Medical image segmentation system based on dual-scale consistency

Abstract

The invention belongs to the field of medical image analysis and processing systems, and discloses a medical image segmentation system based on dual-scale consistency, and the specific system scheme is as follows: taking a breast ultrasound image as an example, adopting a semi-supervised learning mechanism, carrying out supervised training on a small amount of labeled data, and carrying out unsupervised training on a large amount of unlabeled data; the backbone network is a dual-scale consistent teacher-student network, and accurate segmentation of a medical target is realized through feature extraction and feature interaction of a dual-scale medical image; a common constraint network of supervision loss, scale consistency loss and reliable pseudo-label loss is adopted for training, the scale consistency loss is used for implementing constraint on images before and after disturbance, more reliable pseudo-labels are obtained from prediction probability distribution of two teacher networks, and the reliable pseudo-label loss can guarantee smooth implementation of unsupervised training. According to the method, the dependence of the model on the annotation data can be reduced on the premise of ensuring the model segmentation precision.

Description

Technical Field

[0001] The present invention belongs to the field of medical image analysis and processing systems, and in particular relates to a medical image segmentation system based on dual-scale consistency. Background Art

[0002] Medical image segmentation refers to the division of an image into different regions based on the feature information of image pixels, each region representing a different tissue, organ, or other anatomical structure of interest. As a core component of medical image analysis, the intelligent segmentation system for breast lesions aims to accurately locate abnormal tissue in ultrasound images using computer vision algorithms. With the iterative upgrade of artificial intelligence systems, image parsing methods based on deep learning have shown revolutionary potential in the medical field, especially in the identification of lesion regions. This type of algorithm relies on a deep convolutional network architecture and extracts features through autonomous learning. Compared with traditional segmentation models that rely on manually designed features, it exhibits better generalization and robustness in medical image processing. The acquisition of breast ultrasound image datasets is somewhat difficult, and the complexity of breast tissue structure and the diversity of lesion morphology also increase the difficulty of segmentation.

[0003] The current methodological system in the field of medical image segmentation can be divided into two categories: segmentation methods based on traditional image processing algorithms and data-driven deep learning segmentation methods. Traditional image processing algorithms cover a series of classic algorithms such as grayscale thresholds, feature clustering, and active contours. These methods have established important benchmarks in early medical image segmentation.

[0004] Threshold-based breast tumor segmentation method: Threshold-based segmentation method is one of the classic image segmentation algorithms. The core idea of ​​this segmentation method is to segment the image based on a specific threshold without considering spatial location information. Therefore, in threshold-based segmentation methods, the main factor affecting the segmentation results is the selection of an appropriate threshold.

[0005] Clustering-based breast tumor segmentation method: Clustering methods have been widely used in breast ultrasound image segmentation. The basic principle is to group pixels according to pixel features in the image (such as grayscale value, texture, etc.). Common clustering algorithms include the K-means algorithm and the fuzzy C-means (FCM) algorithm.

[0006] Breast tumor segmentation method based on active contours: The active contours method is an effective breast ultrasound image segmentation system. By defining dynamic contours, it can adaptively adjust its shape in the image to fit the target boundary.

[0007] The above-mentioned methods based on traditional image processing algorithms usually have the advantages of low computational complexity, simple implementation and wide applicability. However, when processing complex scenes and irregular shapes, they are often not as robust and accurate as deep learning-based methods.

[0008] Deep Learning-Based Segmentation Algorithms: Image segmentation methods based on convolutional neural networks (CNNs): In recent years, deep learning-based medical image segmentation algorithms have become a key research direction in the field of deep learning due to their outstanding performance in accuracy and automation. Within deep learning models, convolutional neural networks can automatically extract image features, improving image segmentation accuracy.

[0009] Image segmentation methods based on generative adversarial networks (GANs): Generative adversarial networks (GANs) are a classic generative model in deep learning. Their core architecture consists of two complementary neural networks: a generator (G) and a discriminator (D). The core of the generator is to generate virtual samples that approximate the distribution of real data by learning the underlying distribution characteristics of real data, while the discriminator is responsible for distinguishing whether the input data is real data or synthesized by the generator. The two networks are trained adversarially through a dynamic game mechanism, ultimately driving the generator to produce samples that are highly similar to real data. Addressing the widespread scarcity of labeled data in the field of medical image segmentation, GANs have demonstrated unique application value. Many researchers have proposed semi-supervised segmentation algorithms based on GANs that have low requirements for labeled data and improve their segmentation performance by utilizing large amounts of unlabeled data.

[0010] Denoising Diffusion Probabilistic Models (DDPMs) are widely used in a variety of fields, including image synthesis, super-resolution reconstruction, object detection, and image segmentation, due to their powerful data generation capabilities. For example, Wang et al. used a diffusion probability model to refine the output of an existing segmentation model, improving segmentation performance compared to a baseline model. DDPMs have also attracted widespread attention in the field of medical image segmentation. Wu et al. first applied the diffusion probability model to medical image segmentation. By adding noise to the real mask through a reverse process, the generated noise map and the original image were used as input to the diffusion probability model, achieving accurate segmentation of the target area.

[0011] Image segmentation method based on mean teacher network: Mean-teacher is a semi-supervised learning framework proposed by Antti et al. Unlike the teacher-student network based on knowledge distillation, in the mean teacher network, the parameter update of the teacher model is derived from the exponential moving average (EMA) of the student model parameters. In this way, the teacher model can give more accurate guidance to the student model without additional training. Unlike the uncertainty-aware self-ensembling model teacher network (UAMT) framework, the uncertainty-guided co-mean teacher network (UCMT) uses two student models and a shared teacher model to accurately segment medical images. The parameter update of the teacher model is realized by weighted sum of the exponential moving average of the parameters of the two student models, and the UCMT adopts a two-stage way to train the network model, the first stage obtains the uncertainty map by training the UCMT and generates new samples, and the second stage trains the model with the uncertainty map as the sample, which effectively utilizes the unlabeled sample images to realize the accurate segmentation of skin cancer and polyps.

[0012] Through the research on various deep learning network architectures in the field of medical image segmentation, it is found that the image segmentation method based on deep learning has shown surprising performance in its own task, and has provided more powerful system support for the analysis and research of medical images. However, the existing method still has certain limitations: although the CNN-based method has strong feature extraction ability, its segmentation performance is often limited when dealing with medical images with serious noise interference and high data complexity; the GAN-based method performs excellently in image generation ability, but the instability of its training process limits its promotion in practical application. SUMMARY

[0013] To solve the problem of excessive dependence on labeled data in constructing large-scale and high-quality medical image segmentation data sets in existing systems, a medical image segmentation system based on double-scale consistency is provided, which can effectively utilize unlabeled samples to improve the performance of the segmentation network.

[0014] To achieve the above purpose, the system scheme adopted is as follows: the medical image segmentation system based on double-scale consistency, the specific structure is as follows:

[0015] A dual-scale consistency semi-supervised learning framework is constructed in the segmentation system. The dual-scale consistency semi-supervised framework consists of two pairs of teacher-student networks. First, the dual scale is reflected in the fact that the segmentation system processes medical images of two different scales at the same time: the medical images of the first scale are labeled medical image pairs of the original scale (i.e., each medical image to be segmented and its corresponding segmentation label image) and unlabeled medical images (containing only medical images to be segmented and no segmentation label images); the medical images of the second scale are labeled medical image pairs of a specific scale (i.e., each medical image to be segmented and its corresponding segmentation label image) and unlabeled medical images (containing only medical images to be segmented and no segmentation label images). The specific-scale medical image is obtained by bilinear upsampling of the original-scale medical image and its perturbation. Taking breast ultrasound images as an example, the medical image of the original scale is upsampled to the original scale. Input the first teacher network T1 and process it to obtain the perturbed medical image , This is the input of the first student network S1. and After bilinear interpolation, the corresponding upsampled medical image pair of a specific size can be obtained. and , and is the input of the second pair of teacher-student networks; secondly, the semi-supervised learning framework is reflected in the use of teacher-student models for training, and the training process includes supervised training and unsupervised training. First, supervised training is used for labeled data, and the perturbation image of the original scale is and the corresponding specific size upsampling They are fed into the student network S1 and the student network S2 respectively, and the outputs of both are constrained by supervised loss. The constraint objects of supervised loss are the prediction results of the two student networks and the segmentation labels (original size segmentation labels and its upsampled segmentation labels ), the parameters of the teacher network are updated by the exponential moving average of the student network parameters, and the calculation process is as follows:

[0016] (1)

[0017] in, Represents the parameters of the student model at the current iteration number, is the teacher model parameter of the previous iteration, is the learning rate, Represents the parameters of the teacher model at the current iteration. The supervised learning loss function is defined as follows:

[0018] (2)

[0019] Where L is the cross entropy loss function and A combination of loss functions, is the number of images in a batch. represent student networks S1 and S2 respectively, Represents the original size segmentation label and its upsampled segmentation label respectively Secondly, unsupervised training is used for unlabeled data. Unlabeled medical images of two scales (each medical image to be segmented has no corresponding segmentation label) and their perturbation versions are input into two pairs of teacher-student networks respectively. The teacher network TI and the student network S1 input the images of the original scale, and the teacher network T2 and the student network S2 input the upsampled images of a specific scale. The training process is constrained by unsupervised loss. In particular, this system proposes a training strategy with joint scale consistency constraints. The joint scale consistency loss includes the consistency loss before and after the perturbation. and scale consistency loss , which aims to enhance the feature representation ability under different scale inputs and thus improve the segmentation performance of the model. The constraints of the consistency loss before and after perturbation are the prediction result map output by the student network S1 (or S2) and the pseudo label map output by the teacher network T1 (or T2); the constraints of the scale consistency loss are the prediction results output by the two student networks S1 and S2.

[0020] Specifically, the reliable pseudo-label generation process works as follows: Softmax operation is performed on the output of the teacher network TI and T2 to obtain the output features, and the uncertainty graphs of the two teacher networks are calculated respectively. and , a more reliable prediction category can be obtained. The input of fusion module 1 includes the probability distribution map of TI output And T2 output probability distribution map The downsampling diagram of the fusion module 2 includes the probability distribution diagram of the T2 output. And T1 output probability distribution map The upsampling graph of , the probability graph fusion process can be expressed as:

[0021] U (3)

[0022] (4)

[0023] (5)

[0024] (6)

[0025] (7)

[0026] in, is a very small constant to avoid singularities, is the probability that pixel i belongs to the cth category, U is the obtained uncertainty map. I is a function that determines the size, is the uncertainty map obtained from the first teacher network T1, is the uncertainty map obtained from the second teacher network. Less than If true, the judgment function returns 1, otherwise it returns 0; represents the more reliable pseudo-label obtained by fusion module 1, represents the probability distribution of the first decoder output, represents the probability distribution of the second decoder output, represents the downsampling function, for Less than The mask image, represents the more reliable pseudo-label obtained by fusion module 2, represents the upsampling function, for Less than The mask image.

[0027] The unsupervised loss function consists of two parts, namely the consistency loss before and after the perturbation and scale consistency loss , the specific unsupervised learning loss function is defined as follows:

[0028] (8)

[0029] The consistency loss before and after perturbation is defined as follows:

[0030] (9)

[0031] in, represent teacher networks T1 and T2 respectively, is the pseudo label output by fusion module 1, is the pseudo label output by fusion module 2;

[0032] The scale consistency loss is specifically defined as follows:

[0033] (10)

[0034] in, is the predicted segmentation result of the student network S1, is the predicted segmentation result of the student network S2, the total loss of training Loss due to supervision and unsupervised loss It consists of two parts:

[0035] (11)

[0036] (12)

[0037] in, is the regularization parameter, is the maximum value of the weighted function, is the current iteration number, is the maximum number of iterations;

[0038] This paper innovatively proposes a simple and efficient semi-supervised medical image segmentation method based on dual-scale consistency. By exploring appropriate training mechanisms and loss function constraints, the method improves the network's segmentation performance for medical targets (e.g., breast tumors). Firstly, a semi-supervised learning mechanism is employed, with supervised training performed on a small amount of paired data and unsupervised training performed on a large amount of unpaired data. This design enables the segmentation model to operate even with severely limited labeled datasets. Secondly, the backbone network is designed as a dual-scale consistency teacher-student network. By extracting and interacting features from medical images at two different scales (e.g., breast ultrasound images), the network achieves accurate segmentation of medical targets (e.g., breast tumors). Furthermore, the network is trained using both supervised and unsupervised losses. Specifically, the unsupervised loss consists of a joint scale consistency loss and a reliable pseudo-label loss. The joint scale consistency loss constrains the perturbed image, deriving more reliable pseudo-labels from the predicted probability distributions of the two teacher networks. The reliable pseudo-label loss ensures smooth unsupervised training. Experimental results demonstrate that the proposed model overcomes the strong dependence of existing segmentation models on labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the network framework for semi-supervised medical image segmentation based on dual-scale consistency (taking breast ultrasound images as an example).

[0040] Figure 2 Schematic diagram of the strategy for obtaining more reliable pseudo-labels for fusion module 1.

[0041] Figure 3 Schematic diagram of the strategy for obtaining more reliable pseudo-labels for fusion module 2.

[0042] Figure 4 The ablation experiment results of this system and the manually annotated mask label map are shown when the amount of labeled data accounts for 5% of the total data.

[0043] Figure 5 The ablation experiment results of this system and the manually annotated mask label map are shown when the amount of annotated data accounts for 10% of the total data.

[0044] Figure 6 The figure shows the qualitative experimental results of this system and several existing medical image segmentation methods when the amount of labeled data accounts for 50% of the total data.

[0045] Figure 7 The following is a diagram showing the qualitative experimental results of several existing medical image segmentation methods in this system when the amount of labeled data accounts for 70% of the total data.

[0046] Figure 8 The following is a diagram showing the qualitative experimental results of several existing medical image segmentation methods in this system when the amount of labeled data accounts for 90% of the total data volume.

[0047] exist Figure 6 、 Figure 7 、 Figure 8 In the figure, the green line represents the manually annotated breast tumor outline, and the red line represents the breast tumor outline predicted by the model. DETAILED DESCRIPTION

[0048] In order to make the system problems, system solutions and beneficial effects to be solved more clearly understood, the following is a further detailed description of the system with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explanation and not for limitation.

[0049] The medical image segmentation system based on dual-scale consistency was experimented with a public dataset. During the experiment, the images in the BUSI dataset, UDIAT dataset, and RODTOOK dataset were merged to construct a dataset containing 977 breast ultrasound images. On this basis, the dataset was divided into 80% of the data for training and 20% of the data for testing. For the training data, different proportions of data were selected as labeled samples as needed, and the remaining data were used as unlabeled samples to explore the impact of different labeling ratios on the performance of breast ultrasound image segmentation. During the training process, Adam with 0.5 momentum and 0.0001 weight decay was used. The optimizer was trained for 150 epochs. All experiments were performed on a 24GB NVIDIA RTX3090 GPU using the PyTorch platform. The resolution of all images was uniformly adjusted to 256. 256 pixel resolution. For fair comparison, all models use the same batch size of 4 and the same evaluation method during testing. The initial learning rate is . Use data augmentation during training, including random horizontal and vertical flips, rotations, random crops, etc.

[0050] The Dice coefficient is used to evaluate the accuracy of the model in the segmentation task. The accuracy of segmentation is measured by comparing the overlap between the predicted segmentation result and the true label. The calculation process is shown in formula (1):

[0051] (1)

[0052] Where FP, FN, and TP represent false positive, false negative, and true positive, respectively. The value of the Dice coefficient ranges from 0 to 1. When the value of the Dice coefficient is closer to 1, it means that the segmentation result of the model is better, i.e., the overlap between the predicted result and the true label region is higher. When the value of the Dice coefficient is closer to 0, it means that the segmentation result of the model is worse.

[0053] The Jaccard index is used to evaluate the accuracy of the model in the segmentation task. The value of the Jaccard index ranges from 0 to 1. When the value of the Jaccard index is closer to 1, the segmentation result of the model is better. When the value of the Jaccard index is closer to 0, the segmentation result of the model is worse. The specific calculation formula is as follows:

[0054] (2)

[0055] The Hausdorff distance (HD) is used to evaluate the accuracy of the model in the segmentation task. HD95 is the result of multiplying the Hausdorff distance value between the predicted boundary and the true boundary by 95%. The smaller the value of HD95, the smaller the difference between the segmentation result and the true label, and the better the segmentation effect. The calculation process of the Hausdorff distance is represented by formula (3):

[0056] (3)

[0057] Where H is the Hausdorff distance, A and B represent the sets of points on the boundaries of the segmentation mask of the true label and the segmentation mask predicted by the model, respectively, a and b represent the points in sets A and B, respectively, and d is the Euclidean distance.

[0058] The precision is used to evaluate the accuracy of the model in the segmentation task. The precision represents the proportion of true positive samples in all samples predicted as positive by the model. The calculation formula is as follows:

[0059] (4)

[0060] Accuracy (ACC) is used to evaluate the accuracy of the model in the segmentation task. It is used to evaluate the proportion of pixels that are correctly classified in the prediction results of the segmentation model. The specific calculation formula is as follows:

[0061] (5)

[0062] TP (True Positives) represents the number of pixels correctly predicted as positive; TN (True Negatives) represents the number of pixels correctly predicted as negative; FP (False Positives) represents the number of pixels incorrectly predicted as positive; and FN (False Negatives) represents the number of pixels incorrectly predicted as negative. The accuracy ranges from 0 to 1. Values ​​closer to 1 indicate better overall model performance, while values ​​closer to 0 indicate poorer performance.

[0063] In medical image segmentation, sensitivity (SE) is also an important indicator for evaluating model performance. It is used to measure the performance of the model on positive samples. Its calculation process is:

[0064] (6)

[0065] Figure 1 The figure shows the medical image segmentation system (taking breast ultrasound images as an example) proposed by the present invention, which is a dual-scale consistency semi-supervised learning framework, consisting of two pairs of teacher-student networks. First, the dual scale is reflected in the fact that the segmentation system processes medical images of two different scales at the same time, pairs of labeled medical images of the original scale (i.e., each medical image to be segmented and its corresponding segmentation label image) and unlabeled medical images (containing only the medical image to be segmented, no segmentation label image), and pairs of labeled medical images of specific scale (i.e., each medical image to be segmented and its corresponding segmentation label image) and unlabeled medical images (containing only the medical image to be segmented, no segmentation label image). The specific scale medical image is obtained by bilinear upsampling operation on the original scale medical image and its perturbation. Taking breast ultrasound images as an example, the medical image of the original scale is upsampled. Input the first teacher network T1 and process it to obtain the perturbed medical image , This is the input of the first student network S1. and After bilinear interpolation, the corresponding upsampled medical image pair of a specific size can be obtained. and , and is the input of the second pair of teacher-student networks; secondly, the semi-supervised learning framework is reflected in the fact that the model training process includes supervised training and unsupervised training. First, supervised training is used for labeled data, and the perturbation image of the original scale is used. and the corresponding specific size upsampling They are fed into the student network S1 and the student network S2 respectively, and the outputs of both are constrained by supervised loss. The constraint objects of supervised loss are the prediction results of the two student networks and the segmentation labels (original size segmentation labels and its upsampled segmentation labels ), the parameters of the teacher network are updated by the exponential moving average of the student network parameters, and the calculation process is as follows:

[0066] (7)

[0067] in, Represents the parameters of the student model at the current iteration number, is the teacher model parameter of the previous iteration, is the learning rate, Represents the parameters of the teacher model at the current iteration. The supervised learning loss function is defined as follows:

[0068] (8)

[0069] Where L is the cross entropy loss function and A combination of loss functions, is the number of images in a batch. represent student networks S1 and S2 respectively, Represents the original size segmentation label and its upsampled segmentation label respectively Secondly, unsupervised training is used for unlabeled data. Unlabeled medical images of two scales (each medical image to be segmented has no corresponding segmentation label) and their perturbation versions are input into two pairs of teacher-student networks respectively. The teacher network TI and the student network S1 input the images of the original scale, and the teacher network T2 and the student network S2 input the upsampled images of a specific scale. The training process is constrained by unsupervised loss. In particular, this system proposes a training strategy with joint scale consistency constraints. The joint scale consistency loss includes the consistency loss before and after the perturbation. and scale consistency loss , which aims to enhance the feature representation ability under different scale inputs and thus improve the segmentation performance of the model. The constraints of the consistency loss before and after perturbation are the prediction result map output by the student network S1 (or S2) and the pseudo label map output by the teacher network T1 (or T2); the constraints of the scale consistency loss are the prediction results output by the two student networks S1 and S2.

[0070] Specifically, the reliable pseudo-label generation process works as follows: Softmax operation is performed on the output of the teacher network TI and T2 to obtain the output features, and the uncertainty graphs of the two teacher networks are calculated respectively. and , a more reliable prediction category can be obtained. Figure 2 As shown, the input of fusion module 1 includes the probability distribution map of TI output And T2 output probability distribution map The downsampled graph of . Figure 3 As shown, the input of fusion module 2 includes the probability distribution map of T2 output And T1 output probability distribution map The upsampling graph of , the probability graph fusion process can be expressed as:

[0071] U (9)

[0072] (10)

[0073] (11)

[0074] (12)

[0075] (13)

[0076] in, is a very small constant to avoid singularities, is the probability that pixel i belongs to the cth category, and U is the obtained uncertainty map. I is a function that determines the size. is the uncertainty map obtained from the first teacher network T1, is the uncertainty map obtained from the second teacher network. Less than If true, the judgment function returns 1, otherwise it returns 0; represents the more reliable pseudo-label obtained by fusion module 1, represents the probability distribution of the first decoder output, represents the probability distribution of the second decoder output, represents the downsampling function, for Less than The mask image of represents the more reliable pseudo-label obtained by fusion module 2, represents the upsampling function, for Less than The mask image.

[0077] The unsupervised loss function consists of two parts, i.e., consistency loss before and after perturbation and scale consistency loss The specific unsupervised learning loss function is defined as follows:

[0078] (14)

[0079] The consistency loss before and after perturbation is defined as follows:

[0080] (15)

[0081] wherein, T1 and T2 represent the teacher network, is the pseudo label output by the fusion module 1, is the pseudo label output by the fusion module 2;

[0082] The scale consistency loss is specifically defined as follows:

[0083] (16)

[0084] wherein, is the predicted segmentation result of the student network S1, is the predicted segmentation result of the student network S2. The total loss of training consists of supervised loss and unsupervised loss Two parts:

[0085] (17)

[0086] (18)

[0087] wherein, is a regularization parameter, is the maximum value of the weighting function, is the current iteration number, is the maximum iteration number.

[0088] Figure 4 , Figure 5Table 1 and Table 1 show the ablation experiment results. To obtain unlabeled image data, labeled data was selected from the dataset at ratios of 5% and 10%. For convenience, Model1 represents the baseline model (i.e., a single teacher-student network); Model2 represents the baseline model with a teacher-student network added and using both scale consistency loss and perturbation consistency loss; Model3 represents the baseline model with a teacher-student network added and using only perturbation consistency loss; and Ours represents the framework proposed in this paper. In the table, Label represents the amount of labeled data, and Unlabel represents the amount of unlabeled data. Table 1 shows the experimental results for cases where labeled data accounts for 5% and 10% of the total training data. When labeled data accounts for 5% of the total data, the Dice index for Model1 is 74.95%, for Model2 and Model3 is 75.24% and 75.52%, respectively, and for our system (Ours) is 75.70%. It can be seen that the combined learning of the dual-scale consistency loss and the more reliable pseudo-label generation mechanism achieves better experimental results than learning each method separately. When the labeled data accounts for 10% of the total data volume, the experimental results obtained by this system achieve the maximum value in the four indicators of Dice, IOU, ACC, and HD95. Figure 4 and The segmentation results of different ablation networks are shown when the labeled data accounts for 5% and 10% of the total dataset. When the amount of labeled data is limited, Model 1's segmentation performance for larger lesion areas is not ideal, indicating its high dependence on data. With the introduction of the dual-scale consistency strategy and a more reliable pseudo-labeling mechanism, the model's prediction results for larger lesion areas have been significantly improved. When the amount of labeled data increases, Model 1's segmentation accuracy for larger samples in the lesion area has been significantly improved. After introducing the dual-scale consistency loss and a more reliable pseudo-labeling mechanism, the segmentation results predicted by the model are closer to the segmentation labels.

[0089]

[0090] Table 2 shows the segmentation results obtained using different medical image segmentation algorithms on a breast tumor dataset. The proposed method is comprehensively compared with six advanced medical image segmentation methods, including two semi-supervised methods and four fully supervised methods. The semi-supervised methods include the uncertainty-rectified pyramid consistency model (SSL4MIS) and the uncertainty-guided collaborative mean-teacher model (UCMT). The fully supervised methods include UNet, CENet, CPFNet, and MSCANet. The performance of the different segmentation models on the test set was compared when the amount of labeled data accounted for 5%, 10%, and 15% of the total data. The results show that when the amount of labeled data is only 5%, the fully supervised learning method's lack of ability to utilize unlabeled data leads to poor segmentation results. Compared with the fully supervised method, the semi-supervised learning method can effectively utilize unlabeled data, significantly improving the model's segmentation performance. The proposed system achieved the best results across all evaluation metrics, with a Dice coefficient of 75.70% and a Hausdorff distance of 28.40. This result demonstrates the particularly strong advantages of semi-supervised methods when very little labeled data is available. When the amount of labeled data increases to 10%, the segmentation performance indicators of all models improve. The experimental results of this system maintain their leading position across all indicators, with the Dice coefficient increasing by 1.94% to 77.64% and the Hausdorff distance decreasing by 1.52 to 26.88. This result further validates the positive impact of increasing the amount of labeled data on model performance. When the proportion of labeled data further increases to 15%, the Dice coefficients of all segmentation models exceed 60%. At the same time, this system continues to perform well across all indicators, with the Dice coefficient increasing by 0.31% to 77.95%, and the Hausdorff distance increasing slightly to 27.79. This phenomenon demonstrates that increasing the amount of labeled data enhances the stability and robustness of the model.

[0091]

[0092] to The visual segmentation results of this system and six existing medical image segmentation models are shown for 50%, 70%, and 90% of the labeled data. These three visual segmentation results show that malignant tumors have complex morphology, making it difficult for most segmentation models to roughly segment the tumor region. However, this system and UCMT can relatively accurately segment the general outline of the tumor region. However, at 50% of the labeled data ratio, UCMT has a large number of over-segmented regions for the benign tumors in the first column. In comparison, this method achieves more accurate segmentation results for this sample. Qualitative comparisons of the results with different amounts of labeled data clearly demonstrate the feasibility and efficiency of this method for breast tumor segmentation.

Claims

1. A medical image segmentation system based on dual-scale consistency, characterized by: The specific structure is as follows: A dual-scale consistency semi-supervised learning framework is constructed in the segmentation system. The dual-scale consistency semi-supervised learning framework consists of two pairs of teacher-student networks. First, the dual-scale is reflected in the fact that the segmentation system processes medical images of two different scales simultaneously. Second, the semi-supervised learning framework is reflected in the use of teacher-student models for training, and the training process includes supervised training and unsupervised training. The segmentation system simultaneously processes medical images of two different scales. The medical images of the first scale are pairs of labeled medical images and unlabeled medical images at the original scale. The medical images of the second scale are pairs of labeled medical images and unlabeled medical images at a specific scale. The medical images of the second scale are obtained by bilinear upsampling of the original scale medical images and their perturbations. The original scale medical image x u Input the first teacher network T1 and process it to obtain the perturbed medical image x p , x p That is the input of the first student network S1, for x u and x p After performing bilinear interpolation, obtain the upsampled medical image pair corresponding to a specific size. and , and is the input of the second teacher-student network; The total loss L during training total By the supervision loss L sup and the unsupervised loss L un It consists of two parts, where the unsupervised loss L un Including the consistency loss L before and after the perturbation p and scale consistency loss L s The constraints of the consistency loss before and after the perturbation are the prediction result graph output by the student network S1 or S2 and the pseudo label graph output by the teacher network T1 or T2; the constraints of the scale consistency loss are the prediction results output by the two student networks S1 and S2. A pseudo label generation mechanism is constructed, and the fusion module is used to obtain pseudo labels from the output probability distribution of the two teacher networks to constrain the outputs of the two student networks of different scales.

2. The medical image segmentation system based on dual-scale consistency according to claim 1, characterized in that: The semi-supervised learning framework performs supervised training on a small amount of labeled data and unsupervised training on a large amount of unlabeled data in model training; First, supervised training is used for labeled data, and the perturbation image x of the original scale is p and the corresponding specific size upsampling They are fed into student networks S1 and S2 respectively, and the outputs of both are constrained by supervised loss. The constraints of the supervised loss are the prediction results and segmentation labels of the two student networks. The parameters of the teacher network are updated by the exponential moving average of the student network parameters. The calculation process is as follows: （1） in, Represents the parameters of the student model at the current iteration number, is the teacher model parameter of the previous iteration, is the learning rate, Represents the parameters of the teacher model at the current iteration. The supervised learning loss function is defined as follows: （2） Where L is the cross entropy loss function and A combination of loss functions, is the number of images in a batch, represent student networks S1 and S2 respectively, Represents the original size segmentation label and its upsampled segmentation label respectively ; Unsupervised training is used for unlabeled data. Unlabeled medical images of two scales and their perturbed versions are input into two pairs of teacher-student networks respectively. The teacher network TI and the student network S1 are input with images of the original scale, and the teacher network T2 and the student network S2 are input with upsampled images of a specific scale. An unsupervised loss constraint training process is adopted.

3. The medical image segmentation system based on dual-scale consistency according to claim 1, characterized in that: Construct a pseudo-label generation mechanism and use the fusion module to obtain pseudo-labels from the output probability distribution of the two teacher networks to constrain the output of the two student networks of different scales; The reliable pseudo-label generation process works as follows: Softmax operation is performed on the output of the teacher network TI and T2 to obtain the output features, and the uncertainty graphs of the two teacher networks are calculated respectively. and , the input of fusion module 1 includes the probability distribution map of TI output And T2 output probability distribution map The downsampling diagram of the fusion module 2 includes the probability distribution diagram of the T2 output. And T1 output probability distribution map The upsampling graph of , the probability graph fusion process is expressed as: U （3） （4） （5） （6） （7） in, is a constant, is the probability that pixel i belongs to the cth category, U is the obtained uncertainty map, and I is a function to judge the size. is the uncertainty map obtained from the first teacher network T1, is the uncertainty map obtained from the second teacher network; when Less than If true, the judgment function returns 1, otherwise it returns 0; represents the pseudo label obtained by fusion module 1, represents the probability distribution of the first decoder output, represents the probability distribution of the second decoder output, represents the downsampling function, for Less than The mask image, represents the pseudo label obtained by fusion module 2, represents the upsampling function, for Less than The mask image.

4. The medical image segmentation system based on dual-scale consistency according to claim 1, characterized in that: The unsupervised loss function consists of two parts, namely the consistency loss before and after the perturbation and scale consistency loss , the specific unsupervised learning loss function is defined as follows: （8） The consistency loss before and after perturbation is defined as follows: （9） in, represent teacher networks T1 and T2 respectively, is the pseudo label output by fusion module 1, is the pseudo label output by fusion module 2; The scale consistency loss is specifically defined as follows: （10） in, is the predicted segmentation result of the student network S1, is the predicted segmentation result of the student network S2; The total loss during training Loss due to supervision and unsupervised loss It consists of two parts: （11） （12） in, is the regularization parameter, is the maximum value of the weighted function, is the current iteration number, is the maximum number of iterations.

Images