阅读说明：本技术 面向多种ct肺部纹理识别的非监督保内容域适应方法 (Unsupervised content protection domain adaptation method for multiple CT lung texture recognition ) 是由 徐睿 叶昕辰 丛臻 于 2020-06-15 设计创作，主要内容包括：本发明公开了一种面向多种CT肺部纹理识别的非监督保内容域适应方法,属于图像处理和计算机视觉领域。该方法能使在一种CT数据上(源域上)事先训练的肺部纹理识别的深层网络模型,在应用到另一种CT图像上(目标域上)时,在仅获得目标域的CT图像并无需人工标注典型肺部阴影的前提下,运用对抗学习机制和特殊设计的内容一致性网络模块,就可以对深层网络模型进行微调,使目标域上的肺部纹理识别保持较高性能。此方法不仅能节省开发的人力和时间成本,而且易于实现并具有较高的实用性。(The invention discloses an unsupervised content protection domain adaptation method for multi-CT lung texture recognition, and belongs to the field of image processing and computer vision. The method can enable a deep network model of lung texture recognition trained in advance on CT data (on a source domain) to be applied to another CT image (on a target domain), and can be finely adjusted by applying an antagonistic learning mechanism and a specially designed content consistency network module on the premise of only obtaining the CT image of the target domain and not manually marking typical lung shadows, so that the lung texture recognition on the target domain keeps higher performance. The method can save the development labor and time cost, is easy to realize and has higher practicability.)

1. The unsupervised content protection domain adaptation method for the identification of various CT lung textures is characterized by comprising the following steps of:

1) training and test data preparation: two groups of CT images of different types are collected, and a typical lung texture region is marked on the two groups of CT images by using a manual mode; then, randomly appointing one group of images as source domain data and the other group of images as target domain data; processing the CT image on the source domain and the artificially marked lung texture region into small labeled CT blocks for supervising and training a deep network model on the source domain; processing data of a target domain into CT small blocks with labels and CT small blocks without labels, wherein the CT small blocks without labels are used for carrying out fine adjustment on a pre-trained deep network model in an unsupervised mode, and the CT small blocks with labels are used for testing the final result of the technical scheme provided by the invention;

2) and (3) construction and supervision training of the recognition network on the source domain: constructing a deep network model by using a residual error network, wherein the structure of the deep network model comprises an encoder and a classifier; the encoder extracts and inputs CT lung texture image feature representation, and the classifier uses the feature representation to generate a recognition result; using the labeled CT small blocks in the source domain, and training the deep network in a supervision mode to ensure that the network model achieves good identification performance in the data of the source domain;

3) deep model fine-tuning on the target domain: aiming at the source domain deep layer network model obtained in the step 2), using CT small blocks without labels of a target domain, carrying out unsupervised domain adaptation by using a loss function based on an antagonistic learning mechanism, simultaneously carrying out content keeping constraint on a target domain encoder by using a content consistency module and a content consistency loss function, combining supervised classification training in the source domain, and carrying out fine adjustment on the deep layer model of the target domain by using the CT small blocks with labels in the source domain again, so that the deep layer network model can keep good lung texture recognition performance in the target domain.

2. The unsupervised content-preserving domain adaptation method for multi-CT lung texture recognition according to claim 1, wherein the deep model fine-tuning on the target domain in step 3) comprises the following steps:

3-1) constructing a deep network with the same structure as the source domain network model for the target domain data, enabling encoders and classifiers of the two networks to share the same network parameter weight, and carrying out fine tuning on the network model on the target domain by taking the parameter weight of the network model trained in the source domain data in the step (2) as an initial value;

3-2) constructing a discriminator by using a countermeasure learning mechanism, performing domain adaptation by optimizing a countermeasure loss function, and reducing the characteristic representation domain deviation of the encoder of the source domain and the target domain; the discriminator consists of a convolution module and a full connection layer, the characteristics of the source domain and the target domain encoder are represented as input, the characteristics of the source domain encoder are represented to be judged as a source domain result, the label is 1, the characteristics of the target domain encoder are represented to be judged as a target domain result, and the label is 0; the penalty function is formulated as follows:

in the formula, L_adv(. cndot.) represents the countermeasure loss function value, D represents the discriminator, f represents the encoder,representing a mathematical expectation, x_sSource domain CT image data matrix, x, representing a single batch of training participation_tRepresenting a target field CT image data matrix, X, of a single batch of training participants_sRepresenting a set of source-domain CT image matrices, X_tRepresenting a target domain CT image matrix set, and log (-) representing a logarithm operation;

3-3) using a content consistency module to constrain the feature representation of the target domain encoder and the input target domain CT lung texture image through a content consistency loss function, and keeping the content consistency of the target domain; the content consistency module comprises a convolution module and a residual error module, the target domain encoder feature representation is reconstructed into a single-channel image, and the single-channel image and the input target domain CT lung texture image are constrained through an L1 norm; the content consistency loss function is formulated as follows:

in the formula, L_cp(. cndot.) represents a content conformance loss function value, f represents an encoder, g represents a content conformance module,

3-4) calculating the sum of the countervailing loss function, the content consistency loss function and the classification cross-entropy loss function in the source domain as the overall loss function for network fine tuning by using the unlabeled CT patches in the target domain and again using the labeled CT patches in the source domain, with the following specific formula:

L_total(f,h,g,D)＝L_adv(D,f)+λ_cpL_cp(f,g)+λ_taskL_task(f,h)

in the formula, L_total(. h) represents the overall loss function value for unsupervised content domain adaptation, f represents the encoder, h represents the classifier, g represents the content consistency module, D represents the discriminator, L_advValue representing the function of the resistance loss, λ_cpCoefficient representing content consistency loss function, L_cpFor content consistency loss function values, λ_taskRepresenting the coefficients of the categorical cross-entropy loss function, L_taskRepresenting a classification cross entropy loss function value; the calculation formula of the classification cross entropy loss function is as follows:

in the formula, L_task(. cndot.) denotes the cross entropy loss function value, f denotes the encoder, h denotes the classifier,representing a mathematical expectation, x_sRepresenting a matrix of source-domain CT image data, y, in a single batch of training_sDenotes x_sCorresponding class label matrix, X_sRepresenting a set of source-domain CT image matrices, Y_sRepresents X_sCorresponding class label matrix set, ∑ denotes the sum operator, K denotes the number of classification classes, log (-) denotes the log operation.

Technical Field

The invention belongs to the field of medical image processing and computer vision, and particularly relates to an unsupervised content protection domain adaptation method for various CT lung texture recognition.

Background

Diffuse lung disease refers to the general term of abnormal texture of the pulmonary interstitium caused by inflammation or injury in the lung region. A Computer-Aided Diagnosis (CAD) system of the diffuse lung disease can assist a radiologist to diagnose the diffuse lung disease, wherein accurate identification of the texture of the diffuse lung disease in a CT image is a key step for constructing the CAD system. At present, a method based on a deep convolutional neural network is invented and applied to the identification of the texture of diffuse lung diseases, and a better result is obtained.

However, due to the difficulty of acquiring medical data, when deep-network-based algorithms are developed, only one type of CT image data is generally collected, and the algorithm is designed and verified on the data. In the algorithm research and development stage, the developed deep network model can only obtain better recognition results on the same type of CT image data; in the practical application stage, when the model is directly applied to different types of CT image data, the accuracy of identifying the lung shadow is greatly reduced. This is because different types of CT images differ in terms of noise, imaging, and the like, which leads to a problem that when a network model is designed for only one type of CT data and is directly applied to other types of CT data, the generalization of the network model is insufficient.

For this problem, the generalization of the network model to the CT image data of different types can be improved by acquiring the CT image data of different types and mixing them together to train the network model (Yongjun Chang, journal cum Lim and et al, "a short vector machine classifier recycling in the scanner variation of the read classification of the regional distance mapping errors," medical to physics, vol.40, No.5, pp.912, 2013.). However, this approach requires not only the collection of other different kinds of CT image data, but also the cumbersome and time-consuming manual labeling of typical lung shadow regions in these new CT images. Therefore, a more convenient and efficient technique for solving the problem of generalization of deep network models of different types of CT data is needed.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides an unsupervised content protection domain adaptation method for identifying various CT lung textures. The method can enable a deep network model which is trained in advance on CT data (on a source domain) to comprise a CT image and artificial labeling of a typical lung shadow region, and when the method is applied to another CT image (on a target domain), the deep network model can be finely adjusted by applying a counterstudy mechanism and a specially designed content consistency network module on the premise of only obtaining the CT image of the target domain and not needing artificial labeling of the typical lung shadow, so that the lung texture recognition on the target domain keeps high performance.

The specific technical scheme of the invention is that an unsupervised content protection domain adaptation method for various CT lung texture identification comprises the following steps:

1) training and test data preparation: two sets of CT images of different types are acquired and a typical lung texture region is manually labeled on the two sets of CT images. Thereafter, one set of images is randomly designated as source domain data and the other set of images is designated as target domain data. The CT image on the source domain and the artificially labeled lung texture region are processed into CT small blocks with labels (texture classes) for supervising and training the deep network model on the source domain. Processing the data of the target domain into small CT blocks with labels and CT blocks without labels, wherein the small CT blocks without labels are used for carrying out fine adjustment on a pre-trained deep network model in an unsupervised mode, and the small CT blocks with labels are used for testing the final result of the technical scheme provided by the invention;

2) and (3) construction and supervision training of the recognition network on the source domain: a Deep network model is constructed by using a residual network (Kaming He, XiangyuZhang, and et. al., "Deep residual learning for image Recognition," in computer Vision and Pattern Recognition,2016, pp.770-778.), and the structure of the Deep network model comprises two parts, namely an encoder and a classifier. The encoder extracts the input CT lung texture image feature representation, and the classifier uses the feature representation to generate a recognition result. Using the labeled CT small blocks in the source domain, and training the deep network in a supervision mode to ensure that the network model achieves good identification performance in the data of the source domain;

3) deep model fine-tuning on the target domain: aiming at the source domain deep layer network model obtained in the step (2), using CT small blocks without labels of a target domain, carrying out unsupervised domain adaptation by using a loss function based on an antagonistic learning mechanism, simultaneously using a content consistency module and a content consistency loss function to carry out content keeping constraint on a target domain encoder, combining supervised classification training in the source domain, and needing to use the CT small blocks with labels in the source domain again to carry out fine adjustment on the deep layer model of the target domain together, so that the deep layer network model can keep good lung texture recognition performance in the target domain;

4) and (3) performance test of the deep network model: and (3) testing the performance of the finally obtained deep network model by using the marked CT small blocks on the target domain and calculating common indexes for measuring the recognition performance, such as correct recognition accuracy and an F value.

The method comprises the following steps of constructing and supervising training of a recognition network on a source domain:

2-1) constructing an identification network by using a residual error network, wherein the identification network comprises an encoder and a classifier. The encoder is composed of a plurality of convolution modules and extracts and inputs CT lung texture image characteristic representation. The classifier comprises a plurality of convolution modules, a global average pooling layer and a full-connection layer, and generates an identification result by using feature representation;

2-2) each convolution module consists of a convolution layer, a batch normalization layer and a rectification linear unit layer, which are all general structures of a deep convolutional neural network;

2-3) except the first convolution module, remaining convolution modules in the identification network are connected through jumping to form a residual error module. The residual module is also a general network structure, and can refer to the existing documents (Kaiming He, Xiangyu Zhang, and et al., "Deep residual learning for image Recognition," in Computer Vision and pattern Recognition,2016, pp.770-778.);

2-4) carrying out supervised network training on the deep network by using the labeled CT small blocks of the source domain. Specifically, a classification cross entropy loss function is calculated in a small batch mode, and the classification cross entropy loss function is optimized through a random gradient descent algorithm to obtain a deep network model of a source domain. The calculation formula of the classification cross entropy loss function is as follows:

in the formula, L_task(. cndot.) represents the cross-entropy loss function value,f denotes an encoder, h denotes a classifier,

representing a mathematical expectation, x_sRepresenting a matrix of source-domain CT image data, y, in a single batch of training_sDenotes x_sCorresponding class label matrix, X_sRepresenting a set of source-domain CT image matrices, Y_sRepresents X_sCorresponding to the class label matrix set, ∑ represents a summation operator, K represents the number of classification classes, K is 6 in the invention, and log (-) represents a logarithm operation.

The fine tuning of the deep model on the target domain specifically comprises the following steps:

3-1) constructing a deep network with the same structure as the source domain network model for the target domain data, enabling the encoders and the classifiers of the two networks to share the same network parameter weight, and carrying out fine tuning on the network model on the target domain by taking the parameter weight of the network model obtained by training in the source domain data in the step (2-4) as an initial value;

3-2) constructing a discriminator by using a countermeasure learning mechanism, carrying out domain adaptation by optimizing a countermeasure loss function, and reducing the deviation of the encoder feature representation domain of the source domain and the target domain. The discriminator is composed of a convolution module and a full connection layer, and takes the characteristics of the source domain encoder and the target domain encoder as input, the characteristics of the source domain encoder are represented and judged as a source domain result (label is 1), and the characteristics of the target domain encoder are represented and judged as a target domain result (label is 0). The penalty function is formulated as follows:

in the formula, L_adv(. cndot.) represents the countermeasure loss function value, D represents the discriminator, f represents the encoder,

representing a mathematical expectation, x_sSource domain CT image data matrix, x, representing a single batch of training participation_tRepresenting a target field CT image data matrix, X, of a single batch of training participants_sRepresenting a set of source-domain CT image matrices, X_tRepresenting a target domain CT image matrix set, and log (-) representing a logarithm operation;

3-3) using a content consistency module to constrain the feature representation of the target domain encoder and the input target domain CT lung texture image through a content consistency loss function, and keeping the content consistency of the target domain. The content consistency module comprises a convolution module and a residual error module, reconstructs the characteristic representation of the target domain encoder into a single-channel image, and is constrained with the input target domain CT lung texture image through an L1 norm. The content consistency loss function is formulated as follows:

in the formula, L_cp(. cndot.) represents a content conformance loss function value, f represents an encoder, g represents a content conformance module,

representing a mathematical expectation, x_tRepresenting a target field CT image data matrix, X, of a single batch of training participants_tRepresenting a set of target domain CT image matrices, | |)₁Represents the L1 norm;

3-4) calculating the sum of the countervailing loss function, the content consistency loss function and the classification cross-entropy loss function in the source domain as the overall loss function for network fine tuning by using the unlabeled CT patches in the target domain and again using the labeled CT patches in the source domain, with the following specific formula:

L_total(f，h，g，D)＝L_adv(D，f)+λ_cpL_cp(f，g)+λ_taskL_task(f，h)

in the formula, L_total(. h) represents the overall loss function value for unsupervised content domain adaptation, f represents the encoder, h represents the classifier, g represents the content consistency module, D represents the discriminator, L_advValue representing the function of the resistance loss, λ_cpDenotes the content consistency loss function coefficient, λ in the present invention_cpIs 1.0, L_cpFor content to be consistentValue of the sexual loss function, λ_taskRepresenting the coefficients of a categorical cross-entropy loss function, lambda in the present invention_taskIs 100.0, L_taskThe classification cross-entropy loss function value (see the formula in step (2-4) for its definition) is expressed.

And 3-5) optimizing the overall optimization loss function in the step (3-4) through a random gradient descent algorithm to obtain a deep network model which is finally fine-tuned for the target domain.

The invention has the beneficial effects that:

the invention relates to an unsupervised content protection domain adaptation method for various CT lung texture identification. The method can enable a deep network model trained in advance on one CT data (source domain) to be applied to another CT image (target domain), and can be finely adjusted by applying a counterstudy mechanism and a specially designed content consistency network module on the premise of only obtaining a new type of CT image and not manually marking typical lung shadows, so that the lung texture recognition on the target domain keeps high performance. The system has the following characteristics:

1. network fine adjustment of the target domain is realized, complicated and time-consuming manual marking on target domain data is not needed, labor and time cost are saved, and the practicability of the method is improved;

2. the method is simple and easy to realize;

3. reducing the deviation of the encoder feature representation domain of the source domain and the target domain by using an unsupervised domain adaptation mode based on an antagonistic learning mechanism;

4. and designing a content consistency module, and keeping the content constraint on the target domain encoder through a content consistency loss function to ensure the consistency of the target domain content in the domain network fine tuning.

Drawings

FIG. 1 is a flow chart of an embodiment.

Fig. 2 is a diagram of an unsupervised content domain adaptation network architecture.

Fig. 3 is a diagram of a structure of a discriminator based on a counterlearning mechanism.

Fig. 4 is a diagram of a content consistency module structure.

Detailed Description

The invention is described in detail below with reference to the accompanying drawings and embodiments:

the invention provides an unsupervised content protection domain adaptation method for multi-CT lung texture recognition. The method can enable a deep network model trained in advance on one CT data (on a source domain) to be applied to another CT image (on a target domain), and can be used for finely adjusting the deep network model by using a counterstudy mechanism and a specially designed content consistency network module on the premise of only obtaining the CT image of the target domain and not manually marking typical lung shadows, so that the lung texture recognition on the target domain keeps high performance. The specific implementation flow is shown in fig. 1, and the method comprises the following steps:

1) training and test data preparation: two sets of CT images of different types are acquired and a typical lung texture region is manually labeled on the two sets of CT images. Thereafter, one set of images is randomly designated as source domain data and the other set of images is designated as target domain data. The CT image on the source domain and the artificially labeled lung texture region are processed into CT small blocks with labels (texture classes) for supervising and training the deep network model on the source domain. And processing the data of the target domain into small CT blocks with labels and small CT blocks without labels, wherein the small CT blocks without labels are used for carrying out fine adjustment on the pre-trained deep network model in an unsupervised mode, and the small CT blocks with labels are used for testing the final result of the technical scheme provided by the invention. The method comprises the following specific steps:

1-1) collecting two groups of CT images of different types, wherein the two groups of CT images comprise 6 common typical textures of lungs, namely, integrated, ground glass, honeycomb, emphysema, nodular and normal lung textures;

1-2) on the two sets of CT images, having an experienced radiologist select 3 coronal slices per CT image and manually delineate the lung region containing the above-mentioned typical texture on these slices;

1-3) during algorithm design and test, randomly selecting one group of image data as data on a source domain, and selecting the other group of image data as target domain data;

1-4) processing the CT image on the source domain and the labeled typical lung texture region to generate a plurality of labeled (texture class) CT image small blocks with the size of 32 x 32. Specifically, on the labeled CT coronal plane fault, the upper left corner starts to move a 32 x 32 scanning frame by taking the fixed step as 16 pixels in the horizontal and vertical directions, and when the center point of the searching frame is positioned in the labeled typical texture area, the CT image in the scanning frame at the moment is intercepted and the category of the texture is recorded. These labeled CT patches in the source domain will be used for supervised training of the deep network model over the source domain;

1-5) the CT image of the target field is divided into two parts for generating several 32 x 32 CT patches with and without labels, respectively. Wherein the labeled CT small blocks are generated in the same way as in the step (1-4). The generation method of CT small blocks without labels is that a lung region automatic segmentation algorithm (Rui, Xu, Jiano Pan and et al, "A pilot stereo trajectory estimation a depth dependent volume network to segment distances with complete strategies," in 2017Chinese Automation Consistency (CAC), IEEE,2017, pp.3291-3295.) is used to automatically segment the lung region of a CT image, then a plurality of faults on the coronal axis are randomly selected, and scanning is carried out by using a 32 x 32 scanning frame on each fault from the upper left corner in the horizontal and vertical directions at a fixed step of 16 pixels, when the center of the scanning frame is in the lung, the 32 x 32 CT small blocks covered by the scanning frame are cut out as CT small blocks without labels.

1-6) CT small blocks without labels on the target domain are used for fine tuning of a deep network model based on an unsupervised training mode, and the CT small blocks with labels are used for performance testing of a final model.

2) And (3) construction and supervision training of the recognition network on the source domain: and constructing a deep network model by using the residual error network, wherein the structure of the deep network model comprises an encoder and a classifier, and the specific structure is shown in the upper half dashed box of FIG. 2. The encoder extracts the input CT lung texture image feature representation, and the classifier uses the feature representation to generate a recognition result. The method comprises the following steps of training the deep network by using labeled CT small blocks in a source domain and adopting a supervision mode to ensure that a network model achieves good identification performance in data of the source domain, and specifically comprises the following steps:

2-1) constructing an identification network by using a residual error network, wherein the identification network comprises an encoder and a classifier. The encoder includes 7 convolution modules that extract the input CT lung texture image feature representation. The classifier comprises 12 convolution modules, 1 global average pooling layer and a full-connection layer, and generates a recognition result by using feature representation;

2-2) each convolution module consists of a convolution layer, a batch normalization layer and a rectification linear unit layer, which are all general and common structures of a deep convolutional neural network;

2-3) except the 1 st convolution module, identifying every two residual convolution modules in the network as a group, and forming 9 residual modules through jumping connection, wherein the encoder is provided with 3 residual modules, and the classifier is provided with 6 residual modules. The residual module is also a general network structure, and can be referred to the existing literature (Kaiming He, "iterative learning for image recognition," in Computer Vision and pattern recognition,2016, pp.770-778.);

2-4) carrying out supervised network training on the deep network by using the labeled CT small blocks of the source domain. Specifically, a classification cross entropy loss function is calculated in a small batch mode, and the classification cross entropy loss function is optimized through a random gradient descent algorithm to obtain a deep network model of a source domain. The calculation formula of the classification cross entropy loss function is as follows:

in the formula, L_task(. cndot.) denotes the cross entropy loss function value, f denotes the encoder, h denotes the classifier,representing a mathematical expectation, x_sRepresenting a matrix of source-domain CT image data, y, in a single batch of training_sDenotes x_sCorresponding class label matrix, X_sRepresenting a set of source-domain CT image matrices, Y_sRepresents X_sCorresponding category label matrix set, ∑ represents summation operator, K represents number of classification categories, K is 6 in the inventionAnd log (·) denotes a logarithmic operation.

3) Deep model fine-tuning on the target domain: aiming at the source domain deep layer network model obtained in the step (2), using a CT small block without a label of a target domain, carrying out unsupervised domain adaptation by using a loss function based on a counterstudy mechanism, simultaneously using a content consistency module and a content consistency loss function to carry out content keeping constraint on a target domain encoder, and combining supervised classification training in the source domain (needing to reuse the CT small block with the label in the source domain), carrying out fine adjustment on the deep layer model of the target domain together, so that the deep layer network model can keep good lung texture recognition performance in the target domain, and the specific steps are as follows:

3-1) as shown by a dotted frame at the lower half part of FIG. 2, constructing a deep network with the same structure as that of the source domain network model for the target domain data, enabling encoders and classifiers of the two networks to share the same network parameter weight, and carrying out fine tuning on the network model on the target domain by taking the parameter weight of the network model trained in the source domain data in (2-4) as an initial value;

3-2) constructing a discriminator as shown in the figure 3 by using a confrontation learning mechanism, and carrying out domain adaptation by optimizing a confrontation loss function to reduce the deviation of the encoder characteristic representation domain of the source domain and the target domain. The discriminator is composed of 4 convolution modules and 3 full-connection layers, and takes the characteristics of the source domain encoder and the target domain encoder as input, the characteristics of the source domain encoder are represented and judged as a source domain result (the label is 1), and the characteristics of the target domain encoder are represented and judged as a target domain result (the label is 0). The penalty function is formulated as follows:

in the formula, L_adv(. cndot.) represents the countermeasure loss function value, D represents the discriminator, f represents the encoder,representing a mathematical expectation, x_sSource domain CT image data matrix, x, representing a single batch of training participation_tIndicating parameters in a single batchWith a trained target field CT image data matrix, X_sRepresenting a set of source-domain CT image matrices, X_tRepresenting a target domain CT image matrix set, and log (-) representing a logarithm operation;

3-3) using a content consistency module to constrain the feature representation of the target domain encoder and the input target domain CT lung texture image through a content consistency loss function, and keeping the content consistency of the target domain. The content consistency module, as shown in fig. 4, includes 2 convolution modules and 3 residual modules, reconstructs the target domain encoder feature representation into a single-channel image, and is constrained by the L1 norm with the input target domain CT lung texture image. The content consistency loss function is formulated as follows:

in the formula, L_cp(. cndot.) represents a content conformance loss function value, f represents an encoder, g represents a content conformance module,

representing a mathematical expectation, x_tRepresenting a target field CT image data matrix, X, of a single batch of training participants_tRepresenting a set of target domain CT image matrices, | |)₁Represents the L1 norm;

3-4) calculating the sum of the countervailing loss function, the content consistency loss function and the classification cross-entropy loss function in the source domain as the overall loss function for network fine tuning by using the unlabeled CT patches in the target domain and again using the labeled CT patches in the source domain, with the following specific formula:

L_total(f，h，g，D)＝L_adv(D，f)+λ_cpL_cp(f，g)+λ_taskL_task(f，h)

in the formula, L_total(. h) represents the overall loss function value for unsupervised content domain adaptation, f represents the encoder, h represents the classifier, g represents the content consistency module, D represents the discriminator, L_advValue representing the function of the resistance loss, λ_cpCoefficient of function representing content consistency lossIn the present invention, lambda_cpIs 1.0, L_cpFor content consistency loss function values, λ_taskRepresenting the coefficients of a categorical cross-entropy loss function, lambda in the present invention_taskIs 100.0, L_taskThe classification cross-entropy loss function value (see the formula in step (2-4) for its definition) is expressed.

And 3-5) optimizing the overall optimization loss function in the step (3-4) through a random gradient descent algorithm to obtain a deep network model which is finally fine-tuned for the target domain.

4) And (3) performance test of the deep network model: and (3) testing the performance of the finally obtained deep network model by using the marked CT small blocks on the target domain and calculating common indexes for measuring the recognition performance, such as correct recognition accuracy and an F value. The test results and comparison results with two other recognized unsupervised domain adaptation methods of the present invention are shown in table 1, wherein (a) the correct Recognition accuracy and F-value based on the methods of ADDA (Eric tzing, Judy Hoffman and et al, "adaptive differential Recognition," in Proceedings of the IEEE Conference on Computer vision and Pattern Recognition,2017, pp.7167-7176); (b) correct recognition accuracy and F value based on Cycle-GAN (Jun-Yan Zhu, Taesung Park and et al, 'Ungained image-to-image transformation using Cycle-dependent adaptation network,' in Proceedings of the IEEE International Conference on Computer Vision,2017, pp.2223-2232.); (c) the invention (CPDA-Net) correctly identifies the accuracy and the F value. Group 1 → group 2 means that the deep network model is trained and fine-tuned with the first set of CT data as the source domain and the second set of CT data as the target domain. Group 2 → group 1 means that the deep network model is trained and fine-tuned with the second set of CT data as the source domain and the first set of CT data as the target domain.

TABLE 1 evaluation of the Performance of the method of the invention and comparison with other methods

The two methods are based on ADDA and cycleGAN, respectively, and although the methods are not proposed for lung texture recognition of different kinds of CT images, they are recognized as effective methods for deep network field adaptation. The technical scheme provided by the invention is more effective than the two methods.