Tumor detection method and system based on image recognition

Abstract

The invention provides a tumor detection method based on image recognition. The method comprises the following steps: acquiring a human body medical image; identifying a gray level abnormal region in the CT image through a preset first identification model, and performing image identification on the gray level abnormal region to obtain a first feature; performing feature extraction on the MR image according to a preset second recognition model to obtain a second feature; and respectively giving different weights to the first feature and the second feature through a linear regression model, carrying out feature fusion on the first feature and the second feature according to the different weights based on a random forest model to obtain a comprehensive feature, and identifying a tumor according to the comprehensive feature. According to the method, multiple features of different medical images are combined for tumor recognition, and the feature extraction precision is improved through the network model.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a tumor detection method and system based on image recognition. Background Art

[0002] With the development of artificial intelligence, the application of image recognition models is becoming more and more widespread. It can be used not only in life, but also in medical technology. For example, it can recognize medical images of patients and assist doctors in diagnosing tumors.

[0003] However, the current recognition of medical images only recognizes a single feature of the area to be recognized in the medical image, ignoring other relevant information of the image and the subtle features of the local image, resulting in insufficient recognition accuracy and precision. Therefore, it is very necessary to design a tumor detection method and system based on image recognition. Summary of the invention

[0004] The purpose of the present invention is to provide a tumor detection method and system based on image recognition, so as to identify tumors according to the combination of multiple features of different medical images and improve the accuracy of feature extraction according to a network model.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A tumor detection method based on image recognition comprises the following steps:

[0007] Acquire human medical images; human medical images include CT images and MR images;

[0008] Identify the grayscale abnormal area in the CT image through a preset first recognition model, and perform image recognition on the grayscale abnormal area to obtain a first feature; the grayscale abnormal area includes: a black area, a white area, and a gray area;

[0009] The feature extraction of the MR image is performed according to the preset second recognition model to obtain the second feature; the second feature includes: morphological feature, edge feature, tissue feature and microvascular feature; the feature extraction of the MR image is performed according to the preset second recognition model to obtain the second feature, specifically including:

[0010] The MR images of various sequences are registered by image registration technology to obtain a registered image;

[0011] The morphological features are obtained by registering the signal differences between pixels in the image;

[0012] Extract edge features in the registration image using an edge detection algorithm built into the second recognition model;

[0013] The tissue features in the registered image are identified by a semantic segmentation network model built into the second recognition model;

[0014] The vascular characteristics are calculated by registering the vascular density, vascular morphology, vascular perfusion parameters and vascular texture in the image; vascular morphology includes: length, diameter, number of branches and curvature of the vessel; vascular perfusion parameters include: perfusion peak and enhancement rate; vascular texture includes: uniformity, symmetry and complexity;

[0015] Different weights are assigned to the first feature and the second feature respectively through the linear regression model, and based on the random forest model, the first feature and the second feature are fused according to different weights to obtain a comprehensive feature, and the tumor is identified based on the comprehensive feature.

[0016] Optionally, identifying the grayscale abnormal area in the CT image by using a preset first recognition model, and performing image recognition on the grayscale abnormal area to obtain the first feature specifically includes:

[0017] Perform data enhancement operations on the CT images to obtain a CT data set; data enhancement includes: image rotation, scaling, and flipping;

[0018] Grayscale difference recognition and image aggregation are performed on each image in the CT data set to obtain a first recognition result set;

[0019] Training the convolutional neural network model according to the first recognition result set and the first loss function to obtain a first recognition model;

[0020] The black area, white area and gray area in the CT image are divided into regions according to the adaptive threshold segmentation algorithm built into the first recognition model, the size of the divided area is identified according to the color, and the area size is used as the first feature.

[0021] Optionally, training a convolutional neural network model according to the first recognition result set and the first loss function to obtain a first recognition model specifically includes:

[0022] Perform a convolution operation on the first recognition result set through a 3*3 convolution kernel to obtain a feature map;

[0023] Perform average pooling operation on the feature map to obtain a reduced-dimensional feature map;

[0024] Calculate the predicted value of the reduced-dimensional feature map according to the built-in softmax function of the convolutional neural network model;

[0025] The loss value of the predicted value is calculated according to the cross entropy loss, and the convolutional neural network model is trained according to the loss value to obtain the first recognition model.

[0026] Optionally, multiple sequences of MR images are registered by image registration technology to obtain registered images, specifically including:

[0027] Extract the feature points of each image in multiple sequences of MR images respectively;

[0028] Pair the feature points by similarity measurement to obtain matching feature point pairs;

[0029] Any one of the MR images is used as the reference image, and the remaining images are used as the images to be registered, and the spatial coordinate transformation parameters of the images to be registered are obtained according to the feature point pairs;

[0030] The images to be registered are transformed and stitched according to their respective spatial coordinate transformation parameters to obtain the registered images.

[0031] Optionally, the morphological features are obtained by registering the signal differences between pixels in the image, specifically including:

[0032] Calculate the signal difference between each pixel in the registered image;

[0033] When the signal difference exceeds the preset signal difference threshold, the two pixel points generating the signal difference are marked as morphological boundary points;

[0034] The morphological boundary points are summarized to obtain the morphological boundary, and the area surrounded by the morphological boundary is used as the morphological feature.

[0035] Optionally, edge features in the registration image are extracted by using an edge detection algorithm built into the second recognition model, specifically including:

[0036] Perform image segmentation on the boundaries of the morphological features to obtain a segmented image; the size of the segmented image is 5*5;

[0037] Calculate the average signal intensity of all pixels in the segmented image;

[0038] The edge feature is obtained based on the comparison result between the average value and the preset boundary threshold.

[0039] Optionally, the edge feature is obtained according to the comparison result between the average value and a preset boundary threshold, specifically including:

[0040] If the average value exceeds the preset boundary threshold, the corresponding segmented image is marked as clear;

[0041] If the average value does not exceed the preset boundary threshold, the corresponding segmented image is marked as blurred;

[0042] When the segmented images marked as blurred exceed 30% of the total segmented images, the edge feature is defined as abnormal; otherwise, the edge feature is defined as normal.

[0043] Optionally, identifying tissue features in the registered image through a semantic segmentation network model built into the second recognition model specifically includes:

[0044] The signal intensity of each pixel in the registration image is formed into a first matrix;

[0045] Performing an encoding operation on the first matrix through an encoding layer of a semantic segmentation network model to obtain an encoding matrix;

[0046] The encoding matrix is ​​decoded through the decoding layer of the semantic segmentation network model to obtain a decoding matrix, and the decoding matrix is ​​used as the organizational feature.

[0047] Optionally, the calculation formula of the blood vessel characteristics is: Among them, F is the vascular characteristic value, D is the vascular density, L is the vascular length, and D V is the vessel diameter, B is the number of vascular branches, C is the vascular curvature, P v is the perfusion peak, R is the enhancement rate, U is the uniformity of vascular texture, S is the symmetry of vascular texture, C t is the complexity of blood vessel texture, w 1 、w 2 、w 3 and w 4 are all weight ratios, and α, β, γ, δ, ∈ and ζ are all weighting parameters.

[0048] A tumor detection system based on image recognition, comprising:

[0049] A data acquisition module is used to acquire human medical images; human medical images include CT images and MR images;

[0050] A first recognition module, used to recognize grayscale abnormal regions in the CT image by using a preset first recognition model, and perform image recognition on the grayscale abnormal regions to obtain a first feature;

[0051] A second recognition module is used to extract features from the MR image according to a preset second recognition model to obtain a second feature;

[0052] The comprehensive judgment module is used to assign different weights to the first feature and the second feature respectively through a linear regression model, and based on a random forest model, the first feature and the second feature are fused according to different weights to obtain a comprehensive feature, and the tumor is identified based on the comprehensive feature.

[0053] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: a tumor detection method based on image recognition, comprising: acquiring a human medical image; identifying the grayscale abnormal area in the CT image through a preset first recognition model, and performing image recognition on the grayscale abnormal area to obtain a first feature; performing feature extraction on the MR image according to a preset second recognition model to obtain a second feature; assigning different weights to the first feature and the second feature respectively through a linear regression model, and based on a random forest model, performing feature fusion on the first feature and the second feature according to different weights to obtain a comprehensive feature, and identifying the tumor based on the comprehensive feature. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0055] Figure 1 is a flow chart of a tumor detection method based on image recognition according to an embodiment of the present invention;

[0056] Figure 2 A flowchart of obtaining a first feature according to an embodiment of the present invention;

[0057] Figure 3 A flowchart of obtaining a second feature according to an embodiment of the present invention;

[0058] Figure 4 This is a flow chart of image registration according to an embodiment of the present invention;

[0059] Figure 5 A flow chart of obtaining morphological features according to an embodiment of the present invention;

[0060] Figure 6 This is a flow chart of obtaining edge features according to an embodiment of the present invention;

[0061] Figure 7 This is a flow chart of obtaining tissue characteristics according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] like Figure 1 As shown, the present invention provides a tumor detection method based on image recognition, comprising the following steps:

[0065] Step 100: Acquire a human medical image; the human medical image includes a CT image and an MR image.

[0066] Step 200: Identify the grayscale abnormal area in the CT image through a preset first recognition model, and perform image recognition on the grayscale abnormal area to obtain a first feature.

[0067] Specifically, an X-ray beam is used to scan the human body layer by layer. The detector receives the X-rays that pass through the human tissue, generates a digital matrix after computer processing, and then generates pixels with different grayscales through digital-to-analog conversion to form a CT image. Among them, the grayscale abnormal area includes: black area, white area and gray area. The black area indicates that the area is a low-density area, which may be a cystic tumor; the white area indicates that the area is a high-density area, which may be a hard tumor; the gray area indicates that the area is a mixed density.

[0068] Furthermore, the grayscale abnormal area in the CT image is identified by a preset first recognition model, and image recognition is performed on the grayscale abnormal area to obtain a first feature, as shown in FIG. Figure 2 As shown, including:

[0069] Step 201: performing data enhancement operation on the CT image to obtain a CT data set; data enhancement includes: image rotation, scaling and flipping;

[0070] Specifically, the CT images are first rotated by 90, 180, and 270 degrees, and cubic interpolation is used to reduce the aliasing effect during rotation. Secondly, the CT images are scaled within a scaling range of 0.7 to 1.3, and linear interpolation is used during the scaling process to prevent the image from blurring. The CT images are then flipped horizontally and vertically. Finally, the rotated, scaled, and flipped images are spatially aligned with the original CT images to obtain the CT dataset.

[0071] It should be noted that data augmentation of images increases the diversity of the data set, allowing subsequent models to better adapt to images of different perspectives and shapes, reducing the risk of overfitting and improving the generalization ability of the model. Using cubic interpolation and linear interpolation methods in the data augmentation process can improve the visual quality of the image, thereby maintaining the edges and details of the image and reducing blur.

[0072] Step 202: performing grayscale difference recognition and image aggregation on each image in the CT data set to obtain a first recognition result set;

[0073] Specifically, the grayscale values ​​of all pixels in the image are obtained, the difference between the grayscale value of any target pixel and the grayscale value of an adjacent pixel is calculated, and all the difference results calculated in all images are integrated to obtain a first recognition result set.

[0074] Step 203: training the convolutional neural network model according to the first recognition result set and the first loss function to obtain a first recognition model;

[0075] Specifically, the convolution operation is performed on the first recognition result set through a 3*3 convolution kernel in the convolution layer of the convolution neural network model to obtain a feature map. The feature map is averaged pooled through the pooling layer of the convolution neural network model to obtain a reduced-dimensional feature map. In the fully connected layer of the convolution neural network model, the predicted value of the reduced-dimensional feature map is calculated through the built-in softmax function, and then the loss value of the predicted value is calculated based on the cross-entropy loss. Based on the stochastic gradient descent algorithm, the convolution neural network model is iteratively trained according to the loss value, and the training stops when the number of iterations reaches the preset number to obtain the first recognition model.

[0076] Step 204: Divide the black area, white area and gray area in the CT image according to the adaptive threshold segmentation algorithm built into the first recognition model, identify the size of the divided area according to the color, and use the area size as the first feature.

[0077] Specifically, the CT image is input into the first recognition model, the CT image is smoothed by Gaussian smoothing operation, and the adaptive threshold is calculated by the smoothed image and the preset ratio. In this embodiment, the preset ratio is 0.15. The adaptive threshold is compared with the original image, and local threshold segmentation is performed to obtain the segmented black area, white area and gray area, and the regional contour and contour size are obtained according to the regional growing method.

[0078] It should be noted that Gaussian smoothing can effectively remove noise from images and enhance signal quality. The adaptive threshold segmentation method can dynamically adjust the threshold according to the local features of the image, making the segmentation result more accurate. The region growing method can gradually expand the region according to the grayscale value, effectively extract the contour of irregular shapes, and improve the accuracy of contour acquisition. The automated image processing process significantly reduces human intervention and manual operation, greatly improves work efficiency, and reduces the error rate.

[0079] Step 300: Extract features from the MR image according to the preset second recognition model to obtain second features; the second features include: morphological features, edge features, tissue features and microvascular features; the specific steps are as follows Figure 3 As shown, including:

[0080] Step 310: align multiple sequences of MR images using image registration technology to obtain a registered image; the specific steps are as follows: Figure 4 As shown below:

[0081] Step 311: extracting feature points of each image in multiple sequences of MR images respectively;

[0082] Specifically, convolution is performed on adjacent images through different Gaussian kernel functions to obtain a Gaussian difference pyramid, and then the difference between adjacent images in the Gaussian difference pyramid is calculated to obtain the key value points. Taylor expansion is performed on the pixel values ​​of the image in the Gaussian difference pyramid to obtain the sub-pixel position of the key point. The key points are unified in direction according to the gradient of the area around the key points. Based on the SIFT algorithm, the descriptor of the key point is calculated according to the position and direction of the key point, and the descriptor is normalized to obtain the feature point.

[0083] Step 312: Pairing the feature points by similarity measurement to obtain matching feature point pairs;

[0084] Specifically, the feature points are paired according to the Hamming distance, and the matching feature points in different images are found by the Fast Approximate Nearest Neighbor (FLANN) search algorithm. The RANSAC algorithm is used to eliminate mismatches.

[0085] Step 313: taking any one of the MR images as a reference image and the remaining images as images to be registered, and obtaining the spatial coordinate transformation parameters of the images to be registered according to the feature point pairs;

[0086] Specifically, this embodiment uses the least squares fitting technology to obtain the transformation matrix of the correspondence relationship between the feature points, and uses the transformation matrix as the space coordinate transformation parameter.

[0087] Step 314: performing image transformation and image stitching on the images to be registered according to their respective space coordinate transformation parameters to obtain a registered image.

[0088] Specifically, this embodiment performs a coordinate transformation operation on the image to be registered according to the transformation matrix, and uses bilinear interpolation to resample the pixels of the transformed image, and finally splices the resampled images to generate a unified registered image.

[0089] It should be noted that by using image registration techniques such as feature point extraction, pairing, and spatial coordinate transformation, multiple sequences of MR images can be efficiently integrated into a consistent view, enhancing the spatial consistency between images of different sequences and greatly improving the accuracy and comprehensiveness of image recognition.

[0090] Step 320: Obtain morphological features by registering the signal differences between pixels in the image; the specific steps are as follows: Figure 5 As shown, including:

[0091] Step 321: Calculate the signal difference between each pixel in the registration image;

[0092] Step 322: When the signal difference exceeds a preset signal difference threshold, both two pixel points generating the signal difference are marked as morphological boundary points.

[0093] Specifically, the preset signal difference threshold in this embodiment is 130 gray levels.

[0094] Step 323: Summarize the morphological boundary points to obtain the morphological boundary, and use the area surrounded by the morphological boundary as the morphological feature.

[0095] Specifically, adjacent morphological boundary points are connected to obtain a closed boundary. The regular shape of the enclosed area can be obtained through the closed boundary. Different regular shapes can be used as one of the judgment conditions for whether the area is a tumor area.

[0096] Step 330: extract edge features in the registration image using the edge detection algorithm built into the second recognition model; the specific steps are as follows: Figure 6 As shown, including:

[0097] Step 331: Segment the boundary of the morphological feature to obtain a segmented image; the size of the segmented image is 5*5;

[0098] Specifically, this embodiment detects the boundary of the morphological feature by using the Sobel operator, and then uses the watershed algorithm to implement image segmentation.

[0099] Step 332: Calculate the average signal intensity of all pixels in the segmented image;

[0100] Specifically, the segmented 5x5 image is traversed to extract the signal strength values ​​of all pixels, and the average value is calculated using the average formula. The average formula is: Where A is the average value, N is the number of pixels, and I(i) is the signal intensity of the i-th pixel.

[0101] Step 333: Obtain edge features based on the comparison result between the average value and the preset boundary threshold.

[0102] Specifically, the preset boundary threshold in this embodiment is 85. If the average value exceeds the preset boundary threshold, the corresponding segmented image is marked as clear; if the average value does not exceed the preset boundary threshold, the corresponding segmented image is marked as blurred. When the segmented images marked as blurred exceed 30% of the total segmented images, the edge feature is defined as abnormal; otherwise, the edge feature is defined as normal.

[0103] It should be noted that effectively segmenting the morphological features and calculating the average value of the signal intensity can accurately extract and analyze the edge features in the image, simplify the image feature analysis process, improve the automation and accuracy of segmentation and feature extraction, and provide effective support for subsequent image recognition, lesion detection and medical diagnosis. At the same time, the set threshold comparison mechanism can make the determination of edge features more objective and consistent.

[0104] Step 340: Identify tissue features in the registration image using the semantic segmentation network model built into the second recognition model; the specific steps are as follows: Figure 7 As shown, including:

[0105] Step 341: compose a first matrix with the signal strength of each pixel in the registration image;

[0106] Specifically, the signal strength of the pixels is arranged into a matrix according to the spatial position. For example, if the image size is H*W (height × width), the size of the matrix is ​​also (H×W), and the matrix is: Among them, a HW Represents the signal strength of the pixel at the Hth row and the Wth column in the image.

[0107] Step 342: performing an encoding operation on the first matrix through the encoding layer of the semantic segmentation network model to obtain an encoding matrix;

[0108] Specifically, the encoding layer includes a convolution layer and a pooling layer, and convolution and average pooling operations are performed on the convolution layer and the pooling layer to reduce the spatial dimension of the first matrix and generate an encoding matrix. The encoding process is optimized by adjusting the hyperparameters of the convolution layer (such as the convolution kernel size, stride, etc.).

[0109] Step 343: The encoding matrix is ​​decoded by the decoding layer of the semantic segmentation network model to obtain a decoding matrix, and the decoding matrix is ​​used as an organizational feature.

[0110] Specifically, the encoding matrix is ​​restored to the size of the original image through deconvolution and upsampling operations in the decoding layer, and the softmax activation function maps the signal intensity of each pixel to obtain the decoding matrix.

[0111] It should be noted that the use of semantic segmentation networks for encoding and decoding operations can enhance the recognition ability of complex structural features and improve the generalization performance of the model by compressing and restoring high-dimensional data.

[0112] Step 350: Calculate the vascular features by registering the vascular density, vascular morphology, vascular perfusion parameters and vascular texture in the image.

[0113] Specifically, vascular morphology includes: length, diameter, number of branches and curvature of the blood vessel; vascular perfusion parameters include: perfusion peak and enhancement rate; vascular texture includes: uniformity, symmetry and complexity. The calculation formula of vascular characteristics is:

[0114]

[0115] Among them, F is the vascular characteristic value, D is the vascular density, L is the vascular length, and D V is the vessel diameter, B is the number of vascular branches, C is the vascular curvature, P v is the perfusion peak, R is the enhancement rate, U is the uniformity of vascular texture, S is the symmetry of vascular texture, C t is the complexity of blood vessel texture, w 1 、w 2 、w 3 and w 4 All are weight ratios, and α, β, γ, δ, ∈ and ζ are all weight parameters. The weight ratios and weight parameters in this embodiment are set differently according to different human body detection parts, and this embodiment does not limit this.

[0116] Step 400: Different weights are assigned to the first feature and the second feature respectively through a linear regression model, and based on a random forest model, the first feature and the second feature are fused according to the different weights to obtain a comprehensive feature, and the tumor is identified based on the comprehensive feature.

[0117] It should be noted that the linear regression model assigns different weights to different features, so that the model can better understand the relative importance of each feature in tumor identification and ensure the rationality of feature fusion. Random forest effectively improves the stability and anti-interference ability of the output results by integrating multiple decision trees.

[0118] The present invention also provides a tumor detection system based on image recognition, comprising:

[0119] A data acquisition module is used to acquire human medical images; human medical images include CT images and MR images;

[0120] A first recognition module, used to recognize grayscale abnormal regions in the CT image by using a preset first recognition model, and perform image recognition on the grayscale abnormal regions to obtain a first feature;

[0121] A second recognition module is used to extract features from the MR image according to a preset second recognition model to obtain a second feature;

[0122] The comprehensive judgment module is used to assign different weights to the first feature and the second feature respectively through a linear regression model, and based on a random forest model, the first feature and the second feature are fused according to different weights to obtain a comprehensive feature, and the tumor is identified based on the comprehensive feature.

[0123] The beneficial effects of the present invention are as follows:

[0124] 1) The feature extraction of CT images is performed through convolutional neural networks, which improves the accuracy of the extracted features;

[0125] 2) The encoding and decoding mechanism of the semantic segmentation network model improves the accuracy and efficiency of extracting tissue features;

[0126] 3) The present invention obtains the features of various MR images through various methods, thereby improving the diversity of feature extraction;

[0127] 4) The multiple features of the two images are integrated through linear regression model and random forest model to identify tumor images, which enhances the accuracy and precision of recognition, reduces manual labor consumption, and greatly improves work efficiency.

[0128] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0129] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A tumor detection method based on image recognition, characterized in that: The steps include: Acquire a human medical image; the human medical image includes a CT image and an MR image; Identify the grayscale abnormal area in the CT image by using a preset first recognition model, and perform image recognition on the grayscale abnormal area to obtain a first feature; The grayscale abnormal area includes: a black area, a white area and a gray area; Extracting features from the MR image according to a preset second recognition model to obtain a second feature; the second feature includes: morphological features, edge features, tissue features, and microvascular features; extracting features from the MR image according to the preset second recognition model to obtain the second feature specifically includes: Registering the MR images of various sequences by image registration technology to obtain a registered image; Obtaining the morphological feature through the signal difference between the pixels in the registered image; Extracting the edge features in the registered image by using an edge detection algorithm built into the second recognition model; Identify the tissue features in the registered image using a semantic segmentation network model built into the second recognition model; The vascular features are obtained by calculating the vascular density, vascular morphology, vascular perfusion parameters and vascular texture in the registered image; the vascular morphology includes: the length, diameter, number of branches and curvature of the blood vessel; the vascular perfusion parameters include: perfusion peak and enhancement rate; the vascular texture includes: uniformity, symmetry and complexity; Different weights are respectively assigned to the first feature and the second feature through a linear regression model, and based on a random forest model, the first feature and the second feature are fused according to the different weights to obtain a comprehensive feature, and the tumor is identified based on the comprehensive feature.

2. The tumor detection method based on image recognition according to claim 1, characterized in that: Identifying the grayscale abnormal region in the CT image by using a preset first recognition model, and performing image recognition on the grayscale abnormal region to obtain a first feature, specifically including: Performing data enhancement operation on the CT image to obtain a CT data set; the data enhancement includes: rotating, scaling and flipping the image; Performing grayscale difference recognition and image aggregation on each image in the CT data set to obtain a first recognition result set; Training a convolutional neural network model according to the first recognition result set and the first loss function to obtain the first recognition model; The black area, the white area and the gray area in the CT image are divided into regions according to the adaptive threshold segmentation algorithm built into the first recognition model, the size of the divided regions is identified according to the colors, and the region size is used as the first feature.

3. The tumor detection method based on image recognition according to claim 2, characterized in that: Training a convolutional neural network model according to the first recognition result set and the first loss function to obtain the first recognition model specifically includes: Performing a convolution operation on the first recognition result set through a 3*3 convolution kernel to obtain a feature map; Performing an average pooling operation on the feature map to obtain a dimension-reduced feature map; Calculating the predicted value of the dimensionality reduction feature map according to the softmax function built into the convolutional neural network model; The loss value of the predicted value is calculated according to the cross entropy loss, and the convolutional neural network model is trained according to the loss value to obtain the first recognition model.

4. The tumor detection method based on image recognition according to claim 1, characterized in that: The MR images of various sequences are registered by image registration technology to obtain registered images, which specifically includes: Extracting feature points of each image in the MR images of multiple sequences respectively; Pairing the feature points by similarity measurement to obtain matching feature point pairs; Taking any one of the MR images as a reference image and the remaining images as images to be registered, and obtaining spatial coordinate transformation parameters of the images to be registered according to the feature point pairs; The images to be registered are transformed and stitched according to their respective spatial coordinate transformation parameters to obtain the registered images.

5. The tumor detection method based on image recognition according to claim 4, characterized in that: The morphological feature is obtained by using the signal difference between the pixels in the registered image, specifically including: Calculating the signal difference between each pixel in the registered image; When the signal difference exceeds a preset signal difference threshold, both the two pixel points generating the signal difference are marked as morphological boundary points; The morphological boundary points are summarized to obtain a morphological boundary, and the area surrounded by the morphological boundary is used as the morphological feature.

6. The tumor detection method based on image recognition according to claim 1, characterized in that: Extracting the edge features in the registered image by using the edge detection algorithm built into the second recognition model specifically includes: Perform image segmentation on the boundary of the morphological feature to obtain a segmented image; the size of the segmented image is 5*5; Calculating the average value of the signal intensities of all pixels in the segmented image; The edge feature is obtained according to a comparison result between the average value and a preset boundary threshold.

7. The tumor detection method based on image recognition according to claim 1, characterized in that: The edge feature is obtained according to the comparison result between the average value and the preset boundary threshold, specifically including: If the average value exceeds the preset boundary threshold, the corresponding segmented image is marked as clear; If the average value does not exceed the preset boundary threshold, the corresponding segmented image is marked as blurred; When the segmented images marked as blurred exceed 30% of the entire segmented images, the edge feature is defined as abnormal; otherwise, the edge feature is defined as normal.

8. The tumor detection method based on image recognition according to claim 1, characterized in that: Identifying the tissue features in the registered image by using a semantic segmentation network model built into the second recognition model specifically includes: The signal intensity of each pixel in the registered image is used to form a first matrix; Performing an encoding operation on the first matrix through the encoding layer of the semantic segmentation network model to obtain an encoding matrix; The encoding matrix is ​​decoded by the decoding layer of the semantic segmentation network model to obtain a decoding matrix, and the decoding matrix is ​​used as the organizational feature.

9. The tumor detection method based on image recognition according to claim 1, characterized in that: The calculation formula of the vascular characteristics is: Among them, F is the vascular characteristic value, D is the vascular density, L is the vascular length, and D V is the vessel diameter, B is the number of vascular branches, C is the vascular curvature, P v is the perfusion peak, R is the enhancement rate, U is the uniformity of vascular texture, S is the symmetry of vascular texture, C t is the complexity of vascular texture, w1, w2, w3 and w4 are weight ratios, and α, β, γ, δ, ∈ and ζ are weighting parameters.

10. A tumor detection system based on image recognition, characterized in that: include: A data acquisition module, used for acquiring human medical images; The human medical images include CT images and MR images; A first recognition module, used for recognizing a grayscale abnormal region in the CT image by using a preset first recognition model, and performing image recognition on the grayscale abnormal region to obtain a first feature; A second recognition module, used for extracting features from the MR image according to a preset second recognition model to obtain a second feature; A comprehensive judgment module is used to assign different weights to the first feature and the second feature respectively through a linear regression model, and based on a random forest model, perform feature fusion on the first feature and the second feature according to the different weights to obtain a comprehensive feature, and identify the tumor based on the comprehensive feature.

Images