Multi-modal sign classification method suitable for Lee's artificial liver

Abstract

The invention relates to the technical field of image recognition, and discloses a multi-modal sign classification method suitable for Lee's artificial liver application, which comprises the following steps: acquiring CT image data and physiological examination index data, in an upper computer, acquiring CT image feature vectors from the CT image data through a CT image feature extraction network, and classifying the CT image feature vectors into the physiological examination index data; the physiological examination index input vector obtains a physiological examination index feature vector through a physiological examination index feature extraction network, and then the CT image feature vector and the physiological examination index feature vector obtain a dichotomy result through a feature fusion output network. According to the method, collected CT image data and physiological examination index data are fully utilized, semantics and features contained in different modal data are extracted through a deep learning method, adaptive fusion of different modal features is achieved by introducing a cross attention method, multi-modal combined signs of CT-physiological examination indexes are constructed, and the multi-modal combined signs of the CT-physiological examination indexes are obtained. And the representation capability of the network is further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a multimodal sign classification method suitable for Lee's artificial liver application. Background Art

[0002] The concept of artificial liver was proposed in the late 1950s. In the more than 60 years since its development, it has undergone many technical improvements and achieved a lot of results. In the 1980s, the team of Academician Li Lanjuan in China began to study non-biological artificial liver technology. After years of hard work, they created an artificial liver system suitable for clinical application and providing individualized treatment according to the patient's condition, namely Li's Artificial Liver System (Li-ALS). However, non-biological artificial liver treatment also has certain limitations in improving liver failure. Considering many issues such as the biocompatibility of materials and media used in artificial liver treatment, it is necessary to extract and classify the combined signs of CT images and physiological examination indicators of patients before surgery.

[0003] Before receiving artificial liver treatment, patients will consult with doctors and record their medical records. Then, based on the chief complaint and medical history, they will make a preliminary attempt to determine the cause of the disease. They will then undergo a number of routine examinations to obtain more objective physiological information. After integrating the above information, professional doctors will formulate a treatment plan. In order to improve the efficiency and accuracy of diagnosis and treatment, it is necessary to explore ways to quickly mine and analyze medical information through computer technology and help doctors make judgments to improve the efficiency of clinical treatment. Multimodal fusion refers to the integration and fusion of information from two or more modalities to obtain a more comprehensive, accurate and rich joint semantic representation. Its goal is to effectively integrate information from different modalities to improve system performance and enhance the ability to understand and express data. The invention number CN116467440A, "Multi-level semantic text classification method based on Li's artificial liver medical records", collects text data of current medical history, chief complaint and past medical history in medical records, and then obtains the chief complaint word vector matrix, past medical history word vector matrix and current medical history word vector matrix through the text representation network based on multi-level semantics, and then outputs the classification results through the feature fusion prediction network after passing through the feature extraction network, which fully and effectively utilizes the semantic information of the three texts of current medical history, chief complaint and past medical history, and improves the accuracy of prediction. The invention number CN202210343260.3, "A preoperative diagnosis and treatment system and use method of Li's artificial liver based on electronic medical records", includes a medical record management subsystem and an intelligent analysis subsystem, and uses a deep learning network to analyze the pure text category data and test numerical data in the electronic medical record information and give classification results. Since the analysis and prediction of whether it is suitable for Li's artificial liver are only based on the patient's medical record text analysis and processing, and are not combined with the patient's CT image data, the obtained prediction results are inevitably limited. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a multimodal sign classification method suitable for the application of Lee's artificial liver, which is used to realize the combined sign classification suitable for the application of Lee's artificial liver by extracting and fusing features of CT image data and physiological examination index data.

[0005] In order to solve the above technical problems, the present invention provides a multimodal sign classification method suitable for Lee's artificial liver application, comprising the following steps: collecting CT image data and physiological examination index data, and then in a host computer, the CT image data is enhanced by histogram equalization, and the physiological examination index data is normalized to construct a physiological examination index input vector, which are used together as inputs of a multimodal sign classification network to obtain a binary classification result, and the result is displayed in the host computer;

[0006] The multimodal sign classification network includes a CT image feature extraction network, a physiological examination index feature extraction network and a feature fusion output network. The CT image data obtains a CT image feature vector through the CT image feature extraction network, and the physiological examination index input vector obtains a physiological examination index feature vector through the physiological examination index feature extraction network. Then, the CT image feature vector and the physiological examination index feature vector are passed through the feature fusion output network to obtain a binary classification result.

[0007] As an improvement of the multimodal sign classification method suitable for Lee's artificial liver application of the present invention:

[0008] The CT image feature extraction network is based on the ResNet50 network, and the original fully connected layer is deleted. It includes a starting convolution layer, a starting maximum pooling layer, four residual convolution layers and an average pooling layer in sequence, and then the basic residual structure in the four residual convolution layers is replaced by a branched convolution block.

[0009] As a further improvement of the multimodal sign classification method suitable for Lee's artificial liver application of the present invention:

[0010] The physiological examination index feature extraction network is based on a multi-layer perceptron network, including two fully connected layers of hidden layers connecting the physiological examination index one-dimensional input vector, and using Relu as a nonlinear transformation activation function to obtain an output feature vector.

[0011] As a further improvement of the multimodal sign classification method suitable for Lee's artificial liver application of the present invention:

[0012] The specific processing steps of the feature fusion classification network are:

[0013] (1) Calculate the cross-attention of CT image feature vector and physiological examination index feature vector

[0014]

[0015] Among them, X is the input feature vector, M and N represent the CT image feature vector and the physiological examination index feature vector respectively, and W Q , W K , W V is a trainable matrix;

[0016] (2) Compute new multimodal input features based on cross-attention:

[0017]

[0018] (3) The feature vector M in and N in Perform concatenation to obtain a multimodal feature input vector;

[0019] (4) The multimodal feature input vector is classified into two categories through a fully connected neural network and softmax. The fully connected neural network includes three fully connected hidden layers to process the multimodal feature input vector.

[0020] As a further improvement of the multimodal sign classification method suitable for Lee's artificial liver application of the present invention:

[0021] The operation of the branch convolution block is specifically as follows:

[0022] (1) The operations of each branch are:

[0023] z i =G(T i (x)), i∈{1.2,...,C} (1)

[0024] Among them, x is the input feature, G represents the operation of the GC block module, and T i Indicates the dimension reduction, transformation and dimension increase operations on the input x, i∈{1.2,…,C}, where C represents the number of branches;

[0025] (2) Add the processing results of each branch, expressed as:

[0026]

[0027] (3) Using cross-layer connections, the original input x is added to the output of the branch addition to obtain the output of each branch convolution block, which is expressed as:

[0028] y=x+F(x) (3).

[0029] As a further improvement of the multimodal sign classification method suitable for Lee's artificial liver application of the present invention:

[0030] The training process of the multimodal sign classification network is:

[0031] (1) CT imaging data and physiological examination index data of patients with liver failure were obtained through Internet search and from cooperative hospitals. Each sample was classified into two categories and then divided into training set and test set. The CT images in the training set and test set were subjected to data enhancement processing, including random mirroring, random rotation, random cropping and noise addition, and the physiological examination index data were normalized.

[0032] (2) The training set is input into the multimodal sign classification network, and the Adam optimizer is used to optimize the network. The learning rate is decayed in the form of cosine decay. The training is repeated for 100 iterations in total, and then the test set is used to test the multimodal sign classification network after training.

[0033] The beneficial effects of the present invention are mainly reflected in:

[0034] 1. The present invention improves the feature capture capability at different levels by introducing grouped convolution and global attention modules into the ResNet50 network, optimizes the feature extraction capability of the original ResNet50 network for images, makes it more suitable for classification tasks in this scenario, simplifies the network on the basis of improving the network capability, and reduces the number of parameters and the amount of computation.

[0035] 2. The present invention makes full use of the CT image data and physiological examination index data collected during the preoperative examination of Li's artificial liver treatment, extracts the semantics and features contained in the data of different modalities through deep learning methods, introduces the cross-attention method to achieve adaptive fusion of different modal features, and innovatively constructs a multi-modal joint sign of CT-physiological examination indicators, further improving the representation ability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0037] Figure 1 A schematic diagram of a module of a multimodal sign classification method suitable for Lee's artificial liver application of the present invention;

[0038] Figure 2 It is a schematic diagram of the structure of the branch convolution block in the CT image feature extraction network of the present invention;

[0039] Figure 3 It is a structural schematic diagram of the CT image feature extraction network of the present invention;

[0040] Figure 4 Schematic diagram of the structure of the global context attention module;

[0041] Figure 5 This is a schematic diagram of the network structure for extracting physiological examination index features according to the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto:

[0043] Embodiment 1: A multimodal sign classification method suitable for Lee's artificial liver application, comprising the following steps:

[0044] Step S1: constructing a multimodal sign classification network suitable for Lee's artificial liver application (referred to as a multimodal sign classification network)

[0045] The multimodal sign classification network is applied to the preoperative scenario of Lee's artificial liver, such as Figure 1 As shown, CT image data is processed by a CT image feature extraction network, and physiological examination index data is processed by a physiological examination index feature extraction network to obtain a multimodal feature vector, and the multimodal feature vector is output through a feature fusion network to obtain a joint sign classification result.

[0046] S101 builds CT image feature extraction network

[0047] The CT image feature extraction network is based on the ResNet network theory and takes the ResNet50 network as the basic model to construct a CT image feature extraction network in medical scenarios. From the ResNet50 network, the fully connected layer for output classification is deleted, and the network structure is sequentially a starting convolution layer, a starting maximum pooling layer, 4 residual convolution layers, an average pooling layer, and a fully connected layer for classification, wherein the 4 residual convolution layers sequentially contain 3, 4, 6, and 3 basic residual structures, and then the basic residual structure is replaced by a branch convolution block. The CT image feature extraction network structure of the present invention improved based on the ResNet50 network is as follows: Figure 3 shown.

[0048] The basic residual structure (bottleneck) in the four residual convolution layers of the ResNet50 network is replaced by a branched convolution block with an isomorphic branch structure. The structure of the branched convolution block is as follows: Figure 2 In order to avoid the huge amount of computation caused by directly performing convolution operations on high-dimensional feature maps, the branch convolution block reduces the dimensionality of multi-channel inputs through 1x1 convolution kernels before performing convolution operations, optimizing the network's ability to express features while controlling the number of parameters and the amount of computation. In addition, since the parameters of each branch are independent, different branches can learn feature representations of different scales by processing the same feature map in parallel.

[0049] Each branch of the isomorphic branch structure contains the residual structure of the input x for dimensionality reduction, transformation, and dimensionality increase, denoted as T i(x), i∈{1.2,...,C}, where C represents the number of branches; the GCblock module embedded after the residual structure of each branch is used to capture the global attention of the output of the residual structure, improving the network's ability to capture features at different spatial locations when processing images;

[0050] The output of each branch can be expressed as:

[0051] z i =G(T i (x)), i∈{1.2,...,C} (1)

[0052] Among them, G i () indicates the operation of the GC block module.

[0053] Add the processing results of each branch to realize the aggregation transformation of the branch sub-network, which is expressed as:

[0054]

[0055] The final output of each branch convolution block is obtained by adding the original input x to the output of the branch addition using cross-layer connections, which is expressed as:

[0056] y=x+F(x) (3)

[0057] The GC block module passes 1x1 convolution W k Change the number of channels of the input feature map, and then use the softmax function to calculate the similarity between pixel neighborhoods to obtain the attention weight a of the input feature map j , and then perform attention pooling to obtain the global context feature f i , expressed as:

[0058]

[0059] Where N p is the size of the input feature map; x is the pixel value at each position of the feature map.

[0060] Then, the bottleneck transform module is used to transform the global context features to capture channel dependencies. The transformation process can be expressed as:

[0061] S(·)=W v2 ReLU(LN(W v1 O)) (6)

[0062] Where W v1 , W v2represents 1x1 convolution, LN represents layer normalization operation; the global context features are aggregated to the features of each position by element accumulation, and the output of GC block is:

[0063] z i =x i +W v2 ReLU(LN(W v1 f)) (7)

[0064] S102 constructs a physiological examination index feature extraction network;

[0065] Based on the multi-layer perceptron network theory, a physiological examination index feature extraction network in medical scenarios was constructed, such as Figure 5 As shown in the figure, the one-dimensional vector constructed after Z-Score normalization of 55 physiological examination indicators such as total bilirubin, C-reactive protein, and average hemoglobin content is used as input. The network as a whole connects the physiological examination indicator input vector through a hidden layer containing two fully connected layers. The two hidden layers include 100 neurons and 55 neurons respectively to extract features from the input vector, and Relu is used as the nonlinear transformation activation function to obtain the output 55x1 physiological examination indicator feature vector.

[0066] S103 builds a feature fusion classification network

[0067] The feature fusion classification network is used to combine the CT image feature vector and the physiological examination index feature vector to perform multi-scale evaluation and classification of the CT-physiological examination index combined signs during the patient's visit. The feature fusion classification network includes the following processing steps:

[0068] (1) Calculate the cross-attention of CT image feature vector and physiological examination index feature vector

[0069]

[0070] Among them, X is the input feature vector, M and N represent the CT image feature vector and the physiological examination index feature vector respectively, and W Q , W K , W V is a trainable matrix;

[0071] (2) Compute new multimodal input features based on cross-attention:

[0072]

[0073] (3) The two sets of feature vectors M obtained after attention calculation in and N inThe multi-modal feature input vector is obtained by splicing. Since the size of the CT image feature vector and the physiological examination index feature vector are not uniform, feature selection and adjustment are performed before splicing the feature vector. Feature vector M in The size is 7x7x2048. The number of channels is changed to 100 through a set of convolutions containing 100 1x1 convolution kernels, and then the vectors are horizontally spliced ​​to obtain a 49x100 CT image feature vector; the feature vector N in The size is 55x1. The number of channels is changed to 100 through a set of 1x1 convolution kernels to obtain a 55x100 physiological examination index feature vector; the 49x100 CT image feature vector and the 55x100 physiological examination index feature vector are concatenated to obtain a 104x 100 multimodal feature input vector.

[0074] (4) A fully connected neural network was constructed to achieve the combined physical sign classification task of whether the patient is suitable for the Lee's artificial liver application. The network as a whole processed the input multimodal feature input vector through a hidden layer consisting of three fully connected layers, and finally obtained the output prediction result through softmax.

[0075] Step S2: training a multimodal sign classification network

[0076] Collection and construction of the S201 dataset;

[0077] In order to study whether the signs of patients with liver failure are suitable for the use of Li's artificial liver, data collection and construction were carried out. The data were obtained mainly through two channels: one is to collect relevant CT images and physiological examination indicators through network resources such as Google, Baidu, and Sogou; the other is to cooperate with relevant hospitals in Zhejiang Province to obtain a real data set of medical records of patients with liver failure in the hospital. Each case of data contains the patient's CT image data and physiological examination indicator data, among which the physiological examination indicator data include 55 indicators such as total bilirubin, C-reactive protein, and average hemoglobin content. The constructed data set mainly contains two categories to be classified: suitable for application and not suitable for application. The specific data set information is shown in Table 1 below.

[0078] Table 1. Training dataset

[0079] Dataset Is it suitable for treatment with Lee's artificial liver? Number of categories 2 (positive and negative samples) Category Details Suitable for application, not suitable for application Number of patients 1157 Suitable for treatment 569 Not suitable for treatment 588

[0080] S202 dataset preprocessing,

[0081] The constructed dataset consists of 1,157 samples. The CT image data of each patient corresponds to the physiological examination index data one by one. Then the CT image data and the physiological examination index data are divided into a training set and a test set, with 926 cases in the training set and 231 cases in the test set. In order to improve the robustness and generalization ability of the model, the CT images in the training set and the test set were enhanced, including random mirroring, random rotation (-10, 10) degrees, random cropping and noise addition. The physiological examination index data in the training set and the test set were normalized.

[0082] S203 training environment configuration

[0083] The training of the multimodal sign classification network was carried out on a CentOS operating system server, using Tesla P4 GPU to accelerate the experiment, and the development environment was based on the PyTorch deep learning framework. The specific software and hardware configuration is shown in Table 2:

[0084] Table 2. Training hardware and software environment configuration

[0085]

[0086]

[0087] The training set is input into the multimodal sign classification network established in step S1, and the Adam optimizer is used to optimize the network. The learning rate can be dynamically adjusted to adapt to more optimization scenarios. The momentum coefficient is 0.9 and the weight decay is 0.0001. The initial learning rate (lr) of the optimizer is set to 0.01, and the learning rate is decayed in the form of cosine decay. Due to the problem of video memory, the input image size of a batch is 32, and a total of 100 iterations are trained. Then the test set is used to test the trained multimodal sign classification network, so as to obtain a multimodal sign classification network that can be used online.

[0088] Practical use of the S3 multimodal sign classification network model

[0089] The real data of the medical records of patients with liver failure are obtained through cooperative hospitals in Zhejiang Province, including CT image data and physiological examination index data, and the image data is enhanced by histogram equalization, and the physiological examination index input vector is constructed by Z-Score normalization. Then, in the host computer, the multimodal sign classification network with online production capability obtained in step S2 is input, the CT image data is obtained by the CT image feature extraction network to obtain the CT image feature vector, and the physiological examination index input vector is obtained by the physiological examination index feature extraction network to obtain the physiological examination index feature vector, and then the joint sign classification result is obtained through the feature fusion output network, including the binary classification result of whether it is suitable for the application of Lee's artificial liver or not, and the result is displayed and stored in the host computer.

[0090] experiment:

[0091] In order to further verify the performance improvement effect of the method of the present invention on the multimodal sign classification network suitable for Lee's artificial liver application, this experiment compares the multimodal sign classification network of the present invention with the algorithm models of the other two documents on the multimodal data set before Lee's artificial liver surgery constructed in step 2 of Example 1. Document 1 (see Wu X, Hong D, Chanussot J. Convolutional neural networks for multimodal remote sensing data classification [J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60) proposes a multimodal remote sensing data classification framework CCR-Net, which extracts image modal features through multi-layer convolution operations and splices modal features to obtain the final fusion feature representation. Reference 2 (Zhang R, Qiu L, Zhang W, et al. Vt-clip: Enhancing vision-language models with visual-guided texts [J]. arXiv preprint arXiv: 2112.02399, 2021) uses visually guided text to enhance the performance of CLIP (Contrastive Language-Image Pre-training). After extracting image and text features separately, it uses the attention mechanism to guide the text to aggregate visual features.

[0092] The experimental evaluation indicators are as follows:

[0093] In the classification task of the detection problem, the classification results of the model and the true labels of the samples are represented by a confusion matrix to represent four combinations, namely true positive (TP), false positive (FP), true negative (TN), and false negative (FN). The confusion matrix is ​​shown in Table 3 below:

[0094] Table 3. Confusion matrix

[0095]

[0096] Accuracy refers to the percentage of correctly predicted samples in the total samples, which is defined as follows:

[0097]

[0098] Precision (for positive / negative samples) refers to the proportion of samples that are actually positive / negative among all samples predicted to be positive / negative samples, and is defined as follows:

[0099]

[0100] The recall rate is used to describe the proportion of positive samples judged as true in the classifier prediction results to the total samples, and is defined as follows:

[0101]

[0102] The F1 value is a harmonic mean of precision and recall, which takes into account both indicators and can more objectively reflect the effect of the model. It is defined as follows:

[0103]

[0104] The comparative experimental results are shown in Table 4:

[0105] Table 4. Comparative experimental results

[0106] Model Accuracy Accuracy Recall F1 Reference 1 0.8902 0.8823 0.8963 0.8893 Reference 2 0.9092 0.8973 0.9209 0.9089 The present invention 0.9127 0.9007 0.9244 0.9124

[0107] The results show that compared with the model in the reference, the model of the present invention improves the ResNet50 structure by introducing grouped convolution blocks and global attention modules, enhances the CT image feature extraction capability, effectively realizes the combination of contextual information and shallow detail features, and completes the adaptive fusion of CT image features and physiological examination index features through the cross-attention mechanism, and makes a multimodal comprehensive analysis and classification of whether the patient's current physical condition is suitable for treatment. Therefore, the model performs better and has the best performance in various indicators.

[0108] In summary, the present invention improves the feature capture capability at different levels by introducing grouped convolution and global attention modules, optimizes the original ResNet50 network's feature extraction capability for images, makes it more suitable for classification tasks in this scenario, simplifies the network on the basis of improving network capabilities, and reduces the number of parameters and computations. In addition, the present invention makes full use of the CT image data and physiological examination index data collected during the preoperative examination of the Lee's artificial liver, extracts the semantics contained in different modal data through deep learning methods, and introduces the cross-attention method to achieve adaptive fusion of different modal features, further improving the network's representation capabilities.

[0109] Finally, it should be noted that the above examples are only some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A multimodal sign classification method suitable for Lee's artificial liver application, characterized by: The method includes the following steps: collecting CT image data and physiological examination index data, and then in a host computer, the CT image data is enhanced by histogram equalization, and the physiological examination index data is normalized to construct a physiological examination index input vector, which are used together as inputs of a multimodal sign classification network to obtain a binary classification result, and the result is displayed in the host computer; The multimodal sign classification network includes a CT image feature extraction network, a physiological examination index feature extraction network and a feature fusion output network. The CT image data obtains a CT image feature vector through the CT image feature extraction network, and the physiological examination index input vector obtains a physiological examination index feature vector through the physiological examination index feature extraction network. Then, the CT image feature vector and the physiological examination index feature vector are passed through the feature fusion output network to obtain a binary classification result.

2. A multimodal sign classification method suitable for Lee's artificial liver application according to claim 1, characterized in that: The CT image feature extraction network is based on the ResNet50 network, and the original fully connected layer is deleted. It includes a starting convolution layer, a starting maximum pooling layer, four residual convolution layers and an average pooling layer in sequence, and then the basic residual structure in the four residual convolution layers is replaced by a branched convolution block.

3. A multimodal sign classification method suitable for Lee's artificial liver application according to claim 2, characterized in that: The physiological examination index feature extraction network is based on a multi-layer perceptron network, including two fully connected layers of hidden layers connecting the physiological examination index one-dimensional input vector, and using Relu as a nonlinear transformation activation function to obtain an output feature vector.

4. A multimodal sign classification method suitable for Lee's artificial liver application according to claim 3, characterized in that: The specific processing steps of the feature fusion classification network are: (1) Calculate the cross-attention of CT image feature vector and physiological examination index feature vector Among them, X is the input feature vector, M and N represent the CT image feature vector and the physiological examination index feature vector respectively, and W Q , W K , W V is a trainable matrix; (2) Compute new multimodal input features based on cross-attention: (3) The feature vector M in and N in Perform concatenation to obtain a multimodal feature input vector; (4) The multimodal feature input vector is classified into two categories through a fully connected neural network and softmax. The fully connected neural network includes three fully connected hidden layers to process the multimodal feature input vector.

5. A multimodal sign classification method suitable for Lee's artificial liver application according to claim 4, characterized in that: The operation of the branch convolution block is specifically as follows: (1) The operations of each branch are: z i =G(T i (x)),i∈{1.2,...,C} (1) Among them, x is the input feature, G represents the operation of the GC block module, and T i Indicates the dimension reduction, transformation and dimension increase operations on the input x, i∈{1.2,…,C}, where C represents the number of branches; (2) Add the processing results of each branch, expressed as: (3) Using cross-layer connections, the original input x is added to the output of the branch addition to obtain the output of each branch convolution block, which is expressed as: y=x+F(x) (3).

6. A multimodal sign classification method suitable for Lee's artificial liver application according to claim 5, characterized in that: The training process of the multimodal sign classification network is: (1) CT imaging data and physiological examination index data of patients with liver failure were obtained through Internet search and from cooperative hospitals. Each sample was classified into two categories and then divided into training set and test set. The CT images in the training set and test set were subjected to data enhancement processing, including random mirroring, random rotation, random cropping and noise addition, and the physiological examination index data were normalized. (2) The training set is input into the multimodal sign classification network, and the Adam optimizer is used to optimize the network. The learning rate is decayed in the form of cosine decay. The training is repeated for 100 iterations in total, and then the test set is used to test the multimodal sign classification network after training.

Images