-target slices and hard pseudo label masks are evenly segmented based on the image block, and the non-target image block projection set is obtained by nonlinear projection MLP operation. and background image patch projection collection The target category slice is guided by the hard pseudo-label mask, and the mask average pooling MAP operation is performed to obtain the foreground image block projection set Given an image block size (D H ,D H ) and a feature embedding MAP is implemented by the following formula:
[0038]
[0039] Where (h, w) is the spatial coordinate of each image block, is the image block d j The projection vector obtained after mask average pooling, is the index of the image block.
[0040] Furthermore, the contrastive learning loss is calculated using the following formula:
[0041]
[0042]
[0043] Where N is the number of anchor samples selected in each iteration, v i is the prototype vector of the i-th anchor sample, is the positive sample prototype vector, is the negative sample prototype vector, τ is the temperature hyperparameter, and sim(·) represents the cosine similarity;
[0044] The total loss function is:
[0045]
[0046] where λ 1-3 is the weight value of each part of the loss to balance different supervision losses.
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] The present invention provides a graffiti supervised medical image segmentation method based on multi-scale contrastive learning. A dual-branch network structure is proposed. The network shares an encoder, the main branch adopts a CNN decoder, and the auxiliary branch adopts an ASPP architecture decoder. By incorporating non-target slices into the segmentation task, the model's ability to distinguish between the target category foreground and the non-target category background is enhanced. The present invention introduces a threshold-based hard pseudo-label mechanism to supervise unlabeled pixels so that the two branches can learn from each other. In this way, the local feature details obtained by the main branch can be integrated with the global representation of the auxiliary branch at multiple scales, and the global shape prior knowledge can be implicitly learned without a fully labeled unpaired mask.
[0049] A multi-scale contrastive learning method based on image patches is proposed to obtain high-quality pseudo-labels while learning feature representations. Under the guidance of pseudo-labels, the sampling of positive and negative samples for contrastive learning is refined, and the contrast loss and graffiti supervision loss are jointly optimized. In addition, the background tissue information of non-target slices is fully utilized as negative samples, which improves the robustness and generalization ability of the segmentation model and reduces the interference of noise labels. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 It is a schematic diagram of the overall method flow of the present invention;
[0051] Figure 2 A schematic diagram of a graffiti tag of the present invention;
[0052] Figure 3 It is a schematic diagram of the framework flow of the graffiti supervised medical image segmentation model of the present invention. DETAILED DESCRIPTION
[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0054] Example 1
[0055] like Figures 1 to 3 As shown, a graffiti supervised medical image segmentation method based on multi-scale contrastive learning, the method comprising:
[0056] A graffiti supervised medical image segmentation method based on multi-scale contrastive learning, comprising the following steps:
[0057] S1. Obtain a medical image to be segmented and annotate the medical image at the graffiti level;
[0058] S2, preprocessing the medical image to be segmented and the corresponding annotated image, and dividing the preprocessed data set into a training set, a validation set, and a test set according to the ratio;
[0059] In this embodiment, in order to effectively train, optimize and evaluate the weakly supervised medical image segmentation model, the data set is divided into a training set, a validation set and a test set according to a certain ratio, and the images in the training set, the validation set and the test set are divided into target category slices and non-target slices; general image segmentation only focuses on the target area to be segmented, that is, the position marked by the label. The present invention focuses on and mines the tissue structure anatomical information of the background area. Therefore, the images in the data set can be divided into two categories: target category slices and non-target slices, and the background area in the target category slices also belongs to the non-target.
[0060] S3. Use the divided training set and validation set to train the graffiti supervised medical image segmentation model, and use the divided test set to test the model. During training, a dual-branch network and a hybrid supervised learning mechanism are used for model training.
[0061] In this embodiment, according to the extracted target category features and non-target features, a dual-branch network decoder composed of a CNN decoder and an ASPP decoder is used to predict and segment the medical image, and two prediction results are output respectively through the dual-branch network decoder. The graffiti label of the target category slice is used to calculate the graffiti supervision loss, and then the two prediction results are fused to generate a hard false label and a corresponding uncertainty map. The hard false label supervision loss is calculated according to the generated hard false label and the prediction results output by the two branches;
[0062] Graffiti supervision belongs to the category of weak supervision. In weak supervision labels, take the semantic segmentation of two categories as an example. These two categories are background and foreground. The given weak supervision labels only annotate a small part of the foreground. There are some problems in directly using such weak supervision labels for training, because most of the foreground labels are not annotated, resulting in a large suppression of the foreground category. Therefore, this example uses a graffiti supervision loss based on partial cross entropy:
[0063]
[0064]
[0065] in, is the graffiti supervision loss, s represents the one-hot graffiti annotation of the input image, is the partial cross entropy loss, where C is the category set of the input image x, Ω s is the set of marked pixels for the graffiti label s, and They represent the probability that the i-th pixel in the graffiti annotation belongs to the c-th class and the probability that the i-th pixel in the prediction result belongs to the c-th class respectively;
[0066] The present invention uses a dual-branch network to generate two sets of predictions with different focuses, namely local fine-grained features and multi-scale contextual global information. Then the dynamic mixture y CNN and ASPP Generate hard pseudo labels to supervise the two predictions. In order to reduce errors, hard pseudo labels are generated using a threshold-based method, which is obtained according to the following formula:
[0067] Y t =argmax[α×(y CNN >t)×y CNN +(1-α)×(y ASPP >t)×y ASPP ]
[0068] where Y t represents a hard false label, α is randomly generated in each iteration, the value range of α is [0,1], t is the threshold, the value of the threshold is 0.5, y CNN is the prediction result output by the CNN decoder of the CNN branch, y ASPP The prediction result output by the ASPP decoder of the ASPP branch;
[0069] In the graffiti supervision scenario, in the early segmentation stage, since the Dice loss function is region-based, it can have a significant effect on irregularly shaped structures such as the right ventricle. However, in the later stages of segmentation, when it comes to the overlapping area between the left ventricular myocardium and the left ventricle, it is often difficult to further improve the accuracy of segmentation by relying solely on regional loss. In this case, it may be necessary to introduce a boundary loss function that captures contour features to solve the problem. The present invention combines boundary loss and regional loss, and assigns a weight value β greater than 0 but relatively small to the boundary loss. As the training progresses, the weight of the boundary loss becomes larger and larger, while the weight of the regional loss becomes smaller and smaller. When calculating the hard pseudo-label supervision loss, it is obtained according to the following formula:
[0070]
[0071]
[0072] in, is the Dice loss function, is the hard pseudo-label supervision loss, β is the boundary loss weight, and the value range of β is [0,1]. Its initial value is greater than 0 but relatively small. As the training progresses, β becomes larger and larger.
[0073]
[0074] in, and They are the boundary representations of the prediction results and hard pseudo labels respectively. Dist represents the distance measurement in the region boundary space. θ :Ω→[0,1] represents the probability output of the prediction result after softmax, the level set function According to Y t By calculating their dot product and integrating over the entire domain Ω, we can measure the distance between the prediction result and the hard false label in the boundary space. When the softmax predictions of all pixels in the region are 1) and all positive values are discarded (that is, the model's softmax predictions for the background are all 0), the boundary loss reaches the global minimum;
[0075] When sampling image blocks for the prediction head of the ASPP auxiliary branch, the features obtained by the target category slices, non-target slices and hard pseudo label masks are evenly segmented based on the image blocks, and the non-target image block projection set is obtained through nonlinear projection MLP operation. and background image patch projection collection The target category slice is guided by the hard pseudo-label mask, and the mask average pooling MAP operation is performed to obtain the foreground image block projection set Given an image block size (D H ,D H ) and a feature embedding MAP is implemented by the following formula:
[0076]
[0077] Where (h, w) is the spatial coordinate of each image block, is the image block d j The projection vector obtained after mask average pooling, is the index of the image block;
[0078] When constructing the contrast loss, the present invention projects the foreground image block into a set As positive sample key, background image block projection set and non-target image patch projection set As negative sample key. This scheme can generate enough negative samples, which helps to learn more discriminative feature representations. When calculating the contrast learning loss, it is implemented by the following formula:
[0079]
[0080]
[0081] Where N is the number of anchor samples selected in each iteration, v i is the prototype vector of the i-th anchor sample, is the positive sample prototype vector, is the negative sample prototype vector, τ is the temperature hyperparameter, and sim(·) represents the cosine similarity;
[0082] The total loss function is:
[0083]
[0084] where λ 1-3 is the weight value of each part of the loss to balance different supervision losses.
[0085] S4. Use the tested graffiti supervised medical image segmentation model to perform image segmentation on the medical image.
[0086] After the graffiti supervised medical image segmentation model training is completed, an independent test set is used to test the graffiti supervised medical image segmentation model to evaluate its performance on unseen data.
[0087] Example 2
[0088] like Figure 3 As shown, when the graffiti supervised medical image segmentation model is working, the target category slice x1 is input into the model, and the important structural information in the image is first captured by a shared encoder. Then the main branch further processes the feature map F1 output by the encoder to generate a fine-grained segmentation result Predictiony1; the auxiliary branch captures multi-scale context information through convolutions with different void rates, and generates a segmentation result Predictiony2 with global information. The two branches are trained and work together at the same time. The segmentation results of each branch are compared with the graffiti annotation through partial cross entropy loss, and then pseudo-supervised learning is performed with the hard pseudo labels generated by dynamically mixing the two predictions y1 and y2. In order to enhance the feature extraction capability of the model, the present invention adds a Dropout layer after each group of convolution layers of the encoder and decoder, thereby introducing feature perturbations in the two branches;
[0089] In this embodiment, the main branch uses the U-Net architecture as the backbone network to improve the CNN decoder, and its downsampling factor is 16. In medical images, all features are considered to be crucial, so the Skip Connection structure can play a better role. In addition, in each decoder layer of the main branch, the present invention introduces a gated attention mechanism, which can automatically model the foreground and background and calculate the weights that should be assigned to pixels at different positions, which enables the model to focus more on the area of concern, thereby improving the segmentation quality. Before the last layer of the prediction head of the main branch, the present invention uses a plug-and-play channel attention SE module, which automatically obtains the importance of the four category channels of the final output through learning, thereby enhancing useful features and suppressing features that are not very useful for the segmentation task;
[0090] The auxiliary branch uses the ASPP decoder architecture as the backbone network. Specifically, ASPP consists of a 1x1 convolution layer, three 3x3 dilated convolution layers with expansion rates of 6, 12, and 18, and a global average pooling layer (GAP). In the model framework of this scheme, the deepest feature receptive field extracted by the encoder is 140, which is far from enough for the resolution of the input image 256x256. The receptive fields of the five parallel branches of ASPP are 140, 152, 164, and 176, and the global receptive field provided by GAP can capture contextual information at different scales. In addition, the present invention splices the high-level semantic features obtained by fusing the shallow low-level semantic feature layer with the deep ASPP, so the final output of this branch contains more local details and global contextual information.
[0091] For CT and MRI datasets, most of the original scans contain some slices without target category organs, and their masks are all black. The present invention defines these slices as non-target slice set D nt Since the contrast of medical images is low, the boundary between the target category and the background tissue is not obvious, which makes it difficult to accurately segment the target organ. These non-target slices are rich in anatomical knowledge of irrelevant organs or tissues and can naturally distinguish the target category;
[0092] In order to effectively utilize this information, the present invention designs a non-target slice processing flow. First, all dataset samples are screened and D nt During training, each batch randomly selects K non-target slices with the same number of target category slices;
[0093] For the target category slice x1, a non-target slice x2 is simultaneously input into the shared encoder during training, and the non-target feature F2 is extracted and then sent to the ASPP branch. Finally, at the position of the prediction head, the present invention projects F1 and F2 rich in multi-scale information into the vector space through the MAP and MLP modules respectively, and then constructs positive and negative sample pairs for comparative learning.
[0094] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0095] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.
[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A graffiti supervised medical image segmentation method based on multi-scale contrastive learning, It is characterized in that The following steps are involved: S1. Obtain a medical image to be segmented and annotate the medical image at the graffiti level; S2, preprocessing the medical image to be segmented and the corresponding annotated image, and dividing the preprocessed data set into a training set, a validation set, and a test set according to the ratio; S3. Use the divided training set and validation set to train the graffiti supervised medical image segmentation model, and use the divided test set to test the model. During the training, a dual-branch network and a hybrid supervised learning mechanism are used for model training. S4. Use the tested graffiti supervised medical image segmentation model to perform image segmentation on the medical image.
2. According to claim 1, the graffiti supervised medical image segmentation method based on multi-scale contrastive learning, It is characterized in that The graffiti supervised medical image segmentation model specifically includes: A shared parameter encoder to extract global features of medical images in the training set; The dual-branch network decoder, which includes a CNN decoder and an ASPP decoder, is used to map the medical image from graffiti annotations to the target object segmentation according to the global features to obtain the segmentation result.
3. According to claim 1, the graffiti supervised medical image segmentation method based on multi-scale contrastive learning, It is characterized in that The graffiti supervised medical image segmentation model is trained as follows: Divide the images in the training set, validation set, and test set into target category slices and non-target slices; According to different medical image training samples in a batch selected from the training set, a shared parameter encoder is used to extract global features, and the extracted global features are target category features and non-target features; According to the extracted target category features and non-target features, a dual-branch network decoder consisting of a CNN decoder and an ASPP decoder is used to predict and segment the medical image. Two prediction results are output by the dual-branch network decoder respectively. The graffiti label of the target category slice is used to calculate the graffiti supervision loss. Then, the two prediction results are fused to generate a hard pseudo label and the corresponding uncertainty map. The hard pseudo label supervision loss is calculated based on the generated hard pseudo label and the prediction results output by the two branches. The prediction head of the ASPP auxiliary branch samples the image blocks, and the predicted segmentation mask obtained by the auxiliary branch guides the target category slice. The image blocks containing the target category are taken as positive samples, and mask average pooling MAP is performed to obtain positive sample projections. The image blocks in the background area and all image blocks in the non-target slices are taken as negative samples. The negative sample projections are obtained through nonlinear projection operation MLP, and the contrastive learning loss is calculated for the positive sample projections and negative sample projections; The trained graffiti supervised medical image segmentation model is verified using the validation set to complete the training of the graffiti supervised medical image segmentation model.
4. According to claim 3, the graffiti supervised medical image segmentation method based on multi-scale contrastive learning, It is characterized in that The graffiti supervision loss is calculated according to the following formula: in, is the graffiti supervision loss, s represents the one-hot graffiti annotation of the input image, is the partial cross entropy loss, where C is the category set of the input image x, Ω s The set of marked pixels for the graffiti label s, and They represent the probability that the i-th pixel in the graffiti annotation belongs to the c-th class and the probability that the i-th pixel in the prediction result belongs to the c-th class respectively.
5. The graffiti supervised medical image segmentation method based on multi-scale contrastive learning according to claim 3, It is characterized in that The hard fake label is generated using a threshold-based method, according to the following formula: AND t =argmax[α×(y CNN >t)×y CNN +(1-α)×(and ASPP >t)×y ASPP ] where Y t represents a hard false label, α is randomly generated in each iteration, the value range of α is [0,1], t is the threshold, the value of the threshold is 0.5, y CNN is the prediction result output by the CNN decoder of the CNN branch, y ASPP It is the prediction result output by the ASPP decoder of the ASPP branch.
6. The graffiti supervised medical image segmentation method based on multi-scale contrastive learning according to claim 3, Features: The hard pseudo-label supervision loss is calculated according to the following formula: in, is the Dice loss function, is the hard pseudo-label supervision loss, β is the boundary loss weight, and the value range of β is [0,1]. Its initial value is greater than 0 but relatively small. As the training progresses, β becomes larger and larger. in represents the boundary loss, and They are the boundary representations of the prediction results and hard pseudo labels respectively. Dist representation adopts the form of distance measurement in the regional boundary space.
7. The graffiti supervised medical image segmentation method based on multi-scale contrastive learning according to claim 3, It is characterized in that When the image block is sampled, the features obtained by the target category slices, non-target slices and hard pseudo label masks are evenly segmented based on the image block, and a non-target image block projection set is obtained by nonlinear projection MLP operation. and background image patch projection collection The target category slice is guided by the hard pseudo-label mask, and the mask average pooling MAP operation is performed to obtain the foreground image block projection set Given an image block size (D H ,D H ) and a feature embedding MAP is implemented by the following formula: Where (h, w) is the spatial coordinate of each image block, is the image block d j The projection vector obtained after mask average pooling, is the index of the image block.
8. The graffiti supervised medical image segmentation method based on multi-scale contrastive learning according to claim 7, It is characterized in that The contrastive learning loss is calculated using the following formula: Where N is the number of anchor samples selected in each iteration, v i is the prototype vector of the i-th anchor sample, is the positive sample prototype vector, is the negative sample prototype vector, τ is the temperature hyperparameter, and sim(·) represents the cosine similarity; The total loss function is: where λ 1-3 is the weight value of each part of the loss to balance different supervision losses.
Citation Information
Cited By
Pseudo-supervised learning traffic image quality enhancement method based on edge information guidance
CN119559075A
Weak supervision medical image segmentation method and system based on causal intervention
CN119693642A
Semi-supervised polyp segmentation method based on hierarchical granularity contrast learning
CN120107961A
Ultrasonic artery segmentation method, system and equipment based on positioning knowledge driving
CN121147243A
Multi-organ medical image segmentation method and device based on weak supervision
CN121304709A
