Prostate cancer lesion detection model construction method based on extraction of contrast image features
By using a prostate cancer lesion detection model based on imaging feature extraction, and employing time-intensity curve correction algorithm and three-dimensional deep learning, the problem of poor robustness of traditional detection methods is solved, and high-accuracy detection and precise localization of prostate cancer lesions are achieved.
Patent Information
- Application Number
- CN202310142454.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-21
- Publication Date
- 2026-02-17
- Estimated Expiration
- 2043-02-21
AI Technical Summary
Traditional methods for detecting prostate cancer lesions are difficult to extract effective features, have poor robustness and low stability, and cannot meet the needs of imaging in the accurate identification and precise puncture-targeted localization of prostate cancer.
A prostate cancer lesion detection model based on contrast imaging feature extraction automatically extracts the time segment from the appearance to the peak of arterial microbubbles in ultrasound contrast imaging videos through a time-intensity curve correction algorithm. It combines a spatiotemporal feature extraction network and a target detection network, and uses three-dimensional deep learning and YOLOv7 network to improve detection accuracy.
It improves the applicability and predictive accuracy of prostate cancer lesion detection, can adapt to lesion areas of different sizes, and utilizes the dynamic characteristics of ultrasound contrast imaging to achieve sensitive detection and precise localization of lesions.
Smart Images
Figure CN116109603B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image processing, in particular to a method for establishing a prostate cancer lesion intelligent detection model. BACKGROUND
[0002] Prostate cancer is the second most common and fifth most lethal male cancer worldwide. In current clinical applications, its diagnosis mainly relies on biopsy methods, and the pursuit of puncturing to the highest Gleason score, the puncture result as close as possible to the surgical pathology, while minimizing the number of puncture needles, has always been the goal of its research. Unlike thyroid and breast, prostate can directly use contrast ultrasound video for classification to determine whether there is a lesion area. Prostate not only needs to determine whether there is a lesion, but also needs to divide the lesion area. The traditional prostate cancer lesion area detection method is difficult to extract effective features, and its robustness is poor and its stability is low, which puts higher requirements on the precise identification of prostate cancer and accurate puncture targeting positioning by imaging. At the same time, with the development of imaging, it is also hoped that new technologies can be used to achieve this demand. SUMMARY
[0003] Therefore, the purpose of the present application is to provide a prostate cancer lesion detection model construction method based on contrast image feature extraction, so as to solve the technical problem of providing a more suitable and better prediction accuracy intelligent detection model for prostate cancer lesion detection.
[0004] The prostate cancer lesion detection model construction method based on contrast image feature extraction of the present application comprises the following steps:
[0005] S1: Establish a training sample set:
[0006] S11: Perform time selection based on a time-intensity curve correction algorithm on the prostate ultrasound contrast video to obtain a video segment of the ultrasound contrast microbubbles in the prostate arterial blood vessels from emergence to peak value;
[0007] S12: Group division is performed on the video frames contained in the selected video segment: group determination is performed according to the microbubble density and size change in the same target area, and the video frames with small changes in microbubble density and size before and after are included in the same group;
[0008] S13: Select one ultrasound contrast image from each video frame group for manual annotation: find the prostate cancer lesion target area in the selected image, find the heterogeneous microbubble vascular architecture area in the target area, draw the region of interest in the heterogeneous microbubble vascular architecture area, generate a detection box, including the coordinates and offset of the detection box, and perform significant and non-significant binary classification on the region of interest according to the pathological result;
[0009] S14: sharpening and pixel value normalization processing are performed on the images in all video frame groups;
[0010] S2: a prostate cancer lesion detection model is established, and the prostate cancer lesion detection model comprises a space-time feature extraction network and a target detection network;
[0011] The space-time feature extraction network structure comprises an encoding part, a direct connection part and a decoding part;
[0012] The encoding part takes a video frame group as input, extracts a feature map from the input video frame group through a first three-dimensional convolution module, and takes the extracted feature map as a first encoding block; then the encoding part further performs down-sampling on the first encoding block through a second three-dimensional convolution module, extracts a feature map from the output of the second three-dimensional convolution module through a three-dimensional multi-branch stacking module, and takes the extracted feature map as a second encoding block; then the encoding part further performs down-sampling on the second encoding block through a third three-dimensional convolution module, extracts a feature map from the output of the third three-dimensional convolution module through the three-dimensional multi-branch stacking module, and takes the extracted feature map as a third encoding block; then the encoding part further performs down-sampling on the third encoding block through a fourth three-dimensional convolution module, extracts a feature map from the output of the fourth three-dimensional convolution module through the three-dimensional multi-branch stacking module, and takes the extracted feature map as a fourth encoding block; finally, the encoding part further performs down-sampling on the fourth encoding block through a dimension reduction module to output a three-dimensional feature map;
[0013] The direct connection part connects the first encoding block, the second encoding block, the third encoding block and the fourth encoding block with the decoding part;
[0014] The decoding part up-samples the three-dimensional feature map output by the dimension reduction module through a first two-dimensional convolution module, stacks the up-sampled feature map with the fourth encoding block, extracts a feature map from the stacked feature map through a two-dimensional multi-branch stacking module, and takes the extracted feature map as a first decoding block; then the decoding part up-samples the first decoding block through a second two-dimensional convolution module, stacks the up-sampled feature map with the third encoding block, extracts a feature map from the stacked feature map through the two-dimensional multi-branch stacking module, and takes the extracted feature map as a second decoding block; then the decoding part up-samples the second decoding block through a third two-dimensional convolution module, stacks the up-sampled feature map with the second encoding block, extracts a feature map from the stacked feature map through the two-dimensional multi-branch stacking module, and takes the extracted feature map as a third decoding block; then the decoding part up-samples the third decoding block through a fourth two-dimensional convolution module, stacks the up-sampled feature map with the first encoding block, extracts a feature map from the stacked feature map through the two-dimensional multi-branch stacking module, and takes the extracted feature map as a fourth decoding block;
[0015] The decoding part outputs a fourth decoding block to a two-dimensional convolution module, the output of the two-dimensional convolution module is input to a sigmoid function, and the output of the sigmoid function is input to a target detection network;
[0016] S3: training the prostate cancer lesion detection model: input the video frame group in the training sample set into the prostate cancer lesion detection model, and input the output image of the spatio-temporal feature extraction network into the target detection network; the target detection network extracts and learns the lesion features of the prostate cancer, and finally obtains a qualified prostate cancer lesion detection model.
[0017] Further, the time-intensity curve correction algorithm is defined as:
[0018]
[0019] In the formula, [] is an integral function, N Cor is a sequence value of the corrected peak frame number, N Max and T Max are a sequence value of the frame number corresponding to the maximum peak value and a time node value, respectively, j is the number of peak values before the maximum peak value, y j is the gray value of the last peak value in the peak values, and y i and T i are the gray value and the time node value corresponding to the corresponding peak value, respectively.
[0020] Further, the target detection network is a YOLOv7 network.
[0021] The beneficial effects of the present application are:
[0022] The prostate cancer lesion detection model construction method based on contrast image feature extraction has a self-adaptive function, can automatically extract the time segment from the appearance to the peak value of the microbubble of the arterial blood vessel in the ultrasound contrast video, and becomes more sensitive to the prostate tumor detection. The established prostate cancer lesion detection model has the advantages of three-dimensional deep learning, can richly extract the spatio-temporal features and prostate cancer lesions in the training sample, can adapt to different sizes of prostate cancer lesions, and utilizes the dynamic characteristics of ultrasound contrast, thereby improving the prediction accuracy of the lesions. BRIEF DESCRIPTION OF DRAWINGS
[0023] Figure 1 is a technical route schematic diagram of the prostate cancer lesion detection model construction method;
[0024] Figure 2 is a spatio-temporal feature extraction network structure schematic diagram;
[0025] Figure 3This is a schematic diagram of the YOLOv7 network structure;
[0026] Figure 4 This is a schematic diagram of the structure of a three-dimensional multi-branch stacked module;
[0027] Figure 5 This is a schematic diagram of a two-dimensional multi-branch stacked module. Detailed Implementation
[0028] The present invention will be further described below with reference to embodiments.
[0029] The method for constructing a prostate cancer lesion detection model based on contrast imaging feature extraction in this embodiment includes the following steps:
[0030] S1: Establish the training sample set:
[0031] S11: For videos of clinically preserved contrast-enhanced ultrasound (CEUS) pathological samples of the prostate, the entire video contains segments showing the emergence of arterial microbubbles to their maximum peak value, as well as segments showing the decline and re-peak value; the time segments have not been processed. However, the effective segment for feature extraction is only the segment from the emergence of arterial microbubbles to the arterial peak value. Furthermore, the maximum peak value of the curve includes venous microbubbles. Therefore, the maximum peak value of the time-intensity curve of the ultrasound contrast-enhanced video cannot be directly selected as the end point of the video segment.
[0032] To address the issue of automatic time point selection in ultrasound contrast imaging videos, a time-intensity curve correction algorithm is proposed. By performing grayscale value analysis on prostate ultrasound contrast imaging videos using this algorithm, the sequence values of frames where the grayscale value begins to rise and the sequence values of frames reaching the peak are initially extracted. This allows for the determination of the time points from the emergence of microbubbles in prostate artery angiography to the peak value in the video segment. The time-intensity curve correction algorithm is defined as follows:
[0033]
[0034] In the formula, [] represents the floor function, and N Cor This is the corrected peak frame count sequence value, which is the sequence value extracted from the end frame of the video segment. N Max and T Max These represent the frame sequence value and time node value corresponding to the maximum peak, respectively, where j is the number of peaks before the maximum peak point, and y is the number of peaks before the maximum peak point. j Let y be the grayscale value of the last peak among these peaks, and y be the grayscale value of the last peak among these peaks. i and T iThese represent the grayscale value and time node value corresponding to the respective peak values. Here, a maximum value is selected every 5 units (frames) in the curve. This is because each fluctuation in grayscale value indicates the influence of venous microbubbles, and the influence of venous vessels needs to be weakened. Simultaneously, the frame rate of the ultrasound contrast-enhanced video is 10fps, and the changes in arterial microbubbles are significant every 5 frames, which can be used as a dividing point for extraction.
[0035] S12: Group the video frames contained in the selected video segments: Group them according to the changes in microbubble density and size in the same target area, and include video frames with small changes in microbubble density and size in the same group; each group preferably includes more than 16 ultrasound contrast images.
[0036] S13: Select one ultrasound contrast image from each video frame group for manual annotation: Locate the target area of the prostate cancer lesion in the selected image, find the heterogeneous microvesicle vascular architecture region in the target area, draw the region of interest in the heterogeneous microvesicle vascular architecture region, generate a detection box, including the coordinates and offset of the generated detection box, and perform binary classification of the region of interest as significant and non-significant according to the pathological results.
[0037] S14: Sharpen and normalize the pixel values of the images in all video frame groups.
[0038] Of course, in practice, some images can also be symmetrically, rotated, and translated to expand the training sample set.
[0039] S2: Establish a prostate cancer lesion detection model, which includes a spatiotemporal feature extraction network and a target detection network.
[0040] The spatiotemporal feature extraction network structure includes an encoding part, a direct connection part, and a decoding part.
[0041] The encoding section takes video frame groups as input. It extracts feature maps from the input video frame groups using a first 3D convolutional module, and records these extracted feature maps as the first encoding block. Then, it downsamples the first encoding block using a second 3D convolutional module and extracts feature maps from the output of the second 3D convolutional module using a 3D multi-branch stacking module, recording these extracted feature maps as the second encoding block. Next, it downsamples the second encoding block using a third 3D convolutional module and extracts feature maps from the output of the third 3D convolutional module using a 3D multi-branch stacking module, recording these extracted feature maps as the third encoding block. Then, it downsamples the third encoding block using a fourth 3D convolutional module and extracts feature maps from the output of the fourth 3D convolutional module using a 3D multi-branch stacking module, recording these extracted feature maps as the fourth encoding block. Finally, the encoding section downsamples the fourth encoding block into a 3D feature map using a dimensionality reduction module. The direct connection section connects the first, second, third, and fourth encoding blocks to the decoding section.
[0042] The decoding part upsamples the 3D feature map output by the dimensionality reduction module using a first 2D convolutional module. The upsampled feature map is stacked with the fourth encoding block. The stacked feature map is then used by a 2D multi-branch stacking module to extract the feature map, which is recorded as the first decoding block. Next, the decoding part upsamples the first decoding block using a second 2D convolutional module. The upsampled feature map is stacked with the third encoding block. The stacked feature map is then used by a 2D multi-branch stacking module to extract the feature map, which is recorded as the second decoding block. Then, the decoding part upsamples the second decoding block using a third 2D convolutional module. The upsampled feature map is stacked with the second encoding block. The stacked feature map is then used by a 2D multi-branch stacking module to extract the feature map, which is recorded as the third decoding block. Finally, the decoding part upsamples the third decoding block using a fourth 2D convolutional module. The upsampled feature map is stacked with the first encoding block. The stacked feature map is then used by a 2D multi-branch stacking module to extract the feature map, which is recorded as the fourth decoding block.
[0043] like Figure 4As shown, taking an image size of 640 (image width) * 640 (image height) * 3 (image channel count) * 16 (video frame sequence length) as an example, the encoding part generates a 640*640*64*16 feature map through 3D convolution, denoted as encoding block 1. Next, 3D convolution is used sequentially for dimensionality reduction and a 3D multi-branch stacking module to extract features, repeated four times, resulting in feature maps of sizes 320*320*128*8, 160*160*256*4, 80*80*512*2, and 40*40*1024*1, respectively. The first three feature maps are denoted as encoding block 2, encoding block 3, and encoding block 4. Each encoding block is connected to the decoding block through a direct connection. The direct connection extracts a 3D feature map of the encoding block along the video frame sequence length dimension, achieving stacking with the feature map of the decoding part. For example, encoding block 1 extracts a feature map of 640 (image width) * 640 (image height) * 64 (image channel count) and connects it to decoding block 1, and so on. The 4D feature map obtained from the encoding part is 40*40*1024*1, which is then dimensionality-reduced to obtain an output feature map of 40 (image width) * 40 (image height) * 1024 (image channel count). This feature map is input into the encoding part. The encoding part can be regarded as the feature extraction part, while the decoding block is the feature reconstruction part. Next, two-dimensional convolution is used sequentially for dimensionality upscaling and two-dimensional multi-branch stacking module for feature extraction, repeated four times to obtain feature maps of sizes 80*80*512, 160*160*256, 320*320*128, and 640*640*64, respectively denoted as decoding block 1, decoding block 2, decoding block 3, and decoding block 4. This achieves one-to-one feature stacking between decoding blocks and encoding blocks, thus realizing the pixel set feature extraction of training samples. This completes the spatiotemporal feature extraction of prostate ultrasound contrast video frame groups.
[0044] In both the encoding and decoding sections, the activation function for the multi-branch stacked modules uses the Leaky ReLU activation function, specifically:
[0045]
[0046] Among them, Leaky_ReLU(x n ) represents the input x of the nth neuron in the same layer of the artificial neural network. n The output after the Leaky ReLU activation function has a fixed parameter 'a', where a ≥ 1, and typically a takes the value 100. An activation function is a function added to an artificial neural network to help the network learn complex patterns in the data. The Leaky ReLU function is a variant of the ReLU activation function, which improves upon the input by giving a very small linear component to the negative input (e.g., a = 100, input is 0.01x). nThis addresses the zero gradient problem in negative values by adjusting for gradients, which reduces the emergence of silent neurons and allows for gradient-based learning. It solves the problem of ReLU functions failing to learn when negative inputs are zero in the negative range.
[0047] The decoding section outputs the fourth decoding block to the two-dimensional convolution module. The output of the two-dimensional convolution module is used as the input of the sigmoid function, and the output of the sigmoid function is used as the input of the object detection network.
[0048] S3: Training the Prostate Cancer Lesion Detection Model: Video frame sets from the training sample set are input into the prostate cancer lesion detection model. The output images from the spatiotemporal feature extraction network are fed into the target detection network. The target detection network extracts and learns the lesion features of prostate cancer, ultimately obtaining a successfully trained prostate cancer lesion detection model. The target detection network is a YOLOv7 network. The YOLOv7 network consists of three parts: Backbone, FPN, and YOLO Head.
[0049] The backbone is the main feature extraction network of YOLOv7. The input feature image is first processed for feature extraction in the backbone network. The extracted features can be called feature layers, which are the feature sets of the input image. Three feature layers are obtained in the backbone for the next step of network construction. I call these three feature layers the effective feature layers.
[0050] FPN is an enhanced feature extraction network for YOLOv7. The three effective feature layers obtained in the backbone are fused in this part to combine feature information from different scales. In the FPN section, the already obtained effective feature layers are used to further extract features.
[0051] Yolo Head is a classifier and regressor in YoloV7. Through Backbone and FPN, it obtains three enhanced effective feature layers. Each feature layer has width, height, and number of channels. We can then view the feature map as a collection of feature points, each with three prior boxes, and each prior box has a number of channels and features. The job of Yolo Head is to determine whether there is an object corresponding to the prior boxes at each feature point.
[0052] For YOLOv7's SPPC module: SPP's role is to increase the receptive field, allowing the algorithm to adapt to images of different resolutions. This is achieved through max pooling to obtain different receptive fields. In the first branch, there are four branches undergoing max pooling: 5, 9, 13, and 1. These four different max pooling operations represent the different objects it can handle. In other words, these four different scales of max pooling provide four different receptive fields to distinguish between large and small targets. In contrast-enhanced ultrasound images, the scale of microvessels varies in different cases and target areas; some are relatively small, and some are relatively large. Different max pooling methods can better distinguish between small and large targets.
[0053] A 640*640 feature image is input into the backbone network. Based on the output of the three layers in the backbone network, the head layer continues to output three feature maps of different sizes through the backbone network. After passing through the classification network module and the convolutional network module, the three types of image detection tasks (classification, foreground and background classification, and bounding boxes) are predicted, and the final result is output.
[0054] With the goal of converging the error between the predicted output and the true value of the prostate cancer lesion detection model, the loss function consists of three parts:
[0055] (1) Reg part: Obtain the prior box corresponding to each ground truth box. After obtaining the prior box corresponding to each prostate cancer lesion box, extract the predicted box corresponding to the prior box, and calculate the CIOU loss using the ground truth box and the predicted box, which is used as the Loss component of the Reg part.
[0056] (2) In the Obj part, the prior boxes corresponding to each ground truth box are obtained. All prior boxes corresponding to ground truth boxes are positive samples, and the remaining prior boxes are negative samples. That is, the prior boxes corresponding to prostate cancer lesion boxes are positive samples, and the remaining prior boxes that do not contain lesion areas are negative samples. The cross-entropy loss is calculated based on the positive and negative samples and the prediction results of whether the feature points contain objects, and it is used as the Loss component of the Obj part.
[0057] (3) In the Cls part, the prior box corresponding to each ground truth box is obtained. After obtaining the prior box corresponding to each box, the category prediction result of the prior box is extracted. Here, the prediction result includes significant lesions and non-significant lesions. The cross-entropy loss is calculated based on the category of the ground truth box and the category prediction result of the prior box, which is used as the Loss component of the Cls part.
[0058] The Adam optimizer was used during the training of the prostate cancer lesion detection model until convergence.
[0059] Based on ultrasound contrast imaging, since ultrasound contrast imaging is a dynamic video, detection on a 2D plane ignores its temporal information. The prostate cancer lesion detection model constructed in this embodiment uses a spatiotemporal feature extraction network that takes a group of video frames containing the time dimension as input. Through 3D convolution, it simultaneously extracts information from the video frame group in both the temporal and spatial dimensions. This ensures that the extracted features contain information from both dimensions, achieving the fusion of sequential features from consecutive frames of the dynamic ultrasound contrast imaging image. Consequently, the features extracted by the spatiotemporal feature extraction network can more accurately reflect the characteristics of the prostate cancer lesion region, and thus enable the target detection network trained based on the feature images output by the spatiotemporal feature extraction network to achieve higher detection accuracy.
[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for constructing a prostate cancer lesion detection model based on contrast imaging feature extraction, characterized in that: Includes the following steps: S1: Establish the training sample set: S11: The time selection is performed on the prostate ultrasound contrast video based on the time-intensity curve correction algorithm to obtain the video segment from the emergence of ultrasound contrast microbubbles in the prostate artery to the peak value. S12: Group the video frames contained in the selected video segment: Group the video frames according to the changes in microbubble density and size in the same target area, and include video frames with small changes in microbubble density and size in the same group. S13: Select one ultrasound contrast image from each video frame group for manual annotation: Locate the target area of prostate cancer lesions in the selected image, find the heterogeneous microvesicle vascular architecture region in the target area, draw the region of interest in the heterogeneous microvesicle vascular architecture region, generate a detection box, including the coordinates and offset of the generated detection box, and perform binary classification of the region of interest as significant and non-significant according to the pathological results. S14: Sharpen and normalize the pixel values of the images in all video frame groups; S2: Establish a prostate cancer lesion detection model, which includes a spatiotemporal feature extraction network and a target detection network; The spatiotemporal feature extraction network structure includes an encoding part, a direct connection part, and a decoding part; The encoding section takes video frame groups as input. It extracts feature maps from the input video frame groups using a first 3D convolutional module, and records the extracted feature maps as the first encoding block. Then, it downsamples the first encoding block using a second 3D convolutional module and extracts feature maps from the output of the second 3D convolutional module using a 3D multi-branch stacked module, recording the extracted feature maps as the second encoding block. Next, it downsamples the second encoding block using a third 3D convolutional module and extracts feature maps from the output of the third 3D convolutional module using a 3D multi-branch stacked module, recording the extracted feature maps as the third encoding block. Then, it downsamples the third encoding block using a fourth 3D convolutional module and extracts feature maps from the output of the fourth 3D convolutional module using a 3D multi-branch stacked module, recording the extracted feature maps as the fourth encoding block. Finally, the encoding section downsamples the fourth encoding block into a 3D feature map using a dimensionality reduction module. The direct connection section connects the first coding block, the second coding block, the third coding block, and the fourth coding block to the decoding section; The decoding part upsamples the three-dimensional feature map output by the dimensionality reduction module through the first two-dimensional convolution module. The upsampled feature map is stacked with the fourth coding block. The stacked feature map is then extracted by the two-dimensional multi-branch stacking module. The extracted feature map is counted as the first decoding block. Then, the decoding part upsamples the first decoding block through the second two-dimensional convolution module. The upsampled feature map is stacked with the third coding block. The stacked feature map is extracted by the two-dimensional multi-branch stacking module. The extracted feature map is counted as the second decoding block. Then, the decoding part upsamples the second decoding block through the third two-dimensional convolution module. The upsampled feature map is stacked with the second coding block. The stacked feature map is extracted by the two-dimensional multi-branch stacking module. The extracted feature map is counted as the third decoding block. Then, the decoding part upsamples the third decoding block through the fourth two-dimensional convolution module. The upsampled feature map is stacked with the first coding block. The stacked feature map is extracted by the two-dimensional multi-branch stacking module. The extracted feature map is counted as the fourth decoding block. The decoding part outputs the fourth decoding block to the two-dimensional convolution module, and the output of the two-dimensional convolution module is used as the input of the sigmoid function, and the output of the sigmoid function is used as the input of the object detection network. S3: Training the prostate cancer lesion detection model: Input the video frame group from the training sample set into the prostate cancer lesion detection model, put the output image of the spatiotemporal feature extraction network into the target detection network, and the target detection network extracts and learns the lesion features of prostate cancer, finally obtaining a qualified prostate cancer lesion detection model.
2. The method for constructing a prostate cancer lesion detection model based on contrast imaging feature extraction according to claim 1, characterized in that: The time-intensity curve correction algorithm is defined as follows: In the formula, [] represents the floor function, and N Cor N represents the corrected peak frame count sequence value. Max and T Max These represent the frame sequence value and time node value corresponding to the maximum peak, respectively, where j is the number of peaks before the maximum peak point, and y is the number of peaks before the maximum peak point. j Let y be the grayscale value of the last peak among these peaks, and y be the grayscale value of the last peak among these peaks. i and T i These represent the grayscale value and time node value corresponding to the respective peak values.
3. The method for constructing a prostate cancer lesion detection model based on contrast imaging feature extraction according to claim 1, characterized in that: The target detection network is the YOLOv7 network.