Prostate MRI image segmentation method and system based on improved U-Net network, and storage medium

Abstract

The invention discloses a prostate MRI image segmentation method and system based on an improved U-Net network, and a storage medium, and relates to the technical field of medical image processing. The invention discloses a prostate MRI image segmentation method based on an improved U-Net network. The method comprises the steps of obtaining a prostate MRI image data set, performing slicing processing on each piece of 3D image data in the prostate MRI image data set, and setting a pixel effective area to obtain a prostate input image; a basic U-Net network model is constructed and improved, an up-sampling module symmetrical to down-sampling is added in an encoder, a new down-sampling module is added after the added up-sampling, and a double-branch structure of first down-sampling and then up-sampling and first up-sampling and then down-sampling is formed; meanwhile, an efficient additive attention module is added between the encoder module and the decoder module, so that an O-Net network model is obtained; the prostate input image is input into the O-Net network model, a prostate MRI image segmentation result is obtained, and the segmentation precision of edge details is remarkably improved while deep semantic information is kept.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, system, and storage medium for prostate MRI image segmentation based on an improved U-Net network. Background Technology

[0002] The prostate gland, located in the lower part of the male pelvis in front of the rectum, surrounds the urethra like a "guard" and is an indispensable part of the male reproductive system. Prostate cancer, a common malignant tumor in men, has a high incidence rate and a tendency to metastasize, making it a major challenge in global public health. Statistics from the International Agency for Research on Cancer show that approximately 1.4 million new cases of prostate cancer are diagnosed globally each year. The 5-year survival rate for prostate cancer patients in China is 56.4%, significantly lower than the 99% in developed countries in Europe and America. The main reason for this is the inadequacy of early screening programs and systems for prostate cancer, resulting in a low early diagnosis rate and causing many patients to miss the optimal treatment window.

[0003] Currently, clinical methods for prostate cancer screening include digital rectal examination, prostate-specific antigen (PSA) testing, and prostate biopsy. However, these methods all have limitations. Digital rectal examination is difficult to accurately determine the location of prostate cancer; PSA testing may produce false positive and false negative results; and prostate biopsy may cause complications such as infection, threatening the patient's life. Magnetic Resonance Imaging (MRI) is a non-invasive, non-radioactive imaging technique that, with its excellent tissue contrast and clear anatomical details, has become an important tool for prostate cancer screening and diagnosis. To improve the consistency and accuracy of prostate MRI images in prostate cancer screening and diagnosis, the European Society of Urology, the American College of Radiology, and the European Society of Radiology jointly developed the Prostate Imaging Reporting and Data System (PIRS). This system classifies the risk level of prostate cancer based on factors such as the structure, shape, and size of the prostate tissue, and the basis for this classification is the segmentation of the prostate tissue in the prostate MRI image.

[0004] Current technologies for segmenting prostate MRI images often rely on experienced physicians for manual work. However, manual segmentation is not only time-consuming and labor-intensive, but the results are also subject to subjective factors such as the operator's personal cognition, experience, and skills, leading to errors like missed segments, missegments, and duplicate segments, thus affecting the accuracy and reliability of the results. With the continuous development of deep learning, in recent years, researchers have utilized the U-Net convolutional neural network model, proposing a skip connection operation to fuse low-resolution and high-resolution feature maps. This effectively avoids the spatial information loss caused by downsampling operations and achieves good performance with limited training data. Furthermore, some researchers have proposed U-Net based on U-Net. U-Net++ redesigns skip connections. Traditional U-Net skip connections directly concatenate the encoder's feature maps to the corresponding decoder layers. U-Net++, however, employs a dense skip connection structure. In U-Net++, each layer's feature map from the encoder is connected to multiple layers at different depths in the decoder. This dense connection approach allows the decoder to access feature information from different levels of abstraction in the encoder. During the decoder's upsampling process, U-Net++'s skip connections, combined with progressive upsampling, perform feature fusion. Convolutional operations are used to fuse the concatenated feature maps, enabling the feature maps to recover spatial resolution while incorporating feature information from different levels in the encoder, thus improving segmentation accuracy and detail capture.

[0005] U-Net can be used for end-to-end image segmentation of prostate MRI images. However, U-Net and its enhancement networks often neglect the complementary information between edge features at different scales. In addition, due to the semantic differences between feature maps of different resolutions, the global context information of deep feature maps has a stronger semantic expressive power, while shallow detail information is easily masked during the fusion process, resulting in the loss of detail information and affecting the segmentation accuracy and detail preservation ability of the model. Summary of the Invention

[0006] To overcome the shortcomings of existing networks in segmenting prostate MRI images, which tend to overlook shallow image information and thus result in low segmentation accuracy, this invention proposes a prostate MRI image segmentation method based on an improved U-Net network.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a prostate MRI image segmentation method based on an improved U-Net network, comprising:

[0008] S1: Obtain the prostate MRI image dataset and slice each 3D image data in the prostate MRI image dataset to obtain 2D images;

[0009] S2: Extract the effective area of ​​each pixel from each 2D image to obtain the prostate input image;

[0010] S3: Construct the basic U-Net network model and improve it to obtain the O-Net network model;

[0011] The basic U-Net network model includes an encoder module, a decoder module, a skip connection layer, and an output layer. The encoder module includes N downsampling modules connected in sequence; the decoder module includes N upsampling modules connected in sequence; the encoder module and the decoder module form an inverted U-shaped structure.

[0012] The basic U-Net network model is improved by adding an upsampling module symmetrical to the downsampling module in the encoder, and adding a new downsampling module after the added upsampling module. The new downsampling module forms a symmetrical structure with the upsampling module of the decoder and an inverted U-shaped structure with the new upsampling module, ultimately forming a dual-branch structure of downsampling followed by upsampling and upsampling followed by downsampling. At the same time, an efficient additive attention module is added between the encoder module and the decoder module.

[0013] S4: Input the prostate input image into the O-Net network model to obtain the prostate MRI image segmentation results.

[0014] Preferably, each downsampling module and each upsampling module includes a convolutional layer, a batch normalization layer, a ReLU activation function, and a pooling layer connected in sequence; the encoder module extracts image features step by step through the convolutional layer and the pooling layer to obtain a first feature image; the decoder module is used to decode the first feature image to obtain a second feature image; the skip connection layer is used to fuse the first feature image of the encoder's corresponding layer sampling module with the second feature image of the decoder's corresponding layer sampling module; the output layer is used to perform a sigmoid operation on the fused feature image and divide it to obtain the prostate MRI image segmentation result.

[0015] Preferably, the processing steps of the efficient additive attention module include:

[0016] SS1: The output feature image of each sampling module in the encoder module is patched, linearly transformed, and positional encoding is added to obtain the input sequence Z0;

[0017] SS2: Map the input sequence Z0 to the query space and key space to obtain the query Q and key value K;

[0018] SS3: Divide the query Q into channels to obtain M query components Q. i Obtain the attention query vector a based on each query component. i ;

[0019] SS4: The global query vector q is obtained by weighted summation of all attention query vectors;

[0020] SS5: Obtaining efficient additive attention features based on global query vector q and key value K.

[0021] SS6: Features of Highly Effective Additive Attention Perform inverse positional encoding to obtain the final efficient additive attention feature X′; the final efficient additive attention feature X′ is input into the skip connection layer for feature fusion.

[0022] Preferably, in step SS1, the expression for the input sequence Z0 is:

[0023]

[0024] in, Let E be the feature vector of the i-th patch; E is the linear transformation matrix; E pos This is the position encoding matrix; is the embedding vector of the i-th patch after projection; is the sequence concatenation operator.

[0025] Preferably, in step SS2, the formulas for calculating query Q and key value K are as follows:

[0026] Q = Z0·W q

[0027] K = Z0·W k

[0028] Among them, W q For query weight; W k Key-value weights;

[0029] d model d is the dimension of the input sequence. k For the dimensions of Q and K.

[0030] Preferably, in step SS3, the attention query vector a i The calculation formula is:

[0031]

[0032] Among them, W d Let d be the learnable parameter vector, and d be the dimension of the key value K.

[0033] Preferably, in step SS4, the formula for calculating the global query vector q is:

[0034]

[0035] Where M is the total number of channels and i is the channel index number.

[0036] Preferably, in step SS5, the efficient additive attention feature The calculation formula is:

[0037]

[0038] Where vec(·) flattens the three-dimensional tensor into a two-dimensional matrix; W is a trainable weight matrix; For performing product operations; T(·) indicates performing a linear transformation; Norm(·) indicates performing normalization; Q is the key value.

[0039] A prostate MRI image segmentation system based on an improved U-Net network includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor executes the computer program to implement a prostate MRI image segmentation method based on an improved U-Net network.

[0040] A storage medium having a computer program stored thereon, which, when executed, implements a prostate MRI image segmentation method based on an improved U-Net network.

[0041] The advantages of this invention are:

[0042] (1) This invention adopts a dual-branch structure (positive U-shape + inverted U-shape), adds a symmetrical upsampling module to the encoder and a downsampling module to the corresponding position in the decoder, so that the network can learn multi-scale features at the same time and avoid the loss of shallow image detail information; through the efficient additive attention module (EAAM) to dynamically adjust feature weights, enhance attention to key areas such as prostate edge and small lesions, and reduce the problem of missed segmentation or missegmentation; compared with traditional U-Net and U-Net++, this method significantly improves the segmentation accuracy of edge details while maintaining deep semantic information, which is more in line with the needs of clinical diagnosis.

[0043] (2) This invention introduces an efficient additive attention module (EAAM) in skip connections, which adaptively adjusts feature weights through global query vectors, making the decoder pay more attention to high-value regions (such as tumor edges); it uses patch partitioning + positional encoding to preserve spatial information, avoiding the high complexity problem of traditional attention calculation; it effectively alleviates the problem of deep features masking shallow details, improves the model's ability to identify small lesions and blurred boundaries, and reduces the risk of false negatives / false positives.

[0044] (3) This invention uses 2D slicing to reduce computational complexity, while ensuring segmentation accuracy through a dual-branch structure and attention mechanism; the efficient additive attention module reduces the number of parameters through channel partitioning and linear transformation, which is more efficient than traditional Transformer-type attention calculation; while ensuring accuracy, it reduces computational resource consumption and is more suitable for hospital deployment and real-time assisted diagnosis.

[0045] (4) The EAAM module of this invention suppresses irrelevant noise (such as interference from adjacent tissues such as bladder and rectum) through dynamic weight adjustment, thereby improving the robustness of the model; it performs more stably on cross-device and cross-hospital datasets, and reduces segmentation errors caused by differences in image quality. Attached Figure Description

[0046] Figure 1 Diagram of the O-Net network model;

[0047] Figure 2 This is a flowchart of the method steps of the present invention;

[0048] Figure 3 A structural diagram of an efficient additive attention module;

[0049] Figure 4 This is a comparison chart of the experimental results of different models. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] like Figure 2 As shown, this invention proposes a prostate MRI image segmentation method based on an improved U-Net network, comprising:

[0052] S1: Obtain the prostate MRI image dataset and slice each 3D image data in the prostate MRI image dataset to obtain 2D images.

[0053] S2: Extract the set effective pixel area from each 2D image to obtain the prostate input image.

[0054] In this embodiment, the publicly available prostate MRI dataset PROMISE12 is selected. This dataset was used in the Prostate MRI Image Segmentation Challenge held by the Medical Image Computing and Computer Assisted Intervention Society (MICCAI) in 2012. It contains a total of 100 prostate MRI images, all of which have undergone a standardized preprocessing workflow. The first step is to slice each 3D image and convert it into a 2D image. The second step is to extract a valid region of 512×512 pixels from the 2D image, which is used as input data for subsequent models. If the image size in the 2D image is smaller than 512×512, edge padding is used to maintain the consistency of image size across the dataset.

[0055] Meanwhile, to ensure that the samples are evenly distributed across different sets, these prostate input image data were randomly divided into three parts: training set, test set, and validation set, in a ratio of 6:2:2.

[0056] S3: Construct and improve the basic U-Net network model to obtain the O-Net network model. The O-Net network model diagram is shown below. Figure 1 As shown.

[0057] The basic U-Net network model includes an encoder module, a decoder module, skip connection layers, and an output layer. The encoder module extracts image features stepwise through convolutional and pooling layers to obtain a first feature image. Shallow convolutional layers capture low-level visual features, while deeper layers extract high-level semantic information. Through multi-scale feature extraction and spatial-semantic balance design, it provides the decoder with a feature foundation that combines detail sensitivity and global understanding. The decoder module decodes the first feature image to obtain a second feature image. Skip connection layers fuse the first feature image from the corresponding layer of the encoder with the second feature image from the corresponding layer of the decoder. The output layer performs a sigmoid operation on the fused feature image and segments it to obtain the prostate MRI image segmentation result.

[0058] The encoder module includes N downsampling modules connected in sequence; the decoder module includes N upsampling modules connected in sequence; each downsampling module and each upsampling module includes a convolutional layer, a batch normalization layer, a ReLU activation function, and a pooling layer; the encoder module and the decoder module form an upright U-shaped structure.

[0059] Figure 1In the code, DERB Convolution is a detail enhancement convolution; Convolution 1×1 is a 1×1 convolution operation; MaxPooling 2×2 is a 2×2 max pooling operation; Skip connectionwith EAA is a combination of skip connection and efficient additive attention module (EAA); Up Sampling2×2 is a 2×2 upsampling operation; and Block copied is a feature map copied from the encoder to the decoder.

[0060] The basic U-Net network model is improved by adding an upsampling module symmetrical to the downsampling module in the encoder, and then adding a new downsampling module after the added upsampling module. The new downsampling module forms a symmetrical structure with the upsampling module of the decoder, and an inverted U-shaped structure with the new upsampling module. Finally, a dual-branch structure is formed, which first downsamples and then upsamples, and then upsamples and then downsamples. The dual-branch structure is used to improve the fusion accuracy of shallow and deep features. At the same time, an efficient additive attention module (EAA) is added between the encoder module and the decoder module. The EAA is used to connect each element in the first feature image of the encoder, thereby aiming to dynamically adjust the feature weights so that the model focuses on key information and suppresses irrelevant noise.

[0061] like Figure 3 As shown, the processing steps of the efficient additive attention module include:

[0062] SS1: The output feature image of each sampling module in the encoder module is patched, linearly transformed, and positional encoding is added to obtain the input sequence Z0. The expression of the input sequence Z0 is:

[0063]

[0064] in, Let E be the feature vector of the i-th patch, which is used to divide local region data units from the input feature map; E is a linear transformation matrix with dimension P. 2 ×D is used to compress a high-dimensional patch (P×P) into a low-dimensional feature vector D; E pos This is the position encoding matrix; is the embedding vector of the i-th patch after projection, which is essentially a token; is the sequence concatenation operator, used to combine the embedding vectors of n patches into an input sequence of length n. The output feature image of each sampling module in the encoder module has dimensions w×h×d.

[0065] The output feature image of each sampling module in the encoder module is divided and encoded into an input sequence Z0 to extract the global context. Each patch is P×P in size. Then, a linear transformation and positional encoding are performed on each patch. Each element in Z0... An embedding vector representation corresponding to a patch.

[0066] SS2: Map the input sequence Z0 to the query space and key space, obtain the query Q and key value K, calculated using the following formula:

[0067] Q = Z0·W q

[0068] K = Z0·W k

[0069] Among them, W q To determine the query weight, the input sequence Z0 is mapped to the query space, representing the information currently being searched; W k As key-value weights, the input sequence Z0 is mapped to the key space, representing each word as a feature of the search object; d model d is the dimension of the input sequence. k For the dimensions of Q and K;

[0070] SS3: Divide the query Q into channels to obtain M query components Q. i Obtain the attention query vector a based on each query component. i The calculation formula is:

[0071]

[0072] Among them, W d Let K be a learnable parameter vector, and d be the dimension of the key K.

[0073] SS4: Summing the corresponding query components based on all attention query vectors yields the global query vector q. The calculation formula is as follows:

[0074]

[0075] Where M is the total number of channels, and i is the channel index number;

[0076] SS5: Obtaining efficient additive attention features based on global query vector q and key value K. The calculation formula is:

[0077]

[0078] Where vec(·) flattens the three-dimensional tensor into a two-dimensional matrix; W is a trainable weight matrix; For product operations; T(·) denotes linear transformation; Norm(·) denotes normalization; Q is the key value;

[0079] The interaction between the global query vector q and the key K is encoded using element-wise multiplication to form a global context. The global context matrix extracts information from each token and can flexibly learn the relevance in the input sequence. To reduce model complexity and computational cost, a linear transformation T is introduced, T:

[0080] SS6: Features of Highly Effective Additive Attention The inverse positional encoding operation is performed to obtain the final efficient additive attention feature X′. By simulating the selective attention mechanism of humans, the model's ability to process key information is improved. The final efficient additive attention feature X′ is then input into the skip connection layer for feature fusion.

[0081] Figure 3 In this context, Positional Encoding is a key operation used to inject positional information into sequence data, aiming to solve the problem that self-attention mechanisms cannot naturally capture sequence order; Point-wise addition is an element-wise addition operation that adds two or more tensors (such as vectors, matrices, or feature maps) of the same dimension; Element-wise refers to the operation of independently calculating the elements at the same position in two tensors. When the two tensors have different shapes, the system automatically expands the dimension of the smaller tensor to match the shape of the larger tensor, thereby realizing the element-wise operation; Multiplicationwise summation is a composite operation that refers to the calculation process of first performing element-wise multiplication and then summing.

[0082] S4: Input the prostate image into the O-Net network model to obtain the prostate segmentation results.

[0083] In this embodiment, the MRI scanner detection results are converted into a single-channel image of size 512×512 pixels, and then fed into the O-Net model. After processing by O-Net, a two-dimensional tensor of size 512×512 is finally generated. The value of the prostate region in the tensor is 1, and the value of other places is 0. The tensor is then converted into a NumPy array and saved as a PNG image for doctors and relevant personnel to help interpret the MRI imaging results.

[0084] A prostate MRI image segmentation system based on an improved U-Net network includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor executes the computer program to implement a prostate MRI image segmentation method based on an improved U-Net network.

[0085] A storage medium having a computer program stored thereon, which, when executed, implements a prostate MRI image segmentation method based on an improved U-Net network.

[0086] To verify the effectiveness of the method of the present invention, O-Net, U-Net and U-Net++ were tested and compared under the same training conditions. The comparison results are shown in Table 1 below.

[0087] Table 1 Comparison of Experimental Results

[0088] Model Dice (%) HD95 / mm U-Net 70.14 15.76 U-Net++ 71.23 16.43 O-Net 76.90 14.02

[0089] The Dice coefficient is used to calculate the degree of overlap between two sample sets. It is a geometric similarity metric, ranging from 0 to 1. A higher Dice coefficient indicates better segmentation accuracy.

[0090] Hausdorff distance is a distance defined between any two sets in a metric space. To eliminate the influence of unreasonable distances caused by outliers, Hausdorff distance is multiplied by 95% to obtain the final result HD95. The smaller the HD95, the more accurate the segmentation.

[0091] Visualization of the segmentation results of each network is as follows: Figure 4 As shown, the first column is the input image of the network, the second column is the corresponding ground truth segmentation label, and the third, fourth, and fifth columns are the inference results of U-Net, U-Net++, and O-Net, respectively.

[0092] The quantitative and qualitative comparative experimental results above show that O-Net extracts image detail information more effectively, verifying the effectiveness of the model.

[0093] Of course, those skilled in the art will recognize that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0094] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0095] The technologies, shapes, and structures not described in detail in this invention are all known technologies.

Claims

1. A prostate MRI image segmentation method based on an improved U-Net network, characterized in that, include: S1: Obtain the prostate MRI image dataset and slice each 3D image data in the prostate MRI image dataset to obtain 2D images; S2: Extract the effective area of ​​each pixel in each 2D image to obtain the prostate input image; S3: Construct the basic U-Net network model and improve it to obtain the O-Net network model; The basic U-Net network model includes an encoder module, a decoder module, a skip connection layer, and an output layer. The encoder module includes N downsampling modules connected in sequence. The decoder module consists of N upsampling modules connected in sequence; the encoder module and the decoder module form an upright U-shaped structure. The basic U-Net network model is improved by adding an upsampling module symmetrical to the downsampling module in the encoder, and adding a new downsampling module after the added upsampling module. The new downsampling module forms a symmetrical structure with the upsampling module of the decoder and an inverted U-shaped structure with the new upsampling module, ultimately forming a dual-branch structure of downsampling followed by upsampling and upsampling followed by downsampling. At the same time, an efficient additive attention module is added between the encoder module and the decoder module. S4: Input the prostate input image into the O-Net network model to obtain the prostate MRI image segmentation results.

2. The prostate MRI image segmentation method based on an improved U-Net network as described in claim 1, characterized in that, Each downsampling module and each upsampling module includes a convolutional layer, a batch normalization layer, a ReLU activation function, and a pooling layer connected in sequence. The encoder module extracts image features step by step through the convolutional and pooling layers to obtain a first feature image. The decoder module decodes the first feature image to obtain a second feature image. The skip connection layer fuses the first feature image from the corresponding layer of the encoder with the second feature image from the corresponding layer of the decoder. The output layer performs a sigmoid operation on the fused feature image and divides it to obtain the prostate MRI image segmentation result.

3. The prostate MRI image segmentation method based on an improved U-Net network as described in claim 1, characterized in that, The processing steps of the efficient additive attention module include: SS1: The output feature image of each sampling module in the encoder module is patched, linearly transformed, and positional encoding is added to obtain the input sequence Z0; SS2: Map the input sequence Z0 to the query space and key space to obtain the query Q and key value K; SS3: Divide the query Q into channels to obtain M query components Q. i Obtain the attention query vector a based on each query component. i ; SS4: The global query vector q is obtained by weighted summation of all attention query vectors; SS5: Obtaining efficient additive attention features based on global query vector q and key value K. SS6: Features of Highly Effective Additive Attention Perform inverse positional encoding to obtain the final efficient additive attention feature X′; the final efficient additive attention feature X′ is input into the skip connection layer for feature fusion.

4. The prostate MRI image segmentation method based on an improved U-Net network as described in claim 3, characterized in that, In step SS1, the expression for the input sequence Z0 is: in, Let E be the feature vector of the i-th patch; E is the linear transformation matrix; E pos This is the position encoding matrix; is the embedding vector of the i-th patch after projection; is the sequence concatenation operator.

5. The prostate MRI image segmentation method based on an improved U-Net network as described in claim 3, characterized in that, In step SS2, the formulas for calculating the query Q and key value K are as follows: Q=Z0·W q K=Z0·W k Among them, W q For query weight; W k Key-value weights; d model It is the dimension of the input sequence, d k For the dimensions of Q and K.

6. The prostate MRI image segmentation method based on an improved U-Net network as described in claim 3, characterized in that, In step SS3, the attention query vector a i The calculation formula is: Among them, W d Let d be the learnable parameter vector, and d be the dimension of the key value K.

7. The prostate MRI image segmentation method based on an improved U-Net network as described in claim 3, characterized in that, In step SS4, the formula for calculating the global query vector q is: Where M is the total number of channels and i is the channel index number.

8. The prostate MRI image segmentation method based on an improved U-Net network as described in claim 3, characterized in that, In step SS5, efficient additive attention features The calculation formula is: Where vec(·) flattens the three-dimensional tensor into a two-dimensional matrix; W is a trainable weight matrix; For performing product operations; T(·) indicates performing a linear transformation; Norm(·) indicates performing normalization; Q is the key value.

9. A prostate MRI image segmentation system based on an improved U-Net network, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, the processor is connected to the memory, and the processor is used to execute the computer program to implement a prostate MRI image segmentation method based on an improved U-Net network as described in any one of claims 1-8.

10. A storage medium, characterized in that, It stores a computer program that, when executed, implements the prostate MRI image segmentation method based on an improved U-Net network as described in any one of claims 1-8.

Images