Central NSCLC thoracotomy exploration risk prediction method and system

Abstract

The invention belongs to the technical field of medical image analysis, and particularly discloses a central NSCLC thoracotomy exploration risk prediction method and system, and the method comprises the following steps: collecting clinical data of a sample and a chest CT image, carrying out the three-dimensional reconstruction of the CT image, selecting a key coronal layer, and calculating the tumor area and the blood vessel compression deformation degree; performing multi-factor regression risk analysis on the clinical data, the tumor area and the blood vessel compression deformation degree to determine independent risk factors; constructing a prediction model based on the independent risk factors; independent risk factors of a patient are collected and input into the prediction model, and a risk score is generated. By adopting the technical scheme, the thoracotomy exploration risk prediction is realized by utilizing the prediction model.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image analysis and relates to a method and system for predicting the risk of central NSCLC thoracotomy. Background Art

[0002] Lung cancer is a major health threat worldwide, and surgical resection is the main treatment. However, due to surgical limitations, the prognosis of advanced non-small cell lung cancer is very poor. Ideally, the goal of surgery should be radical (R0) resection, because residual tumor tissue significantly increases the risk of death (hazard ratio: 1.5-8.2). However, thoracotomy is unresectable due to local tumor invasion of key anatomical structures (such as large blood vessels and carina), and the surgery is terminated. During surgery for patients with central NSCLC (Non-Small Cell Lung Cancer, non-small cell lung cancer), approximately 2%-17% require thoracotomy (without radical resection) due to tumor invasion of large blood vessels or carina, which significantly affects the prognosis.

[0003] Currently, surgeons rely primarily on subjective CT (computed tomography) imaging to assess resectability, which results in significant individual variability. Inexperienced surgeons, in particular, can make significant errors in their judgment of the resectability of complex tumors. This subjective and unquantifiable assessment poses significant challenges to clinicians.

[0004] Few studies have focused on accurately and quantifiably predicting the risks associated with exploratory thoracotomy. Preoperative assessment of these radiographic features is crucial for optimizing neoadjuvant therapy strategies and determining the appropriate timing of surgical intervention. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for predicting the risk of thoracotomy for central NSCLC to achieve risk prediction of thoracotomy.

[0006] To achieve the above objectives, the basic scheme of the present invention is: a method for predicting the risk of central NSCLC thoracotomy, comprising the following steps:

[0007] Collect clinical data and chest CT images of the samples;

[0008] Perform three-dimensional reconstruction of CT images, select key coronary planes, and calculate the tumor area and degree of vascular compression and deformation;

[0009] Multivariate regression risk analysis was performed on clinical data, tumor area, and degree of vascular compression and deformation to identify independent risk factors;

[0010] Based on independent risk factors, a prediction model was constructed;

[0011] The patient's independent risk factors are collected and input into the prediction model to generate a risk score.

[0012] The working principle and beneficial effects of this basic scheme are: This technical scheme is based on the clinical data and chest CT images of the samples to screen independent risk factors and build a prediction model to evaluate the risk of thoracotomy and its postoperative survival.

[0013] Furthermore, the CT image is reconstructed in three dimensions and the method for selecting key coronal planes is as follows:

[0014] A four-stage encoder-decoder framework is used to segment the 3D tumor mask of CT images.

[0015] The 3D tumor mask is fed into a lightweight post-processing module to calculate the tumor area layer by layer in the coronal, axial, and sagittal planes.

[0016] The slice index with the largest recorded area is the key fault.

[0017] The key coronal planes were selected to facilitate subsequent analysis.

[0018] Furthermore, a four-stage encoder-decoder framework is used to segment the 3D tumor mask of the CT image as follows:

[0019] In the encoder, QuadVSS Block is used to replace standard convolution, and standard VSS Block is stacked to build a multi-layer visual state modeling unit to capture global semantics;

[0020] The spatial size is halved between each stage to gradually extract deep features;

[0021] Set up large-core attention gate LGAG to model local channel attention;

[0022] The frequency fusion module FreqFusion combines DCT / FFT frequency domain transform with low-frequency residual guidance to guide the model to focus on strong structural boundaries and texture details, enhancing the fusion expression between multi-scale features. Specifically:

[0023] The FreqFusion module decomposes features into low-frequency structures and high-frequency detail components through frequency domain transformation, uses low-frequency residuals to generate attention weights to enhance the subject contour, preserves texture details through high-frequency pathways, and finally fuses multi-frequency features through adaptive weights;

[0024] Based on skip connections and multi-level fusion, the context consistency of the upsampling path is enhanced through residual weighted connections of different scales. Specifically:

[0025] Through multi-scale skip connections between the encoder and decoder, residual convolution is performed on the encoder and decoder features of each level respectively, and learnable dynamic attention weights are introduced to achieve adaptive weighted fusion of cross-layer features;

[0026] Through pyramid-style cross-level interaction, a multi-level residual fusion path is formed from local details to global semantics, and finally multi-scale context is aggregated through pyramid pooling;

[0027] A deep supervision mechanism is introduced into the decoder. The decoder uses two consecutive VSS Block decoding features, combined with layer-by-layer 1×1 convolution and deep supervision to improve training stability and gradient transfer. The final output of the decoder is the tumor segmentation mask, which also supports the attachment of the "maximum cross-section extraction module" to calculate the tumor area and the degree of vascular compression and deformation. First, based on the 3D tumor mask, the key coronal plane with the largest tumor area is selected through multi-plane slicing, and its actual area (number of pixels × physical size) is calculated. At the same time, the vascular area is segmented at this plane, and the degree of vascular compression and deformation is quantified by comparing the cross-sectional area ratio of compressed and normal vessels and the boundary displacement distance.

[0028] A four-stage encoder-decoder framework is adopted to implement a medical segmentation process with multi-scale and global-local information joint enhancement by introducing Quadtree-Vision State Space Block (QuadVSS), frequency fusion module (FreqFusion), large kernel attention gating (LGAG) and deep supervision mechanism.

[0029] Furthermore, the prediction model adopts a Nomogram model.

[0030] The nomogram model is useful for predicting the risk score for exploratory thoracotomy.

[0031] Furthermore, we set the risk threshold to determine the patient's risk level:

[0032] Low risk level: if the risk score is ≤30 points;

[0033] Medium risk level: if the risk score is 30-60 points;

[0034] High risk level: If the risk score is ≥60 points.

[0035] Using risk scores to classify patients into risk levels helps determine the patient's physical condition.

[0036] Furthermore, the patient's chest CT image is collected and checked to see if it is qualified. Specifically:

[0037] S1, check whether the DICOM file of the chest CT image is complete. If so, proceed to step S2; otherwise, determine that the chest CT image is unqualified, reacquire the chest CT image and execute step S1;

[0038] S2, verify whether the scanning range covers the entire lung. If so, proceed to step S3; otherwise, determine that the chest CT image is unqualified, reacquire the chest CT image and execute step S1;

[0039] S3, calculating the artifact area in the chest CT image. If the artifact area is less than 5%, proceed to step S4; otherwise, the chest CT image is judged to be unqualified, and the chest CT image is re-acquired and step S1 is executed;

[0040] S4, compare the chest CT image with the preset qualified image, calculate the similarity based on body position and key anatomical structures, and if the similarity meets the preset value, output the chest CT image for 3D reconstruction; otherwise, judge the chest CT image as unqualified, reacquire the chest CT image and execute step S1.

[0041] Check whether the chest CT image is qualified, obtain high-quality image information, extract more accurate body feature information, and obtain more accurate prediction results.

[0042] Furthermore, similarity is calculated based on body position and key anatomical structures, specifically:

[0043] Let the posture similarity be S p , the anatomical structure similarity is S a , then the comprehensive similarity S t for:

[0044] S t =w1S p +w2s a

[0045] Wherein, w1 and w2 are weighting coefficients, satisfying w1+w2=1;

[0046] Mutual information MI is used to measure body position similarity. Let R be the rotation matrix, which represents the rotation transformation from chest CT image I1 to chest CT image I2; T is the translation vector, which represents the displacement from chest CT image I1 to chest CT image I2; the image similarity after rotation and translation correction is:

[0047] B=MI(I1,I2)

[0048] Among them, MI(I1, I2) represents the mutual information, which is used to evaluate the registration quality of two images:

[0049]

[0050] Where p(I1, I2) is the joint probability distribution, which represents the joint occurrence probability of pixel pairs in chest CT images I1 and I2; p(I1) and p(I2) are the marginal probability distributions of chest CT images I1 and I2, respectively;

[0051] Use deep learning algorithms to segment chest CT images, extract the lung regions L1 and L2, and the heart regions C1 and C2 from the chest CT images, and use the structural similarity index to calculate the similarity between the two regions R1 and R2:

[0052]

[0053] Where μ1 and μ2 are the means of regions R1 and R2, σ1 and σ2 are the variances, and σ 1,2 is the covariance of regions R1 and R2, and c1 and c2 are constants used to prevent the denominator from being zero.

[0054] The similarity is calculated based on body position and key anatomical structures, with simple calculation and easy to use.

[0055] The present invention also provides a central NSCLC thoracotomy risk prediction system, comprising a bed, a human-computer interaction module, a CT image acquisition module, and a processing module;

[0056] The human-computer interaction module and the CT image acquisition module are both installed on the bed, the patient lies flat on the bed, and the human-computer interaction module is used to collect clinical data of the sample;

[0057] The CT image acquisition module is used to acquire chest CT images of the sample and transmit them to the processing module;

[0058] The processing module executes the method of the present invention to complete the risk prediction of thoracotomy.

[0059] This system uses a human-computer interaction module to input the corresponding clinical data. The CT image acquisition module collects sample chest CT images and inputs them into the processing module, which analyzes the data, effectively predicts the risk of thoracotomy, and outputs the prediction results.

[0060] Furthermore, it also includes auxiliary mechanisms, which include a conveyor belt, an air bag, an elastic restraint belt, and a suction cup type fixing member;

[0061] A groove is provided from the chest examination position to the foot of the bed, and the conveyor belt is arranged in the groove, and the upper surface of the conveyor belt is flush with the upper surface of the bed;

[0062] The airbag is arranged at the chest examination position of the bed, and the CT image acquisition module is located above the chest examination position of the bed. The airbag is made of elastic deformable material and is connected to an inflation and deflation mechanism. The inflation and deflation mechanism is installed on the bed. The airbag is initially in an inflated state.

[0063] Both ends of the elastic binding belt are connected with suction cup type fixing parts, which can be adsorbed and positioned with the upper surface of the bed. During examination, the patient's arm is placed between the elastic binding belt and the upper surface of the bed.

[0064] The conveyor belt is set in the groove. When the patient sits on the conveyor belt and then lies down on the bed, if the patient's chest does not correspond to the position of the CT image acquisition module above the bed, the conveyor belt can be started to move forward or reverse, driving the patient to move a certain distance to the head or foot of the bed to assist the patient in moving his body.

[0065] The airbag is set at the chest examination position of the bed. When the patient lies on the bed, the back corresponding to the chest presses on the airbag, which is convenient for positioning. At the same time, the airbag has a certain raised structure, which can help the patient expand the chest and prevent the patient from holding the chest and affecting image acquisition.

[0066] The elastic restraint belt can be connected to the upper surface of the bed through a suction cup type fixing part. When the patient is examined, the patient's arm is fixed on the bed to prevent the patient from moving accidentally and achieve positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 2. This is a front view structural diagram of the central NSCLC thoracotomy risk prediction system of the present invention;

[0068] Figure 2 1 is a schematic top view of the structure of the central NSCLC thoracotomy risk prediction system of the present invention.

[0069] The reference numerals in the drawings of the specification include: bed 1 , human-computer interaction module 2 , CT image acquisition module 3 , airbag 4 , and conveyor belt 5 . DETAILED DESCRIPTION

[0070] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0071] In the description of the present invention, it should be understood that the terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention.

[0072] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal communication between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0073] The present invention discloses a method for predicting the risk of thoracotomy for central NSCLC. A prediction model is constructed based on CT imaging features after neoadjuvant therapy to assess the risk of thoracotomy and postoperative survival. The method for predicting the risk of thoracotomy for central NSCLC includes the following steps:

[0074] Clinical data and chest CT images of the samples were collected; clinical data included variables such as age, gender, smoking status, tumor location, tumor pathological type, pathological stage, neoadjuvant treatment and corresponding tumor response.

[0075] Using Materialise Mimics Medical (Version 21.0), 3D reconstruction of the CT images facilitated visualization of the internal lung structures and their interrelationships. Key coronal slices (which clearly visualize the tumor and surrounding mediastinal vessels) were selected, and the tumor area and degree of vascular compression and deformation were calculated to quantify the relationship between the tumor and the vessels.

[0076] Multivariate regression risk analysis was performed on clinical data, tumor area, and degree of vascular compression and deformation to identify independent risk factors, such as tumor area ≥250 mm at key CT scans. 2 (HR: 2.777, 95% CI: 0.983-8.137) and adenocarcinoma type (HR: 2.334, 95% CI: 0.921-5.914);

[0077] Based on independent risk factors, a prediction model is constructed; preferably, the prediction model adopts the nomogram model.

[0078] The patient's independent risk factors are collected and input into the prediction model to generate a risk score.

[0079] In a preferred embodiment of the present invention, a method for performing three-dimensional reconstruction of a CT image and selecting key coronal slices is as follows:

[0080] A four-stage encoder-decoder framework is used to segment the 3D tumor mask of CT images.

[0081] The 3D tumor mask is fed into a lightweight post-processing module (which can use existing common modules to calculate the coronal / axial / sagittal area of the 3D tumor mask through multi-plane slicing, layer-by-layer pixel statistics, and physical size calibration, with the characteristics of no parameters and high efficiency). The tumor area is calculated layer by layer in the coronal / axial / sagittal planes.

[0082] The slice index with the largest recorded area is the key fault.

[0083] The core process of the lightweight post-processing module is divided into three steps:

[0084] 1. Multi-plane slicing: Automatically slice the 3D tumor mask along three orthogonal planes: coronal, sagittal, and axial planes to generate a series of 2D images.

[0085] 2. Area calculation: Pixel statistics are performed on the tumor area of each slice, and combined with the physical resolution of the CT image (such as 0.5mm / pixel), the actual tumor area at each level is calculated.

[0086] 3. Key level determination: Compare the area calculation results of all slices and automatically mark the layer with the largest tumor area as the key diagnostic section.

[0087] In a preferred embodiment of the present invention,

[0088] Using a four-stage encoder-decoder framework, the method for segmenting the 3D tumor mask of CT images is as follows:

[0089] In the encoder, QuadVSS Block is used to replace standard convolution, and standard VSS Block is stacked to build a multi-layer visual state modeling unit to capture global semantics;

[0090] The spatial size is halved between each stage to gradually extract deep features;

[0091] Set a large kernel (such as a 3x3 convolution kernel to enhance spatial context modeling capabilities) and attention gate LGAG to perform local channel attention modeling;

[0092] The frequency fusion module FreqFusion combines DCT / FFT frequency domain transform with low-frequency residual guidance to guide the model to focus on strong structural boundaries and texture details, enhancing the fusion expression between multi-scale features. Specifically:

[0093] The FreqFusion module decomposes features into low-frequency structure and high-frequency detail components through frequency domain transforms (DCT / FFT). It uses the low-frequency residuals to generate attention weights to enhance the subject's outline, while preserving texture details through high-frequency pathways. Finally, it fuses multi-frequency features using adaptive weighting. Specifically, low-frequency components use lightweight convolution to generate spatial attention maps to enhance boundary areas, while high-frequency components undergo dynamic convolution to highlight subtle features. Combined with a multi-scale frequency domain alignment strategy, the model synergistically optimizes structural integrity and texture refinement, significantly improving segmentation accuracy.

[0094] Based on skip connections and multi-level fusion, the context consistency of the upsampling path is enhanced through residual weighted connections of different scales. Specifically:

[0095] Through multi-scale skip connections between the encoder and decoder, residual convolution processing (Conv+Skip) is performed on the encoder features and decoder features of each level respectively, and learnable dynamic attention weights are introduced to achieve adaptive weighted fusion of cross-layer features; at the same time, through pyramid-style cross-level interaction (such as weighted addition of upsampled deep features and shallow features), a multi-level residual fusion path from local details to global semantics is formed, and finally multi-scale context is aggregated through pyramid pooling, significantly improving the coherence of feature expression and boundary accuracy.

[0096] A deep supervision mechanism is introduced into the decoder. The decoder uses two consecutive VSS Block decoding features, combined with layer-by-layer 1×1 convolution and deep supervision mechanism to improve training stability and gradient transfer. The final output of the decoder is the tumor segmentation mask, and it also supports the attachment of the "maximum section extraction module."

[0097] In a preferred embodiment of the present invention, a risk threshold is set to determine the patient's risk level:

[0098] Low risk level: if the risk score is ≤30 points;

[0099] Medium risk level: if the risk score is 30-60 points;

[0100] High risk level: If the risk score is ≥60 points.

[0101] Using risk scores to classify patients into risk levels helps determine the patient's physical condition.

[0102] In a preferred embodiment of the present invention, a chest CT image of a patient is collected and checked to see whether the chest CT image is qualified, specifically:

[0103] S1, check whether the DICOM file of the chest CT image is complete. If so, proceed to step S2; otherwise, determine that the chest CT image is unqualified, reacquire the chest CT image and execute step S1;

[0104] S2, verify whether the scanning range covers the entire lung. If so, proceed to step S3; otherwise, determine that the chest CT image is unqualified, reacquire the chest CT image and execute step S1;

[0105] S3, calculating the artifact area in the chest CT image. If the artifact area is less than 5%, proceed to step S4; otherwise, the chest CT image is judged to be unqualified, and the chest CT image is re-acquired and step S1 is executed;

[0106] S4, compare the chest CT image with the preset qualified image, calculate the similarity based on body position and key anatomical structures, and if the similarity meets the preset value, output the chest CT image for 3D reconstruction; otherwise, judge the chest CT image as unqualified, reacquire the chest CT image and execute step S1.

[0107] Check whether the chest CT image is qualified, obtain high-quality image information, extract more accurate body feature information, and obtain more accurate prediction results.

[0108] In a preferred embodiment of the present invention, similarity is calculated based on body position and key anatomical structures, specifically:

[0109] Let the posture similarity be S p , the anatomical structure similarity is S a , then the comprehensive similarity S t for:

[0110] S t =w1S p +w2S a

[0111] Wherein, w1 and w2 are weighting coefficients, satisfying w1+w2=1;

[0112] Mutual information MI is used to measure body position similarity. Let R be the rotation matrix, which represents the rotation transformation from chest CT image I1 to chest CT image I2; T is the translation vector, which represents the displacement from chest CT image I1 to chest CT image I2; the image similarity after rotation and translation correction is:

[0113] B=MI(I1,I2)

[0114] Among them, MI(I1, I2) represents the mutual information, which is used to evaluate the registration quality of two images:

[0115]

[0116] Where p(I1, I2) is the joint probability distribution, which represents the joint occurrence probability of pixel pairs in chest CT images I1 and I2; p(I1) and p(I2) are the marginal probability distributions of chest CT images I1 and I2, respectively;

[0117] Use deep learning algorithms to segment chest CT images, extract the lung regions L1 and L2, and the heart regions C1 and C2 from the chest CT images, and use the structural similarity index to calculate the similarity between the two regions R1 and R2:

[0118]

[0119] Where μ1 and μ2 are the means of regions R1 and R2, σ1 and σ2 are the variances, and σ 1,2 is the covariance of regions R1 and R2, and c1 and c2 are constants used to prevent the denominator from being zero.

[0120] The present invention also provides a central NSCLC thoracotomy risk prediction system. Figure 1 As shown, it includes a bed 1, a human-computer interaction module 2, a CTCT image acquisition module 3 and a processing module.

[0121] The human-computer interaction module 2 and the CTCT image acquisition module 3 are both installed on the bed 1. The human-computer interaction module 2 uses a computer, smart touch screen, etc. The patient lies flat on the bed 1, and the human-computer interaction module 2 is used to collect clinical data of the sample.

[0122] The CTCT image acquisition module 3 is used to acquire chest CT images of the sample and transmit them to the processing module. The processing module executes the method of the present invention to complete the risk prediction of thoracotomy.

[0123] The system uses the human-computer interaction module 2 to input the corresponding clinical data. The CTCT image acquisition module 3 collects the chest CT images of the sample and inputs them into the processing module. The processing module analyzes the data, effectively predicts the risk of thoracotomy, and outputs the prediction results.

[0124] In a preferred embodiment of the present invention, Figure 2 As shown, the central NSCLC thoracotomy risk prediction system also includes an auxiliary mechanism, which includes a conveyor belt 5, an air bag 4, an elastic restraint belt, and a suction cup fixing part.

[0125] A groove is provided from the chest examination position to the foot of the bed 1 , and the conveyor belt 5 is arranged in the groove, and the upper surface of the conveyor belt 5 is flush with the upper surface of the bed 1 .

[0126] The airbag 4 is arranged at the chest examination position of the bed 1, and the CTCT image acquisition module 3 is located above the chest examination position of the bed 1. The airbag 4 is made of elastic deformable material. The airbag 4 is connected to an inflation and deflation mechanism, and the inflation and deflation mechanism is installed on the bed 1. The airbag 4 is initially in an expanded state.

[0127] Both ends of the elastic strap are connected with suction cup type fixing parts, which can be adsorbed and positioned with the upper surface of the bed 1. During the examination, the patient's arm is placed between the elastic strap and the upper surface of the bed 1.

[0128] The conveyor belt 5 is set in the groove. When the patient sits on the conveyor belt 5 and then lies down on the bed 1, if the patient's chest does not correspond to the position of the CTCT image acquisition module 3 above the bed 1, the conveyor belt 5 can be started to move forward or reverse, driving the patient to move a certain distance to the head or foot of the bed to assist the patient in moving his body.

[0129] The airbag 4 is set at the chest examination position of the bed 1. When the patient lies on the bed 1, the back corresponding to the chest presses on the airbag 4, which is convenient for positioning. At the same time, the airbag 4 has a certain convex structure, which can help the patient expand his chest and prevent the patient from holding his chest and affecting image acquisition.

[0130] The elastic restraint belt can be connected to the upper surface of the bed 1 through a suction cup type fixing piece. When the patient is examined, the patient's arm is fixed on the bed 1 to prevent the patient from moving accidentally and achieve positioning.

[0131] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0132] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A method for predicting the risk of central NSCLC thoracotomy, characterized by: The steps include: Collect clinical data and chest CT images of the samples; Perform three-dimensional reconstruction of CT images, select key coronary planes, and calculate the tumor area and degree of vascular compression and deformation; Multivariate regression risk analysis was performed on clinical data, tumor area, and degree of vascular compression and deformation to identify independent risk factors; Based on independent risk factors, a prediction model was constructed; The patient's independent risk factors are collected and input into the prediction model to generate a risk score.

2. The method for predicting the risk of central NSCLC thoracotomy according to claim 1, wherein: The method for selecting key coronal slices for 3D reconstruction of CT images is as follows: A four-stage encoder-decoder framework is used to segment the 3D tumor mask of CT images. The 3D tumor mask is fed into a lightweight post-processing module to calculate the tumor area layer by layer in the coronal, axial, and sagittal planes. The slice index with the largest recorded area is the key fault.

3. The method for predicting the risk of central NSCLC thoracotomy according to claim 2, wherein: Using a four-stage encoder-decoder framework, the method for segmenting the 3D tumor mask of CT images is as follows: In the encoder, QuadVSS Block is used to replace standard convolution, and standard VSS Block is stacked to build a multi-layer visual state modeling unit to capture global semantics; The spatial size is halved between each stage to gradually extract deep features; Set up large-core attention gate LGAG to model local channel attention; The frequency fusion module FreqFusion combines DCT / FFT frequency domain transform with low-frequency residual guidance to guide the model to focus on strong structural boundaries and texture details, enhancing the fusion expression between multi-scale features. Specifically: The FreqFusion module decomposes features into low-frequency structures and high-frequency detail components through frequency domain transformation, uses low-frequency residuals to generate attention weights to enhance the subject contour, preserves texture details through high-frequency pathways, and finally fuses multi-frequency features through adaptive weights; Based on skip connections and multi-level fusion, the context consistency of the upsampling path is enhanced through residual weighted connections of different scales. Specifically: Through multi-scale skip connections between the encoder and decoder, residual convolution is performed on the encoder and decoder features of each level respectively, and learnable dynamic attention weights are introduced to achieve adaptive weighted fusion of cross-layer features; Through pyramid-style cross-level interaction, a multi-level residual fusion path is formed from local details to global semantics, and finally multi-scale context is aggregated through pyramid pooling; A deep supervision mechanism is introduced into the decoder. The decoder uses two consecutive VSS Block decoding features, combined with layer-by-layer 1×1 convolution and deep supervision to improve training stability and gradient transfer. The final output of the decoder is the tumor segmentation mask, and it also supports the attachment of the "maximum section extraction module." 4. The method for predicting the risk of central NSCLC thoracotomy according to claim 1, wherein: The prediction model adopts the Nomogram model.

5. The method for predicting the risk of central NSCLC thoracotomy according to claim 1, wherein: Set risk thresholds to determine the patient's risk level: Low risk level: if the risk score is ≤30 points; Medium risk level: if the risk score is 30-60 points; High risk level: If the risk score is ≥60 points.

6. The method for predicting the risk of central NSCLC thoracotomy according to claim 1, wherein: Collect the patient's chest CT image and check whether the chest CT image is qualified, specifically: S1, check whether the DICOM file of the chest CT image is complete. If so, proceed to step S2; otherwise, determine that the chest CT image is unqualified, reacquire the chest CT image and execute step S1; S2, verify whether the scanning range covers the entire lung. If so, proceed to step S3; otherwise, determine that the chest CT image is unqualified, reacquire the chest CT image and execute step S1; S3, calculating the artifact area in the chest CT image. If the artifact area is less than 5%, proceed to step S4; otherwise, the chest CT image is judged to be unqualified, and the chest CT image is re-acquired and step S1 is executed; S4, compare the chest CT image with the preset qualified image, calculate the similarity based on body position and key anatomical structures, and if the similarity meets the preset value, output the chest CT image for 3D reconstruction; otherwise, judge the chest CT image as unqualified, reacquire the chest CT image and execute step S1.

7. The method for predicting the risk of central NSCLC thoracotomy according to claim 6, wherein: Similarity is calculated based on body position and key anatomical structures, specifically: Let the posture similarity be S p , the anatomical structure similarity is S a , then the comprehensive similarity S t for: S t =w1S p +w2s a Wherein, w1 and w2 are weighting coefficients, satisfying w1+w2=1; Mutual information MI is used to measure body position similarity. Let R be the rotation matrix, which represents the rotation transformation from chest CT image I1 to chest CT image I2; T is the translation vector, which represents the displacement from chest CT image I1 to chest CT image I2; the image similarity after rotation and translation correction is: B=MI(I1,I2) Among them, MI(I1, I2) represents the mutual information, which is used to evaluate the registration quality of two images: Where p(I1, I2) is the joint probability distribution, which represents the joint occurrence probability of pixel pairs in chest CT images I1 and I2; p(I1) and p(I2) are the marginal probability distributions of chest CT images I1 and I2, respectively; Use deep learning algorithms to segment chest CT images, extract the lung regions L1 and L2, and the heart regions C1 and C2 from the chest CT images, and use the structural similarity index to calculate the similarity between the two regions R1 and R2: Where μ1 and μ2 are the means of regions R1 and R2, σ1 and σ2 are the variances, and σ 1,2 is the covariance of regions R1 and R2, and c1 and c2 are constants used to prevent the denominator from being zero.

8. A central NSCLC thoracotomy risk prediction system, characterized by: It includes bed, human-computer interaction module, CT image acquisition module and processing module; The human-computer interaction module and the CT image acquisition module are both installed on the bed, the patient lies flat on the bed, and the human-computer interaction module is used to collect clinical data of the sample; The CT image acquisition module is used to acquire chest CT images of the sample and transmit them to the processing module; The processing module executes the method according to any one of claims 1 to 7 to complete the risk prediction of thoracotomy.

9. The central NSCLC thoracotomy risk prediction system according to claim 8, wherein: It also includes an auxiliary mechanism, which includes a conveyor belt, an air bag, an elastic strap, and a suction cup type fixing member; A groove is provided from the chest examination position to the foot of the bed, and the conveyor belt is arranged in the groove, and the upper surface of the conveyor belt is flush with the upper surface of the bed; The airbag is arranged at the chest examination position of the bed, and the CT image acquisition module is located above the chest examination position of the bed. The airbag is made of elastic deformable material and is connected to an inflation and deflation mechanism. The inflation and deflation mechanism is installed on the bed. The airbag is initially in an inflated state. Both ends of the elastic binding belt are connected with suction cup type fixing parts, which can be adsorbed and positioned with the upper surface of the bed. During examination, the patient's arm is placed between the elastic binding belt and the upper surface of the bed.

Images