A Semi-Supervised MRI Brain Tumor Segmentation Method Based on Contrast-Guided Diffusion Model

Abstract

The present invention proposes a semi-supervised MRI brain tumor segmentation method based on a contrast-guided diffusion model, belonging to the field of image segmentation. In step S1, the T1 sequence data of the BraST2018 dataset is sliced, and the slice data without tumors is partially occluded; in step S2, pairs of lesion and healthy image data are generated, and the minimum bounding rectangle occlusion is constructed using the brain tumor sample mask, and the lesion area is restored to healthy tissue; in step S3, through the data pairs obtained in S2, the contrast information is used to guide the diffusion model to denoise and generate lesion labels. Thereafter, pre-training is performed with labeled data, and pseudo-labels are generated for unlabeled data through the pre-trained model, and then brought into the model together with the labeled data for the re-training process; in step S4, the structural contrast loss is used to improve the information mining ability of the model when the confidence of the pseudo-labels is insufficient; in step S5, only a small amount of labeled data is used for training and validation. The present invention realizes the lesion segmentation of MRI brain tumors by constructing a semi-supervised segmentation method based on a contrast-guided diffusion model, on the premise of only requiring a small amount of labeled data. This method solves the problem that traditional deep learning segmentation methods overly rely on a large amount of labeled data and improves the segmentation performance of the model under the condition of a small amount of labeled data.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and designs a semi-supervised MRI brain tumor segmentation method based on a contrast-guided diffusion model, belonging to the field of image segmentation. Background Art

[0002] Brain tumor research is of great significance at the clinical and social levels. According to the latest statistical data of The Lancet, in 2020, the number of new brain tumor cases in China was approximately 106,000, and the number of deaths was approximately 59,000, with both the incidence and mortality rates ranking first in the world. Brain tumors are a highly harmful disease, posing a great threat to human health. Especially malignant brain tumors, such as glioblastoma (GBM), due to their complexity and high invasiveness, the average survival period of patients is only 1 year, and the five-year survival rate is less than 5%. Brain tumors are characterized by high incidence, high mortality, and high recurrence rate. Therefore, in-depth research on the diagnosis and treatment methods of brain tumors is of great significance for reducing mortality and improving the quality of life of patients.

[0003] In the application of automatic brain tumor segmentation, deep learning has significantly improved the efficiency and accuracy of research. Deep learning, especially convolutional neural networks (CNNs), has become an important tool for medical image segmentation due to its insensitivity to image noise and contrast changes. Automatic brain tumor segmentation through deep learning can significantly improve the accuracy and efficiency of diagnosis, reducing the time and error of manual operations. Classic deep learning models such as AlexNet, SegNet, and ResNet have been widely used in computer vision tasks and demonstrated great potential in the field of medical images. Medical lesion image segmentation plays a crucial role in medical image analysis. Its main goal is to accurately locate and segment the lesion area in the image, thereby providing auxiliary support for clinical diagnosis, treatment plan formulation, and efficacy evaluation. Efficient lesion segmentation can significantly improve the accuracy and efficiency of diagnosis, reducing the time and error of manual operations.

[0004] However, in the field of brain tumor image segmentation, obtaining fully annotated medical image data is a highly challenging task. The annotation process of medical images is not only costly but also requires annotators to have profound medical expertise. This process usually relies on the subjective judgment of experts, so the knowledge differences among evaluators may affect the reliability and consistency of the annotation results. In addition, the performance of deep learning algorithms depends to a large extent on large-scale high-quality data for training, and the high cost and complexity of medical image annotation limit the acquisition of large-scale, high-quality annotated datasets, which poses a severe challenge to deep learning-based brain tumor image segmentation algorithms.

[0005] Semi-supervised learning provides a feasible direction for solving the problem of high data annotation costs. This method is a way to train deep learning models under the condition of a small number of labeled samples and a large amount of unlabeled data. This method can effectively utilize the rich information of unlabeled data, thereby improving the performance and generalization ability of the model when the labeled data is limited. Among them, semi-supervised learning with pseudo-labels is one of the commonly used methods in this field. This method first uses a segmentation model to predict unlabeled images and generate pseudo-labels, and then uses these pseudo-labels as new examples for further model training. However, the output of the segmentation model trained with limited labeled data often contains noise. If these noisy outputs are directly used as pseudo-labels for subsequent training, it may lead to instability in the training process and even damage the performance of the model. To address this issue, Yao et al. proposed a confidence-aware cross-pseudo-supervised network to improve the quality of pseudo-labels for unlabeled images from unknown distributions. Wang et al. re-evaluated the pseudo-labels in the model output by introducing a trust module and set a threshold to select high-confidence pseudo-labels. These methods effectively improve the stability and performance of the model in semi-supervised learning by enhancing the reliability of pseudo-labels.

[0006] Given these limitations, the importance of semi-supervised MRI brain tumor segmentation technology has become increasingly prominent. This technology can automatically segment brain tumor lesions in the absence of sufficient labeled brain tumor samples. It not only reduces the need and cost of data annotation, but also improves the efficiency of medical research and diagnosis. Summary of the Invention

[0007] Aiming at the defects in the prior art, the present invention provides a semi-supervised MRI brain tumor segmentation method based on a contrast-guided diffusion model. This method can train an MRI brain tumor segmentation model using only a small amount of labeled data to complete automatic segmentation. The specific solution of the present invention is as follows:

[0008] A semi-supervised MRI brain tumor segmentation method based on a contrast-guided diffusion model includes the following steps:

[0009] S1: Data acquisition and preprocessing of MRI brain tumor samples;

[0010] S2: Construct a healthy tissue generation model corresponding to the brain tumor samples;

[0011] S3: Construct an MRI brain tumor semi-supervised segmentation framework based on a contrast-guided diffusion model;

[0012] S4: Construct a new structural similarity contrast loss function;

[0013] S5: Train and validate the semi-supervised MRI brain tumor segmentation framework based on the contrast-guided diffusion model.

[0014] Furthermore, in the experiment of step S1, the BraTS2018 dataset, The Multimodal Brain Tumor Segmentation Challenge 2018) dataset was used. This dataset provides multi-modal MRI images (T1, T1c, T2, FLAIR) and professionally annotated tumor regions (enhanced tumor, tumor core, whole tumor region), covering three-dimensional data of 285 patients, and preprocess this dataset.

[0015] Furthermore, step S2 constructs a healthy tissue generation model corresponding to the brain tumor sample, specifically:

[0016] S21: In the brain tumor sample, select the slices without lesions as the training samples of the healthy tissue generation model, and randomly perform mask occlusion on the brain tissue parts of these samples as the condition of the diffusion model, with the aim of restoring the healthy tissue in the occluded area;

[0017] S22: According to the model trained in S21, in the generation stage, construct the minimum bounding rectangle mask for occlusion according to the mask area of the brain tumor sample as the input of the condition generation, bring it into the model trained in S21, restore its lesion area to healthy tissue, and obtain the paired lesion image and healthy image data pair.

[0018] Furthermore, in step S3, construct an MRI brain tumor semi-supervised segmentation framework based on the contrast-guided diffusion model. This model uses the mask label of the labeled data as the initial input image x0, and transforms the noise image x t |x t-1 ) through the forward diffusion process q(x T . In the reverse denoising process, two conditions are introduced: the paired lesion image c1 and its corresponding healthy image c2 data pair obtained in step S2. Therefore, the reverse denoising process is p θ (x t-1 |x t , (c1, c2)). In each step of the denoising process, the lesion image and the corresponding healthy image are connected to the input of the denoising network channel by channel. Use U-net as the denoising network, and add a spatial attention mechanism (Spartial Attention Module, SAM) in the downsampling and upsampling respectively, so that the network can adaptively focus on the lesion area in the image globally, thereby enhancing the feature expression ability. Furthermore, in step S4, construct a new structural similarity contrast loss function, specifically as follows:

[0019] Adopt a contrastive learning algorithm based on positive and negative sample pairs. By learning the differences in lesion regions between paired lesion images and healthy images, as well as the similarities in lesion regions between unpaired lesion images, the algorithm performance is further optimized. On the one hand, by maximizing the differences in lesion regions between positive sample pairs (i.e., paired lesion images and healthy images) and minimizing the similarities in lesion regions between negative sample pairs (i.e., unpaired lesion images), the algorithm's ability to handle the uncertainty of the supervised loss can be enhanced in the absence of precisely labeled data. On the other hand, minimizing the similarities in lesion regions of negative sample pairs can better utilize global information, thereby improving the overall performance of the algorithm. Structural similarity has significant advantages in measuring local image similarity. By comprehensively considering the similarities in brightness, contrast, and structure, it can more accurately reflect the perception of image quality by the human visual system. Therefore, structural similarity is used as the similarity metric function for this method.

[0020] Furthermore, in step S5, the semi-supervised MRI brain tumor segmentation framework based on the contrast-guided diffusion model is trained and verified as follows:

[0021] The semi-supervised MRI brain tumor segmentation framework based on the contrast-guided diffusion model is trained and verified. The T1 modality of the BraTS2018 dataset is used as the dataset, and the model is trained and verified with only a small amount of labeled data. Description of the Drawings

[0022] Figure 1 Schematic diagram of the healthy tissue generation model based on the diffusion model

[0023] Figure 2 Schematic diagram of the semi-supervised MRI brain tumor segmentation framework based on the contrast-guided diffusion model

[0024] Figure 3 Schematic diagram of the segmentation result Detailed Implementation Manner

[0025] S1: Data acquisition and preprocessing of MRI brain tumor samples;

[0026] S11. Data acquisition stage:

[0027] In the experiments of the present invention, the BraTS2018 dataset was used as the basis of the experimental dataset. This dataset provides multi-modal MRI images (T1, T1c, T2, FLAIR) and professionally labeled tumor regions (enhanced tumor, tumor core, whole tumor region). This study mainly focuses on using the T1 modality image data and the whole tumor region, where the size of each image is 240×240×155.

[0028] S12. Data preprocessing stage:

[0029] When preprocessing the data set, since the construction of the healthy tissue generation model corresponding to the brain tumor sample in step S2 requires healthy tissue images as input, and since the brain tumor data set is three-dimensional data, some slices on the depth axis do not contain tumor lesions, or only contain extremely small tumor lesions. Therefore, the brain tumor data is sliced on the depth axis and divided into two parts, namely the tumor-containing slice data set and the tumor-free slice data set, and both are uniformly cropped to 160×160. Random-sized rectangular masks are generated on the tumor-free slice data images to occlude the tissue, which is used as the conditional input for the generation model in step S2.

[0030] S2. Construct a healthy tissue generation model corresponding to the brain tumor sample;

[0031] S21. During the training process, the diffusion model is divided into a forward diffusion process and a reverse denoising process. In the forward diffusion process, the data is gradually transformed into random noise, expressed as the conditional Gaussian distribution q(x t |x t-1 ). In the reverse denoising process, the neural network is trained to recover the original data through the reverse denoising process p θ (x t-1 |x t ), where θ is the parameter of the reverse process. The data starts from the Gaussian noise sample x T ~p(x t ) and is generated by iterative sampling through p θ (x t-1 |x t ) until the image x0 is restored, where t = T - 1, …, 0. The forward diffusion process can be expressed as where α t = 1 - β t is the variance of the given preset noise β t , and

[0032] Use the tumor-free slice data image as the input of the original image x0, and use the randomly occluded image in S1 as the condition c and add it to the denoising process of each step. For example, Figure 1 , the forward diffusion process q(x t |x t-1 ) remains unchanged, but the reverse denoising process becomes p θ (x t-1 |x t , c), and at the same time, the log-likelihood estimation also becomes the conditional log p(x0|c).

[0033] In a generative model, the log-likelihood is usually not easy to calculate directly. We can maximize the ELBO (Evidence Lower Bound) instead of maximizing the likelihood function to optimize the latent variable model. The expression of ELBO is as follows:

[0034]

[0035] ELBO can be approximately optimized by training a neural network ∈ θ (x t , t) to predict the noise ∈ added to each data point x0 at each time step t. The loss function can be written as:

[0036]

[0037] S22. According to the model trained in S21, in the generation stage, the minimum bounding rectangle mask is constructed based on the mask region of the brain tumor sample for occlusion, which is used as the input for conditional generation and brought into the model trained in S21 to restore the lesion area to healthy tissue, obtaining a paired lesion image and healthy image data pair.

[0038] S3. Construct an MRI brain tumor semi-supervised segmentation framework based on the contrast-guided diffusion model;

[0039] The semi-supervised medical lesion image segmentation framework of the contrast-guided diffusion model is as Figure 2 shown, consisting of two stages. (1) Step S22 generation stage: The diffusion model is used to process the lesion regions in the labeled and unlabeled data. The occlusion strategy of the minimum bounding rectangle is calculated through the labels and pseudo-labels to obtain a paired lesion image and healthy image data pair. These data are then used to assist in the training of the segmentation model to improve the model's recognition ability at the boundary between healthy and diseased tissues. This method specifically solves the problem of inaccurate classification of pseudo-labels at the lesion edge and effectively utilizes the contrast information between healthy and diseased tissues through the contrast-guided strategy in a limited labeled data environment, enhancing the segmentation accuracy. (2) Segmentation stage: We establish a segmentation model based on the diffusion model, using the mask label of the labeled data as the initial input image x0, and transforming the noise image x t |x t-1 ) through the forward diffusion process q(x T . In the reverse denoising process, two conditions are introduced: the paired lesion image c1 and its corresponding healthy image c2 data pair obtained in step S22. Therefore, the reverse denoising process is p θ (x t-1 |x t, (c1, c2)), in each denoising step, the lesion image and the corresponding healthy image are concatenated channel by channel into the input of the denoising network. Using U-net as the denoising network, and a spatial attention mechanism (Spartial Attention Module, SAM) is added during downsampling and upsampling respectively, enabling the network to adaptively focus on the lesion area in the image globally, thus enhancing the feature expression ability. The contrast information is used to guide the diffusion model for denoising to generate lesion labels. After pre-training with labeled data, pseudo-labels are generated for unlabeled data through the pre-trained model, and then they are brought into the model together with the labeled data for the retraining process.

[0040] S4. Construct a new structural similarity contrast loss function;

[0041] Traditional supervised losses such as cross-entropy loss, Dice loss, etc. can effectively solve the problem of the difference between the segmentation result and the true label. However, in semi-supervised learning, in the face of a large number of unlabeled data, pseudo-labels usually cannot well represent the true label, thus affecting the accuracy of the model. Therefore, how to better utilize the contrast information between the lesion area and the corresponding healthy tissue area becomes a key issue. To solve this problem, the present invention proposes a new structural similarity contrast (SSC) loss function. Specifically, inspired by the contrast loss method, we adopt a contrast learning algorithm based on positive and negative sample pairs. By learning the differences in the lesion areas between paired lesion images and healthy images, as well as the similarities in the lesion areas between unpaired lesion images, the algorithm performance is further optimized. On the one hand, by maximizing the differences in the lesion areas between positive sample pairs (i.e., paired lesion images and healthy images) and minimizing the similarities in the lesion areas between negative sample pairs (i.e., unpaired lesion images), the ability of the algorithm to handle the uncertainty of the supervised loss can be enhanced in the absence of accurately labeled data. On the other hand, minimizing the similarities in the lesion areas of negative sample pairs can better utilize the global information, thus improving the overall performance of the algorithm.

[0042] However, current contrastive learning usually uses image embedding vectors and cosine similarity for representation learning, which is not applicable to segmentation tasks. First, using image embedding vectors for contrastive representation learning cannot directly learn the semantic information of segmentation targets. Second, cosine similarity focuses on the similarity of vector directions in high-dimensional space, ignoring the spatial distribution and local features of pixels in the image, and cannot fully capture the local structure and context information of the image. In contrast, structural similarity has significant advantages in measuring local similarity of images. By comprehensively considering the similarity in terms of brightness, contrast, and structure, SSIM can more accurately reflect the perception of image quality by the human visual system. Therefore, the present invention uses SSIM instead of cosine similarity as the similarity metric function.

[0043] Next, the proposed Structural Similarity Contrast (SSC) loss will be described in detail. When constructing positive and negative sample pairs, the lesion area is defined based on the segmentation prediction result map. The paired lesion image and healthy image are used as positive sample pairs, and the unpaired lesion images are used as negative sample pairs. Since the diffusion model has a high computational complexity and slow prediction result generation, in order to save the computational time cost, only a small batch of data is randomly selected for processing when calculating the contrast loss. Specifically, K paired samples are selected as the input of the contrast loss, where the lesion area of each paired sample is used as a positive sample pair, so there are K positive sample pairs, defined as the set K + , and the unpaired lesion areas are used as negative sample pairs, so there are K - 1 negative sample pairs, defined as the set K-, so the description of the positive and negative sample sets is as follows:

[0044]

[0045] where x′ i represents the lesion image, x″ i represents the corresponding healthy image, and M(·) represents the lesion area, that is, the mask area of the prediction result. Next, in order to measure the similarity of different regions, SSIM is used as the similarity metric function. However, since the contrast loss usually minimizes the positive sample distance and maximizes the negative sample distance, but in the present invention, it is necessary to maximize the similarity of the positive samples, so the calculation method of SSIM is simply modified as follows:

[0046]

[0047] where μ x and μ y are the means of images x and y, and represent variances, σ xy is the covariance, and C1 and C2 are stable constants to avoid the denominator being zero. The SSC loss for each positive pair is defined as follows:

[0048]

[0049] Among them is a positive sample pair in K+, and h is in K - and corresponding negative sample, but this is only the SSC loss calculation for one of the negative samples. Then the total SSC loss for K positive samples is as follows:

[0050]

[0051] S5. Train and validate the semi-supervised MRI brain tumor segmentation framework based on the contrast-guided diffusion model.

[0052] The present invention mainly uses the Dice similarity coefficient and the Hausdorff distance as evaluation indicators, which are specifically as follows:

[0053] (1) Dice

[0054] Dice is mainly used to evaluate the overlap degree between the segmentation result and the true label, and the range is [0, 1]. The closer this value is to 1, the better the segmentation cutting, and vice versa. Its algorithm is as follows:

[0055]

[0056] where ∩ is the logical AND operator, |·| is the area or size of the region, P1 is the region of the segmentation result predicted by the model, and T1 represents the true segmentation region.

[0057] (2) Hausdorff distance

[0058] The Hausdorff distance is used to measure the similarity or difference between two sets, and is widely used especially in shape comparison and evaluation of image segmentation results. It is defined as the farthest minimum distance from one set to another set, reflecting the distance between the farthest points in the two sets. The specific calculation formula is:

[0059]

[0060] where sup represents the maximum value in the set, inf represents the minimum value in the set, and ||a - b|| represents the Euclidean distance between point a and point b. The Hausdorff distance reflects the maximum dissimilarity between the two sets, and the smaller the distance, the more similar the sets.

[0061] In the T1 modality of the BraTS2018 dataset, with only a small amount of labeled data, perform training and validation as Figure 3 , and the results of its ablation experiment are as follows:

[0062] Table 1 Comparison of segmentation results of different methods added to the diffusion model

[0063]

[0064] As can be seen from the data in Table 1, when paired lesion images and healthy image data pairs are added as conditions to the model, the segmentation ability is significantly improved. Specifically, the Dice coefficient of the model increases from 0.5479 to 0.7612, and the Hausdorff distance decreases significantly from 19.51 to 12.83. This indicates that introducing the contrast guidance mechanism has a great promoting effect on the model performance. In addition, on this basis, further adding pseudo-labels for retraining, the Dice coefficient of the model is further improved to 0.8173, and the Hausdorff distance is reduced to 8.25, indicating that pseudo-label retraining effectively enhances the learning ability of the model. Finally, after combining the structural contrast loss, the Dice coefficient of the model reaches 0.8221, and the Hausdorff distance drops to 7.96, further improving the information mining ability of the model in the case of insufficient pseudo-label confidence. The gradual introduction of these improvement methods significantly improves the segmentation performance of the model, verifying the effectiveness of the semi-supervised MRI brain tumor segmentation method based on the contrast-guided diffusion model of the present invention.

Claims

1. A semi-supervised MRI brain tumor segmentation method based on a contrast-guided diffusion model, characterized in that, The method includes: S1: Data acquisition and preprocessing of MRI brain tumor samples; S2: Construct a healthy tissue generation model corresponding to the brain tumor sample; S3: Construct a semi-supervised segmentation framework for MRI brain tumors based on a contrast-guided diffusion model; S4: Construct a new structural similarity contrast loss function; S5: Train and validate the semi-supervised MRI brain tumor segmentation framework based on the contrast-guided diffusion model; S3. Construct a semi-supervised segmentation framework for MRI brain tumors based on a contrast-guided diffusion model, including: The semi-supervised medical lesion image segmentation framework based on the contrast-guided diffusion model consists of two stages: (1) step S22 generation stage: the diffusion model is used to process the lesion area in the labeled and unlabeled data, and the occlusion strategy of the minimum bounding rectangle is calculated through the label and pseudo-label to obtain the paired lesion image and healthy image data pairs; (2) segmentation stage: a segmentation model is established based on the diffusion model, and the mask label of the labeled data is used as the initial input image x0. t |x t-1 ) Transform the noisy image x T In the reverse denoising process, two conditions are introduced: the paired lesion image c1 and its corresponding healthy image c2 data obtained in step S22 are paired, so the reverse denoising process is p θ (x t-1 |x t ,(c1,c2)), in each denoising process, the lesion image and the corresponding healthy image are connected to the denoising network input channel by channel; U-net is used as the denoising network, and the spatial attention mechanism is added in downsampling and upsampling respectively, so that the network can adaptively focus on the lesion area in the image globally; the contrast information is used to guide the diffusion model denoising and generate lesion labels; after pre-training with labeled data, pseudo labels are generated for the unlabeled data through the pre-training model, and then brought into the model together with the labeled data for retraining; In step S4, a new structural similarity contrast loss function is constructed, using SSIM instead of cosine similarity as the similarity metric function; the contrast loss is usually calculated by minimizing the distance between positive samples and maximizing the distance between negative samples, specifically as follows: where μ x and μ y are the means of images x and y, and denote the variances, σ xy is the covariance, and C1 and C2 are stability constants to avoid a zero denominator.

2. The semi-supervised MRI brain tumor segmentation method based on a contrast-guided diffusion model according to claim 1, wherein: In step S1, the experimental dataset uses the BraTS2018 dataset, which provides multi-modal MRI images and professionally annotated tumor regions, covering three-dimensional data of 285 patients; preprocess this dataset.

3. A semi-supervised MRI brain tumor segmentation method based on a contrast-guided diffusion model according to claim 1, characterized in that: Step S2 constructs a healthy tissue generation model corresponding to the brain tumor sample, specifically: S21. In the brain tumor sample, select the slices without lesions as the training samples of the healthy tissue generation model, and randomly perform masking on the brain tissue parts of these samples as the condition for the diffusion model, with the aim of restoring the healthy tissue in the masked area; S22. According to the model trained in S21, in the generation stage, construct a minimum bounding rectangle mask for masking based on the masked area of the brain tumor sample, and use this as the input for conditional generation. Bring it into the model trained in S21 to restore the lesion area to healthy tissue, and obtain a paired comparison of lesion image and healthy image data.

4. A semi-supervised MRI brain tumor segmentation method based on a contrast-guided diffusion model according to claim 1, characterized in that: In step S5, the semi-supervised MRI brain tumor segmentation framework based on the contrast-guided diffusion model is trained and validated. The T1 modality of the BraTS2018 dataset is used as the dataset, and the model is trained and validated with only a small amount of labeled data.

Images