Chest cross-sectional CT tilt recognition method and system based on key point detection

Abstract

The present invention proposes a chest cross-sectional CT tilt recognition method and system based on key point detection, which relates to the field of image processing technology. It includes obtaining a chest cross-sectional CT image and the corresponding key point coordinates; based on the key point coordinates, generating a label heat map matrix for each chest cross-sectional CT image; building a convolutional neural network model, using the chest cross-sectional CT image to train the convolutional neural network model, obtaining a predicted heat map matrix, calculating the model loss, and completing the training of the model; inputting the unlabeled chest cross-sectional CT image into the trained convolutional neural network model, obtaining a predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image; finding the coordinates of the point with the highest heat in the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image, which are the predicted key point coordinates. The present invention uses the calculated positioning deviation angle as a benchmark to dynamically remind technicians to adjust the positioning and improve inspection efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a chest cross-sectional CT tilt recognition method and system based on key point detection. Background Art

[0002] A CT scan (computed tomography) is a clinical imaging procedure. As one of the most common imaging modalities used in clinical diagnosis, CT helps healthcare providers detect diseases and injuries in patients. It uses a series of X-rays and computer software to create detailed images of bones and soft tissues, and is widely used in screening and diagnosing major diseases.

[0003] Before a CT scan, a technician typically positions the patient manually, relying on their experience. Cross-sectional deviations caused by positioning errors or the patient's body tilt during organ scanning can severely impact image quality, hindering the doctor's diagnosis. Substandard CT images require post-processing correction by the technician. Severe deviations that cannot be corrected require a rescan, increasing the patient's exposure to radiation and reducing diagnostic efficiency. Summary of the Invention

[0004] To overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a chest cross-sectional CT tilt recognition method and system based on key point detection. A convolutional neural network is used to extract the key point features of the chest cross-sectional CT. The angle of the patient's positioning deviation is calculated based on the position information of the key points in the image. Based on the calculated positioning deviation angle, the system dynamically reminds the technician to adjust the positioning, thereby improving examination efficiency.

[0005] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0006] A first aspect of the present invention provides a chest cross-sectional CT tilt recognition method based on key point detection.

[0007] The method for identifying tilt in chest cross-sectional CT images based on key point detection includes the following steps:

[0008] Obtain the original chest cross-sectional CT image, manually mark the key points therein, and perform preprocessing to obtain the preprocessed chest cross-sectional CT image and the corresponding key point coordinates;

[0009] Generate a label heat map matrix for each chest cross-sectional CT image based on the key point coordinates;

[0010] Build a convolutional neural network model and train it using chest cross-sectional CT images to obtain a predicted heat map matrix. Calculate the model loss based on the predicted heat map matrix and the label heat map matrix until the loss converges, completing the model training.

[0011] The unlabeled chest cross-sectional CT image is input into the trained convolutional neural network model to obtain the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image;

[0012] Find the coordinates of the point with the highest heat in the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image, which are the predicted key point coordinates;

[0013] Based on the predicted key point coordinates, the tilt angle during chest cross-sectional CT scanning is calculated to complete the tilt angle recognition.

[0014] The second aspect of the present invention provides a chest cross-sectional CT tilt recognition system based on key point detection.

[0015] The chest cross-sectional CT tilt recognition system based on key point detection includes:

[0016] The preprocessing module is configured to: obtain an original chest cross-sectional CT image, manually annotate key points therein, and perform preprocessing to obtain a preprocessed chest cross-sectional CT image and corresponding key point coordinates;

[0017] The label heat map matrix generation module is configured to: generate a label heat map matrix for each chest cross-sectional CT image based on the key point coordinates;

[0018] The training module is configured to: build a convolutional neural network model, train the convolutional neural network model using chest cross-sectional CT images to obtain a predicted heat map matrix, and calculate the model loss based on the predicted heat map matrix and the label heat map matrix until the loss value converges, thereby completing the model training;

[0019] The predicted heat map matrix calculation module is configured to: input the unlabeled chest cross-sectional CT image into the trained convolutional neural network model to obtain the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image;

[0020] The key point coordinate determination module is configured to: find the coordinates of the point with the highest heat value in the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image, and use them as the predicted key point coordinates;

[0021] The tilt angle calculation module is configured to calculate the tilt angle during chest cross-sectional CT scanning based on the predicted key point coordinates, thereby completing tilt angle recognition.

[0022] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the chest cross-sectional CT tilt identification method based on key point detection as described in the first aspect of the present invention.

[0023] The fourth aspect of the present invention provides an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the chest cross-sectional CT tilt identification method based on key point detection as described in the first aspect of the present invention are implemented.

[0024] One or more of the above technical solutions have the following beneficial effects:

[0025] The present invention provides a method and system for identifying chest cross-sectional CT tilt based on key point detection. A convolutional neural network is used to extract key point features of chest cross-sectional CT. The angle of patient positioning deviation is calculated based on the position information of the key points in the image. The calculated positioning deviation angle is used as a reference to dynamically remind technicians to adjust the positioning, thereby improving examination efficiency.

[0026] Unlike most existing solutions that connect convolutional blocks in series, the convolutional neural network used in this paper connects high- to low-resolution sub-networks in parallel. This convolutional neural network design preserves high-resolution features rather than restoring resolution through a low-to-high-resolution process. Accordingly, multi-scale fusion is used to enhance high-resolution representations, resulting in rich feature estimates for high-resolution representations. As a result, the heatmaps predicted by this convolutional neural network design are more accurate.

[0027] The present invention uses key points as a benchmark to detect patient positioning deviation. It uses chest clinical cross-sectional CT examination sequence images, annotates the sternum, vertebral body, and lamina as key points for training labels of the deep learning algorithm, and calculates the model loss based on the predicted heat map matrix and the label heat map matrix until the loss value converges. The trained model has higher accuracy.

[0028] Based on the predicted key point coordinates, the present invention divides the key points into two and three cases, designs calculation formulas respectively, calculates the inclination angle during chest cross-sectional CT scanning, completes the inclination angle recognition, and makes the calculation more targeted.

[0029] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0031] Figure 1 This is a flow chart of the method of the first embodiment.

[0032] Figure 2 This is the overall structure diagram of the convolutional neural network model.

[0033] Figure 3 Schematic diagram of the predicted heat map matrix for predicting 2 key points.

[0034] Figure 4 Schematic diagram of the predicted heat map matrix for predicting three key points. DETAILED DESCRIPTION

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0036] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0037] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0038] Example 1

[0039] This embodiment discloses a chest cross-sectional CT tilt recognition method based on key point detection.

[0040] like Figure 1 As shown, the chest cross-sectional CT tilt recognition method based on key point detection includes the following steps:

[0041] Obtain the original chest cross-sectional CT image, manually mark the key points therein, and perform preprocessing to obtain the preprocessed chest cross-sectional CT image and the corresponding key point coordinates;

[0042] Generate a label heat map matrix for each chest cross-sectional CT image based on the key point coordinates;

[0043] Build a convolutional neural network model and train it using chest cross-sectional CT images to obtain a predicted heat map matrix. Calculate the model loss based on the predicted heat map matrix and the label heat map matrix until the loss converges, completing the model training.

[0044] The unlabeled chest cross-sectional CT image is input into the trained convolutional neural network model to obtain the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image;

[0045] Find the coordinates of the point with the highest heat in the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image, which are the predicted key point coordinates;

[0046] Based on the predicted key point coordinates, the tilt angle during chest cross-sectional CT scanning is calculated to complete the tilt angle recognition.

[0047] With the widespread application of artificial intelligence technology in the medical field, research in this field is progressing rapidly. If artificial intelligence technology can be used to detect whether the patient's positioning is correct during CT scans, it can dynamically remind technicians to adjust the positioning, thereby improving examination efficiency.

[0048] Therefore, this embodiment proposes a chest cross-sectional CT tilt recognition method based on key point detection, which specifically includes the following steps:

[0049] a) Obtain the chest cross-sectional CT image sequence data set ct_imgs, and the key point data of each CT image corresponding to the annotation personnel are obtained as the annotation set keypoint_jsons.

[0050] b) The CT image data converted in step a) is further preprocessed to produce the CT image ct_imgs , with dimensions (batch_size, chanels, img_height, img_width), where batch_size is the training data batch size, the number of image chanels is n, and CT images are multi-channel images. img_height and img_width are the dimensions of a single CT image, 512×512. The annotation set keypoint_jsons is preprocessed to produce the training annotation keypoint_annotation.

[0051] The preprocessing steps are as follows:

[0052] Each image ct in b1.ct_imgs i Using different window widths and window positions will produce images that focus on different organs and tissues, so three windows are selected here: the lung window, the soft tissue window, and the bone window. The lung window focuses on showing the texture of the lungs, the soft tissue window is suitable for showing the condition of the mediastinum, and the bone window is suitable for showing the bone tissue related to the ribs.

[0053] b2. ct of each image i Using the lung window, soft tissue window, and bone window through linear mapping, three single-channel image CT images are generated respectively. i-lung ,ct i-soft ,ct i-bone .

[0054] b3. ct i-lung ,ct i-soft ,ct i-bone Three single-channel images are fused into a multi-channel image, ct i =concat([cti-lung ,ct i-soft ,ct i-bone ]), the final ct i becomes a three-channel image.

[0055] c) Initialize model parameters and training parameters, set model training strategy, load training annotation keypoint_annotation, and load CT image data set ct_imgs.

[0056] d) Based on the labeled data read in step c), generate a heat map matrix label_mask for each CT image, whose feature dimension is (1, img_height, img_width).

[0057] e) After loading CT images in batches, the model performs four stages of feature extraction using a convolutional neural network. The tensor features obtained at each stage are called stage1_conv, stage2_conv, stage3_conv, and stage4_conv. The tensor features at each stage vary in size.

[0058] f) After processing in step e), tensor features at different levels are obtained, and these features are fused to obtain richer information.

[0059] This step includes feature map downsampling (feature downsample), feature map skip connection (feature residual connection), feature map compression (feature compress), and feature map fusion (feature fusion) operations.

[0060] g) Calculate the prediction loss loss using the prediction heatmap predictionheatmap generated during model training and the label heatmap matrix label_mask for the next round of optimization iteration.

[0061] h) Use the optimizer to backpropagate the learning parameters of the convolutional neural network based on the prediction loss.

[0062] i) Repeat steps b) to h) until the loss reaches a stable convergence state, at which point the optimal model weight M is obtained.

[0063] j) Load the model weights obtained in step i) and input the chest cross-sectional CT image to obtain the model prediction result result, whose dimensions are (3, x, y, prob), where prob is the confidence level of the prediction point.

[0064] k) Calculate the deviation angle of the cross-sectional CT using the key points predicted by the model.

[0065] More specifically:

[0066] 1. The experimental data in step a) uses real clinical CT image data. Data preprocessing is required before training, which mainly includes two steps:

[0067] a-1) Image format conversion to obtain CT image data set ct_imgs. The original CT image data storage format is a dicom file, which is converted into a jpg file with default window width and window position. The default window width and window position are obtained from the dicom metadata, which are the window width field '0028|1051', the window position field

[0068] '0028|1050'.

[0069] a-2) Annotators mark key points in each CT image and generate corresponding annotation files. Key points in cross-sectional CT images are labeled as the sternum, vertebral body, and lamina, and their coordinates (x, y) in the image are annotated. All annotation data is reviewed by a professional physician and used for further processing.

[0070] 2. The data preprocessing in step b) mainly includes two parts: CT image and annotation.

[0071] b-1) CT image. Rename the CT image to the format of 'CT examination number_serial number_image number.jpg', such as CT01931088_SN3_000011.jpg.

[0072] b-2) CT annotation set. Read the annotation file of each image, convert it and store it in the annotation format of the COCO dataset to obtain keypoint_jsons. The converted annotation file contains 5 parts: info, licenses, annotations, categories, and image. Info stores the version information of the dataset, licenses is the copyright statement, categories are the categories of CT images, and images is all the image information in the dataset. Annotations is all the annotation information in the dataset, where image_id is the image id corresponding to the annotation, id is the unique identifier of the annotation file, and keypoints is a list of key point annotations in the format of (x, y, visible), that is, the coordinates of the annotation point and whether it is visible (occluded).

[0073] 3. Initialize model parameters and training parameters in step c) and set the model training strategy.

[0074] c-1) Load the model parameters based on the ImageNet pre-trained weight initialization.

[0075] Set the model to a 4-stage convolutional neural network, such as Figure 2 As shown, the feature map name and size of each stage change as follows:

[0076] Stage 1: This stage consists of 4 residual units, each of which consists of a bottleneck of width 64, followed by a 3×3 convolutional feature map with width reduced to 32.

[0077] The second stage consists of a swap block containing 4 residual units, each of which contains 2 3×3 convolutions at each resolution and a swap unit at each resolution.

[0078] Stage 3: This stage contains 4 exchange blocks. Each exchange block contains 4 residual units, where each unit contains 2 3×3 convolutions at each resolution and a resolution exchange unit.

[0079] Stage 4: This stage contains 3 exchange blocks. Each exchange block contains 4 residual units, where each unit contains 2 3×3 convolutions at each resolution and a resolution exchange unit.

[0080] c-2) Initialize training parameters. Set the batch size BATCH_SIZE_PER_GPU to 128, set the data shuffle to true, set the total number of training epochs to 210, set the learning rate LR to 0.0001, and decay the learning rate after 170 epochs. Set the optimizer OPTIMIZER to Adam, set the GAMMA1 parameter to 0.99, and set the GAMMA2 parameter to 0.

[0081] c-3) Set the model training strategy. Use sklearn.model_selection to divide the dataset into a training set: validation set ratio of 9:1. During training, use image flipping and affine transformation to enhance the data.

[0082] The keypoint threshold OKS_THRE is set to 0.9. When loading images for training, normalization is performed, with a mean of [0.485, 0.456, 0.406] and a variance of [0.229, 0.224, 0.225].

[0083] 4. Step d) Generate a heatmap matrix label_mask for each CT image. For each keypoint (x, y) in each CT image, generate a Gaussian heatmap matrix centered at that point. You need to set the standard deviation parameter σ of the Gaussian matrix generated for each keypoint and the heatmap matrix size (M*N). σ, M, and N are all hyperparameters.

[0084] 5. Tensor features at each stage in step e).

[0085] e-1) stage1_conv. The first stage of the model contains 4 residual units, each of which consists of a bottleneck with a width of 64, followed by a 3×3 convolution, and the width of the feature map is reduced to 32. i Input to stage1_conv. The size of the feature map output by this stage is 32×32×64, and this feature map is called stage1_conv.

[0086] e-2) stage2_conv. The second stage of the model consists of a convolution block containing four residual units, each of which contains two 3×3 convolutions at each resolution and a convolution unit at each resolution. The output feature map of this stage is 16×16×96 and is called stage2_conv.

[0087] e-3) stage3_conv. The third stage of the model consists of four convolution blocks. Each convolution block contains four residual units, each of which includes two 3×3 convolutions at each resolution and one convolution unit at each resolution. The output feature map of this stage is 8×8×192 and is called stage3_vonv.

[0088] e-4) stage4_conv. The fourth stage of the model consists of three convolution blocks. Each convolution block contains four residual units, each of which includes two 3×3 convolutions at each resolution and one convolution unit at each resolution. The output feature map of this stage is 4×4×320 and is called stage4_conv.

[0089] 6. Feature fusion operation in step f). The feature fusion stage mainly extracts features at different scales and fuses these features to obtain richer information.

[0090] f-1) Downsampling.

[0091] This step is achieved through a series of convolution operations and downsampling operations to convert the high-resolution feature map into a low-resolution feature map, so that the high-resolution feature map can be aligned with the low-resolution feature map in spatial size.

[0092] f-2) Skip connection.

[0093] At each stage of a convolutional neural network, skip connections are used to connect low-resolution feature maps with high-resolution feature maps. These connections are usually implemented by pixel-by-pixel addition.

[0094] f-3) Feature compression.

[0095] This step is usually achieved through a series of convolution operations, the purpose of which is to compress the feature map so that its size becomes smaller, so that the high-resolution feature map can be connected with the low-resolution feature map.

[0096] f-4) Feature fusion.

[0097] In the feature fusion stage, feature maps of different scales need to be fused. These feature maps of different scales are usually fused by pixel-by-pixel addition.

[0098] 7. The loss function used in step g) is the weighted fine-tuned MSELoss, and the specific calculation expression is as follows.

[0099]

[0100] Among them, k is the number of key points, kp_pred is the model prediction heat map matrix, kp_label is the label truth matrix, weight i is the weight vector for each key point.

[0101] 8. The optimizer in step h) is set to Adam optimizer, and its momentum parameter beta1 is initialized to 0.99, the quadratic momentum parameter beta2 is initialized to 0, the numerical stability parameter epsilon is 1e-8, and the learning rate is set to 0.0001.

[0102] 9. In step k), the angle calculation is divided into two cases, namely, the CT image contains 2 key points and 3 key points.

[0103] k-1)2 key points, such as Figure 3 The inclination angle θ is calculated based on the coordinates of the vertebral body and lamina using the following formula:

[0104]

[0105] k-2) 3 key points, such as Figure 4 The angle θ1 is calculated based on the coordinates of the sternum and vertebral body, and the angle θ2 is calculated based on the coordinates of the vertebral body and lamina. The average value is the inclination angle θ, and the calculation formula is:

[0106]

[0107] Among them, (x1, y1) is the coordinate of the vertebral arch; (x2, y2) is the coordinate of the vertebral body; (x3, y3) is the coordinate of the sternum.

[0108] Example 2

[0109] This embodiment discloses a chest cross-sectional CT tilt recognition system based on key point detection.

[0110] The chest cross-sectional CT tilt recognition system based on key point detection includes:

[0111] The preprocessing module is configured to: obtain an original chest cross-sectional CT image, manually annotate key points therein, and perform preprocessing to obtain a preprocessed chest cross-sectional CT image and corresponding key point coordinates;

[0112] The label heat map matrix generation module is configured to: generate a label heat map matrix for each chest cross-sectional CT image based on the key point coordinates;

[0113] The training module is configured to: build a convolutional neural network model, train the convolutional neural network model using chest cross-sectional CT images to obtain a predicted heat map matrix, and calculate the model loss based on the predicted heat map matrix and the label heat map matrix until the loss value converges, thereby completing the model training;

[0114] The predicted heat map matrix calculation module is configured to: input the unlabeled chest cross-sectional CT image into the trained convolutional neural network model to obtain the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image;

[0115] The key point coordinate determination module is configured to: find the coordinates of the point with the highest heat value in the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image, and use them as the predicted key point coordinates;

[0116] The tilt angle calculation module is configured to calculate the tilt angle during chest cross-sectional CT scanning based on the predicted key point coordinates, thereby completing tilt angle recognition.

[0117] Example 3

[0118] The purpose of this embodiment is to provide a computer-readable storage medium.

[0119] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the chest cross-sectional CT tilt identification method based on key point detection as described in Example 1 of the present disclosure.

[0120] Example 4

[0121] The purpose of this embodiment is to provide an electronic device.

[0122] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the chest cross-sectional CT tilt recognition method based on key point detection as described in Example 1 of the present disclosure are implemented.

[0123] The steps involved in the apparatuses of Examples 2, 3, and 4 above correspond to those of Method Example 1. For detailed implementations, please refer to the relevant description of Example 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any method of the present invention.

[0124] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0125] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A chest cross-sectional CT tilt recognition method based on key point detection, characterized in that: The following steps are involved: Obtain the original chest cross-sectional CT image, manually mark the key points therein, and perform preprocessing to obtain the preprocessed chest cross-sectional CT image and the corresponding key point coordinates; Generate a label heat map matrix for each chest cross-sectional CT image based on the key point coordinates; Build a convolutional neural network model and train it using chest cross-sectional CT images to obtain a predicted heat map matrix. Calculate the model loss based on the predicted heat map matrix and the label heat map matrix until the loss converges, completing the model training. The unlabeled chest cross-sectional CT image is input into the trained convolutional neural network model to obtain the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image; Find the coordinates of the point with the highest heat in the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image, which are the predicted key point coordinates; Based on the predicted key point coordinates, the tilt angle of the chest cross-sectional CT scan is calculated to complete the tilt angle recognition; The convolutional neural network model includes four stages of residual convolution modules: The residual convolution module of the first stage includes four first residual units, each of which includes a bottleneck and a subsequent convolutional layer; The residual convolution module of the second stage includes 1 exchange block, which includes 4 second residual units, which include 2 convolution layers and a resolution exchange unit; The residual convolution module in the third stage includes 4 exchange blocks; The residual convolution module in the fourth stage includes 3 exchange blocks; Input the unlabeled chest cross-sectional CT image into the trained convolutional neural network model: Four stages of residual convolution modules are used to extract the first, second, third, and fourth feature maps of different scales. In each residual convolution module, the low-resolution feature map is connected to the high-resolution feature map through a jump connection. Downsampling and feature compressing the first feature map, the second feature map, the third feature map, and the fourth feature map; The first feature map, the second feature map, the third feature map and the fourth feature map after feature compression are fused to obtain the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image.

2. The chest cross-sectional CT tilt recognition method based on key point detection according to claim 1, characterized in that: The coordinates of the sternum, vertebral body, and lamina in the original chest cross-sectional CT image are manually annotated, and the format of the original chest cross-sectional CT image and the manually annotated key points are converted to complete the preprocessing process.

3. The chest cross-sectional CT tilt recognition method based on key point detection according to claim 1, characterized in that: The label heat map matrix of each chest cross-sectional CT image is generated using Gaussian distribution, including: Define the parameters of the Gaussian distribution; Identify the coordinates of the marked key points and calculate the standard deviation of the Gaussian distribution of the key points; Generates a label heatmap matrix based on the standard deviation of a Gaussian distribution.

4. The chest cross-sectional CT tilt recognition method based on key point detection according to claim 1, characterized in that: The loss function of the convolutional neural network model is calculated as follows: in, is the number of key points, is the prediction heat map matrix, is the label heat map matrix, is the weight vector for each key point.

5. The chest cross-sectional CT tilt recognition method based on key point detection according to claim 1, characterized in that: Based on the predicted key point coordinates, the tilt angle of the chest cross-sectional CT scan is calculated, including: First, distinguish the names of the key points; When two key points are predicted, the inclination angle is calculated based on the coordinates of the vertebral body and lamina. , the calculation formula is: in, is the coordinate of the lamina; are the coordinates of the vertebra; When the three key points are predicted, the angle is calculated based on the coordinates of the vertebral body and lamina , calculate the angle based on the coordinates of the sternum and vertebrae ,Pick and The average value of the tilt angle , the calculation formula is: in, are the coordinates of the sternum.

6. A chest cross-sectional CT tilt recognition system based on key point detection, characterized by: include: The preprocessing module is configured to: obtain an original chest cross-sectional CT image, manually annotate key points therein, and perform preprocessing to obtain a preprocessed chest cross-sectional CT image and corresponding key point coordinates; The label heat map matrix generation module is configured to: generate a label heat map matrix for each chest cross-sectional CT image based on the key point coordinates; The training module is configured to: build a convolutional neural network model, train the convolutional neural network model using chest cross-sectional CT images to obtain a predicted heat map matrix, and calculate the model loss based on the predicted heat map matrix and the label heat map matrix until the loss value converges, thereby completing the model training; The predicted heat map matrix calculation module is configured to: input the unlabeled chest cross-sectional CT image into the trained convolutional neural network model to obtain the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image; The key point coordinate determination module is configured to: find the coordinates of the point with the highest heat value in the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image, and use them as the predicted key point coordinates; The tilt angle calculation module is configured to: calculate the tilt angle during chest cross-sectional CT scanning based on the predicted key point coordinates to complete tilt angle recognition; The convolutional neural network model includes four stages of residual convolution modules: The residual convolution module of the first stage includes four first residual units, each of which includes a bottleneck and a subsequent convolutional layer; The residual convolution module of the second stage includes 1 exchange block, which includes 4 second residual units, which include 2 convolution layers and a resolution exchange unit; The residual convolution module in the third stage includes 4 exchange blocks; The residual convolution module in the fourth stage includes 3 exchange blocks; Input the unlabeled chest cross-sectional CT image into the trained convolutional neural network model: Four stages of residual convolution modules are used to extract the first, second, third, and fourth feature maps of different scales. In each residual convolution module, the low-resolution feature map is connected to the high-resolution feature map through a jump connection. Downsampling and feature compressing the first feature map, the second feature map, the third feature map, and the fourth feature map; The first feature map, the second feature map, the third feature map and the fourth feature map after feature compression are fused to obtain the predicted heat map matrix corresponding to the unlabeled chest cross-sectional CT image.

7. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the chest cross-sectional CT tilt recognition method based on key point detection as described in any one of claims 1 to 5 are implemented.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the chest cross-sectional CT tilt recognition method based on key point detection according to any one of claims 1 to 5 are implemented.

Images