Brain MRI (Magnetic Resonance Imaging) image segmentation method combining Poisson denoising and Mamba architecture

Abstract

The invention discloses a brain MRI (Magnetic Resonance Imaging) image segmentation method combining Poisson denoising and Mamba architecture, which comprises the following steps of: 1, carrying out Poisson denoising on a brain MRI image to be segmented; step 2, constructing an image segmentation model MambaNet, wherein the image segmentation model MambaNet is used for segmenting the brain MRI image; step 3, the image segmentation model MambaNet is trained and optimized; and step 4, segmenting the brain MRI image after Poisson denoising in the step 1 by using the image segmentation model MambaNet trained in the step 2, and completing the brain MRI image segmentation combining Poisson denoising and Mamba architecture. The multi-level feature capturing capability of the model is improved, excessive smoothness and generation of artifacts are reduced, retention of cross-scale space information is ensured, and the segmentation performance of the method is superior to that of an existing method.

Description

Technical Field

[0001] The present invention relates to an MRI image segmentation method, in particular to a brain MRI image segmentation method combining Poisson denoising with Mamba architecture. Background Art

[0002] This section merely provides background information related to the present disclosure and is not necessarily prior art.

[0003] With the development of artificial intelligence, deep learning has been widely used in medical image denoising, segmentation and detection. Due to the influence of imaging sensors and the environment, medical images often have noises such as Poisson noise, Gaussian noise and impulse noise. The presence of these noises leads to false detection and missed detection in downstream segmentation and detection tasks, which in turn leads to low model accuracy. Therefore, denoising medical images before performing downstream segmentation tasks can ensure the accuracy of segmentation and detection. In the past medical image denoising methods, it is usually assumed that the noise is independent of the signal, and the noise is often modeled as zero-mean additive Gaussian noise. However, this assumption is not applicable to medical image systems because the sensor noise source is proportional to the signal strength, which can be modeled as a Poisson process. Therefore, modeling noise as additive Gaussian noise in medical images is physically impossible because the noise varies proportionally with the signal strength and is related to the signal.

[0004] Noise removal is one of the key steps to improve the performance of downstream segmentation models. Medical image segmentation models can assist doctors in diagnosing brain and nervous system diseases and improve their diagnostic efficiency. Most existing methods are based on convolutional neural networks and visual Transformer (ViT) (Chen, Zhengsu, et al. "Visformer: The vision-friendly transformer." Proceedings of the IEEE / CVF international conference on computer vision. 2021.), such as UNet (Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III18. Springer International Publishing, 2015.), UNet++ (Zhou, Zongwei, et al. "Unet++: A nested u-net architecture for medical image segmentation." Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop,ML-CDS2018,Held in Conjunction with MICCAI 2018,Granada,Spain,September 20,2018,Proceedings 4.Springer International Publishing,2018.),UNet3+(Huang,Huimin,et al."Unet 3+: A full-scale connected unet for medical image segmentation." ICASSP 2020 - 2020 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2020.) and TransUNet (Chen, Jieneng, et al. "Transunet: Transformers make strong encoders for medical image segmentation." arXiv preprint arXiv:2102.04306 (2021).), Swin-UNet (Cao, Hu, et al. "Swin-unet: Unet-like pure transformer for medical image segmentation." European conference on computer vision. Cham: Springer Nature Switzerland, 2022.), DS-TransUNet (Lin, Ailiang, et al. "Ds-transunet: Dual swin transformer u-net for medical image segmentation." IEEE Transactions on Instrumentation and Measurement 71 (2022): 1 - 15.) and TransFuse (Zhang, Yundong, Huiye Liu, and Qiang Hu. "Transfuse: Fusing transformers and cnns for medical image segmentation." Medical image computing and computer assisted intervention–MICCAI 2021: 24th international conference, Strasbourg, France, September 27–October 1, 2021, proceedings, Part I 24. Springer International Publishing, 2021.) etc. Convolutional neural network-based methods can capture local features well, but cannot effectively utilize global context information, resulting in poor segmentation accuracy. ViT has excellent global context extraction capabilities and is therefore widely used in medical image segmentation, but its quadratic computational complexity leads to high computational costs in medical image segmentation with higher resolutions. This high computational cost makes model deployment difficult in medical environments with limited computing resources. .

[0005] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a brain MRI image segmentation method combining Poisson denoising with Mamba architecture in view of the shortcomings of the prior art.

[0007] In order to solve the above technical problems, the present invention discloses a brain MRI image segmentation method combining Poisson denoising with Mamba architecture, comprising the following steps:

[0008] Step 1, perform Poisson denoising on the brain MRI image to be segmented;

[0009] Step 2, constructing an image segmentation model MambaNet for segmenting the brain MRI image;

[0010] Step 3, training and optimizing the image segmentation model MambaNet;

[0011] Step 4: Segment the brain MRI image after Poisson denoising in step 1 using the image segmentation model MambaNet trained in step 2 to complete the brain MRI image segmentation combining the Poisson denoising and Mamba architecture.

[0012] Furthermore, the image segmentation model PDMambaNet is based on the segmentation model MambaNet and adds a dual-branch structure TPD.

[0013] Furthermore, the image segmentation model PDMambaNet specifically includes:

[0014] encoder, bottleneck layer, first decoder and second decoder; wherein,

[0015] The encoder extracts features from the input brain MRI image to obtain 4-level features;

[0016] The output of the encoder passes through the bottleneck layer as the input of the first decoder;

[0017] The first decoder combines the features at each level extracted by the encoder to obtain a fine segmentation result;

[0018] The second decoder obtains a rough segmentation result according to the first two level features extracted by the encoder;

[0019] The fine segmentation results and the coarse segmentation results are fused by element-by-element addition to generate the final segmentation result.

[0020] Furthermore, the brain MRI image input to the encoder is the image after Poisson denoising in step 1, and a two-dimensional grayscale MRI image of size H×W×1 is obtained after preprocessing and feature extraction initialization.

[0021] Furthermore, the encoder comprises:

[0022] The input two-dimensional grayscale MRI image is divided into a plurality of small blocks through a block partitioning layer and converted into a one-dimensional sequence;

[0023] The one-dimensional sequence is embedded into a high-dimensional feature space through a linear embedding layer to obtain a first feature with a dimension of H / 4*W / 4*C; the above feature is input into the first visual Mamba module of the encoder for feature extraction to obtain a second feature with a dimension of H / 8*W / 8*2C;

[0024] The second feature reduces the resolution and increases the channel depth through the first fusion layer, and extracts features through the second visual Mamba module to obtain the third feature with dimensions of H / 16*W / 16*4C;

[0025] The third feature is extracted through the second fusion layer and the third visual Mamba module, and through the third fusion layer, a fourth feature with a dimension of H / 32*W / 32*8C is obtained.

[0026] Furthermore, the bottleneck layer includes:

[0027] 1 Visual Mamba module, i.e. two VSS modules;

[0028] After the fourth feature is processed by the bottleneck layer, the output is used as the input of the first decoder.

[0029] Furthermore, the first decoder comprises:

[0030] 3 sub-decoders, block expansion layer and linear mapping layer; among them,

[0031] Each sub-decoder is composed of a visual Mamba module and a block expansion layer; the three sub-decoders are used to receive the first feature, the second feature and the third feature generated in the encoder respectively;

[0032] The input of the first decoder is the output of the bottleneck layer, which is sequentially combined with the third feature, the second feature and the first feature through three sub-decoders, and then passes through a block expansion layer and a linear mapping layer, and the output is a fine segmentation result.

[0033] Furthermore, the second decoder comprises:

[0034] 2 sub-decoders, , block expansion layer and linear mapping layer; among them,

[0035] Each sub-decoder is composed of a visual Mamba module and a block expansion layer; the two sub-decoders are used to receive the first feature and the second feature generated in the encoder respectively;

[0036] The input of the second decoder is the second feature of the encoder, which is combined with the second feature and the first feature in sequence through two sub-decoders, and then passes through a block expansion layer and a linear mapping layer, and the output is a coarse segmentation result.

[0037] Furthermore, the visual Mamba module consists of 2 VSS blocks.

[0038] Furthermore, in step 3, the image segmentation model MambaNet is trained and optimized, including:

[0039] Step 3-1, collect the data set, preprocess it, and divide it into training set and test set;

[0040] Step 3-2, inputting the training set into the image segmentation model MambaNet;

[0041] Step 3-3, calculating the loss function according to the segmentation result, and optimizing the image segmentation model MambaNet;

[0042] Step 3-4, repeat and iterate steps 3-2 to 3-3 until the preset conditions are met and the optimization is completed.

[0043] Beneficial effects:

[0044] 1. The present invention introduces a Poisson denoising module to remove noise in MRI images to improve the performance of downstream segmentation tasks.

[0045] 2. The present invention designs a Twin-Path Decoder (TPD) dual-branch structure to enhance the model's ability to express information at different levels. Through this structure, not only is detail loss reduced, but the model's ability to capture multi-level information is also enhanced. The collaborative work of the TPD dual-branch structure effectively alleviates the problem of image over-smoothing and artifact introduction, ensuring that spatial information across various network scales is maintained. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0047] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0048] Figure 2 It is a schematic diagram of the Poisson denoising process in the present invention.

[0049] Figure 3 This is a comparison diagram before and after Poisson denoising in an embodiment.

[0050] Figure 4 This is a schematic diagram of the overall architecture of the image segmentation model combining Poisson denoising and Mamba architecture proposed in the present invention.

[0051] Figure 5 This is a comparison chart of the results of the present invention and different segmentation models on the OASIS-1 dataset.

[0052] Figure 6 This is a comparison chart of the results of the present invention and different segmentation models on the MRBrainS13 dataset. DETAILED DESCRIPTION

[0053] The overall design idea of ​​the present invention is as follows: In order to effectively remove the noise of MRI images, a Poisson denoising module is first introduced to remove the noise in the MRI images to improve the performance of downstream segmentation tasks. However, Poisson denoising will cause the loss of fine details and blurred edges in MRI images. This loss of details and blurred edge information will cause the segmentation network to be unable to fully utilize multi-level information, especially when processing brain MRI images with complex textures and structures, missed detection and false detection may occur. In addition, excessive smoothing in the denoising process may cause the segmentation network to weaken the ability to distinguish different tissue regions, affecting the final segmentation results. In order to solve the above problems caused by Poisson denoising, the present invention proposes a segmentation model (MambaNet). In MambaNet, the present invention designs a Twin-Path Decoder (TPD) dual-branch structure to enhance the model's ability to express information at different levels. Specifically, one decoder focuses on restoring global structural information, and the other decoder focuses on extracting and retaining detail information. Through this structure, not only the loss of details is reduced, but also the model's ability to capture multi-level information is enhanced. The collaborative work of TPD effectively alleviates the problem of image over-smoothing and artifact introduction, ensuring that spatial information across various network scales is maintained. Combining the advantages of Poisson denoising and the segmentation model MambaNet, the present invention proposes an image segmentation model PDMambaNet that combines Poisson denoising with the Mamba architecture.

[0054] In the field of medical images, especially in the processing of brain MRI images, improving image quality is crucial for accurate diagnosis of brain diseases. This paper proposes a complete solution to the common problems of Poisson noise, Gaussian noise and impulse noise in the process of MRI image segmentation, such as Figure 1 As shown in the figure, the overall process mainly includes two parts: Poisson denoising module and Mamba-based dual-branch segmentation framework. The specific process is as follows:

[0055] Step 1, such as Figure 2 As shown, Poisson denoising is performed, and the specific process is as follows:

[0056] The Poisson denoising process starts with inputting a noisy image and setting the number of iterations. First, the sparse representation of the noisy image is initialized and the maximum number of iterations T and the current iteration count t=0 are set. During the iteration process, the system continuously determines whether the current number of iterations is less than the set maximum number of iterations T. If the condition is met, the image is updated and the denoising result is adjusted through the sparse representation to gradually approach the real noise-free image. The updated image evaluates the current denoising effect by calculating the residual, and the sparse representation is further optimized based on the residual to improve the denoising performance. After each iteration, the iteration count increases by 1, and the loop operation is repeated until the set maximum number of iterations is reached or the denoising result meets expectations. After the entire iteration is completed, the processed denoised image is finally output. Through this process, Poisson denoising gradually reduces the noise in the image, retains important detail information, and obtains high-quality denoising results.

[0057] The specific implementation of Poisson denoising is:

[0058] For a given noisy image and its vectorized form, since the sensor noise source is proportional to the signal strength during medical image acquisition, it is assumed that the pixel value of the noisy image is a random variable that obeys the Poisson distribution, and its parameters are determined by the pixel value x of the real image at the i-th index, that is, x0[i]~P(x[i]), where P is a Poisson distribution process, defined as:

[0059]

[0060] In order to estimate the clean vector after denoising The log-likelihood of (1) needs to be maximized. According to the properties of the Poisson distribution, the maximum log-likelihood estimation of the true image is achieved by minimizing the following optimization problem:

[0061]

[0062] in, is a unit vector. However, this is an ill-posed problem. To solve this problem, a sparse representation method is used, and a dictionary D and a sparse vector α are introduced, so that Then the objective function becomes:

[0063]

[0064] Among them, ‖α‖0 is the l0 norm of α, and s is the preset sparsity. Since the l0 norm will make the problem an NP-hard problem, the l1 norm is used for relaxation, and Dα=exp(Dα) is set to deal with the non-negative constraint problem, and we get:

[0065]

[0066] Next, we iterate to solve α. It is difficult to directly solve min|α|0s.tx=Dα. A common approximation method is to use the l1 relaxation method to achieve Then the soft threshold operator (ISTA) is used to update the To solve, where L≤σ max (D T D), S is the soft threshold operator, which is defined as S(x)=sign(x)max(|x|-τ,0, τ is the pre-threshold parameter.

[0067] In order to solve the problem that the size of the dictionary D depends on the size of the input image, a convolutional sparse coding model is adopted, using convolution operations instead of matrix-vector multiplication: Dα = ∑D j *A j =D*A, where D j It is around the coefficient feature map A j This new form of sparse coding and dictionary application decouples the size of the dictionary from the input image size and eliminates the need to scale the model based on the image size. It has the following update steps:

[0068] A j ←S(A j-1 +D T *(X0-D*A j-1 )) (5)

[0069] Using the iterative soft thresholding algorithm (ISTA), we can eliminate the need to optimize α and rewrite the objective function as:

[0070]

[0071] Where ⊙ is the Hadamard product. In order to solve this problem using the ISTA algorithm, the present invention uses a neural network f θ Denotes D*A. Network f θ Contains a single encoder and decoder that compute a sparse representation A based on the network parameters, allowing the dictionary D to be learned via back-propagation. Mathematically, this can be expressed as D*A=f θ (X0). The modified optimization problem can be rewritten as:

[0072]

[0073] Furthermore, by replacing D and D with the decoder and encoder respectively in the above update step Tto approximate the update step in the ISTA algorithm: A←S(A+Encoder(X0-Decoder(A))). The update uses a recurrent neural network-like approach to iteratively refine the sparse code A over T steps, which is different from traditional sequence processing. This method applies a dictionary to A in the forward pass and enhances A through a predetermined number of iterations, moving away from traditional convergence-centric ISTA methods.

[0074] Step 2, such as Figure 4 As shown, an image segmentation model combining Poisson denoising and Mamba architecture is constructed. The specific process is as follows:

[0075] The MambaNet segmentation framework first preprocesses and initializes the feature extraction of the image after Poisson denoising. The input is a two-dimensional grayscale MRI image of size H×W×1. The image is first divided into several small blocks and converted into a one-dimensional sequence through the Patch Partition module. It is embedded into a high-dimensional feature space through the Linear Embedding layer, and the dimension becomes H / 4*W / 4*C. Subsequently, these features are input into the first layer encoder composed of two visual Mamba modules (VSS Block) for preliminary feature extraction. The output of each layer of the encoder is gradually reduced in resolution and increased in channel depth through the Patch Merging module, which are H / 8*W / 8*2C, H / 16*W / 16*4C, and finally to the bottleneck layer H / 32*W / 32*8C. The bottleneck layer further processes the deepest features through two VSS blocks.

[0076] Each layer of features of the encoder is passed to two decoders through skip connections, which are used to restore global information and retain detail information respectively. The decoder gradually increases the resolution by the Patch Expanding block, and combines it with the features from the encoder for fusion processing. Each decoder extracts multi-level features through multiple VSS blocks, and finally restores it to the same resolution H×W×Class as the input. The feature maps output by the two decoders are converted into coarse segmentation results and fine segmentation results respectively through Linear Projection, and are fused by element-by-element addition to generate the final segmentation result. Through this design, the model can effectively capture global and local information at multiple scales, improve segmentation accuracy and reduce artifacts and over-smoothing problems.

[0077] Example:

[0078] This example uses the OASIS-1 and MRBrainS13 datasets, and adopts a brain MRI image segmentation method combining Poisson denoising and Mamba architecture proposed in the present invention to perform image segmentation on the datasets, as follows:

[0079] The OASIS-1 dataset is from the Open Access Imaging Study Series (OASIS) and consists of 421 subjects aged between 18 and 96 years old, each of whom has a T1-weighted MRI scan with the following imaging acquisition details: gap-free thickness, resolution pixels. The labels of this dataset classify brain tissue into cerebrospinal fluid (CSF), gray matter (GM), and white matter (WM).

[0080] The MRBrainS13 challenge dataset consists of 20 subjects obtained from 3.0TPhilipsAchievaMR scans at the University Medical Center Utrecht, the Netherlands. Multi-sequence MRI brain scans, including T1 (TR: 7.9ms, TE: 4.5ms), T1-IR (TR: 4416ms, TE: 15ms, TI: 400ms), and T2-FLAIR (TR: 11000ms, TE: 125ms, TI: 2800ms), were acquired and used for the challenge. All scans were rigidly registered using Elastix and bias corrected using SPM8. After such preprocessing, the voxel spacing of all provided sequences is 0.96×0.96×3.00mm. At the same time, the dataset also provides manual segmentation labels for cerebrospinal fluid (CSF), gray matter (GM), and white matter (WM).

[0081] After Poisson denoising in step 1, the result is as follows Figure 3 As shown, the first row of three pictures are original images; the second row of three pictures are images after denoising;

[0082] In this embodiment, the two data sets are divided into training set, test set and evaluation set in a ratio of 8:1:1. In order to enable the model to better learn image features, all images are normalized and resized to 224×224. At the same time, data enhancement techniques are used, including vertical flipping, horizontal flipping and random rotation, to increase the diversity of the data set and improve the generalization ability of the model.

[0083] The models of the present invention were trained on an Ubuntu 22.04 system equipped with an Nvidia A40 GPU with 48G memory, using Python 3.8.19, PyTorch 2.2.0 and CUDA11.98.

[0084] PDMambaNet was trained for 40,000 iterations on the OASIS1 dataset and 20,000 iterations on the MRBrainS13 dataset, with a batch size of 12. The stochastic gradient descent (SGD) optimizer was used with a learning rate of 0.01, a momentum of 0.9, and a weight decay of 0.0001. During training, the network performance was evaluated on the validation set every 200 iterations, and the model weights were saved only when a new best performance was achieved on the validation set to ensure that the model could converge to the optimal state, thereby achieving better results in the brain MRI image segmentation task.

[0085] In order to comprehensively evaluate the performance of PDMambaNet of the present invention, seven objective evaluation indicators are used for quantitative comparison, including Dice, Accuracy (Acc), Precision (Pre), Sensitivity (Sen), Specificity (Spe), Hausdorff distance (HD) 95% and average surface distance (ASD). Among them, Dice, Accuracy, Precision, Sensitivity and Specificity belong to similarity measurement indicators, represented by upward arrows (↑), and the closer the value is to 1, the better the performance; Hausdorff distance 95% and average surface distance belong to difference measurement indicators, represented by downward arrows (↓), and the lower the value, the better, indicating that the similarity between the predicted segmentation and the true segmentation is higher.

[0086] like Figure 5As shown, it is a comparison chart of the results of the present invention and different segmentation models on the OASIS-1 dataset, wherein, the first column is the original image; the second column is the label image; the third column is the Mamba-UNet result graph; the fourth column is the UNet result graph; the fifth column is the PNet result graph; the sixth column is the Swin-UNet result graph; the seventh column is the PDMambaNet result graph. On the large-scale OASIS1 dataset, compared with mainstream models such as Mamba-UNet, UNet, PNet, and Swin-UNet, the Dice coefficient of PDMambaNet can reach 0.9266 (Mamba-UNet is 0.9097, UNet is 0.8743, PNet is 0.9262, and Swin-UNet is 0.9248), and the Accuracy is 0.9733 (Mamba-UNet is 0.9739, UNet is 0.9678, PNet is 0.9757, Swin-UNet is 0.9728), Hausdorff distance 95% is 106410 (Mamba-UNet is 2.1684, UNet is 2.4922, PNet is 1.6571, Swin-UNet is 1.4741), and the average surface distance is 0.4700 (Mamba-UNet is 0.6380, UNet is 0.7795, PNet is 0.3753, Swin-UNet is 0.3918).

[0087] like Figure 6The figure shows the comparison of the results of the present invention and different segmentation models on the MRBrainS13 dataset, where the first column is the original image; the second column is the label image; the third column is the Mamba-UNet result graph; the fourth column is the UNet result graph; the fifth column is the PNet result graph; the sixth column is the Swin-UNet result graph; and the seventh column is the PDMambaNet result graph. On the small-scale MRBrainS13 dataset, the performance is also excellent, with a Dice coefficient of 0.7158 (Mamba-UNet is 0.6908, UNet is 0.7048, PNet is 0.7068, and Swin-UNet is 0.7102), and an Accuracy of 0.8319 (Mamba-UNet is 0.8194, UNet is 0.8214, PNet is 0.8232, Swin-UNet is 0.8366). et is 0.8312), Hausdorff distance 95% is 3.0223 (Mamba-UNet is 3.3607, UNet is 3.2435, PNet is 2.9803, Swin-UNet is 2.2916), and the average surface distance is 0.7706 (Mamba-UNet is 0.9981, UNet is 0.9775, PNet is 0.8765, Swin-UNet is 0.7697).

[0088] These data show that PDMambaNet can maintain high segmentation accuracy on data sets of different sizes, effectively utilize multi-level information of MRI images, and excel in detail preservation and overall structure grasp. Compared with other models, it can better deal with the problems of detail loss and edge blur when processing images after Poisson denoising, and has high accuracy in segmenting different brain tissues. It is of great application value in medical imaging diagnosis of brain diseases, and has effectively promoted technological progress and development in this field.

[0089] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of the brain MRI image segmentation method combining Poisson denoising and Mamba architecture provided by the present invention and some or all steps in each embodiment can be run. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0090] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on such an understanding, the technical solutions in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of computer programs, i.e., software products, which can be stored in a storage medium and include several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, an MCU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0091] The present invention provides a brain MRI image segmentation method combining Poisson denoising with Mamba architecture. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A brain MRI image segmentation method combining Poisson denoising and Mamba architecture, characterized in that: The following steps are involved: Step 1, perform Poisson denoising on the brain MRI image to be segmented; Step 2, constructing an image segmentation model MambaNet for segmenting the brain MRI image; Step 3, training and optimizing the image segmentation model MambaNet; Step 4: Segment the brain MRI image after Poisson denoising in step 1 using the image segmentation model MambaNet trained in step 2 to complete the brain MRI image segmentation combining the Poisson denoising and Mamba architecture.

2. According to claim 1, a brain MRI image segmentation method combining Poisson denoising and Mamba architecture is characterized in that: The image segmentation model PDMambaNet is based on the segmentation model MambaNet and adds a dual-branch structure TPD.

3. According to claim 2, a brain MRI image segmentation method combining Poisson denoising and Mamba architecture is characterized in that: The image segmentation model PDMambaNet specifically includes: encoder, bottleneck layer, first decoder and second decoder; wherein, The encoder extracts features from the input brain MRI image to obtain 4-level features; The output of the encoder passes through the bottleneck layer as the input of the first decoder; The first decoder combines the features at each level extracted by the encoder to obtain a fine segmentation result; The second decoder obtains a rough segmentation result according to the first two level features extracted by the encoder; The fine segmentation results and the coarse segmentation results are fused by element-by-element addition to generate the final segmentation result.

4. According to claim 3, a brain MRI image segmentation method combining Poisson denoising and Mamba architecture is characterized in that: The brain MRI image input to the encoder is the image after Poisson denoising in step 1, and a two-dimensional grayscale MRI image of size H×W×1 is obtained after preprocessing and feature extraction initialization.

5. The brain MRI image segmentation method combining Poisson denoising and Mamba architecture according to claim 4 is characterized in that: The encoder comprises: The input two-dimensional grayscale MRI image is divided into a plurality of small blocks through a block partitioning layer and converted into a one-dimensional sequence; The one-dimensional sequence is embedded into the high-dimensional feature space through a linear embedding layer to obtain a feature space with dimensions H / 4*W / 4*C First feature; the above feature is input into the first visual Mamba module of the encoder for feature extraction to obtain a second feature with a dimension of H / 8*W / 8*2C; The second feature reduces the resolution and increases the channel depth through the first fusion layer, and extracts features through the second visual Mamba module to obtain the third feature with dimensions of H / 16*W / 16*4C; The third feature is extracted through the second fusion layer and the third visual Mamba module, and through the third fusion layer, a fourth feature with a dimension of H / 32*W / 32*8C is obtained.

6. The brain MRI image segmentation method combining Poisson denoising and Mamba architecture according to claim 5, characterized in that: The bottleneck layer comprises: 1 Visual Mamba module, i.e. two VSS modules; After the fourth feature is processed by the bottleneck layer, the output is used as the input of the first decoder.

7. The brain MRI image segmentation method combining Poisson denoising and Mamba architecture according to claim 6 is characterized in that: The first decoder comprises: 3 sub-decoders, block expansion layer and linear mapping layer; among them, Each sub-decoder is composed of a visual Mamba module and a block expansion layer; the three sub-decoders are used to receive the first feature, the second feature and the third feature generated in the encoder respectively; The input of the first decoder is the output of the bottleneck layer, which is sequentially combined with the third feature, the second feature and the first feature through three sub-decoders, and then passes through a block expansion layer and a linear mapping layer, and the output is a fine segmentation result.

8. The brain MRI image segmentation method combining Poisson denoising and Mamba architecture according to claim 7, characterized in that: The second decoder comprises: 2 sub-decoders, , block expansion layer and linear mapping layer; among them, Each sub-decoder is composed of a visual Mamba module and a block expansion layer; the two sub-decoders are used to receive the first feature and the second feature generated in the encoder respectively; The input of the second decoder is the second feature of the encoder, which is combined with the second feature and the first feature in sequence through two sub-decoders, and then passes through a block expansion layer and a linear mapping layer, and the output is a coarse segmentation result.

9. The brain MRI image segmentation method combining Poisson denoising and Mamba architecture according to claim 8, characterized in that: The Visual Mamba module consists of 2 VSS blocks.

10. The brain MRI image segmentation method combining Poisson denoising and Mamba architecture according to claim 9, characterized in that: In step 3, the image segmentation model MambaNet is trained and optimized, including: Step 3-1, collect the data set, preprocess it, and divide it into training set and test set; Step 3-2, inputting the training set into the image segmentation model MambaNet; Step 3-3, calculating the loss function according to the segmentation result, and optimizing the image segmentation model MambaNet; Step 3-4, repeat and iterate steps 3-2 to 3-3 until the preset conditions are met and the optimization is completed.

Images