A multi-scale adversarial learning approach for tumor segmentation in medical images
Through the medical image tumor segmentation method of multi-scale adversarial learning, multi-level segmentation is performed using Res-UNet and MSALDS-UNet network models, combined with channel attention and discriminator, the problem of inaccurate tumor segmentation in the existing technology is solved, and precise segmentation of organs, tumors and tumor core areas is achieved.
Patent Information
- Application Number
- CN202211556140.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-06
- Publication Date
- 2025-04-15
- Estimated Expiration
- 2042-12-06
AI Technical Summary
The prior art is difficult to accurately distinguish organs, tumors and tumor core areas from normal tissues, resulting in inaccurate division of kidney and renal tumors.
The medical image tumor segmentation method of multi-scale adversarial learning is adopted, and the Res-UNet network model is roughly segmented through the Res-UNet network model to construct a three-level cascaded MSALDS-UNet adversarial learning network model, combining the channel attention module and discriminator for multi-scale adversarial learning, and using the difficult area supervision loss function for segmentation, and the organs, tumors and tumor core areas are divided layer by layer.
Accurate segmentation of organs, tumors and tumor core areas is achieved, segmentation accuracy is improved, especially the segmentation effect in boundary areas, errors and overfitting are reduced, and the performance of segmentation network is improved.
Smart Images
Figure CN116402835B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a medical image tumor segmentation method based on multi-scale adversarial learning. Background Art
[0002] Cancer is a major urinary disease worldwide. Tumors vary in morphology, are complex in type, and are difficult to locate. It is still difficult for radiologists and surgeons to manually segment tumors and core areas. Currently, medical imaging technologies such as computed tomography (CT) are widely used in tumor-related research. It is of great significance to predict the location, shape, and benign or malignant nature of tumors through deep learning methods. The accuracy of organ segmentation by existing deep learning methods is already very close to the manual level, but there is still a certain gap between the segmentation of tumors and manual segmentation. Due to the large variations in the size, position, intensity, and texture of tumors, small tumors are difficult to distinguish from normal tissues, and accurate tumor segmentation remains a challenging task. Summary of the invention
[0003] The technical problem to be solved by the present invention is that the segmentation technology in the prior art is difficult to accurately distinguish organs, tumors and tumor core areas from normal tissues, making it impossible to achieve accurate segmentation of kidneys and kidney tumors. The purpose is to provide a medical image tumor segmentation method based on multi-scale adversarial learning, which solves the problem that the prior art cannot achieve accurate segmentation of tumors.
[0004] The present invention is achieved through the following technical solutions:
[0005] A first aspect of the present invention provides a medical image tumor segmentation method based on multi-scale adversarial learning, comprising the following steps:
[0006] S1. Obtain a medical image, and use a Res-UNet network model to roughly segment the target area of the medical image to obtain a region of interest;
[0007] S2. Constructing a three-level cascaded MSALDS-UNet adversarial learning network model, inputting the region of interest into the three-level cascaded MSALDS-UNet adversarial learning network model for segmentation, and obtaining an organ region, a tumor region, and a tumor core region;
[0008] S3. Replace the organ region, the tumor region and the tumor core region into the medical image to obtain a segmented image.
[0009] In the above technical solution, the medical image is first roughly segmented through the Res-UNet network model to obtain the target region of interest. On the basis of the rough segmentation, a three-level cascade MSALDS-UNet network model is constructed to perform fine segmentation of the region of interest, transforming the multi-class segmentation problem into multiple simple binary segmentation problems, effectively alleviating the problem of sample imbalance. The MSALDS-UNet network model is a multi-scale adversarial learning difficult supervised network model, which applies adversarial learning strategies at multiple scales of the segmentation network to improve the final segmentation performance, thereby accurately segmenting organs, tumors, and tumor core areas. Organs, tumors, and tumor core areas are replaced in medical images to obtain segmented images, which accurately display the organs, tumors, and tumor cores in medical images.
[0010] In an optional embodiment, a three-level cascaded MSALDS-UNet adversarial learning network model is constructed, and a method for inputting the region of interest into the three-level cascaded MSALDS-UNet adversarial learning network model for segmentation is as follows:
[0011] S21, introduce the channel attention module in the encoding layer, introduce the discriminator in the decoding layer to construct the first-level MSALDS-UNet adversarial learning network model, the second-level MSALDS-UNet adversarial learning network model and the third-level MSALDS-UNet adversarial learning network model respectively;
[0012] S22, inputting the region of interest into the primary MSALDS-UNet adversarial learning network model for fine segmentation to obtain an organ region, and inputting the region of interest into the secondary MSALDS-UNet adversarial learning network model for fine segmentation to obtain a tumor region;
[0013] S23, inputting the tumor region into the three-level MSALDS-UNet adversarial learning network model for fine segmentation to obtain the tumor core region.
[0014] In an optional embodiment, the fine segmentation method is as follows:
[0015] The encoding layer extracts organ image features, and adjusts the organ image features through the channel attention module to obtain enhanced image features;
[0016] The decoding layer decodes the enhanced image features to obtain a feature map and a true label;
[0017] splicing the feature map and the true label with the medical image respectively, and sending the spliced image to the discriminator for processing to obtain a confidence map;
[0018] A difficult region supervision loss model is constructed according to the confidence map, and the difficult region that is difficult to segment is supervised by the difficult region supervision loss model.
[0019] In an optional embodiment, the method for adjusting the organ image features by the channel attention module is as follows:
[0020] The channel attention module uses the squeeze operation to perform feature compression in the spatial dimension to obtain a 1*1*C channel;
[0021] The degree of dependence on the channel is learned by using an excitation operation, and the organ image feature is adjusted according to the degree of dependence.
[0022] In an optional embodiment, before sending the spliced image to the discriminator for processing, the method further includes constructing a discriminator loss function to train the discriminator using a binary cross entropy loss;
[0023] Among them, the discriminator loss function is constructed as follows:
[0024]
[0025] Among them, L D is the discriminator loss, X is the input image, Y is the true mask, P is the decoding layer output, D is the discriminator, and L bce is the binary cross entropy loss.
[0026] In an optional embodiment, before performing fine segmentation, the method further includes constructing a segmentation network loss function to train the segmentation network in the MSALDS-UNet adversarial learning network model;
[0027] The method for constructing the segmentation network loss function is as follows:
[0028] L seg =L dice +L dce +λ1L adv
[0029] Where L seg is the segmentation network training loss function, L dice is the Dice loss, L dce is the supervision loss in the difficult area, L adv is the adversarial loss, and λ1 is the scaling factor of the adversarial learning regularization term.
[0030] In an optional embodiment, the method for constructing the anti-loss is as follows:
[0031] L adv (X) = L bce(D(S(X)),1)
[0032] Among them, S is the segmentation network.
[0033] In an optional embodiment, the method for constructing the difficult region supervision loss is as follows:
[0034] L dce =λL ce (Y,P)
[0035] Among them, λ is the supervision parameter of the difficult area, L ce is the cross entropy loss.
[0036] In an optional embodiment, the method for confirming the difficult area supervision parameter λ is as follows:
[0037] λ=(1-M) 2 +|1-M|
[0038] Among them, M is the confidence map.
[0039] In an optional embodiment, the scaling factor λ1 of the adversarial learning regularization term is set to 0.05.
[0040] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0041] 1. The coarse segmentation in the present invention is to segment the entire area of the target and the tumor (i.e., the entire target area). As for the fine segmentation, considering the anatomical structure of the tumor, the present invention uses a multi-level cascade method to segment the target object layer by layer and each layer uses binary classification segmentation, which not only effectively utilizes the spatial position relationship of the segmentation network, but also the next layer works on the region of interest extracted by the previous layer, which helps to reduce errors. The binary classification task method adopted by the present invention is easier to train than the multi-classification task and can reduce overfitting.
[0042] 2. In addition to using residual convolution and SEAttention modules to optimize Unet in the prior art, the segmentation network model in the present invention adds a discriminator to each decoding layer of the segmentation network to perform multi-scale adversarial learning. Multi-scale adversarial learning is a game between the segmentation network and the discriminator. In this process, the segmentation network and the discriminator continuously improve their capabilities. At the same time, adversarial learning is performed at multiple scales of the segmentation network, making full use of multi-scale information to make the segmentation result more accurate.
[0043] 3. In addition, the present invention proposes a difficult region supervision loss function, which can improve the segmentation effect of difficult regions, such as boundary regions. And because adversarial learning is performed at multiple scales, the loss can supervise difficult-to-segment regions of different scales, which can further improve the segmentation of difficult regions. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:
[0045] Figure 1 A flowchart of a medical image tumor segmentation method using multi-scale adversarial learning provided by an embodiment of the present invention;
[0046] Figure 2 A schematic diagram of the structure of a segmentation framework for a medical image tumor segmentation method using multi-scale adversarial learning provided by an embodiment of the present invention;
[0047] Figure 3 A schematic diagram of the structure of the MSALDS-UNet adversarial learning network model provided in one embodiment of the present invention;
[0048] Figure 4 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.
[0050] Example 1
[0051] Figure 1 A flowchart of a medical image tumor segmentation method using multi-scale adversarial learning provided in Example 1 of the present invention is shown in FIG. Figure 1 As shown, the method comprises the following steps:
[0052] Step S1, obtaining a medical image, and using a Res-UNet network model to roughly segment the target of the medical image to obtain a region of interest.
[0053] Usually, organs in medical images only occupy a small part of the entire medical image, and tumors and tumor cores are relatively smaller. Their segmentation can easily be misled by other irrelevant tissues and organs in the medical image. In addition, the imbalance of categories makes the identification and segmentation of organs as a whole extremely difficult. Considering that tumors and tumor cores are usually contained in a certain organ, cascade segmentation is a good solution.
[0054] Figure 2A schematic diagram of a segmentation framework for a medical image tumor segmentation method provided in Example 2 of the present invention is shown in FIG. Figure 2 As shown in the figure, after acquiring the medical image, the Res-UNet network model is used to roughly segment the target. Then the medical image is cropped according to the position of the target to obtain the region of interest. Among them, the region of interest contains organs and tumors. The medical image is roughly segmented, which facilitates the subsequent targeted fine segmentation of organs, tumors and tumor cores.
[0055] According to the existing data, coarse segmentation is necessary. It can remove unnecessary parts, retain specific areas, reduce background interference, and help improve segmentation efficiency and results.
[0056] Step S2: construct a three-level cascade MSALDS-UNet adversarial learning network model, input the region of interest into the three-level cascade MSALDS-UNet adversarial learning network model for segmentation, and obtain organ regions, tumor regions, and tumor core regions.
[0057] Since there is a problem of sample imbalance in the segmentation process, a cascade segmentation method is adopted in the present invention to transform the multi-classification problem into a binary classification problem. Specifically, the present invention adopts a three-level cascade structure, such as Figure 2 As shown, the first-level MSALDS-UNet adversarial learning network model is used to segment organs, the second-level MSALDS-UNet adversarial learning network model is used to segment tumors, and the third-level MSALDS-UNet adversarial learning network model further segments the tumor to obtain the core area of the tumor based on the segmentation of the tumor by the second-level MSALDS-UNet adversarial learning network model.
[0058] Further, a three-level cascaded MSALDS-UNet adversarial learning network model is constructed, and the region of interest and the mass region of interest are input into the three-level cascaded MSALDS-UNet adversarial learning network model for segmentation as follows:
[0059] Step S21, introduce a channel attention module in the encoding layer, and introduce a discriminator in the decoding layer to respectively construct a first-level MSALDS-UNet adversarial learning network model, a second-level MSALDS-UNet adversarial learning network model, and a third-level MSALDS-UNet adversarial learning network model.
[0060] in, Figure 3 The schematic diagram of the structure of the MSALDS-UNet adversarial learning network model provided in Example 1 of the present invention is as follows: Figure 3As shown in the figure, the leftmost is the encoding layer of the MSALDS-UNet adversarial learning network model. The channel attention module is introduced in the encoding layer. The channel attention module can use the squeeze operation to compress features in the spatial dimension to obtain a 1×1×C channel description, so that the encoding layer has a global receptive field. Then, the excitation operation is used to learn the degree of dependence on each channel, and the feature map is adjusted according to the degree of dependence to achieve the function of enhancing important features and weakening unimportant features.
[0061] To the right of the encoding layer is the decoding layer. A discriminator is introduced in each decoding layer. Figure 3 There is a discriminator on the right side of each decoding layer in the network. The discriminator is a simplified version of UNet, which only retains one convolution layer at each layer and reduces the number of feature maps in the convolution layer of the entire network by half. The discriminator uses the discriminator network to perform multi-scale adversarial learning and supervise difficult areas at each decoding layer. Different from the traditional generative adversarial network, the discriminator only outputs the probability that the input image belongs to the real image. It can provide structural information to regularize the output of the segmentation network and strengthen the image recognition and segmentation of organs, tumors, and tumor core areas.
[0062] It should be noted that the first-level MSALDS-UNet adversarial learning network model, the second-level MSALDS-UNet adversarial learning network model and the third-level MSALDS-UNet adversarial learning network model have the same framework structure.
[0063] Furthermore, a dilated residual module is used in the middle layer between the encoding layer and the decoding layer, which is used to expand the receptive field and capture multi-scale contextual information.
[0064] Step S22, inputting the region of interest into the primary MSALDS-UNet adversarial learning network model for fine segmentation to obtain the organ region, and inputting the region of interest into the secondary MSALDS-UNet adversarial learning network model for fine segmentation to obtain the tumor region.
[0065] Among them, the first-level MSALDS-UNet adversarial learning network model, the second-level MSALDS-UNet adversarial learning network model and the third-level MSALDS-UNet adversarial learning network model have the same framework structure and the steps for fine segmentation are also the same. The MSALDS-UNet adversarial learning network model at each level segments the organs in the area according to the steps of fine segmentation based on the area to be segmented.
[0066] Further, the fine segmentation method is as follows:
[0067] Step A: The encoding layer extracts organ image features, and adjusts the organ image features through the channel attention module to obtain enhanced image features.
[0068] Among them, the method for adjusting the organ image features through the channel attention module is as follows:
[0069] The channel attention module uses the squeeze operation to perform feature compression in the spatial dimension to obtain a 1*1*C channel.
[0070] The degree of dependence on the channel is learned by using an excitation operation, and the organ image feature is adjusted according to the degree of dependence.
[0071] Taking the first-level MSALDS-UNet adversarial learning network model for organ segmentation as an example, after acquiring the image of the region of interest, the encoding layer of the first-level MSALDS-UNet adversarial learning network model extracts image features in the region of interest. The channel attention module uses the squeeze operation on the organ features to compress the organ features in the spatial dimension to obtain 1*1*C channels, and different channels contain different image features.
[0072] The excitation operation is then used to learn the degree of dependence on each channel. For example, the channel where important features are located has a high degree of dependence, while the channel where unimportant features are located has a low degree of dependence. The organ features are adjusted according to the degree of dependence to enhance the features with high organ correlation and weaken the features with low organ correlation, thereby enhancing the accuracy of organ recognition.
[0073] Step B: The decoding layer decodes the enhanced image features to obtain a feature map and a true label.
[0074] Step C: splicing the feature map and the true label with the medical image respectively, and sending the spliced image to the discriminator for processing to obtain a confidence map.
[0075] Among them, the feature map and true label obtained by each decoding layer of the MSALDS-UNet adversarial learning network model are concatenated with the original image and then sent to the discriminator in sequence, and then the discriminator outputs a confidence map of size H×W×1. Each pixel of the discriminator output map indicates whether the pixel comes from the true label or from the MSALDS-UNet adversarial learning network model.
[0076] In an optional embodiment, before sending the spliced image to the discriminator for processing, the discriminator is trained using a binary cross entropy loss.
[0077] Step D: constructing a difficult region supervision loss model according to the confidence map, and performing organ boundary region segmentation using the difficult region supervision loss model.
[0078] Furthermore, before sending the spliced image to the discriminator for processing, the method further includes constructing a discriminator loss function and using a binary cross entropy loss to train the discriminator;
[0079] Among them, the discriminator loss function is constructed as follows:
[0080]
[0081] Among them, L D is the discriminator loss, X is the input image, Y is the true mask, P is the decoding layer output, D is the discriminator, and L bce is the binary cross entropy loss.
[0082] Among them, the above formula is a binary cross entropy loss, which is used to train the discriminator. It should be noted that in the present invention, the traditional supervised segmentation network only has one scale, and the supervised segmentation network has multiple scales, so each decoding layer has a loss, so we use binary cross entropy loss to train the discriminator.
[0083] Furthermore, before fine segmentation, it also includes constructing a segmentation network loss function to train the segmentation network in the MSALDS-UNet adversarial learning network model;
[0084] The method for constructing the segmentation network loss function is as follows:
[0085] L seg =L dice +L dce +λ1L adv
[0086] Where L seg is the segmentation network training loss function, L dice is the Dice loss, L dce is the supervision loss in the difficult area, L adv is the adversarial loss, and λ1 is the scaling factor of the adversarial learning regularization term.
[0087] The above formula is the training loss. The loss function of the training segmentation network is composed of Dice loss, difficult area supervision loss and adversarial learning loss. The present invention uses a multi-task loss function as the training loss in the segmentation network.
[0088] Furthermore, the method for constructing the adversarial loss is as follows:
[0089] L adv (X) = Lbce (D(S(X)),1)
[0090] Among them, S is the segmentation network.
[0091] The above formula indicates that adversarial learning is performed using adversarial loss Ladv. Through adversarial loss, the segmentation network S generates segmentations closer to the true value to deceive the discriminator.
[0092] Furthermore, the construction method of the difficult area supervision loss is as follows:
[0093] L dce =λL ce (Y,P)
[0094] Among them, λ is the supervision parameter of the difficult area, L ce is the cross entropy loss.
[0095] The above formula is a difficult region supervision loss proposed using the predicted probability mask and confidence map.
[0096] Since kidney tumors are irregular in shape, small in size, and similar in grayscale to surrounding tissues and organs, accurate segmentation of their boundary regions has always been a challenging difficulty and hotspot internationally. Boundary regions are difficult areas in kidney tumors that are difficult to segment. The confidence map output by the discriminator in adversarial learning contains information from medical images and predicted probability masks used to segment difficult areas. This information can provide structural information of boundary regions, which is beneficial to the segmentation of organ boundary regions.
[0097] In an optional embodiment, the method for confirming the difficult area supervision parameter λ is as follows:
[0098] λ=(1-M) 2 +|1-M|
[0099] Among them, M is the confidence map. The difficult area supervision parameter is added before the cross entropy loss, which consists of the absolute error and square error between the true probability mask and the confidence map.
[0100] In an optional embodiment, the scaling factor λ1 of the adversarial learning regularization term is set to 0.05.
[0101] Step S23, inputting the tumor region into the three-level MSALDS-UNet adversarial learning network model for fine segmentation to obtain the tumor core region.
[0102] Among them, since the framework structures of the first-level MSALDS-UNet adversarial learning network model, the second-level MSALDS-UNet adversarial learning network model and the third-level MSALDS-UNet adversarial learning network model are the same, step S23 uses the above-mentioned fine segmentation method to finely segment the tumor area, thereby obtaining the tumor core area.
[0103] Step S3: replacing the organ region, the tumor region and the tumor core region into the medical image to obtain a segmented image.
[0104] The organ region, tumor region and tumor core region obtained by the coarse segmentation of step S1 and the fine segmentation of step S2 are replaced into the original medical image to obtain a segmented image of the tumor, thereby achieving accurate segmentation of the organ, tumor and tumor core region.
[0105] Compared with the prior art methods for segmenting kidney and kidney tumor, the present invention has the following advantages:
[0106] 1. The coarse segmentation in the present invention is to segment the whole area of organs and tumors (i.e. the whole target area). As for the fine segmentation, considering the anatomical structure of the tumor, the present invention uses a multi-level cascade method to segment the target object layer by layer and each layer uses binary classification segmentation, which not only effectively utilizes the spatial position relationship of the segmentation network, but also the next layer works on the region of interest extracted by the previous layer, which helps to reduce errors. The binary classification task method adopted by the present invention is easier to train than the multi-classification task and can reduce overfitting.
[0107] 2. In addition to using residual convolution and SEAttention modules to optimize Unet in the prior art, the segmentation network model in the present invention adds a discriminator to each decoding layer of the segmentation network to perform multi-scale adversarial learning. Multi-scale adversarial learning is a game between the segmentation network and the discriminator. In this process, the segmentation network and the discriminator continuously improve their capabilities. At the same time, adversarial learning is performed at multiple scales of the segmentation network, making full use of multi-scale information to make the segmentation result more accurate.
[0108] 3. In addition, the present invention proposes a difficult region supervision loss function, which can improve the segmentation effect of difficult regions, such as boundary regions. And because adversarial learning is performed at multiple scales, the loss can supervise difficult-to-segment regions of different scales, which can further improve the segmentation of difficult regions.
[0109] Example 2
[0110] Based on Example 1, Example 2 of the present invention provides the KITS21 dataset to demonstrate the superiority of a multi-scale adversarial learning medical image tumor segmentation method provided in Example 1 over the existing kidney and renal tumor segmentation methods.
[0111] In Example 2 of the present invention, the KITS21 dataset is used to perform quantitative analysis on the algorithm proposed in the present invention and several other cutting-edge segmentation algorithms. The quantitative analysis results are shown in the following table (the best results are shown in bold).
[0112]
[0113] It can be seen from the above table that all the scores of all networks on the kidney are relatively high. In the above table, the best results are bolded. It can be seen from the above table that the method proposed in Example 1 of the present invention obtains the best DSC and JC values in the segmentation of kidney, tumor and tumor core. In terms of the DSC scores of tumor and tumor core, our method is about 2% to 3% better than other segmentation methods, which reflects the advantage of the method proposed in Example 1 of the present invention in tumor segmentation.
[0114] Example 3
[0115] Figure 4 A schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention is shown in FIG. Figure 4 As shown, the electronic device includes a processor 21, a memory 22, an input device 23 and an output device 24; the number of processors 21 in the computer device can be one or more. Figure 4 A processor 21 is taken as an example; the processor 21, the memory 22, the input device 23 and the output device 24 in the electronic device can be connected by a bus or other means. Figure 4 The example of connecting through bus is taken in the following.
[0116] The memory 22 is a computer-readable storage medium that can be used to store software programs, computer executable programs, and modules. The processor 21 executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory 22, that is, implements a multi-scale adversarial learning medical image tumor segmentation method of embodiment 1.
[0117] The memory 22 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory 22 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 22 may further include a memory remotely arranged relative to the processor 21, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.
[0118] The input device 23 can be used to receive the ID and password input by the user, etc. The output device 24 is used to output the network configuration page.
[0119] Example 4
[0120] Embodiment 4 of the present invention further provides a computer-readable storage medium, wherein the computer-executable instructions, when executed by a computer processor, are used to implement a medical image tumor segmentation method based on multi-scale adversarial learning as provided in Embodiment 1.
[0121] An embodiment of the present invention provides a storage medium containing computer executable instructions, and its computer executable instructions are not limited to the method operations provided in Example 1, and can also execute related operations in a multi-scale adversarial learning medical image tumor segmentation method provided in any embodiment of the present invention.
[0122] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A multi-scale adversarial learning method for medical image tumor segmentation, characterized in that: The steps include: S1. Obtain a medical image, and use a Res-UNet network model to roughly segment the target area of the medical image to obtain a region of interest; S2. Construct a three-level cascade MSALDS-UNet adversarial learning network model, input the region of interest into the three-level cascade MSALDS-UNet adversarial learning network model for fine segmentation, and obtain organ regions, tumor regions, and tumor core regions; S3, replacing the organ region, the tumor region and the tumor core region into the medical image to obtain a segmented image; The fine segmentation method is as follows: Each encoding layer extracts organ image features, and adjusts the organ image features through a channel attention module to obtain enhanced image features; Each decoding layer decodes the enhanced image features to obtain a feature map and a true label; splicing the feature map and the true label with the medical image respectively, and sending the spliced image to the discriminator for processing to obtain a confidence map; Constructing a difficult region supervision loss model according to the confidence map, and supervising the difficult region that is difficult to segment through the difficult region supervision loss model; The method for constructing the segmentation network loss function of the fine segmentation is as follows: The method for constructing the segmentation network loss function is as follows: L seg =L dice +L dce +λ1L adv Where L seg is the segmentation network training loss function, L dice is the Dice loss, L dce is the supervision loss in the difficult area, L adv is the adversarial loss, λ1 is the scaling factor of the adversarial learning regularization term; The construction method of the difficult area supervision loss is as follows: L dce =λL ce (Y,P) Among them, λ is the supervision parameter of the difficult area, L ce is the cross entropy loss; The method for confirming the difficult area supervision parameter λ is as follows: λ=(1-M) 2 +|1-M| Among them, M is the confidence map.
2. The method for medical image tumor segmentation based on multi-scale adversarial learning according to claim 1, characterized in that: The method of constructing a three-level cascaded MSALDS-UNet adversarial learning network model and inputting the region of interest into the three-level cascaded MSALDS-UNet adversarial learning network model for segmentation is as follows: S21, introduce the channel attention module in the encoding layer, introduce the discriminator in the decoding layer to construct the first-level MSALDS-UNet adversarial learning network model, the second-level MSALDS-UNet adversarial learning network model and the third-level MSALDS-UNet adversarial learning network model respectively; S22, inputting the region of interest into the primary MSALDS-UNet adversarial learning network model for fine segmentation to obtain an organ region, and inputting the region of interest into the secondary MSALDS-UNet adversarial learning network model for fine segmentation to obtain a tumor region; S23, inputting the tumor region into the three-level MSALDS-UNet adversarial learning network model for fine segmentation to obtain the tumor core region.
3. The medical image tumor segmentation method based on multi-scale adversarial learning according to claim 1, characterized in that: The method for adjusting the organ image features by the channel attention module is as follows: The channel attention module uses the squeeze operation to perform feature compression in the spatial dimension to obtain a 1*1*C channel; The degree of dependence on the channel is learned by using an excitation operation, and the organ image feature is adjusted according to the degree of dependence.
4. The method for medical image tumor segmentation based on multi-scale adversarial learning according to claim 1, characterized in that: Before sending the spliced image to the discriminator for processing, it also includes constructing a discriminator loss function to train the discriminator using a binary cross entropy loss; Among them, the discriminator loss function is constructed as follows: Among them, L D is the discriminator loss, X is the input image, Y is the true mask, P is the decoding layer output, D is the discriminator, and L bce is the binary cross entropy loss.
5. The medical image tumor segmentation method based on multi-scale adversarial learning according to claim 1, characterized in that: The construction method of the adversarial loss is as follows: L adv (X)=L bce (D(S(X)),1) Among them, S is the segmentation network.
6. The medical image tumor segmentation method based on multi-scale adversarial learning according to claim 1, characterized in that: The scaling factor λ1 of the adversarial learning regularization term is set to 0.05.