A pet / ct image recognition method and device

Abstract

The application provides a PET / CT image recognition method and device, the image recognition method comprises the following steps: target substance extraction is carried out through a PET / CT original image, and a CT data set and a PET data set are obtained; the CT data set and the PET data set are input into a double-view deep learning classifier for feature extraction and feature classification, and CT and PET feature classification is obtained; the double-view deep learning classifier uses an MVPD-Net model; based on a voting recognition mechanism, the CT feature classification and the PET feature classification are fused, and a final recognition result is obtained. The model provided by the application is a PET / CT image classification model based on a multi-view 3D convolutional neural network, and can more reliably perform image recognition. Target substance extraction is used, and the problem of high sensitivity and low specificity of the classification model caused by the heterogeneity of the target tissue due to the similarity of the structure and density of the target tissue and the surrounding tissue is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided medicine, and in particular to a PET / CT image recognition method and device. BACKGROUND

[0002] Lung cancer is one of the most serious diseases in the world today, and reliable diagnosis and early prevention of lung cancer is a very important research topic. As the most important early symptom of lung cancer, lung nodules, the benign and malignant classification of lung nodules has become an important measure for early prevention of lung cancer. Therefore, how to reliably classify lung nodules has become a research hotspot for early diagnosis of lung cancer.

[0003] Medical imaging technology, as an image means for capturing and presenting the morphological structure and functional state change of the body in a non-invasive manner, is different from traditional medical images which only provide structural imaging or functional imaging. PET / CT, as a fusion imaging, can simultaneously image the morphological and functional state changes of lung nodules, and provides a strong basis for early diagnosis of lung cancer.

[0004] Current computer-aided diagnosis technology (Computer Aided Diagnosis, CAD) for lung nodules has achieved good results. However, existing methods mainly target CT images, mainly through manual labeling of lung nodules, feature extraction, and construction of classification networks for recognition. However, the pre-labeling work not only relies on manual work, but also due to the subjectivity of each manual labeling, may result in different degrees of error. At the same time, due to the similarity of the structure and density of the target tissue and the surrounding tissue, the heterogeneity of the target tissue itself leads to high sensitivity and low specificity of the classification model. SUMMARY

[0005] The present application provides a PET / CT image recognition method and device to solve the problem of high sensitivity and low specificity of the classification model due to the similarity of the structure and density of the target tissue and the surrounding tissue, and the heterogeneity of the target tissue itself.

[0006] In a first aspect, the present application provides a PET / CT image recognition method, comprising: performing target parenchyma extraction through a PET / CT original image to obtain a CT data set and a PET data set; inputting the CT data set and the PET data set into a dual-view deep learning classifier for feature extraction and feature classification to obtain CT feature classification and PET feature classification; the dual-view deep learning classifier uses a MVPD-Net model; based on a voting-based recognition mechanism, the CT feature classification and the PET feature classification are fused to obtain a final recognition result.

[0007] Optionally, the step of processing the PET / CT original image to obtain a CT data set and a PET data set comprises: obtaining a denoised CT image; performing resampling processing on the denoised CT image to obtain a resampled CT image; performing threshold segmentation on the resampled CT image based on a K-means clustering algorithm to obtain a K-means segmentation threshold; performing threshold segmentation on the resampled CT image based on an OTSU algorithm to obtain an OTSU segmentation threshold; taking the mean of the K-means segmentation threshold and the OTSU segmentation threshold as a target parenchymal segmentation threshold; performing morphological transformation on the CT original image using the target parenchymal segmentation threshold to obtain a CT target parenchymal mask Mask; performing target parenchymal extraction on the CT original image using the CT target parenchymal mask Mask to obtain the CT data set; performing size scaling on the CT target parenchymal mask Mask according to the ratio of the CT resolution to the PET resolution to obtain a PET target parenchymal mask Mask; and performing target parenchymal extraction on the PET original image using the PET target parenchymal mask Mask to obtain the PET data set.

[0008] Optionally, the step of obtaining the denoised CT image comprises: performing data cleaning on the CT original image to obtain a cleaned CT image; the data cleaning is to filter and remove data with missing information, unclear results, and outdated data; and performing data denoising on the cleaned CT image to obtain the denoised CT image; the data denoising is to denoise the cleaned CT image based on a median filtering method with a filter window of 3x3.

[0009] Optionally, the step of performing resampling processing on the denoised CT image to obtain a resampled CT image comprises: converting the pixel value of the denoised CT image into a CT value using a conversion formula to obtain the resampled CT image.

[0010] The conversion formula is

[0011] HU=pixel_valxrescale_slope+rescale_intercept

[0012] wherein pixel_val is the pixel value of the denoised CT image, and rescale_slope and rescale_intercept are the slope and intercept.

[0013] Optionally, the step of performing threshold segmentation on the resampled CT image based on a K-means clustering algorithm to obtain a K-means segmentation threshold comprises: randomly selecting K as 2 data as initial clustering centers for the resampled CT image; calculating the Euclidean distance of the remaining data from the initial clustering centers, classifying according to the nearest principle to generate clusters; calculating the average value of each cluster center as a new clustering center; if the clustering center changes or the number of iterations reaches a threshold, the clustering result is obtained, and the average value of the clustering result is taken as the K-means segmentation threshold.

[0014] Optionally, the step of performing threshold segmentation on the resampled CT image based on an OTSU algorithm to obtain an OTSU segmentation threshold comprises: calculating the global expectation of the gray image for the resampled CT image; traversing the gray value to divide the gray image into foreground and background; calculating the expectation and probability of the foreground and background; when the threshold inter-class variance is maximum, the threshold is obtained as the OTSU segmentation threshold.

[0015] Optionally, the step of performing morphological transformation on the CT original image using the target parenchyma segmentation threshold to obtain a CT target parenchyma mask Mask comprises: performing binaryzation processing on the CT original image using the target parenchyma segmentation threshold to obtain a first image; performing erosion and expansion operations with a kernel of 3x3 and 8x8 on the first image to obtain a second image; performing a closing operation with a kernel of 25x25 on the second image to obtain the CT target parenchyma mask Mask.

[0016] Optionally, the MVPD-Net model comprises:

[0017] The feature extraction sub-network comprises: a CT feature extractor and a PET feature extractor, which are used to extract features from the CT data set and the PET data set respectively; the CT feature extractor and the PET feature extractor are both constructed by combining 3D convolution layers and 3D pooling layers; the 3D convolution layer of the CT feature extractor comprises a convolution kernel of 7x7x7 and a number of convolution kernels of 64; the 3D pooling layer of the CT feature extractor is 1x3x3; the CT feature extractor extracts high-dimensional features through the stacking of 5 feature extraction blocks; the 3D convolution layer of the PET feature extractor comprises a convolution kernel of 3x3x4 and a number of convolution kernels of 64; the 3D pooling layer of the PET feature extractor is 1x3x3; the PET feature extractor extracts high-dimensional features through the stacking of 4 feature extraction blocks;

[0018] The feature classification sub-network is used to classify the CT features and PET features obtained by the feature extractor; the feature classification sub-network comprises: performing feature vector flattening; performing feature classification through 2 fully connected layers; obtaining classification results based on a Softmax activation function.

[0019] Optionally, the voting-based recognition mechanism comprises:

[0020] y v = MAX(y (v) )

[0021] wherein y (v) represents the recognition results of different view channels, and y v represents the results after voting of different view channels.

[0022] In a second aspect, the application further provides a PET / CT image recognition device, comprising: a controller configured to execute the PET / CT image recognition method of the first aspect.

[0023] The application provides a PET / CT image recognition method and device, the image recognition method comprising: extracting a target substance from a PET / CT original image to obtain a CT data set and a PET data set; inputting the CT data set and the PET data set into a dual-view deep learning classifier for feature extraction and feature classification to obtain CT feature classification and PET feature classification; the dual-view deep learning classifier uses an MVPD-Net model; and a voting-based recognition mechanism is used to fuse the CT feature classification and the PET feature classification to obtain a final recognition result. Figure 3 The MVPD-Net model is a PET / CT image classification model based on a multi-view deep convolutional neural network, which can more reliably recognize PET / CT images and has better interpretability. The mean of the segmentation threshold of K-means and the segmentation threshold of OTSU is taken as the target substance segmentation threshold for target substance extraction, which solves the problem of high sensitivity and low specificity of the classification model due to the structural and density similarity of the target tissue and the surrounding tissue and the heterogeneity of the target tissue itself. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0025] Figure 1 A flowchart of a PET / CT image recognition method according to the application;

[0026] Figure 2 A flowchart of processing a PET / CT original image to obtain a CT data set and a PET data set according to the application;

[0027] Figure 3A flowchart of a process for threshold segmentation based on a K-means clustering algorithm on a resampled CT image to obtain a K-means segmentation threshold according to the present application;

[0028] Figure 4 A flowchart of a process for threshold segmentation based on an OTSU algorithm on a resampled CT image to obtain an OTSU segmentation threshold according to the present application;

[0029] Figure 5 A flowchart of a process for morphological transformation of a CT original image using the target parenchymal segmentation threshold to obtain a CT target parenchymal mask Mask according to the present application;

[0030] Figure 6 A structural framework diagram of the MVPD-Net model according to the present application. DETAILED DESCRIPTION

[0031] The embodiments will be described in detail below with reference to the drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following embodiments are not meant to represent all implementations consistent with the present application. Rather, they are merely examples of systems and methods consistent with some aspects of the present application as detailed in the appended claims.

[0032] PET, which stands for Positron Emission Computed Tomography, is an advanced clinical examination imaging technique in the field of nuclear medicine. CT, which stands for Computed Tomography, is a technique that uses precisely collimated X-ray beams, gamma rays, ultrasound, etc., together with highly sensitive detectors to make one after another cross-sectional scans around a certain part of the human body, and has the characteristics of fast scanning time and clear images, and can be used for the examination of various diseases. PET / CT is a combination of PET and CT, which provides detailed functional and metabolic molecular information of the lesion by PET, and precise anatomical positioning of the lesion by CT. Once the image is obtained, it can provide cross-sectional images of various positions of the whole body, and has the characteristics of sensitivity, accuracy, specificity, and precise positioning.

[0033] With the help of computer aided diagnosis technology (Computer Aided Diagnosis, CAD), CT image data is automatically analyzed, thereby improving processing efficiency. In some embodiments, for CT images, artificial labeling, feature extraction, and classification network construction are performed for recognition. However, the early labeling work not only depends on artificial, but also due to the subjectivity of each artificial labeling, it may cause different degrees of error. At the same time, due to the similarity of the structure and density of the target tissue and the surrounding tissue, the heterogeneity of the target tissue itself leads to high sensitivity and low specificity of the classification model.

[0034] To solve the above problems, the present application provides a PET / CT image recognition method, as shown in Figure 1

[0035] S100: Target parenchyma extraction is performed on the PET / CT original image to obtain a CT data set and a PET data set.

[0036] Traditional labeling only focuses on the extraction of lesion position, for example: the position extraction of lung nodules, but due to the similarity of the structure and density of the lung nodules and the surrounding tissue, the heterogeneity of the lung nodules itself leads to high sensitivity and low specificity of the classification model. In order to overcome the above shortcomings, target parenchyma extraction is adopted, which refers to removing irrelevant areas from the entire PET / CT image and extracting the complete target area, for example: lung parenchyma extraction is to extract the entire lung area completely, and accurate segmentation of lung parenchyma can effectively improve the judgment accuracy of the model for lung nodules.

[0037] In an illustrative embodiment, as shown in Figure 2 The step of processing the PET / CT original image to obtain a CT data set and a PET data set includes:

[0038] S110: Obtain a denoised CT image, including:

[0039] Data cleaning is performed on the CT original image to obtain a cleaned CT image; the data cleaning is to filter and remove the data with missing information, unclear results and expiration. There are some data with missing information, unclear results and expiration in the CT original image data, which need to be filtered and removed, so that the cleaned CT image is all valid data. For example, after data collection, a total of 3145 experimental data are obtained, and after cleaning, a total of 1881 valid data are obtained.

[0040] ​The cleaned CT image is subjected to data denoising to obtain a denoised CT image; the data denoising is a median filtering method based on a 3*3 filtering window to denoise the cleaned CT image. Since the computer tomography imaging of a human body is interfered by various factors, resulting in noise in the CT image, such as non-standard operation, image reconstruction algorithm limitation, etc., the median filtering method based on the 3*3 filtering window is used to denoise the cleaned CT image, so that the obtained CT image is clearer.

[0041] S120: The denoised CT image is subjected to resampling processing to obtain a resampled CT image, including:

[0042] The pixel value of the denoised CT image is converted into a CT value, called Hounsfield unit (HU), by using a conversion formula (1) to obtain the resampled CT image.

[0043] The conversion formula (1) is

[0044] HU=pixel_val*rescale_slope+rescale_intercept (1)

[0045] Wherein, pixel_val is the pixel value of the denoised CT image, and rescale_slope and rescale_intercept are the slope and intercept.

[0046] S130: The resampled CT image is subjected to threshold segmentation based on a K-means clustering algorithm to obtain a K-means segmentation threshold.

[0047] The K-means clustering algorithm is an iterative solution clustering analysis algorithm, and the steps are as follows: the data is divided into K groups, K objects are randomly selected as initial cluster centers, then the distance between each object and each seed cluster center is calculated, and each object is assigned to the cluster center closest to it. The cluster center and the objects assigned to them represent a cluster. The cluster center of the cluster is recalculated according to the existing objects in the cluster after each sample is assigned. This process will be repeated until a certain termination condition is met.

[0048] In an illustrative embodiment, as shown in Figure 3 The step of subjecting the resampled CT image to threshold segmentation based on the K-means clustering algorithm to obtain the K-means segmentation threshold includes:

[0049] S131: randomly select K as 2 data as initial clustering centers after resampling the CT image. Because the OTSU algorithm is a binary algorithm, K is selected as 2 data, in order to cooperate with the OTSU algorithm.

[0050] S132: calculate the Euclidean distance of the remaining data from the initial clustering center, classify according to the nearest principle, and generate clusters.

[0051] S133: calculate the average value of each cluster center as a new clustering center.

[0052] S134: if the clustering center changes, or, the iteration number reaches the threshold, the clustering result is obtained, and the mean value of the clustering result is taken as the K-means segmentation threshold.

[0053] If the clustering center does not change, and, the iteration number does not reach the threshold, jump to step S132 to repeat execution until the conditions are met.

[0054] Using K-means clustering algorithm, the algorithm idea is simple, the convergence speed is fast, the clustering effect is better, and the algorithm is relatively strong in interpretability.

[0055] S140: performing threshold segmentation based on OTSU algorithm on the resampled CT image to obtain the OTSU segmentation threshold.

[0056] OTSU algorithm (common law or maximum between-class variance method) is an image binarization algorithm. The algorithm assumes that there is a threshold T in the image, judges the size relationship of each pixel in the image with T, can divide all pixels into background C0 and foreground C1 two categories, and when the best T threshold is selected, the difference between the background part and the foreground part is the largest, and the OTSU algorithm uses the maximum between-class method to measure this difference. Since the OTSU algorithm uses the idea of maximum between-class method, the algorithm is also called OTSU maximum between-class variance method.

[0057] In an illustrative embodiment, as shown in Figure 4 The step of performing threshold segmentation based on OTSU algorithm on the resampled CT image to obtain the OTSU segmentation threshold comprises:

[0058] S141: calculate the global expectation of the gray scale image for the resampled CT image;

[0059] S142: traverse the gray scale value, and divide the gray scale image into foreground and background;

[0060] S143: calculate the expectation and probability of the foreground and background;

[0061] S144: when the between-class variance of the threshold is the maximum, the threshold is obtained as the OTSU segmentation threshold.

[0062] S150: Take the mean of the K-means segmentation threshold and the OTSU segmentation threshold as the target substantial segmentation threshold.

[0063] Using OTSU algorithm alone is sensitive to image noise and can only segment a single target; when the size ratio of the target and the background is large, the inter-class variance function may present a double peak or a multi-peak, in which case the effect is not good. Using K-means clustering algorithm alone, the clustering result often converges to a local optimum and cannot obtain a global optimal solution, and the recognition effect of non-convex shape clusters is poor and is easily affected by noise, edge points and isolated points. In order to overcome the above shortcomings, the K-means segmentation threshold and the OTSU segmentation threshold obtained are averaged to obtain the final target substantial segmentation threshold, which can improve the accuracy of model recognition.

[0064] S160: Perform morphological transformation on the CT original image using the target substantial segmentation threshold to obtain a CT target substantial mask Mask, as shown in Figure 5 , which includes:

[0065] S161: Perform binaryzation processing on the CT original image using the target substantial segmentation threshold to obtain a first image.

[0066] S162: Perform erosion and expansion operations with a kernel of 3x3 and 8x8 on the first image to obtain a second image. The erosion and expansion operations with a kernel of 3x3 and 8x8 can remove the hollow parts inside and outside the lung parenchyma.

[0067] S163: Perform a closing operation with a kernel of 25x25 on the second image to obtain a CT target substantial mask Mask. Considering the adhesion of nodules at the lung boundary and the nodule inclusion standard (nodule diameter less than 3cm), the closing operation with a kernel of 25x25 is adopted to perform boundary completion, which can better solve the adhesion of large nodules at the lung boundary.

[0068] S170: Perform target substantial extraction on the CT original image using the CT target substantial mask Mask to obtain a CT data set. For example, 1881 valid CT original images are subjected to target substantial extraction to obtain a CT data set including 1881 data.

[0069] S180: The CT target substance mask is scaled according to the ratio of CT resolution to PET resolution to obtain the PET target substance mask. For example, if the CT resolution is 512×512 and the PET resolution is 128×128, the 512×512 CT target substance mask is scaled to 128×128, and then resized based on cubic interpolation to obtain the PET target substance mask.

[0070] S190: Extract the target essence from the original PET images using the PET target essence mask to obtain a PET dataset. For example, extract the target essence from 1881 valid original PET images to obtain a PET dataset containing 1881 data points.

[0071] S200: The CT and PET datasets are input into a dual-view deep learning classifier for feature extraction and classification, respectively, to obtain CT feature classification and PET feature classification. The dual-view deep learning classifier uses the MVPD-Net (multi-view Pulmonary nodules Net) model. The MVPD-Net model is based on multi-view... Figure 3 A classification model for PET / CT images using a D-convolutional neural network.

[0072] A convolutional neural network mainly consists of the following 5 structures:

[0073] The input layer is the input to the entire neural network. In convolutional neural networks that process images, it typically represents the pixel matrix of an image. Starting from the input layer, the convolutional neural network transforms the 3D matrix of the previous layer into the 3D matrix of the next layer through different neural network structures, until the final fully connected layer.

[0074] The convolutional layer is the most important part of a convolutional neural network. Unlike traditional fully connected layers, the input to each node in a convolutional layer is only a small portion of the input from the previous layer, known as the receptive field. Common kernel sizes for convolutions are 3×3 or 5×5. Convolutional layers attempt to analyze each small portion of the neural network more deeply to obtain features with a higher degree of abstraction. Generally, the node matrix processed by a convolutional layer becomes deeper, thus increasing the depth of the node matrix after a convolutional layer.

[0075] Pooling, the pooling layer neural network does not change the depth of the three-dimensional matrix, but it can reduce the size of the matrix, reduce the model operation. Pooling layer operation can be considered as a resolution higher picture into a resolution higher picture into a lower resolution picture. Through the pooling layer, the number of nodes in the last fully connected layer can be further reduced. Thus, the purpose of reducing the parameters in the entire neural network is achieved.

[0076] Fully connected layer, after the processing of multiple rounds of convolutional layer and pooling layer, the last of the convolutional neural network is generally a number of fully connected layers for feature classification and give the final classification result. After several rounds of convolutional layer and pooling layer, it can be considered that the information in the image has been abstracted into higher information content features. Convolutional layer and pooling layer can be considered as an automatic image feature extraction process. After feature extraction is completed, the classification task still needs to use the fully connected layer.

[0077] Softmax activation function, mainly used for classification problems. Through the Softmax activation function, the probability distribution of the current sample belonging to different categories can be obtained.

[0078] In an illustrative embodiment, as shown in Figure 6 The MVPD-Net model includes:

[0079] The feature extraction sub-network includes a CT feature extractor and a PET feature extractor for respectively extracting features from the CT dataset and the PET dataset; the CT feature extractor and the PET feature extractor are both constructed by combining 3D convolutional layers and 3D pooling layers; the 3D convolutional layer of the CT feature extractor includes a convolution kernel of 7x7x7 and a number of convolution kernels of 64; the 3D pooling layer of the CT feature extractor is 1x3x3; the CT feature extractor extracts high-dimensional features through stacking of 5 feature extraction blocks (see Table 1 for details of the CT-view column Conv2, Conv3, Conv4, Conv5, Conv6); the 3D convolutional layer of the PET feature extractor includes a convolution kernel of 3x3x4 and a number of convolution kernels of 64; the 3D pooling layer of the PET feature extractor is 1x3x3; the PET feature extractor extracts high-dimensional features through stacking of 4 feature extraction blocks (see Table 1 for details of the PET-view column Conv2, Conv3, Conv4, Conv5).

[0080] A feature classification subnetwork is used to classify CT and PET features obtained by the feature extractor. The feature classification subnetwork includes: flattening feature vectors; performing feature classification through two fully connected layers; and obtaining the classification result based on the Softmax activation function. If the features to be identified are of two types, such as benign or malignant, the features are classified into two types.

[0081] Table 1 MVPD-Net Parameter Settings

[0082]

[0083]

[0084] S300: Based on a voting-based identification mechanism, the CT feature classification and PET feature classification are fused to obtain the final identification result.

[0085] In one illustrative embodiment, the voting identification mechanism includes:

[0086] y v =MAX(y (v) (2)

[0087] Among them, y (v) Representing the recognition results of different view channels, y v This represents the result of voting from different view channels.

[0088] The identification results obtained from PET and CT are voted on, and the model is output with the highest probability value using formula (2).

[0089] Due to the black-box nature of deep learning models, interpretability has always been a hot research topic, especially in the field of medical image processing, where the lack of model interpretability limits clinicians' trust in such methods. Therefore, this application conducts interpretability analysis of the model based on attention mechanisms and PET images. Specifically, the attention mechanism allows the model to focus more on receptive fields with classification characteristics, where the weights are positively correlated with the importance of the corresponding location information; that is, high-weight input units have a decisive effect on the output results. Based on this, internal model interpretability analysis is performed. Furthermore, considering that PET images are functional imaging, and abnormally bright areas are areas where malignant nodules are frequently found, the combination of attention mechanisms and PET images allows for interpretability analysis from both the model's internal structure and its functional imaging.

[0090] Based on the PET / CT image recognition method provided in the above embodiments, some embodiments of the present application further provide a PET / CT image recognition device, comprising: a controller configured to execute the PET / CT image recognition method provided in the above embodiments.

[0091] The controller can be a computer, a server, an industrial computer, a single-chip microcomputer, a programmable logic controller (PLC), a digital signal processor (DSP), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or other devices with storage and computing functions, and the embodiments of the present application do not limit this.

[0092] The present application provides a PET / CT image recognition method and device, the image recognition method comprising: target substance extraction through PET / CT original images to obtain a CT data set and a PET data set; inputting the CT data set and the PET data set into a dual-view deep learning classifier for feature extraction and feature classification to obtain CT feature classification and PET feature classification; the dual-view deep learning classifier uses a MVPD-Net model; based on a voting-based recognition mechanism, the CT feature classification and the PET feature classification are fused to obtain a final recognition result. Figure 3 The MVPD-Net model proposed in the present application is based on multi-view deep learning, which can more reliably recognize PET / CT images and has better interpretability. The mean of the K-means segmentation threshold and the OTSU segmentation threshold is taken as the target substance segmentation threshold for target substance extraction, which solves the problem of high sensitivity and low specificity of the classification model due to the structural and density similarity of the target tissue and the surrounding tissue and the heterogeneity of the target tissue itself.

[0093] The similar parts between the embodiments provided in the present application can be referred to each other, and the specific embodiments provided above are only several examples under the general concept of the present application, and do not limit the protection scope of the present application. For those skilled in the art, any other embodiments extended according to the present application scheme without creative labor are within the protection scope of the present application.

Claims

1. A PET / CT image recognition method, characterized by, The method comprises the following steps: obtaining a denoised CT image; resampling the denoised CT image to obtain a resampled CT image; performing threshold segmentation on the resampled CT image based on a K-means clustering algorithm to obtain a K-means segmentation threshold; performing threshold segmentation on the resampled CT image based on an OTSU algorithm to obtain an OTSU segmentation threshold; taking the average of the K-means segmentation threshold and the OTSU segmentation threshold as a target parenchymal segmentation threshold; performing morphological transformation on the CT original image using the target parenchymal segmentation threshold to obtain a CT target parenchymal mask Mask; performing target parenchymal extraction on the CT original image using the CT target parenchymal mask Mask to obtain a CT data set; scaling the size of the CT target parenchymal mask Mask according to the ratio of the CT resolution to the PET resolution to obtain a PET target parenchymal mask Mask; performing target parenchymal extraction on the PET original image using the PET target parenchymal mask Mask to obtain a PET data set; inputting the CT data set and the PET data set into a dual-view deep learning classifier for feature extraction and feature classification to obtain CT feature classification and PET feature classification; the dual-view deep learning classifier uses a MVPD-Net model, and the MVPD-Net model comprises: a feature extraction subnetwork comprising a CT feature extractor and a PET feature extractor for extracting features from the CT data set and the PET data set respectively; the CT feature extractor and the PET feature extractor are both constructed by combining 3D convolution layers and 3D pooling layers; the 3D convolution layer of the CT feature extractor comprises a convolution kernel of 7x7x7 and a number of convolution kernels of 64; the 3D pooling layer of the CT feature extractor is 1x3x3; the CT feature extractor extracts high-dimensional features through the stacking of 5 feature extraction blocks; the 3D convolution layer of the PET feature extractor comprises a convolution kernel of 3x3x4 and a number of convolution kernels of 64; the 3D pooling layer of the PET feature extractor is 1x3x3; the PET feature extractor extracts high-dimensional features through the stacking of 4 feature extraction blocks; a feature classification subnetwork for classifying CT features and PET features obtained by the feature extractor; the feature classification subnetwork comprises: feature vector flattening; feature classification through 2 fully connected layers; and classification results based on a Softmax activation function; a voting-based recognition mechanism for fusing the CT feature classification and the PET feature classification to obtain a final recognition result.

2. The PET / CT image recognition method of claim 1, wherein, The step of obtaining a denoised CT image comprises: performing data cleaning on the CT original image to obtain a cleaned CT image; the data cleaning is to filter and remove data with missing information, unclear results, and expired data; performing data denoising on the cleaned CT image to obtain a denoised CT image; the data denoising is to denoise the cleaned CT image based on a median filtering method with a filter window of 3x3.

3. The PET / CT image recognition method of claim 1, wherein, The step of performing resampling processing on the denoised CT image to obtain a resampled CT image comprises: Converting the pixel value of the denoised CT image into a CT value by using a conversion formula to obtain a resampled CT image; The conversion formula is HU=pixel_val×rescale_slope+rescale_intercept Wherein, pixel_val is the pixel value of the denoised CT image, rescale_slope and rescale_intercept are the slope and intercept.

4. The PET / CT image recognition method of claim 1, wherein, The step of performing threshold segmentation on the resampled CT image based on the K-means clustering algorithm to obtain a K-means segmentation threshold comprises: Randomly selecting K as 2 data as the initial clustering center for the resampled CT image; Calculate the Euclidean distance of the remaining data from the initial clustering center, classify according to the nearest principle, and generate clusters; Calculate the average value of each cluster center as a new clustering center; If the clustering center changes, or the number of iterations reaches a threshold, the clustering result is obtained, and the average value of the clustering result is taken as the K-means segmentation threshold.

5. The PET / CT image recognition method of claim 1, wherein, The step of performing threshold segmentation on the resampled CT image based on the OTSU algorithm to obtain an OTSU segmentation threshold comprises: Calculate the global expectation of the gray image for the resampled CT image; Iterate through the gray value to divide the gray image into foreground and background; Calculate the expectation and probability of the foreground and background; When the inter-class variance of the threshold is maximum, the threshold is obtained as the OTSU segmentation threshold.

6. The PET / CT image recognition method of claim 1, wherein, The step of performing morphological transformation on the CT original image by using the target parenchymal segmentation threshold to obtain a CT target parenchymal mask Mask comprises: Performing binaryzation processing on the CT original image by using the target parenchymal segmentation threshold to obtain a first image; Performing erosion and expansion operations on the first image with a kernel of 3×3 and 8×8 to obtain a second image; Performing a closing operation on the second image with a kernel of 25×25 to obtain a CT target parenchymal mask Mask.

7. The PET / CT image recognition method of claim 1, wherein, The voting recognition mechanism comprises: y v = MAX(y (v) ) where y (v) represent the recognition results of different view channels, y v represent the results after voting of different view channels.

8. A PET / CT image recognition apparatus, characterized by, Comprise: The controller is used to execute the PET / CT image recognition method of any one of claims 1-7.

Images