Semi-supervised medical image segmentation method and device based on anatomical connectivity

Abstract

The invention discloses a semi-supervised medical image segmentation method and device based on anatomical connectivity, and belongs to the field of medical image segmentation, and the method comprises the steps: employing a medical image segmentation model to carry out the segmentation of input labeled data and unlabeled data, and obtaining a first feature map and a second feature map; obtaining a pseudo-tag of the maximum connected region of the first feature map and the second feature map through discrimination, and obtaining a deterministic pseudo-tag according to the pseudo-tag of the maximum connected region and the uncertainty map; and combining the deterministic pseudo-label and the labeled data as a supervision signal, and guiding the segmentation model to carry out pixel-level contrast learning so as to improve the feature extraction capability. According to the method, the concept of connectivity judgment is introduced, the understanding of the model for the target structure is enhanced, and the accuracy of the model in the aspect of keeping the continuity of the anatomical structure is also improved.

Description

Technical Field

[0001] The present invention relates to the field of medical image segmentation, and in particular to a semi-supervised medical image segmentation method and device based on anatomical connectivity. Background Art

[0002] In precise medical image analysis, tolerance for tiny geometric deviations is extremely low, as these seemingly insignificant errors may cause changes in the global topological structure, thereby inducing functional misjudgment in subsequent clinical decision-making.

[0003] With the rapid development of deep neural networks, deep learning-based segmentation technology has been able to capture and learn from a set of pixel information from each local receptive field of an image through an encoder-decoder architecture. Existing semi-supervised medical image segmentation algorithms ignore the connectivity of organs in anatomical structures and frame the task as a pure pixel-level classification problem. These models often have topological errors in their predictions for organs such as the heart, such as outlier predictions, disconnected regions, and non-simple connection predictions, which often result in discontinuities in the segmentation results. For example, when using an image segmentation model to segment a cardiac MRI dataset, the distribution of the data is complex and changeable, and due to the similarity of grayscale and texture, it is easy to misidentify redundant regions as target regions, such as Figure 2 As shown in the figure, in the heart segmentation result, unreasonable fractures appear in the segmentation result, while the heart tissue should not have non-connected areas.

[0004] This point-to-point segmentation method ignores the intrinsic connectivity and geometric properties between pixels, which is particularly evident in medical data scenarios with limited sample size and varying quality. Low-quality medical images are often accompanied by high levels of noise and artifacts, which reduce spatial consistency and may cause topological problems. Summary of the invention

[0005] The main purpose of the present invention is to overcome the defect that the existing medical image segmentation model ignores the intrinsic connectivity and geometric properties between pixels, resulting in unreasonable non-connected areas in the segmentation results. A semi-supervised medical image segmentation method and device based on anatomical connectivity is proposed, the concept of connectivity judgment is introduced, the model's understanding of the target structure is enhanced, and the model's accuracy in maintaining the continuity of the anatomical structure is improved.

[0006] The present invention adopts the following technical solution:

[0007] A semi-supervised medical image segmentation method based on anatomical connectivity, comprising:

[0008] Using a medical image segmentation model to segment the input labeled data and unlabeled data to obtain a first feature map and a second feature map;

[0009] Obtaining a pseudo label of a maximum connected area of ​​the first feature map and the second feature map by discrimination, and obtaining a deterministic pseudo label according to the pseudo label of the maximum connected area and an uncertainty map;

[0010] The deterministic pseudo-labels and labeled data are combined as supervisory signals to guide the segmentation model to perform pixel-level contrast learning to improve the feature extraction capability.

[0011] The medical image segmentation model includes a first segmentation unit and a second segmentation unit using a U-Net network as a backbone network; the first segmentation unit segments the input labeled data and the unlabeled data to obtain the first feature map, the first segmentation unit includes an encoder and a first decoder, the first decoder uses a transposed convolutional layer and uses a jump connection with one of the encoders to transfer information between different layers; the second segmentation unit segments the input labeled data and the unlabeled data to obtain the second feature map, the second segmentation unit includes an encoder and a second decoder, the second decoder uses a linear interpolation layer and uses a jump connection with the encoder to transfer information between different layers.

[0012] For labeled data, the CE function and the Dice function are used to jointly calculate the supervision loss of labeled data, as follows:

[0013] ;

[0014] ;

[0015] ;

[0016] Among them, P a represents the first feature map, P b represents the second feature map, represents the label of the labeled data, L sup represents the supervision loss of label data, L a represents the supervision loss of the first feature map, L b represents the supervised loss of the second feature map, Dice(•) represents the Dice loss calculation function, and CE (•) represents the cross entropy loss calculation function.

[0017] For the unlabeled data, by selecting the first feature map P a and the second feature map P b The maximum connected area of ​​each category in the prediction is further improved to improve the spatial consistency of the prediction. The selection process of the connected area is as follows:

[0018] ;

[0019] ;

[0020] in and is the first feature map P a and the second feature map P b The maximum connected area of ​​category c in , ConnComp(Pa==c) means returning an array containing all connected areas belonging to category c, Represents a connected area The number of pixels in; the argmax function means selecting the connected area with the maximum number of pixels;

[0021] The maximum connected area between the two and Take the union and get the pseudo label of the largest connected area , expressed by the following formula:

[0022] ;

[0023] Where N is the total number of categories and ∪ represents the union of regions.

[0024] If some pixels are in the maximum connected area and , the conflict is resolved by comparing the predicted probabilities at the corresponding positions: and Represent the first feature map P a and the second feature map P b The probability of being in category c, the category label of each pixel can be determined as follows:

[0025] ;

[0026] in, represents the position of the pixel, It is a pseudo label In Location The value on; if the position exist Zhongqie Greater than , then the position is marked as category c; if the position exist Zhongqie Greater than or equal to , it is also marked as category c; if the position Not in the largest connected region of any category, then in the pseudo label are marked as background or no label and represented by 0.

[0027] The cross entropy loss is used to measure the consistency loss L between the predicted output of the segmentation model and the pseudo label. mcr , as follows:

[0028] ;

[0029] in, represents the position of the pixel, c represents the category, is a pseudo-label at position and the value on category c, and are the probability values ​​of pixels in the same position and category in the first feature map and the second feature map, respectively.

[0030] The method of combining the deterministic pseudo-labels and the labeled data as a supervisory signal to guide the segmentation model to perform pixel-level contrast learning specifically includes:

[0031] Designing positive samples z+ and negative samples in contrastive learning , for each category , select pixel indexes from the prediction output results of the medical image segmentation model with the same anchor point as positive sample indexes, and filter out the prediction results output by the medical image segmentation model that are different from the current category Different pixels,are used as negative samples;

[0032] Set a fixed-size memory buffer M to store samples from different categories, use anchor points randomly selected from M to calculate the contrast loss, and then update the memory buffer;

[0033] The loss function of the contrastive learning part uses infoNCE to calculate the pixel-level contrastive loss of the retained feature tensor and category:

[0034] ;

[0035] in, represents the pixel-level contrast loss, represents the i-th anchor point vector, and Represent the i-th positive sample vector and the j-th negative sample vector, K represents the number of negative samples, N represents the number of anchor vectors, is the temperature parameter.

[0036] The total loss function Loss of the medical image segmentation model is as follows:

[0037] ;

[0038] in, To monitor losses, is the pixel-level contrastive learning loss, is based on the maximum connected region consistency loss, is the encoder-decoder consistency loss, , and Preset coefficients.

[0039] A semi-supervised medical image segmentation device based on anatomical connectivity, comprising

[0040] A segmentation module, which uses a medical image segmentation model to segment the input labeled data and unlabeled data to obtain a first feature map and a second feature map;

[0041] A deterministic pseudo label acquisition module, which obtains the pseudo label of the maximum connected area of ​​the first feature map and the second feature map by discrimination, and obtains the deterministic pseudo label according to the pseudo label of the maximum connected area and the uncertainty map;

[0042] The pixel-level contrastive learning module combines deterministic pseudo-labels and labeled data as supervisory signals to guide the segmentation model to perform pixel-level contrastive learning to improve the feature extraction capability.

[0043] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:

[0044] Aiming at the common pixel connectivity problem in multi-category heart segmentation tasks, the present invention proposes a connectivity pseudo-label processing method. For unlabeled data, connectivity judgment post-processing is integrated into the prediction results of the model, and then deterministic pseudo-labels with high confidence are selected and combined with labeled data as supervisory signals to guide the segmentation model to perform pixel-level comparative learning, which not only strengthens the model's understanding of the target structure, but also improves the accuracy of the model algorithm in maintaining the continuity of the anatomical structure.

[0045] The segmentation model of the present invention adopts a first segmentation unit and a second segmentation unit composed of an encoder-decoder, and based on the feature consistency loss between the encoder-decoder, strengthens the transmission of feature information and improves the generalization ability of the model. Combined with a consistency module based on the maximum connected area, the maximum connected area of ​​the prediction results of the two decoders is obtained by discrimination, and reliable and logical pseudo labels are further screened for semi-supervised learning, so as to obtain more accurate shape mapping from unlabeled data and avoid discontinuous errors in the prediction results.

[0046] In the contrastive learning part of the invention, labeled data and high-quality pseudo-labels generated based on the maximum connected region are combined as supervisory signals. This strategy not only optimizes the accuracy of the model's classification of each pixel, but also enhances the semantic relationship between pixels, and promotes the model's in-depth learning of features in different regions. A contrastive learning loss function that combines labeled data and pseudo-labels is defined, aiming to maximize the amount of information learned by the model from complex image data. This loss function takes into account direct supervision information from labeled data and unlabeled data guided by pseudo-labels, thereby comprehensively improving the training efficiency and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a model framework diagram of the present invention;

[0048] Figure 2 This is a visualization diagram of the non-connected area;

[0049] Figure 3 (a) and Figure 3 (b) are comparisons of the performance of different methods on the 5% and 10% labeled ACDC datasets, respectively;

[0050] Figure 4 This is a comparison chart of the results under different labeled data;

[0051] Figure 5 The figure is a comparison of the qualitative results of the models on the ACDC dataset.

[0052] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0053] The present invention is further described below through specific implementation modes.

[0054] See also Figure 1 , a semi-supervised medical image segmentation method based on anatomical connectivity, which mainly includes three parts: feature consistency constraint (FC) between encoder and decoder, pseudo-label generation strategy MCR based on maximum connected region, and pixel level contrastive learning (PCL) based on pseudo-label. The specific description is as follows:

[0055] S1 uses a medical image segmentation model to segment the input labeled data and unlabeled data to obtain a first feature map and a second feature map.

[0056] The medical image segmentation model includes a first segmentation unit and a second segmentation unit using a U-Net network as a backbone network; the first segmentation unit segments the input labeled data and unlabeled data to obtain a first feature map, the first segmentation unit includes an encoder and a first decoder, the first decoder uses a transposed convolution layer and uses a jump connection with one of the encoders to transfer information between different layers. The second segmentation unit segments the input labeled data and unlabeled data to obtain a second feature map, the second segmentation unit includes an encoder and a second decoder, the second decoder uses a linear interpolation layer and uses a jump connection with the encoder to transfer information between different layers.

[0057] The encoders of the first segmentation unit and the second segmentation unit of the present invention may be the same, but the decoders are different. The encoder is responsible for capturing image features and gradually reducing the resolution. The encoder includes four downsampling stages, each of which consists of two convolutional layers and a pooling layer. Batch normalization and activation function operations are performed after each convolutional layer to enhance the stability of the network. After each stage of the input image, the size of the output feature map is reduced to half of the original size, and the channel dimension is doubled. The decoder is responsible for gradually restoring the resolution and generating a prediction result of the same size as the input image.

[0058] The invention designs a method to calculate the feature consistency loss between the encoder and decoder, strengthens the transmission of feature information, and improves the generalization ability of the model. First, for the labeled data Xl and its label y, and the unlabeled data Xu, the dual decoder architecture provides two different perspectives on the same data, and obtains the output prediction P a and P b .

[0059] Specifically, for labeled data, the CE function and the Dice function are used to jointly calculate the supervision loss of labeled data, as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] Among them, P a represents the first feature map, P b represents the second feature map, represents the label of the labeled data, L sup represents the supervision loss of label data, L a represents the supervision loss of the first feature map, L b represents the supervised loss of the second feature map, Dice(•) represents the Dice loss calculation function, and CE (•) represents the cross entropy loss calculation function.

[0064] The feature gap loss between encoder and decoder includes attention map generation and gap optimization as follows:

[0065] Attention map generation: First, the corresponding attention map is obtained by applying global average pooling (GAP) to compress the feature map at each layer of the encoder. , the corresponding attention map The generation of can be achieved through the following mapping function:

[0066] ;

[0067] Where p=2 is used to emphasize the sum of squares of elements within a channel, thereby highlighting important features. Represents the number of channels, height and width of the feature map, Represents the height and width of the attention map. Gap optimization: After obtaining the attention maps corresponding to the encoder and decoder, calculate the gap loss between the two attention maps:

[0068] ;

[0069] in, and Represents the decoder and encoder respectively. The attention map pixels,

[0070] is a time-based Gaussian temperature function, defined as:

[0071] ;

[0072] is the current training iteration number, is the maximum number of iterations. Due to the characteristics of the linear interpolation upsampling layer, the second decoder equipped with the linear interpolation upsampling layer is used to calculate the feature gap loss. In this way, the model is encouraged to maintain feature consistency between the encoder and decoder, thereby promoting more accurate learning and prediction of anatomical structures.

[0073] S2 obtains the pseudo label of the largest connected area of ​​the first feature map and the second feature map by discrimination, and obtains the deterministic pseudo label according to the pseudo label of the largest connected area and the uncertainty map.

[0074] For unlabeled data, by selecting the first feature map P a and the second feature map P b The maximum connected area of ​​each category in the prediction is further improved to improve the spatial consistency of the prediction. The selection process of the connected area is as follows:

[0075] ;

[0076] ;

[0077] in, and is the first feature map P a and the second feature map P b The maximum connected area of ​​category c in , ConnComp(Pa==c) means returning an array containing all connected areas belonging to category c, Represents a connected area The number of pixels in; the argmax function means selecting the connected area with the maximum number of pixels;

[0078] The maximum connected area between the two and Take the union and get the pseudo label used to build the consistency of the maximum connected area between the two , if both regions have valid predictions at a certain location, the category with a higher prediction probability is selected as the label of that location, which is expressed by the following formula:

[0079] ;

[0080] Where N is the total number of categories and ∪ represents the union of regions.

[0081] If some pixels are in the maximum connected area and If there is a conflict at ∠(i.e., these pixels are labeled as class c in both predictions), the conflict is resolved by comparing the predicted probabilities at corresponding positions:

[0082] make and Represent the first feature map P a and the second feature map P b The probability of being in category c, the category label of each pixel can be determined as follows:

[0083] ;

[0084] in, represents the position of the pixel, It is a pseudo label In Location The value on; if the position exist Zhongqie Greater than , then the position is marked as category c; if the position exist Zhongqie Greater than or equal to , it is also marked as category c; if the position Not in the largest connected region of any category, then in the pseudo label are marked as background or no label and represented by 0.

[0085] Pseudo Labels By selecting the category with a higher probability in the two prediction results at each pixel position, the probability information is retained, and the maximum connectivity of the categories is maintained through the union operation.

[0086] In this step, the uncertainty graph is used to optimize and screen the pseudo labels of the maximum connected area to obtain deterministic pseudo labels, which can improve the segmentation accuracy of the model and reduce the occurrence of erroneous segmentation to a certain extent.

[0087] The present invention uses cross entropy loss to measure the consistency loss L between the predicted output of the segmentation model and the pseudo label mcr , as follows:

[0088] ;

[0089] in, represents the position of the pixel, c represents the category, is a pseudo-label at position and the value on category c, and are the probability values ​​of pixels in the same position and category in the first feature map and the second feature map, respectively. The overall consistency loss is calculated by summing all positions, categories, and predictions, and then semi-supervised learning is performed using unlabeled data.

[0090] S3 combines the deterministic pseudo-labels and the labeled data as supervisory signals, guides the segmentation model to perform pixel-level contrast learning to improve the ability to extract features, and implements medical image segmentation based on the learned and optimized medical image segmentation model. The labeled data in this step can be different from the labeled data in step S1, or it can be the same data.

[0091] The loss function of the contrastive learning part of the present invention uses infoNCE to calculate the pixel-level contrastive loss of the retained feature tensor and category, aiming to maximize the amount of information learned by the model from complex image data. This loss function takes into account the direct supervision information from the labeled data and the unlabeled data guided by the pseudo-labels, thereby comprehensively improving the training efficiency and generalization ability of the model. First, the similarity between the anchor sample and other samples is calculated, and then the loss is calculated according to the temperature parameter. The loss function is expressed as follows:

[0092] ;

[0093] in, represents the pixel-level contrast loss, represents the i-th anchor point vector, and Represent the i-th positive sample vector and the j-th negative sample vector, K represents the number of negative samples, N represents the number of anchor vectors, is the temperature parameter.

[0094] The total loss function Loss of the medical image segmentation model of the present invention is as follows:

[0095] ;

[0096] in, To monitor losses, is the pixel-level contrastive learning loss, is based on the maximum connected region consistency loss, is the encoder-decoder consistency loss, , and is a preset coefficient and can be set to increase from 0 to 0.1 according to the Gaussian temperature function as the iteration rounds change.

[0097] A semi-supervised medical image segmentation device based on anatomical connectivity, which uses the above-mentioned semi-supervised medical image segmentation method based on anatomical connectivity to achieve medical image segmentation, comprising:

[0098] The segmentation module uses a medical image segmentation model to segment the input labeled data and unlabeled data to obtain a first feature map and a second feature map; the segmentation module is used to execute step S1 of the above method.

[0099] The deterministic pseudo label acquisition module obtains the pseudo label of the maximum connected area of ​​the first feature map and the second feature map by discrimination, and obtains the deterministic pseudo label according to the pseudo label of the maximum connected area and the uncertainty map. The deterministic pseudo label acquisition module is used to execute step S2 of the above method.

[0100] The pixel-level contrast learning module combines the deterministic pseudo-labels and the labeled data as supervisory signals to guide the segmentation model to perform pixel-level contrast learning to improve the feature extraction capability. The pixel-level contrast learning module is used to perform step S3 of the above method.

[0101] In order to verify the effectiveness of the proposed method, the following comparative experiments are used for analysis: the performance of the proposed method is evaluated on the ACDC dataset. The dataset mainly uses human images, including 2D and 3D segmentation tasks, such as the Automated Heart Diagnosis Challenge dataset and the left atrium magnetic resonance image segmentation dataset of patients with atrial fibrillation.

[0102] The experiment selected advanced methods such as UAMT for uncertainty-guided segmentation model, DTC for multi-task consistency proposed for medical image segmentation, and MC-Net+ for mutual consistency learning of cyclic pseudo-labels for comparative analysis, and used U-Net as the backbone network. In addition, the indicators of U-Net at 5% and 10% labeled data were reported as reference performance baseline comparison results.

[0103] To ensure the fairness of the experiment, the comparison methods used in this experiment are all officially disclosed codes and hyperparameters. And the results of a single fully supervised U-Net network are selected as the benchmark for comparison. The present invention has made further improvements on the multi-category heart segmentation task. Table 1 shows the comparative experiments conducted on the ACDC dataset, which clearly show the impact of different algorithms and label data ratios on the heart segmentation task.

[0104] Table 1

[0105]

[0106] The experimental results show that as the proportion of labeled data increases, the performance of all models is significantly improved. When comparing different semi-supervised learning strategies, the proposed method far outperforms existing methods in the four evaluation indicators of Dice, Jaccard, 95HD, and ASD. In particular, under the low proportion of labeled data (5%), the method achieves the best performance in all evaluation indicators. Under the condition of 5% labeled data, the Dice coefficient reaches 83.39%, which is higher than the baseline model U-Net.

[0107] The four indicators of the invented method were improved by 37.06%, 35.45%, 18.83% and 8.54%, far exceeding MT's 52.38% and the performance of other algorithms. This result highlights the superior performance of the invented method in dealing with the situation of scarce labels.

[0108] In addition, as the number of labeled data increases, the performance differences between algorithms gradually decrease. Under the condition of 10% labeled data, the Dice coefficient of the proposed method reaches 89.68%. Compared with the baseline model U-Net, the four indicators of the proposed method are improved by 10.31%, 14.96%, 9.45 and 2.56. After adding connectivity discrimination, the proposed method obtains the best segmentation result. This proves the effectiveness of the method in comprehensively improving segmentation accuracy and reducing prediction errors.

[0109] Figure 3 (a) and Figure 3 (b) more intuitively show the results of all methods in different numbers of labels and the gap between each model and the fully supervised U-Net model. Figure 4 The changes of the method of the present invention with the increase of labeled data are demonstrated. It can be observed that not only good results are achieved with very little labeled data, but also when the amount of labeled data is increased to 20%, the gap between the Dice and Jc indicators of the fully supervised model is narrowed to 1.57% and 2.14%.

[0110] Figure 5 The results of 2D slice visual comparison of the ACDC dataset under 5% and 10% labeled data are shown. The model of the present invention has fewer redundant areas, effectively reduces the prediction of non-connected areas, strengthens the feature map transfer between the encoder and decoder, and significantly improves the accuracy of segmentation. Especially on very little labeled data, the model of the present invention can accurately identify and correctly mark the target area, while the comparison model often produces incorrect segmentation in these challenging areas. The results under 10% labeled data show that the present model has its advantages in preserving the integrity of the cardiac ventricular septum structure. Compared with other methods, it shows significant improvement in segmentation coherence. In addition, by comparing the performance of the model under a low proportion of labeled data, it can be seen that the present model has a stronger adaptability to unlabeled data. Even with only 5% labeled data, the present model can still maintain the coherence and consistency of the segmentation results, thus highlighting its robustness and practicality.

[0111] The present invention conducts ablation experiments on the ACDC dataset with 5% and 10% labeled data to verify the effectiveness of each module of the model of the present invention. The experiment is divided into two groups, using 5% and 10% labeled samples respectively. The results are shown in Table 2. For the case of using 3 labeled samples, the overall performance improvement is more significant. The model L with the encoding-decoding consistency constraint module is introduced. FC Afterwards, compared with the baseline model trained with labeled data only, the Dice coefficient and Jaccard index increased by 5.19% and 4.84%, respectively, while 95HD and ASD decreased by 2.14 and 0.48, respectively.

[0112] Further adding the pseudo label generation strategy L based on the maximum connected area MCR After that, compared with the baseline model trained with only labeled data, the Dice coefficient and Jaccard index increased by 12.41% and 15.97% again, and 95HD and ASD decreased by 4.20 and 1.49, respectively. It can be seen that connectivity discrimination is important for heart segmentation. When all the modules of the present invention are combined (including pixel-level contrast learning based on pseudo labels L PCL ), compared to using only L Seg Compared with the baseline model, the Dice coefficient increased by 22.35%, the Jaccard index increased by 25.29%, and the 95HD and ASD decreased by 6.58 and 3.00, respectively, achieving significant improvements.

[0113] When using 7 labeled samples, the improvement trend of the model is similar to that of using 3 samples. FC After adding L MCR It only improved by 0.30%. However, after combining all modules, compared with the baseline model, the Dice coefficient increased by 3.48%, the Jaccard index increased by 5.38%, and 95HD and ASD decreased by 9.45 and 2.67, respectively, showing a more comprehensive performance improvement. These results clearly show that the addition of different modules has a significant positive impact on model performance, especially in maintaining anatomical continuity and reducing prediction errors, showing the complementarity and effectiveness of each module.

[0114] Table 2

[0115]

[0116] The parameter analysis of the maximum connected region pseudo-label strategy of the present invention is carried out, and only the contrastive learning module is retained. Table 3 shows the comparison results of different contrastive learning methods. By comparing the performance of three different contrastive learning methods on the ACDC dataset, the following results can be observed: the pseudo-label method provides a baseline performance, showing the effect of training with pseudo-labels alone without further optimization, and the obtained Dice coefficient is 87.48%, the Jaccard coefficient is 78.40%, the 95HD is 10.50, and the ASD is 2.84.

[0117] Table 3

[0118]

[0119] When the uncertainty map is used to screen reliable pixels, the pseudo-label method is improved compared to the baseline, with the Dice coefficient increased to 87.99%, the Jaccard coefficient to 79.28%, the 95HD reduced to 7.58, and the ASD reduced to 1.91. This shows that by optimizing the pseudo-label using the uncertainty map, the segmentation accuracy of the model can be improved to a certain extent and the number of incorrect segmentations can be reduced. The uncertainty map of the present invention combined with the connected region discrimination deterministic pseudo-label performs best, with the Dice coefficient further increased to 88.73%, the Jaccard coefficient to 80.38%, the 95HD significantly reduced to 4.85, and the ASD reduced to the lowest 1.43. This result shows that by selecting each category to add connected region discrimination, the reliability of the pseudo-label is improved, the prediction of non-anatomical structures is reduced, and the model's understanding of anatomical structures is enhanced. The noise and error of the pseudo-label during training are effectively reduced, thereby improving the performance of the model in a semi-supervised learning environment.

[0120] In summary, the present invention proposes a connectivity pseudo-label processing method for the common pixel connectivity problem in multi-category heart segmentation tasks. This method is based on the common sense of anatomical connectivity and optimizes the pseudo-label generation process, thereby ensuring the anatomical consistency of the segmentation results. In addition, the consistency of the feature space is combined with the connectivity of the prediction results. This method not only ensures the high accuracy of segmentation, but also significantly improves the accuracy of the model for multi-category heart segmentation.

[0121] Experimental results show that the semi-supervised learning strategy that combines anatomical connectivity with deep learning technology can not only improve the accuracy of specific organ segmentation, but also greatly improve the overall performance and reliability of medical image analysis tasks by ensuring the continuity and consistency of predictions.

[0122] It should be noted that, although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into multiple modules or units to be embodied.

[0123] Through the description of the above implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the implementation of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation of the present disclosure.

[0124] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary technical means in the art that are not disclosed in the present disclosure.

[0125] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A semi-supervised medical image segmentation method based on anatomical connectivity, characterized in that: include: Using a medical image segmentation model to segment the input labeled data and unlabeled data to obtain a first feature map and a second feature map; Obtaining a pseudo label of a maximum connected area of ​​the first feature map and the second feature map by discrimination, and obtaining a deterministic pseudo label according to the pseudo label of the maximum connected area and an uncertainty map; The deterministic pseudo-labels and labeled data are combined as supervisory signals to guide the medical image segmentation model to perform pixel-level contrast learning to improve the feature extraction capability.

2. The semi-supervised medical image segmentation method based on anatomical connectivity as claimed in claim 1, characterized in that: The medical image segmentation model includes a first segmentation unit and a second segmentation unit using a U-Net network as a backbone network; the first segmentation unit segments the input labeled data and unlabeled data to obtain the first feature map, the first segmentation unit includes an encoder and a first decoder, the first decoder uses a transposed convolutional layer and uses a jump connection with one of the encoders to transfer information between different layers; The second segmentation unit segments the input labeled data and the unlabeled data to obtain the second feature map. The second segmentation unit includes an encoder and a second decoder. The second decoder uses a linear interpolation layer and uses a jump connection with the encoder to transfer information between different layers.

3. The semi-supervised medical image segmentation method based on anatomical connectivity as claimed in claim 1, characterized in that: For the labeled data, the CE function and the Dice function are used to jointly calculate the supervision loss of the labeled data, as follows: ； ； ； Among them, P a represents the first feature map, P b represents the second feature map, represents the label of the labeled data, L sup represents the supervision loss of label data, L a represents the supervision loss of the first feature map, L b represents the supervised loss of the second feature map, Dice (•) represents the Dice loss calculation function, and CE (•) represents the cross entropy loss calculation function.

4. The semi-supervised medical image segmentation method based on anatomical connectivity as claimed in claim 1, characterized in that: For the unlabeled data, by selecting the first feature map P a and the second feature map P b The maximum connected area of ​​each category in the prediction is further improved to improve the spatial consistency of the prediction. The selection process of the connected area is as follows: ； ； in and is the first feature map P a and the second feature map P b The maximum connected area of ​​category c in , ConnComp(Pa==c) means returning an array containing all connected areas belonging to category c, Represents a connected area The number of pixels in; the argmax function means selecting the connected area with the maximum number of pixels; The maximum connected area between the two and Take the union and get the pseudo label of the largest connected area , expressed by the following formula: ； Where N is the total number of categories and ∪ represents the union of regions.

5. The semi-supervised medical image segmentation method based on anatomical connectivity as claimed in claim 4, characterized in that: If some pixels are in the maximum connected area and , the conflict is resolved by comparing the predicted probabilities at the corresponding positions: and Respectively represent the first feature map P a And the second feature map P b The probability of being in category c, then the category label of each pixel is determined as follows: ； in, represents the position of the pixel, It is a pseudo label In Location The value on; if the position exist Zhongqie Greater than , then the position is marked as category c; if the position exist Zhongqie Greater than or equal to , it is also marked as category c; if the position Not in the largest connected region of any category, then in the pseudo label are marked as background or no label and represented by 0.

6. The semi-supervised medical image segmentation method based on anatomical connectivity as claimed in claim 5, characterized in that: The cross entropy loss is used to measure the consistency loss L between the predicted output of the segmentation model and the pseudo label. mcr , as follows: ； in, represents the position of the pixel, c represents the category, is a pseudo-label at position and the value on category c, and are the probability values ​​of pixels in the same position and category in the first feature map and the second feature map, respectively.

7. The semi-supervised medical image segmentation method based on anatomical connectivity as claimed in claim 1, characterized in that: The method of combining the deterministic pseudo-labels and the labeled data as a supervisory signal to guide the segmentation model to perform pixel-level contrast learning specifically includes: Designing positive samples z+ and negative samples in contrastive learning , for each category , select pixel indexes from the prediction output results of the medical image segmentation model with the same anchor point as positive sample indexes, and filter out the prediction results output by the medical image segmentation model that are different from the current category Different pixels,are used as negative samples; Set a fixed-size memory buffer M to store samples from different categories, use anchor points randomly selected from M to calculate the contrast loss, and then update the memory buffer; The loss function of the contrastive learning part uses infoNCE to calculate the pixel-level contrastive loss of the retained feature tensor and category: ； in, represents the pixel-level contrast loss, represents the i-th anchor point vector, and Represent the i-th positive sample vector and the j-th negative sample vector, K represents the number of negative samples, N represents the number of anchor vectors, is the temperature parameter.

8. The semi-supervised medical image segmentation method based on anatomical connectivity as claimed in claim 2, characterized in that: The total loss function Loss of the medical image segmentation model is as follows: ； in, To monitor losses, is the pixel-level contrastive learning loss, is based on the maximum connected region consistency loss, is the encoder-decoder consistency loss, , and is the preset coefficient.

9. A semi-supervised medical image segmentation device based on anatomical connectivity, characterized in that: include: A segmentation module, which uses a medical image segmentation model to segment the input labeled data and unlabeled data to obtain a first feature map and a second feature map; A deterministic pseudo label acquisition module, which obtains the pseudo label of the maximum connected area of ​​the first feature map and the second feature map by discrimination, and obtains the deterministic pseudo label according to the pseudo label of the maximum connected area and the uncertainty map; The pixel-level contrastive learning module combines deterministic pseudo-labels and labeled data as supervisory signals to guide the segmentation model to perform pixel-level contrastive learning to improve the feature extraction capability.

Images