CN116542924A - Method, device and storage medium for detecting prostate lesion area - Google Patents

Abstract

The invention discloses a prostate lesion area detection method, device and storage medium, and relates to the field of computer vision. Accurate automatic prostate lesion area segmentation is an important requirement for computer-aided diagnosis and treatment of prostate diseases. However, the lack of clear boundaries and high contrast between the prostate and surrounding tissues makes it difficult to accurately extract the prostate from the background. Due to the scarcity of information in medical images, multiple methods must be used for multi-dimensional information extraction, but current methods cannot fully learn and mine hidden information. In order to deal with the above challenges, the present invention proposes a brand-new prostate lesion area detection method, device and storage medium suitable for the field of computer vision. In order to improve the quality of extracted information at different scales, different volumes are used in the proposed network. Active combinations and various attention mechanisms. So that it can have superior performance in different imaging methods.

Description

Method, device and storage medium for detecting prostate lesion area

technical field

The invention relates to the field of computer vision, in particular to a method, device and storage medium for detecting a prostate lesion area based on a dynamic multi-scale perception self-adaptive integration network.

Background technique

Prostate segmentation refers to detecting the prostate region in an input prostate MR image, and if it is prostate lesion segmentation, it is to separate the lesion from the original image if there is one. Since medical images are very different from conventional natural images, the performance of conventional segmentation algorithms directly applied to medical image segmentation will inevitably decline. This is especially true for the segmentation of prostate lesion regions. Accurate segmentation of the prostate and lesion areas from magnetic resonance (MR) images is crucial for the diagnosis and treatment planning of prostate diseases, especially prostate cancer, prostatitis, benign prostatic hyperplasia and other common prostate diseases in the general diagnosis of the disease , medical images are usually segmented manually by several experts, and this is a time-consuming process. With the development of deep learning, convolutional neural networks have made great progress in the field of prostate and lesion segmentation, which has led to more branch research on prostate and prostate lesion segmentation. And these studies have played a very important role in both the theoretical research on the prostate and its diseases, and the first-line medical practice.

According to different input forms, prostate and prostate lesion segmentation can be divided into two categories: 2D prostate and its lesion area detection and 3D prostate and its lesion area detection. Among them, the 2D prostate and its lesion area detection input is a single MR image, and the number of channels is a grayscale image. The input of 3D detection of prostate and its lesion area is multiple continuous MR images. This is precisely because for a patient, the imaging of the prostate and its lesion area is several consecutive images, that is, in the axial direction Different slices, since the slices of the same patient can be linked, so a 3D detection of the prostate and its lesion area is produced. With the advancement of technology and hardware, the detection of the prostate and its lesion area whose input is video has also emerged recently, that is, more input information is used to achieve more accurate segmentation results.

However, the current state-of-the-art automatic segmentation methods for MR images face several challenges. The lack of a clear border and high contrast between the prostate and surrounding tissues makes it difficult to accurately extract the prostate from the background. In addition, the complexity of the background texture and large variations in the scale, shape, and intensity distribution of the prostate itself complicate segmentation. In addition, the characteristics of the prostate lesion area are more difficult to distinguish visually than the prostate area. Professional doctors also need to check the data in multiple modalities to determine the location of the lesion area. At the same time, due to the scarcity of information in medical images, multiple methods must be used for multi-dimensional information extraction, so that the rich information extracted by the network in various ways can make up for the lack of innate information in the image, but the current methods cannot fully learn and mine images. Hidden information in .

Contents of the invention

Aiming at the current prostate lesion area detection method still using the conventional fixed parameter layer to infer the input prostate image, it is difficult to adapt to the problems of prostate size change and boundary blur in the input image, the present invention provides a dynamic multi-scale perception based self-adaptive The network-integrated prostate lesion area detection method uses the input prostate MR image for lesion area detection, and optimizes and updates through dynamic local pooling and global efficient attention network, achieving high-quality detection in a given prostate image scene. Lesion area detection.

For this reason, the invention provides the following technical solutions:

A method for detecting a prostate lesion area based on a dynamic multi-scale perception adaptive integration network, comprising the following steps:

A. Obtain the input image of the prostate according to the dataset of the prostate and obtain a tensor;

B. Input the tensor into the feature encoder, and obtain the multi-scale encoding features based on each image through the feature encoder;

C. For coding features, a richer feature representation is obtained through the corresponding feature enhancement layer

D. Use the decoder to perform feature decoding on richer feature representations to obtain the final prostate segmentation prediction results, including:

D1. Through the multi-level integration module with convolution, the multi-level feature pyramid is established by using the complementary characteristics of the levels, and the multi-scale information of an image is adaptively encoded into the current pyramid feature vector, and the global and local features are obtained. The feature information integrated at different levels;

The mechanism of hierarchical complementarity includes: the features extracted by the network include local features and global features. Among them, local features are extracted by shallow convolutional networks, which mainly contain information about image details, textures, etc., which do not distinguish whether they belong to the foreground or background; while global features are mainly obtained through deep layers. The convolution module and the attention mechanism are extracted, and the global features mainly include parts related to high-level semantic information, such as the position, the difference between the foreground and the background, and so on. It is precisely because the focus of these two features is local and global respectively, and for the segmentation of prostate lesions, location and detail information are equally important, so the complementarity of global and local features between different levels can be used when integrating features to obtain better prediction results.

The integration module includes: obtaining two consecutive layers of integrated feature maps through convolution and fusion operations; then performing channel transformation and normalization operations on the obtained integrated feature maps to obtain the output of a certain layer of integration module. Among them, the convolution can use multiple convolution kernels of different sizes to fuse neighborhood information or transform the number of channels for better fusion in the next step.

D2. Dynamic fusion of feature layers between different levels in a gradual manner in multiple stages, so as to obtain more accurate and richer fusion feature representations. The fusion weights are automatically learned through convolution operations. Through adaptive dynamic weighted fusion of features between different levels, the final prediction result map of prostate segmentation is obtained.

Further, step A includes:

Divide a fixed number of training sets and test sets according to the prostate data set, where both the training set and the test set have two subsets, namely the input image and the true value;

The first is to perform data augmentation on the data in the prostate training set, including but not limited to, resizing the input prostate image to H×W, and the resizing can be set in use so that it can best match Network model; followed by random flipping and random cropping using random probability; format conversion of the enhanced image to convert it into a tensor that can be processed by the network, thereby obtaining a batchsize tensor. It is worth noting that the above operations are also performed on the ground truth values in the prostate training set.

For the data in the prostate test set, it is different from the above operation. First, the size of the input image will be adjusted, and then the image will be directly tensorized, and the obtained tensor of batch size will be directly sent to the network for testing.

Further, the batchsize is 8; H×W is 224×224.

Further, the feature encoder is a ResNet architecture, and the last two layers are discarded to preserve the spatial structure, and then a global-local complementary module is added after the output of each layer to extract multi-scale context information; and will be except the first The features of all layers other than the layer are stored in the feature pyramid to facilitate the operation of subsequent modules. That is, the feature encoder generates a feature pyramid for each image, including 4 feature maps with different spatial resolutions and channel numbers.

Further, the ResNet architecture is a ResNet-50 architecture, in which the number of input channels of the first convolution module is modified to 1 to adapt to the number of channels of the input image. The global-local complementary module is a module containing two branches, namely the dynamic local pooling branch and the global efficient attention branch.

Dynamic local pooling includes, at each layer of the network, according to the position of the layer in the entire network, dynamically assigning a pooling layer combination of an appropriate size to the layer to better extract local information. That is, in the low layer of the network, a large number of combinations with a large convolution kernel size is used to adapt to the large-sized low-level feature map; while in the high layer of the network, a combination of a small number and a small convolution kernel size is used to adapt to the small Dimensions of high-level feature maps while preventing information loss.

Global efficient attention includes the operation of directly comparing the similarity of the input feature map to obtain a similarity weight map. If the dimension of the feature map of this layer is high, the dimensionality can be reduced first to save computing resources. Then use the combination of the weight map and the input image to calculate which areas should be focused on by the network in the global scope.

Further, the convolution kernels of dynamic local pooling are 1, 3, 5, 7, and 9 in sequence. Among them, those with a size greater than 3 use depth-wise separable convolutions instead of ordinary convolutions to reduce the amount of parameters.

Further, in step D1, if it is used for neighborhood fusion, the size of the dynamic kernel K _t is 3×3; if it is used for channel number integration, the size is 1×1.

Further, step C includes:

In the corresponding feature fusion layer, the features corresponding to the scale of the feature encoder are used as input;

For the features of this scale, the two-stream attention mechanism and the residual-attention mechanism are used to further optimize the feature expression after the feature encoder;

For the feature representations of the same spatial resolution after different attention transformations, pixel-level summation is used to obtain a richer feature representation after fusion.

Furthermore, for the features of each scale, different attention mechanisms are used to make the network focus on different information, thereby improving the richness and accuracy of feature representation, including:

For dual-stream attention, attention is instead used to emphasize the features of different regions. Use spatial attention and channel attention to calculate the foreground area that the network should focus on under normal circumstances, and then normalize and invert the weight map of the foreground attention to make it focus on the background and find out better The foreground information hidden in the background features is then used to obtain the fused feature representation by pixel-level summation and convolution operations.

For the residual-attention mechanism, the combination of the residual module with self-attention and mutual attention is used. That is, a multi-channel residual convolution operation is performed on the input feature representation. Between each branch, the information flow between branches is realized through the combination of self-attention weight and mutual attention weight, so as to obtain richer feature expression.

Further, the number of multi-branches in the residual-attention mechanism is set to 4, and the weight between the self-attention weight map and the mutual-attention weight map is 1.

The above-mentioned technical scheme provided by the present invention has the following beneficial effects:

The present invention proposes a prostate lesion area detection method based on a dynamic multi-scale perception self-adaptive integration network, which considers the coherence between multi-scale information in an input image. First, the multi-scale coding features based on each image are obtained through the feature encoder. In the feature encoder, dynamic pooling convolution is first used to focus on the local information of the image, and the global efficient attention mechanism is used to focus on the image. The global information of the image is integrated through pixel summation and convolution operations to maximize the extraction of effective information at different scales in the input image. Secondly, in order to avoid the negative impact of the noise in the information extraction process on the prediction results, the information enhancement operation is carried out by using multiple attention combinations for the feature maps at each scale. On the one hand, a dual-stream attention mechanism is used to enable the network to pay more attention to foreground information and suppress noise; on the other hand, the residual attention is used to enable the network to extract integrated information while strengthening the relationship between branches. Connections, using the complementary characteristics between branches to make more accurate predictions. Experimental results show that the prostate lesion area detection method based on the dynamic multi-scale perception self-adaptive integration network proposed by the present invention has an accurate prediction effect on prostate lesion area segmentation and prostate segmentation.

Based on the above reasons, the present invention can be widely promoted in the field of prostate and its lesion area.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description These are some embodiments of the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is the schematic diagram of prostate MR input in the embodiment of the present invention;

2 is a flowchart of a method for detecting a prostate lesion area based on a dynamic multi-scale perception adaptive integration network in an embodiment of the present invention;

Fig. 3 is a schematic structural diagram of a dynamic adaptive pooling module in an embodiment of the present invention;

Fig. 4 is a schematic structural diagram of a dual-stream attention module in an embodiment of the present invention;

Fig. 5 is a schematic structural diagram of the internal attention module of the residual-attention module in the embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Referring to FIG. 2 , it shows a flowchart of a method for detecting a prostate lesion area based on a dynamic multi-scale perception adaptive integration network in an embodiment of the present invention. The method includes the following steps:

A. Obtain the input image of the prostate according to the dataset of the prostate and obtain a tensor;

In specific implementation, step A specifically includes:

A1. Obtain the prostate image:

The schematic diagram of the input prostate MR is shown in Figure 1. The prostate data set is divided into a training set and a test set according to a certain ratio, and in the training set and the test set, there are two sub-data sets, namely the input image and the true value.

A2. For the input prostate image, get the tensor T with the number of channels as batchsize

Data enhancement is performed on the input images in the prostate training set and the corresponding true values. First, the input MR original image (that is, the single-channel grayscale image) and the GT image are randomly cropped with a scale of s and a ratio of r, and the size is adjusted. to H×W (the image resolution used in this method is 224×224), and then use the random flip of random probability; then convert the enhanced grayscale image into a tensor that can be processed by the network, and use the data loader to convert The number of channels of the output tensor is adjusted to the batch size, so that the network can converge more quickly, and the value of batch size here is set to 8.

For the input prostate image and the corresponding true value in the prostate test set, the operation of resizing it is used to adjust it to H×W (the image resolution used in this method is 224×224), and then adjust The final grayscale image of the prostate is converted into a tensor that can be processed by the network, and the number of channels of the output tensor is adjusted to the batchsize size by using the data loader, so that the network can converge more quickly. The value of batchsize here is set to 1.

B. Input the tensor into the feature encoder, and obtain the multi-scale encoding features based on each image through the feature encoder

In specific implementation, step B specifically includes:

B1. Input the obtained tensor I _t into the feature encoder:

The feature encoder used is the ResNet-50 architecture, in which the number of input channels of the first convolution module is modified to 1 to adapt to the number of channels of the input image, and the last fully connected layer is removed to adapt to subsequent modules.

B2. Obtain multi-scale coding features

The feature encoder will generate 5 multi-scale feature maps with different spatial resolutions and channel numbers for each image, that is, Its resolution and number of channels (W×H×C) are respectively Since the first layer has too much noise and less useful information, in the subsequent information enhancement, only the 1-5 layer features are operated.

Wherein, step B2 specifically includes:

Each layer contains a residual convolution module and a two-branch global-local complementary module, and these two modules are combined in series.

The residual convolution module is the convolution module with residual connections. The dual-branch global-local complementary module includes a dynamic adaptive local pooling module and a global efficient attention module. As shown in Figure 3, dynamic adaptive local pooling uses convolution kernels of different sizes for convolution. Among them, for low-level features, the more convolutions the dynamic adaptive pooling module contains, the convolution The larger the area involved in the product kernel, it is used to make up for the lack of information at the low level; for high-level features, because the information is already sufficient, more integration and transformation are needed, so the size of the convolution kernel is Smaller and fewer in number. The whole process can be expressed as:

layers(i)∈{RL(j)|j＝0,1}∪{OL(k)|1≤k≤5-i,k∈N ⁺ } (1)

Where layers(i) represents the feature map of the i-th layer. RL represents a mandatory part for all feature layers that contain dynamic adaptive local pooling modules. The RL part includes a global average pooling, upsampling, and a convolution operation with a convolution kernel size of 1. . The OL part means that for different feature layers, the composition of the OL part is different, and the size and number of convolution kernels are different (that is, the value range of k). The size of the convolution kernel is equal to 2k-1.

The global high-efficiency attention module uses the concept of similarity and combines the self-attention mechanism to enhance the ability of the network to focus on the global area, so that the network can better learn more information from the global perspective, so as to make more accurate predictions. And use residual connections to prevent information loss when the network is running. The whole process can be expressed as:

Among them, x represents the input feature map, and Normalize is the normalization operation.

At the end of each layer, the feature maps of the two branches are integrated through pixel-level summation and convolution operations.

C. For coding features, a richer feature representation is obtained through the corresponding feature enhancement layer

In specific implementation, step C specifically includes:

C1, the features of each scale Send it to the dual-stream attention module to enhance the foreground and suppress background noise:

Conventional attention can only focus on the foreground area, while ignoring the hidden foreground information in the background area. Therefore, in the dual-stream attention module in the embodiment of the present invention, the spatial attention mechanism and the channel attention mechanism are combined to make the network in the While paying attention to the foreground area, it can also pay attention to the hidden foreground information in the background area, as shown in Figure 4. At the end of the dual-stream attention module, the features of the foreground branch and the background branch will be integrated in the form of convolution, and the information in it will be learned for more accurate segmentation.

C2, the features of each scale Send it to the residual-attention module to further integrate information in the form of multi-branches, and use the attention mechanism to enhance the association between multiple branches in the form of residuals to jointly highlight foreground targets.

As shown in Figure 5, there are a total of 4 branches in the residual-attention module, and each branch has 2 residual convolution modules. Between each branch, an attention module is used to strengthen the association. The attention module is divided into two parts, one part is self-attention, which is used to generate a weighted matrix about itself; the other part is mutual attention, which uses the similarity of the two branches to generate the two branches The weight weighting matrix when the road communicates with each other. The output of each branch is obtained by the matrix weighted summation of the feature map of the adjacent branch and the feature map of its own branch through the above calculation. Finally, the results of the four branches will be integrated to obtain the output of the residual-attention module.

D. Represent richer features through the decoder Perform feature decoding to obtain the final segmentation prediction result of the prostate lesion area,

The decoder includes an integration module and a prediction module, and the decoder receives a feature pyramid composed of four feature maps of different scales from the previous module. The integration module of the decoder starts from the feature map with the smallest size and the highest number of layers (i.e., 7×7×2048). The final integrated feature map is obtained by gradually integrating the feature maps of two consecutive sizes. The integration operation is mainly divided into two steps: feature map stacking and feature transformation in the channel direction, and the final integrated feature map will be sent to a prediction module. The prediction module will integrate the feature map according to the number of classifications required by the task. category to predict.

Specific steps are as follows:

D1. Through the multi-level integration module with convolution, the multi-level feature pyramid is established by using the complementary characteristics of the levels, and the multi-scale information of an image is adaptively encoded into the current pyramid feature vector, and the global and local features are obtained. The feature information integrated at different levels. The mechanism of hierarchical complementarity includes: the features extracted by the network include local features and global features. Among them, local features are extracted by shallow convolutional networks, which mainly contain information about image details, textures, etc., which do not distinguish whether they belong to the foreground or background; while global features are mainly obtained through deep layers. The convolution module and the attention mechanism are extracted, and the global features mainly include parts related to high-level semantic information, such as the position, the difference between the foreground and the background, and so on. It is precisely because the focus of these two features is local and global respectively, and for the segmentation of prostate lesions, location and detail information are equally important, so the complementarity of global and local features between different levels can be used when integrating features to obtain better prediction results.

The integration module includes: obtaining two consecutive layers of integrated feature maps through convolution and fusion operations; then performing channel transformation and normalization operations on the obtained integrated feature maps to obtain the output of a certain layer of integration module. Among them, the convolution can use multiple convolution kernels of different sizes to fuse neighborhood information or transform the number of channels for better fusion in the next step.

D2. Dynamic fusion of feature layers between different levels in a gradual manner in multiple stages, so as to obtain more accurate and richer fusion feature representations. The fusion weights are automatically learned through convolution operations. Through adaptive dynamic weighted fusion of features between different levels, the final prediction result map of prostate segmentation is obtained. E. Training and optimization of dynamic context-aware filtering network:

The method as a whole can be divided into two stages of training and inference. During training, the tensor of the training set is used as input to obtain the trained network parameters; in the inference stage, the parameters saved in the training stage are used for testing to obtain the final saliency prediction. result.

The embodiment of the present invention is realized under the framework of Pytorch, wherein the ADAMW optimizer is used in the training phase, the learning rate is 1e ^-3 , the weight decay coefficient is 5e ^-2 , and the batch size is 8. During training, the images have a spatial resolution of 224×224, but the model can be applied to any resolution in a fully convolutional manner at test time.

The prostate lesion area detection method based on the dynamic multi-scale perception self-adaptive integration network proposed by the embodiment of the present invention adopts the dynamic local pooling and the global attention mechanism, and compiles the context information in the input prostate MR image into the current feature matrix , to obtain a feature vector containing global information and local detail information, which adapts to the scale change of the target. Secondly, in order to avoid misleading the final saliency results, the present invention adopts multiple attention complementary perceptual fusion methods, and uses multiple attentions to enhance the foreground and suppress background noise for the feature maps generated at each scale. Experimental results show that a prostate lesion area detection method based on a dynamic multi-scale perceptual self-adaptive integration network proposed by the present invention can achieve accurate prediction results for many scenes with prostate size changes and low-quality input images.

Corresponding to the prostate lesion area detection method in the above-mentioned embodiments, an embodiment of the present invention also provides a prostate lesion area detection device based on a dynamic multi-scale perception self-adaptive integration network, including:

The tensor unit is used to obtain the prostate input image and obtain the tensor according to the prostate cancer data set;

The encoding unit is used to input the tensor obtained by the tensor unit into the feature encoder, and use the dynamic pooling convolution to focus on the local information of the image through the feature encoder, and combine the global efficient attention mechanism to focus on the global information of the image, and through pixel summation And the convolution operation integrates local information and global information to obtain multi-scale coding features based on each image;

The enhancement unit is used to obtain the feature representation through the corresponding feature enhancement layer for the coding feature obtained by the coding unit;

The prediction unit is configured to perform feature decoding on the feature representation obtained by the enhancement unit through the decoder to obtain a final prediction result of prostate lesion region segmentation.

For the prostate lesion area detection device in the embodiment of the present invention, because it corresponds to the prostate lesion area detection method in the above embodiment, the description is relatively simple. For related similarities, please refer to the prostate lesion area detection in the above embodiment. The description in the method part is sufficient, and will not be described in detail here.

An embodiment of the present invention also provides a computer-readable storage medium, the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to perform the detection of prostate lesion areas based on the dynamic multi-scale perception adaptive integration network as described above method.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: It is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the various embodiments of the present invention. scope.

Claims (10)

1. A prostate lesion area detection method based on dynamic multi-scale perception self-adaptive integration network, is characterized in that, comprises the steps: Get the prostate input image and get the tensor according to the prostate cancer dataset; The tensor is input into the feature encoder, and the feature encoder uses the dynamic pooling convolution to focus on the local information of the image, and combines the global efficient attention mechanism to focus on the global information of the image, and through the operation of pixel addition and convolution, the Local information and global information are integrated to obtain multi-scale coding features based on each image; Obtaining a feature representation through a corresponding feature enhancement layer for the encoded feature; The feature representation is decoded by a decoder to obtain a final prostate lesion region segmentation prediction result. 2. according to claim 1, a kind of prostate lesion region detection method based on dynamic multi-scale perception self-adaptive integration network, is characterized in that, according to the data set of prostate cancer, obtain prostate input image and obtain tensor, comprising: The dataset partition according to the prostate has a fixed number of training and testing sets, where both the training and testing sets have two subsets, the input image and the ground truth; Perform data augmentation on the data in the prostate training set; Perform format conversion on the enhanced image to convert it into a tensor that the network can handle, and obtain a tensor of batchsize size; For the data in the prostate test set, the size of the input image is adjusted, and the image is directly tensorized to obtain a batchsize tensor. 3. according to claim 1, a kind of prostate lesion area detection method based on dynamic multi-scale perception self-adaptive integration network, is characterized in that: described feature coder is ResNet structure, and discards last two layers to retain spatial structure, then After the output of each layer, a global-local complementary module is added to extract multi-scale context information; and the features of all layers except the first layer are stored in the feature pyramid. 4. A video saliency detection method based on a dynamic context-aware filtering network according to claim 3, wherein the global-local complementary module includes a dynamic local pooling branch and a global high-efficiency attention branch; The dynamic local pooling includes, at each layer of the network, according to the position of the layer in the entire network, dynamically assigning a pooling layer combination of an appropriate size to the layer to better extract local information; The global high-efficiency attention includes the operation of directly comparing the similarity of the input feature map to obtain a similarity weight map; using the combination of the weight map and the input image to calculate the global range, which needs to be focused on area. 5. according to claim 1, a kind of prostate lesion area detection method based on dynamic multi-scale perception self-adaptive integration network, is characterized in that: carry out feature decoding to described feature representation by decoder, obtain final prostate lesion area segmentation prediction Results, including: Through the multi-level integration module with convolution, the multi-level feature pyramid is established by using the complementary characteristics of the levels, and the multi-scale information of an image is adaptively encoded into the current pyramid feature vector, and the global and local differences are obtained. Feature information after hierarchical integration; The feature layers between different levels are adaptively and dynamically weighted and fused in a gradual manner in multiple stages to obtain the final prediction result map of prostate segmentation, in which the fusion weight is automatically learned through convolution operation. 6. according to claim 1, a kind of prostate lesion area detection method based on dynamic multi-scale perception self-adaptive integration network, it is characterized in that, for described encoding feature, obtain feature representation by corresponding feature enhancement layer, comprising: In the corresponding feature fusion layer, the features corresponding to the scale of the feature encoder are used as input; For the features of this scale, the two-stream attention mechanism and the residual-attention mechanism are used to further optimize the feature expression after the feature encoder; For the feature representations of the same spatial resolution after different attention transformations, pixel-level summation is used to obtain a richer feature representation after fusion. 7. according to claim 6, a kind of prostate lesion region detection method based on dynamic multi-scale perception self-adaptive integration network, is characterized in that, described dual-stream attention mechanism uses spatial attention and channel attention to calculate The network should focus on the foreground area, and then normalize and invert the weight map of the foreground attention to make it focus on the background, and then use pixel-level summation and convolution operations to obtain the fused feature representation; The residual-attention mechanism uses the residual module, self-attention, and mutual-attention to perform multi-channel residual convolution operations on the input feature representation, and each branch is passed through the self-attention weight Combining with mutual attention weights, the information flow between branches is realized. 8. according to claim 7, a kind of prostate lesion area detection method based on dynamic multi-scale perception self-adaptive integration network, is characterized in that: the multi-branch number in the residual-attention mechanism is set to 4, self-attention The weight between the force weight map and the mutual attention weight map is 1. 9. A prostate lesion area detection device based on a dynamic multi-scale perception self-adaptive integration network, characterized in that it comprises: The tensor unit is used to obtain the prostate input image and obtain the tensor according to the prostate cancer data set; The encoding unit is configured to input the tensor obtained by the tensor unit into a feature encoder, and use the dynamic pooling convolution to focus on the local information of the image through the feature encoder, and pay attention to the global information of the image in combination with a global efficient attention mechanism, Integrate local information and global information through pixel summation and convolution operations to obtain multi-scale coding features based on each image; An enhancement unit, configured to obtain a feature representation through a corresponding feature enhancement layer for the coding features obtained by the coding unit; The prediction unit is configured to perform feature decoding on the feature representation obtained by the enhancement unit through the decoder to obtain a final prediction result of prostate lesion region segmentation. 10. A computer-readable storage medium, characterized in that, the computer-readable storage medium stores computer instructions, and the computer instructions are used to make a computer execute a computer-based computer program according to any one of claims 1-8. Prostate lesion region detection method with dynamic multi-scale perceptual adaptive integration network.