Prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning

Abstract

The invention discloses a prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning, and relates to the technical field of image segmentation, and the method comprises the steps: carrying out the preprocessing of collected multi-parameter magnetic resonance image original data, and dividing the data into a training set, a test set and a verification set; the method comprises the following steps: respectively constructing an attention-unet network used for prostate segmentation and a Multi-attention-unet network used for prostate cancer segmentation; training and selecting an optimal prostate segmentation model by using the verification set; inputting the multi-parameter magnetic resonance images of the test set into the optimal prostate segmentation model to obtain the multi-parameter magnetic resonance images in the set training set for training after the gland optimal prediction mask is obtained, and selecting the optimal prostate cancer segmentation model by using the verification set; and inputting the multi-parameter magnetic resonance image of the test set and the gland optimal prediction mask into the optimal prostate cancer segmentation model to obtain a prostate cancer segmentation result, thereby improving the segmentation precision and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning. Background Art

[0002] Nowadays, prostate cancer is the second most common cancer in men worldwide, and its incidence and mortality rates are increasing year by year. Early detection, early diagnosis and early treatment can effectively improve the survival rate of prostate cancer patients. Multi-parameter magnetic resonance imaging is the most effective non-invasive diagnostic method for prostate cancer, which can help doctors observe the disease lesions intuitively and comprehensively, but it requires high professional knowledge and ability of doctors. Therefore, in order to assist clinical diagnosis, designing an accurate, efficient and fast automatic segmentation method for prostate cancer to assist doctors in prostate cancer positioning and volume judgment is of great significance for evaluating the malignancy of lesions and guiding biopsy puncture operations.

[0003] At present, the automatic segmentation methods of prostate cancer are mainly divided into traditional machine learning methods and deep learning methods. Among them, the use of traditional machine learning methods to achieve prostate cancer segmentation has improved the prediction accuracy to a certain extent, but the multi-step segmentation process of traditional machine learning methods is time-consuming and prone to errors, and the accuracy of the final result is also strongly dependent on the quality of feature extraction. Compared with traditional machine learning, the system based on deep learning performs outstandingly. Its end-to-end, high generalization, and automatic feature extraction characteristics make deep learning have certain advantages in medical image processing.

[0004] However, the deep learning method is used to directly perform prostate segmentation on multi-parameter magnetic resonance images to remove the influence of other pelvic tissues on lesion segmentation. Only the glandular part image is used as input, so that the network ignores all background information and focuses on the gland for lesion segmentation. Although the influence of other pelvic tissues on the segmentation network is removed, the treatment of the tissues around the prostate is too absolute, resulting in unsatisfactory segmentation results, poor segmentation accuracy and low robust performance. Summary of the invention

[0005] In order to solve the above technical problems, this application proposes the following technical solutions:

[0006] In a first aspect, an embodiment of the present application provides a prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning, comprising: reading multi-parameter magnetic resonance imaging raw data acquired by a low-field magnetic resonance device;

[0007] Preprocessing the multi-sequence data and the label data in the original data respectively and dividing them into a training set, a test set and a validation set;

[0008] Construct an attention-unet network for prostate segmentation and a Mulit-attention-unet network for prostate cancer segmentation respectively;

[0009] Five-fold cross validation was used to input each group of multi-parameter magnetic resonance images belonging to the training set into the attention-unet network for training, and the optimal prostate segmentation model was selected using the validation set to achieve prostate segmentation.

[0010] Inputting the multi-parameter magnetic resonance image of the test set into the optimal prostate segmentation model for testing to obtain the optimal prediction mask of the gland;

[0011] Using five-fold cross validation, the optimal prediction mask of the gland and the multi-parameter magnetic resonance image in the training set are input into the Mulit-attention-unet network for training, and the optimal prostate cancer segmentation model is selected using the validation set;

[0012] The multi-parameter magnetic resonance image of the test set and the gland optimal prediction mask are input into the optimal prostate cancer segmentation model to obtain the prostate cancer segmentation result.

[0013] In a possible implementation, preprocessing is performed on the multi-sequence data and the label data in the raw data respectively, including: performing multi-modality registration of T2WI and DWI / ADC images by rigid transformation according to the coordinate information in the multi-sequence data header file;

[0014] After the registration is completed, all data are resampled according to the voxel information in the multi-sequence data header file, so that the (x, y) voxels are all (1, 1) and the z-axis voxels are the same as T2WI;

[0015] All sequences were centrally cropped to remove blood vessels, bladder, rectum, and seminal vesicle structures, and the image block size was adjusted to (64, 64, 16);

[0016] The image block is enlarged to (256, 256, 16) by using spline interpolation, and normalized, and the preprocessed slice data is saved.

[0017] In a possible implementation, the attention-unet network includes a Conv2D convolution layer, a MaxPool2d pooling layer, an UpsamplingBilinear2d upsampling layer, an attention gating module, and a skip connection module;

[0018] The Conv2D convolution layer is used to complete feature extraction of multi-parameter magnetic resonance imaging, and to increase and decrease the number of channels;

[0019] The MaxPool2d pooling layer is used to downsample the convolutional features;

[0020] The UpsamplingBilinear2d upsampling layer is used to restore the spatial size of the features extracted by downsampling, gradually increasing the size of the feature map, so as to obtain an output of the same size as the input image;

[0021] The attention gating module is used to fuse the features of each convolution layer of the downsampling with the features of the next convolution layer of the corresponding upsampling, adjust the weight of the downsampling features according to the characteristics of the upsampling features, and reduce the semantic gap between the upsampling and downsampling;

[0022] The jump connection module is used to connect the down-sampled features with the features of the corresponding up-sampling layer after they pass through the attention gating module, so as to achieve a refined segmentation result.

[0023] In a possible implementation, the Mulit-attention-unet network introduces a multi-scale module to replace ordinary convolution on the basis of the attention-unet network, uses multi-task learning to classify and calculate the prediction results output by the segmentation network while performing prostate cancer segmentation, and designs a composite loss function based on the segmentation and classification results.

[0024] In a possible implementation, the multi-scale module consists of three 3*3 convolutions and one 1*1 convolution. The output features after the three 3*3 convolutions are connected in series, and the initial input features are subjected to a 1*1 convolution to capture objects of different sizes and detail information.

[0025] In a possible implementation, the calculation formula of the composite loss function is:

[0026] Loss=w1×(loss_seg+loss_extra)+w2×loss_classif

[0027] w1=1―w2

[0028] loss_extra=a×max(p1―t1,0)×10 ―4 +b×max(t1―p1,0)×10 ―4

[0029]

[0030]

[0031] Among them, w1 and w2 are the weights of segmentation loss and classification loss respectively, loss_seg+loss_extra is the segmentation loss used to measure the similarity between the predicted lesion and the gold standard, loss_classif is the classification loss, loss-extra is the additional loss, a and b are control parameters, p1 and t1 are the sum of the grayscale values ​​of the predicted lesion mask and the gold standard image respectively, outpout and target are the predicted lesion mask and the gold standard image respectively, and the size is [H, W].

[0032] In a possible implementation, the five-fold cross validation is used to input each group of multi-parameter magnetic resonance images belonging to the training set into the attention-unet network for training and the optimal prostate segmentation model is selected using the validation set to segment the prostate body, including:

[0033] Divide the preprocessed raw data into training set, test set and validation set;

[0034] The multi-parameter magnetic resonance images in the training set are mirrored, randomly rotated, and center-cropped to enhance the data in the training set;

[0035] The multi-parameter magnetic resonance images belonging to the training set are respectively input into the attention-unet network for prostate segmentation model training;

[0036] Use the corresponding validation set of each group to adjust the parameters, and select the best segmentation model in the model for pre-segmentation of glands.

[0037] In a possible implementation, the step of using five-fold cross validation to input the gland optimal prediction mask and the multi-parameter magnetic resonance image in the training set into a Mulit-attention-unet network for training and using the validation set to select the optimal prostate cancer segmentation model includes:

[0038] Divide the preprocessed raw data into training set, test set and validation set;

[0039] mirroring, randomly rotating, center cropping, and adding noise to the multi-parameter magnetic resonance image data in the training set to enhance the data in the training set;

[0040] Each group of multi-parameter magnetic resonance images belonging to the training set and the gland's optimal prediction mask are input into the Mulit-attention-unet network to train the prostate cancer segmentation model, and the corresponding validation set of each group is used to adjust the parameters to obtain the optimal prostate cancer segmentation model.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] This application uses glandular masks and multi-parameter magnetic resonance images as channel inputs to the prostate segmentation model, which not only retains background information but also makes the network focus more on the glandular part, and at the same time illustrates the importance of glandular tissue and background information around the gland to prostate segmentation.

[0043] This application uses attention-unet combined with a multi-scale module to fully extract global and local information, and designs targeted losses to improve segmentation accuracy. Experiments have shown that this application has the advantages of high calculation accuracy, fast time, good robustness, etc., and can obtain reliable and stable results. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flowchart of a prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning provided in an embodiment of the present application;

[0045] Figure 2 The attention-unet network framework diagram provided for the embodiment of the present application;

[0046] Figure 3 A diagram of the Mulit-attention-unet network framework provided for an embodiment of the present application;

[0047] Figure 4 Prediction graph for prostate cancer segmentation based on different ranges of background information provided in the embodiments of the present application;

[0048] Figure 5 Three glandular masks and background information utilization methods provided in the embodiments of the present application;

[0049] Figure 6 A prediction graph for prostate and prostate cancer segmentation using a model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] The present solution is described below in conjunction with the accompanying drawings and specific implementation methods.

[0051] Figure 1 A flow chart of a prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning provided in an embodiment of the present application, see Figure 1 In this embodiment, a prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning includes:

[0052] S101, reading the original data of multi-parameter magnetic resonance images acquired by a low-field magnetic resonance device.

[0053] In this embodiment, the original image data includes T2WI, DWI and ADC sequences.

[0054] S102, preprocessing the multi-sequence data and label data in the original data respectively and dividing them into a training set, a test set and a validation set.

[0055] When mp-MRI data is collected, due to different acquisition protocols and the displacement of the patient's body or the non-rigid deformation of the internal organs induced by the patient's breathing and systemic circulation, there will be spatial misalignment between the MRI images with different parameters (such as T2w and DWI). If mp-MRI data is used for PCa segmentation and diagnosis, image registration between multi-parameter MRI images is an essential step. In addition, before network training, the image needs to be resampled, center-cropped, and normalized to achieve uniform resolution, remove irrelevant areas, and eliminate noise in the image.

[0056] In this embodiment, according to the coordinate information in the multi-sequence data header file, the T2WI and DWI / ADC images are multimodally registered by rigid transformation. After the registration is completed, all data are resampled according to the voxel information in the multi-sequence data header file, so that the (x, y) voxels are all (1, 1), and the z-axis voxels are the same as T2WI. All sequences are centrally cropped to remove blood vessels, bladder, rectum and seminal vesicle structures, so that the image block size is adjusted to (64, 64, 16), and the image block is enlarged to (256, 256, 16) by spline interpolation, and normalized, and the gray value range is [0, 1], and the pre-processed slice data is saved.

[0057] S103, constructing an attention-unet network for prostate segmentation and a Mulit-attention-unet network for prostate cancer segmentation respectively.

[0058] Since prostate MRI images contain more human tissue structures, the prostate area only occupies a small part of the image. If prostate cancer lesions are segmented directly on prostate MRI images, they are easily interfered by other tissue structure information, resulting in poor segmentation results. Therefore, reliable prostate segmentation plays an important role in the automated diagnosis of prostate cancer. When the auxiliary diagnosis of prostate cancer is limited to prostate images, the network can focus on the region of interest, further improving the accuracy of diagnosis.

[0059] In this embodiment, in order to improve the segmentation effect, the attention-unet network model is constructed by combining 2D U-Net and the attention gating module. Figure 2As shown in the figure, it mainly includes: Conv2D convolution layer, MaxPool2d pooling layer, UpsamplingBilinear2d upsampling layer, jump connection module and attention gating module. Among them, the Conv2D convolution layer is composed of two 3*3 convolutions in series, which is equivalent to a 5*5 convolution operation. It is used to complete the feature extraction of multi-parameter magnetic resonance imaging and increase and decrease the number of channels. Each layer uses the Relu activation function. The MaxPool2d pooling layer is used to downsample the features after convolution. On the one hand, it makes the features smaller and simplifies the network calculation complexity. On the other hand, it compresses the features and extracts the main features. The UpsamplingBilinear2d upsampling layer is used to restore the spatial size of the features extracted after downsampling, gradually increase the size of the feature map, and obtain an output of the same size as the input image. The attention gating module fuses the features of each downsampled convolution layer with the features of the next upsampled convolution layer, and adjusts the weights of the downsampled features according to the characteristics of the upsampled features, so that the network can pay more attention to the features of the downsampled layer that match the current upsampled layer, reducing the semantic gap between upsampling and downsampling. The skip connection module is used to connect the downsampled features with the features of the corresponding upsampled layer after they pass through the AG module, to help refine the segmentation results.

[0060] In addition, due to the small size of prostate cancer lesions, blurred boundaries and differences between different imaging sequences, it is challenging to achieve accurate prostate cancer segmentation. Attention-unet can suppress the activation of irrelevant features and improve the sensitivity and accuracy of the label prediction model. Therefore, in order to more effectively utilize the gland mask and background information, this embodiment introduces a multi-scale module instead of ordinary convolution on the basis of the attention-unet network, and designs a composite loss function to improve the prostate cancer segmentation effect and accuracy. The specific network structure of Mulit-attention-unet is as follows: Figure 3 shown.

[0061] Among them, the multi-scale module is mainly composed of 3 3*3 convolutions and 1 1*1 convolution. After each 3*3 convolution, the features are output and connected in series, which is equivalent to extracting features of 3*3, 5*5, and 7*7 scales respectively, and performing 1*1 convolution on the initial input features to prevent the loss of early information. Through the multi-scale method, receptive fields of different scales can be obtained, and gland and background information can be fully extracted, thereby capturing targets and details of different sizes. Multi-task learning is used to classify and calculate the prediction results output by the segmentation network while performing prostate cancer segmentation, and Loss is designed based on the segmentation and classification results:

[0062] Loss=w1×(loss_seg+loss_extra)+w2×loss_classif

[0063] w1=1―w2

[0064] loss_extra=a×max(p1―t1,0)×10 ―4 +b×max(t1―p1,0)×10 ―4

[0065]

[0066]

[0067] Among them, w1 and w2 are the weights of segmentation loss and classification loss respectively. In this embodiment, w2 is used as a network parameter to train together with the network to achieve the purpose of adaptively adjusting the balance between classification and segmentation losses and improving training stability. Loss_seg+loss_extra is a segmentation loss used to measure the similarity between the predicted lesion and the gold standard. The main loss of Loss-seg is still the sum of Dice loss and cross entropy loss. The purpose of this combination is to allow the model to better capture the boundaries and details of the object while ensuring the segmentation accuracy. In addition, in order to further improve the segmentation accuracy, when the grayscale value of the predicted mask is significantly different from that of the gold standard image, an additional loss loss-extra is introduced. It can be seen in Formula 3 that the size of loss-extra can be adjusted by controlling parameters a and b. In this embodiment, a and b are both set to 1 to ensure that the same penalty value is given for predictions that are too large or too small. p1 and t1 are the sum of the grayscale values ​​of the predicted lesion mask and the gold standard image, respectively. Outpout and target are the predicted lesion mask and the gold standard image, respectively, and the size is [H, W]. In addition, this embodiment adds a classification loss loss-classif to determine whether there is a lesion in the slice, and mainly uses a DL function to control loss-classif, aiming to improve the sensitivity of the lesion and alleviate the imbalance of categories.

[0068] S104, using five-fold cross validation, input each group of multi-parameter magnetic resonance images belonging to the training set into the attention-unet network for training and use the validation set to select the optimal prostate segmentation model to achieve prostate body segmentation.

[0069] Five-fold cross validation is a model evaluation method used to objectively evaluate the performance of the prostate segmentation model and select the optimal model. The data is divided into five folds, one fold is used as the test set, and the other four folds are used as the training set. They are trained on the attention-unet model and the model performance is evaluated on the corresponding validation set. The above steps are repeated five times, and a different fold is selected as the validation set each time. Finally, five models and corresponding performance indicators are obtained. By comprehensively evaluating the performance of these five models, the prostate segmentation model results can be obtained more objectively, the influence of overfitting and contingency can be reduced, and the reliability of model evaluation can be improved.

[0070] In this embodiment, the preprocessed raw data and label data are divided into data sets, and 45 cases are taken as the test set. This set of data has been kept confidential until it is used to verify it. For internal verification, a five-fold cross-validation is used to divide the remaining data into five non-overlapping subsets, and one of the subsets is taken as the verification set. The remaining four subsets are used as training sets, and a total of five sets of training and verification data are obtained. The training data is mirrored, randomly rotated (the rotation angle is within 10°), and center cropped to enhance the data in the training set. The 5 groups of multi-parameter magnetic resonance images belonging to the training set are respectively input into the attention-unet for prostate segmentation model training, and the corresponding verification set of each group is used to adjust the parameters. Finally, the best segmentation model among the five models is selected for pre-segmentation of the gland.

[0071] S105, inputting the multi-parameter magnetic resonance image of the test set into the optimal prostate segmentation model for testing to obtain an optimal prediction mask for the gland.

[0072] S106, using five-fold cross validation, input the gland optimal prediction mask and the multi-parameter magnetic resonance image in the training set into the Mulit-attention-unet network for training and use the validation set to select the optimal prostate cancer segmentation model.

[0073] In this embodiment, the preprocessed raw data is divided into a training set, a test set, and a validation set, and the same data division group is used for prostate segmentation. The training data is mirrored, randomly rotated (the rotation angle is within 10°), center cropped, and noise is added (within 0.001, that is, each pixel value fluctuates randomly within one thousandth) to enhance the data in the training set. Each group of multi-parameter magnetic resonance images belonging to the training set and the corresponding gland mask are input into Mulit-attention-unet to train the prostate cancer segmentation model, and the corresponding validation set of each group is used to adjust the parameters to obtain the optimal lesion segmentation model.

[0074] S107, inputting the multi-parameter magnetic resonance image of the test set and the gland optimal prediction mask together into an optimal prostate cancer segmentation model to obtain a prostate cancer segmentation result.

[0075] The effect of the prostate cancer segmentation method based on multi-parameter magnetic resonance imaging in this embodiment is further illustrated by the following experiment.

[0076] 1. Simulation experiment:

[0077] First, the images and header information of T2WI, ADC, and DWI are read, and then rigid registration is performed based on the position information, and experiments are performed on the processed magnetic resonance images through a pre-written program.

[0078] Experiment 1: Verify the hypothesis that prostate periphery information is useful for prostate cancer segmentation

[0079] Attention-UNET is used to train and test the prostate cancer segmentation model using multi-parameter magnetic resonance images containing only glandular tissue, multi-parameter magnetic resonance images containing partial glandular surrounding information, and multi-parameter magnetic resonance images containing all pelvic information, respectively.

[0080] Experiment 2: Prostate cancer segmentation combining prostate segmentation and background information

[0081] The attention-unet network is first used to segment the glands, and then the gland mask is input into the lesion segmentation network together with the multi-parameter MRI image containing a small part of the gland surrounding information, so as to achieve the purpose of retaining the background information and focusing the network attention on the gland part. The model is compared with the model that only contains the multi-parameter MRI image of the gland as input and the model that uses the multi-parameter MRI image containing a small part of the gland surrounding information as input.

[0082] Experiment 3: Prostate cancer segmentation based on Mulit-attention-unet

[0083] The combined network is improved by combining attention-unet with the multi-scale module to fully extract global and local information, and a targeted loss is designed to improve segmentation accuracy.

[0084] 2. Simulation Experiment Results

[0085] Table 1 Comparison of the effects of different ranges of prostate surrounding information on prostate cancer segmentation model

[0086]

[0087] For the simulation results of Experiment 1, see Figure 4 and Table 1, where Figure 4The first row of images in the figure shows the result of taking only the gland as input. Figure 4 The image in the second row is the original volume 100mm 2 As input, a magnetic resonance image of size Figure 4 The image in the third row is the original volume 150mm 2 As input, a magnetic resonance image of size Figure 4 The image in the fourth row is the original volume 200mm 2 As input, a magnetic resonance image of size Figure 4 The image in the fifth row is the original volume 256mm 2 The input is a magnetic resonance image of size .

[0088] Table 1 compares the results of the effects of different ranges of prostate peripheral information on the prostate cancer segmentation model. Since the network is often more likely to learn the area with a background of 0 when there is no background, the network is insensitive to the recognition of lesion information and the network cannot be fitted. Therefore, for the lesion segmentation experiment with background information removed, this embodiment uses a multi-task learning network to connect the prostate segmentation and prostate cancer segmentation networks in series, and multiply the output of the gland segmentation network with the original image to take only the glandular part of the image for lesion segmentation. It can be seen from these experiments that compared with the experiment without background information, the experiment with background information has higher lesion segmentation accuracy, which is most conducive to improving the network segmentation accuracy when it contains a small amount of glandular peripheral information.

[0089] Table 2 shows the comparison of the results of three gland masks and background information utilization methods when the center crop size is 100*100

[0090]

[0091] For the simulation results of Experiment 2, see Figure 5 and Table 2, where Figure 5 (a) is the prostate cancer segmentation by multiplying the gland mask with the original image. Figure 5 (b) shows prostate cancer segmentation by cropping the original image using only the gland mask. Figure 5 (c) is the use of glandular mask and original Figure 1 The same is used as a channel for prostate cancer segmentation.

[0092] Table 2 is a comparison of the results of three gland masks and background information utilization methods when the center crop size is 100*100. It can be seen from Table 2 that when only gland tissue is used, the accuracy of lesion segmentation can be improved, but compared with the network using glands as input, the accuracy of lesion segmentation is reduced. The network using gland masks together with multi-parameter magnetic resonance images containing a small part of the gland surrounding information can both obtain higher segmentation accuracy and improve segmentation accuracy.

[0093] Table 3 shows the results of prostate cancer segmentation

[0094]

[0095] For the simulation results of Experiment 3, see Figure 6 and Table 3, where Figure 6 The first and second rows of images are the detection of small lesions. Figure 6 The images in the third and fifth rows are large lesion detections, and their dice coefficients are 0.49, 0.88, 0.89, 0.89, and 0.95, respectively. It can be seen that compared with the attention-unet network, the Mulit-attention-unet network designed in this embodiment significantly improves the segmentation precision and accuracy.

[0096] Table 3 shows the results of prostate and prostate cancer segmentation. The experimental results show that compared with the attention-unet network, the Mulit-attention-unet network proposed by combining gland mask with background information can make better use of gland information and background information, and its segmentation dice can reach 0.71 and the segmentation accuracy can reach 0.88.

[0097] 3. Simulation Experiment Analysis

[0098] From the above figures and the results in Tables 1-3, we can draw the following conclusions:

[0099] Experiments have shown that the prostate cancer segmentation algorithm proposed in this application that combines prostate segmentation with background information has the advantages of high calculation accuracy, fast time, good robustness, etc. It has great advantages in prostate cancer segmentation and can obtain reliable and stable results.

[0100] In the embodiments of the present application, "at least one" means one or more, and "more" means two or more. "And / or" describes the association relationship of the associated objects, indicating that three relationships may exist. For example, A and / or B can mean that A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. The character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0101] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0102] The above is only a specific implementation of the present application. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. The protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning, characterized in that: include: Reading the original data of multi-parameter magnetic resonance images acquired by low-field magnetic resonance equipment; Preprocessing the multi-sequence data and the label data in the original data respectively and dividing them into a training set, a test set and a validation set; Construct an attention-unet network for prostate segmentation and a Mulit-attention-unet network for prostate cancer segmentation respectively; Five-fold cross validation was used to input each group of multi-parameter magnetic resonance images belonging to the training set into the attention-unet network for training, and the validation set was used to select the optimal prostate segmentation model to achieve prostate segmentation. Inputting the multi-parameter magnetic resonance image of the test set into the optimal prostate segmentation model for testing to obtain the optimal prediction mask of the gland; Using five-fold cross validation, the optimal prediction mask of the gland and the multi-parameter magnetic resonance image in the training set are input into the Mulit-attention-unet network for training, and the optimal prostate cancer segmentation model is selected using the validation set; The multi-parameter magnetic resonance image of the test set and the gland optimal prediction mask are input into the optimal prostate cancer segmentation model to obtain the prostate cancer segmentation result.

2. The prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning according to claim 1, characterized in that: Preprocessing the multiple sequence data and label data in the original data respectively includes: According to the coordinate information in the multi-sequence data header file, multimodal registration of T2WI and DWI / ADC images was performed through rigid transformation; After the registration is completed, all data are resampled according to the voxel information in the multi-sequence data header file, so that the (x, y) voxels are all (1, 1) and the z-axis voxels are the same as T2WI; All sequences were centrally cropped to remove blood vessels, bladder, rectum, and seminal vesicle structures, and the image block size was adjusted to (64, 64, 16); The image block is enlarged to (256, 256, 16) by using spline interpolation, and normalized, and the preprocessed slice data is saved.

3. The prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning according to claim 1, characterized in that: The attention-unet network includes a Conv2D convolution layer, a MaxPool2d pooling layer, an UpsamplingBilinear2d upsampling layer, an attention gating module, and a skip connection module; The Conv2D convolution layer is used to complete feature extraction of multi-parameter magnetic resonance imaging, and to increase and decrease the number of channels; The MaxPool2d pooling layer is used to downsample the convolutional features; The UpsamplingBilinear2d upsampling layer is used to restore the spatial size of the features extracted by downsampling, gradually increasing the size of the feature map, so as to obtain an output of the same size as the input image; The attention gating module is used to fuse the features of each convolutional layer of the downsampling with the features of the next convolutional layer of the corresponding upsampling, adjust the weight of the downsampling features according to the characteristics of the upsampling features, and reduce the semantic gap between the upsampling and downsampling; The jump connection module is used to connect the down-sampled features with the features of the corresponding up-sampling layer after they pass through the attention gating module, so as to achieve a refined segmentation result.

4. The prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning according to claim 1, characterized in that: The Mulit-attention-unet network introduces a multi-scale module to replace ordinary convolution on the basis of the attention-unet network, uses multi-task learning to classify and calculate the prediction results output by the segmentation network while performing prostate cancer segmentation, and designs a composite loss function based on the segmentation and classification results.

5. The prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning according to claim 4, characterized in that: The multi-scale module consists of three 3*3 convolutions and one 1*1 convolution. The output features after the three 3*3 convolutions are connected in series, and the initial input features are subjected to a 1*1 convolution to capture objects of different sizes and detail information.

6. The prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning according to claim 4 or 5, characterized in that: The calculation formula of the composite loss function is: Loss=w1×(loss_seg+loss_extra)+w2×loss_classif w1=1―w2 loss_extra=a×max(p1―t1,0)×10 ―4 +b×max(t1―p1,0)×10 ―4 Among them, w1 and w2 are the weights of segmentation loss and classification loss respectively, loss_seg+loss_extra is the segmentation loss used to measure the similarity between the predicted lesion and the gold standard, loss_classif is the classification loss, loss-extra is the additional loss, a and b are control parameters, p1 and t1 are the sum of the grayscale values ​​of the predicted lesion mask and the gold standard image respectively, outpout and target are the predicted lesion mask and the gold standard image respectively, and the size is [H, W].

7. The prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning according to claim 1, characterized in that: The method uses five-fold cross validation to input each group of multi-parameter magnetic resonance images belonging to the training set into the attention-unet network for training and uses the validation set to select the optimal prostate segmentation model to segment the prostate body, including: Divide the preprocessed raw data into training set, test set and validation set; The multi-parameter magnetic resonance images in the training set are mirrored, randomly rotated, and center-cropped to enhance the data in the training set; The multi-parameter magnetic resonance images belonging to the training set are respectively input into the attention-unet network for prostate segmentation model training; Use the corresponding validation set of each group to adjust the parameters, and select the best segmentation model in the model for pre-segmentation of glands.

8. The prostate cancer segmentation method based on multi-parameter magnetic resonance imaging and deep learning according to claim 1, characterized in that: The method of using five-fold cross validation to input the gland optimal prediction mask and the multi-parameter magnetic resonance image in the training set into the Mulit-attention-unet network for training and using the validation set to select the optimal prostate cancer segmentation model includes: Divide the preprocessed raw data into training set, test set and validation set; mirroring, randomly rotating, center cropping, and adding noise to the multi-parameter magnetic resonance image data in the training set to enhance the data in the training set; Each group of multi-parameter magnetic resonance images belonging to the training set and the gland's optimal prediction mask are input into the Mulit-attention-unet network to train the prostate cancer segmentation model, and the corresponding validation set of each group is used to adjust the parameters to obtain the optimal prostate cancer segmentation model.

Images