AI-assisted diagnostic method and system for pleural invasion of lung cancer

Abstract

The present invention belongs to the field of pleural invasion diagnosis, and in particular relates to an AI-assisted auxiliary method and system for pleural invasion diagnosis of lung cancer, comprising: first, acquiring scanning images and electrode needle signals, and obtaining an enhanced training set through time synchronization and enhanced annotation; second, constructing a regional prediction model based on the YOLOv8 algorithm and the attention matching network, and training the model using the enhanced training set and the regional electrical signal conversion error function; third, inputting the real-time scanning image into the trained regional prediction model to obtain a suspected exploration area and an electrode needle exploration prediction signal sequence, then configuring a safe exploration boundary, combining the real-time electrode needle signal and the prediction signal to obtain the exploration signal error, inputting the exploration path correction model for training, and obtaining the optimal exploration path through risk assessment; finally, integrating the two models into the electrode needle, and assisting in correcting the electrode needle exploration path based on real-time data to obtain the exploration result of the suspected exploration area and the invasion exploration score.

Description

Technical Field

[0001] The present invention belongs to the field of pleural invasion diagnosis, and in particular relates to an AI-assisted method and system for diagnosing pleural invasion of lung cancer. Background Art

[0002] In the diagnosis and treatment of lung cancer, accurate assessment of pleural invasion is crucial for developing appropriate treatment plans. Traditionally, physicians rely on imaging studies (such as CT and MRI) and pathological biopsies to assess pleural invasion. However, these methods have numerous limitations. Currently, diagnosis relies solely on a single scan image or electrode needle signal, which provides incomplete information and can easily lead to misdiagnosis or missed diagnosis. Furthermore, existing technologies struggle to accurately locate and assess pleural invasion when dealing with the complex relationship between lung structure and lesions. Furthermore, manual interpretation of image and signal data is not only time-consuming and labor-intensive, but also impacts diagnostic accuracy due to individual experience differences.

[0003] For example, Chinese patent application CN116739992B discloses an intelligent assisted interpretation method for thyroid capsule invasion. This method primarily utilizes a biomedical image segmentation method that provides efficient local-global fusion and a parallel Transformer and CNN network. This existing technology, by inputting a thyroid ultrasound image, outputs a localized and segmented image showing suspected thyroid nodules, automatically obtaining a preliminary interpretation of the thyroid nodule.

[0004] For example, Chinese patent application publication number CN115700761A discloses a method, device, and electronic device for predicting microvascular invasion of hepatocellular carcinoma. The method includes: obtaining preoperative imaging images, clinical indicators, and corresponding lesion pathology results of treated patients; training a pre-constructed prediction model based on the preoperative imaging images, clinical indicators, and corresponding lesion pathology results of treated patients, and determining that the prediction model is a microvascular invasion prediction model when the prediction model meets preset standards; obtaining preoperative imaging images and clinical indicators of the patient to be predicted; inputting the preoperative imaging images and clinical indicators of the patient to be predicted into the microvascular invasion prediction model, and outputting a classification result to predict whether the patient to be tested has microvascular invasion of hepatocellular carcinoma.

[0005] The above existing technologies have the following problems: the existing technologies mostly use a single diagnosis and treatment method for invasion diagnosis and treatment, especially electrode needle exploration, which often leads to large errors in diagnosis due to deviations in the exploration path. For this reason, the present invention provides an AI-assisted lung cancer pleural invasion diagnosis auxiliary method and system. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention proposes an AI-assisted auxiliary method and system for the diagnosis of pleural invasion of lung cancer, including: first, obtaining scanning images and electrode needle signals, and obtaining an enhanced training set through preprocessing, time synchronization and enhanced annotation; second, constructing a regional prediction model based on the YOLOv8 algorithm and the attention matching network, and training the model with the enhanced training set and the regional electrical signal conversion error function; third, inputting the real-time scanning image into the trained regional prediction model to obtain the suspected exploration area and the electrode needle exploration prediction signal sequence, and then configuring the safe exploration boundary, combining the real-time electrode needle signal and the prediction signal to obtain the exploration signal error, inputting the configured exploration path correction model for training, and obtaining the optimal exploration path through risk assessment; finally, integrating the regional prediction model and the exploration path correction model into the electrode needle, and assisting in correcting the electrode needle exploration path based on real-time data to obtain the exploration results and invasion exploration score of the suspected exploration area, thereby improving the accuracy and efficiency of the diagnosis of pleural invasion of lung cancer.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] AI-assisted diagnostic methods for pleural invasion of lung cancer include:

[0009] S1. Obtain scanned images and electrode needle signals for preprocessing, synchronize the preprocessed data in time, and enhance and annotate the scanned images and electrode needle signals to obtain an enhanced training set;

[0010] S2. Build a regional prediction model based on the YOLOv8 algorithm and the attention matching network. Input the enhanced training set into the regional prediction model and train it using the built-in regional electrical signal conversion error function to obtain a trained regional prediction model.

[0011] S3, inputting the acquired scan image into the regional prediction model to obtain a suspected detection area and a corresponding electrode needle detection prediction signal sequence;

[0012] S4. Configure a dynamic safe exploration boundary, use the electrode needle signal in the same time dimension as the scan image in S3 and the corresponding electrode needle exploration prediction signal sequence to obtain the exploration signal error, and input the suspected exploration area, safe exploration boundary, and exploration signal error of the scan image into the configured exploration path correction model for training, and obtain simulated multiple exploration paths, and perform exploration risk assessment on the simulated multiple exploration paths to obtain the optimal exploration path;

[0013] S5. Integrate the trained regional prediction model and exploration path correction model into the electrode needle, and perform auxiliary correction on the electrode needle exploration path based on the real-time acquired scanning image and electrode needle signal to obtain the exploration results of the suspected exploration area and the corresponding invasion exploration score.

[0014] Specifically, the steps for obtaining the enhanced training set include:

[0015] S101, preprocessing the acquired scanned image using a grayscale algorithm and a filtering denoising algorithm, while removing low-frequency baseline drift components in the electrode needle signal using a high-pass filter to obtain a preprocessed scanned image and electrode needle signal;

[0016] S102: Based on the pre-processed scanned image, mark the violated areas and potential violated areas using an expert experience method, and mark the levels of the violated areas and potential violated areas to obtain a marked scanned image;

[0017] S103, based on the corresponding marked area in the marked scanned image, using an association algorithm to screen and obtain the electrode signal sequence and corresponding correlation degree corresponding to the same position area from the pre-processed electrode needle signals;

[0018] S104 , pairing the electrode signal sequences corresponding to the same position area with the same position area of the scanned image through the correlation degree to obtain an enhanced training set.

[0019] Specifically, the region prediction model includes an improved backbone layer, an improved bottleneck layer, and a detection layer. The improved backbone layer includes the original first low-dimensional feature extraction sublayer, the second low-dimensional feature extraction sublayer, and the improved third to ninth high-dimensional feature extraction sublayers of the MobileViT network. The improved bottleneck layer includes the original bottleneck sublayer and the ECA channel attention sublayer. The construction and training steps of the region prediction model include:

[0020] S201, inputting the enhanced training set into the improved backbone layer, performing feature extraction on the scanned image in the enhanced training set through the first low-dimensional feature extraction sublayer and the second low-dimensional feature extraction sublayer to obtain initial image extraction features, and simultaneously inputting the electrode signal sequence corresponding to the current input scanned image in the enhanced training set into the one-dimensional convolution sublayer and the dimensionality adjustment sublayer to obtain initial electrode signal features of the same dimension as the initial image extraction features;

[0021] S202, cascade-fusing the initial electrode signal features and the initial image extraction features and inputting them into the third to ninth high-dimensional feature extraction sublayers connected in series to obtain a first high-dimensional image-signal extraction feature;

[0022] S203: Input the first high-dimensional image-signal extraction feature into the original bottleneck sublayer to perform upsampling feature extraction, and fuse the features extracted from the corresponding sub-blocks in the original bottleneck sublayer with the high-dimensional image-signal extraction features obtained from the corresponding sub-layer in the improved backbone layer through the improved fusion path between the backbone layer and the bottleneck layer in the original YOLOv8 to obtain a first fused image-signal extraction feature.

[0023] The improved fusion path is achieved by adding a convolution kernel of size 2 in each fusion path between the backbone layer and the bottleneck layer in the original YOLOv8 network. The two-dimensional convolutional network is constructed.

[0024] Specifically, the steps of building and training the regional prediction model also include:

[0025] S204: Input the first fused image-signal extracted features into the ECA channel attention sublayer, and perform enhanced correlation fusion based on the correlation between the electrode signals and the corresponding regions in the corresponding scanned images to obtain enhanced fusion features;

[0026] S205. Input the enhanced fusion features into the regional prediction sublayer, regional assessment sublayer, and regional electrical signal conversion sublayer in the detection layer, respectively, to obtain the predicted infringed region anchor frame and anchor frame error, regional infringed assessment level and level error assessment probability, and the corresponding predicted anchor frame region electrode signal and regional electrode signal prediction error.

[0027] S206, constructing a regional electrical signal conversion error function using the regional electrode signal prediction error, the anchor frame error, and the level error evaluation probability, and using the regional electrical signal conversion error function for training, while obtaining the loss of each round of training;

[0028] S207: Set a loss threshold and an early stopping value. When the loss of each round of training is less than the loss threshold within a continuous early stopping value range, a trained regional prediction model is obtained.

[0029] Specifically, the steps for constructing the dynamic security exploration boundary include:

[0030] S401, determining an initial safe exploration boundary interval based on the exploration area, and obtaining structural parameters and respiratory parameters corresponding to each exploration sub-area in the exploration area corresponding to different ages;

[0031] S402. Different age groups are divided by expert experience, and a safe exploration boundary adjustment factor is constructed using the structural parameters and respiratory parameters corresponding to each exploration sub-region under the exploration region corresponding to the divided age groups;

[0032] S403: Using the safety exploration boundary adjustment factor and the initial safety exploration boundary interval, obtain dynamic safety exploration boundaries corresponding to different age groups.

[0033] Specifically, the training steps of the exploration path correction model include:

[0034] S411, build an exploration path correction model based on the path algorithm, input the dynamic safety exploration boundary and the infringing area anchor frame predicted by the scanned image into the exploration path correction model, and generate the dynamic safety exploration boundary.N exploration paths and their corresponding electrode signals and the safety assessment score corresponding to each exploration path;

[0035] S412: Set a safety assessment score threshold and a path error threshold, and construct a path error for the exploration area using the electrode signals corresponding to the exploration path and the electrode signals of the anchor frame area predicted by the regional prediction model;

[0036] S413. The exploration path correction model is trained using the safety assessment score threshold, the path error threshold, the safety assessment score corresponding to each exploration path, and the exploration area path error. When all simulated exploration paths satisfy the condition that the exploration area path error is less than the path error threshold and the safety assessment score corresponding to each exploration path is greater than the safety assessment score threshold, a trained exploration path correction model is obtained.

[0037] Specifically, the steps for obtaining the optimal exploration path include:

[0038] S414: Obtain a dynamic safety exploration boundary of the exploration object based on the basic information and physiological status information corresponding to the exploration object;

[0039] S415: Input the dynamic safe exploration boundary of the exploration object, the predicted suspected exploration area of the exploration object, and the corresponding electrode needle exploration prediction signal sequence into the trained exploration path correction model, and obtain the exploration path that meets the threshold conditions of S413 and has the maximum safety assessment score and the minimum exploration area path error as the optimal exploration path;

[0040] S416: Feedback the parameters corresponding to the optimal exploration path to the electrode needle, and correct the real-time exploration path of the electrode needle at the current moment.

[0041] The AI-assisted lung cancer pleural invasion diagnosis auxiliary system includes: data processing module, regional prediction module, exploration path correction module and integrated output module;

[0042] The data processing module includes a data synchronization acquisition unit and a labeling unit;

[0043] The data synchronization acquisition unit is used to synchronously acquire the scan image of lung cancer pleural invasion and the electrode needle signal that matches the scan image area; the labeling unit is used to synchronously label the acquired scan image and electrode needle signal using expert experience to obtain an enhanced training set;

[0044] The regional prediction module includes a model training unit and a prediction unit;

[0045] The model training unit is used to build a regional prediction model based on the YOLOv8 algorithm and the attention matching network. The enhanced training set is input into the regional prediction model and trained using the built-in regional electrical signal conversion error function to obtain a trained regional prediction model.

[0046] The prediction unit is used to input the acquired scan image into the regional prediction model to obtain a suspected detection area and a corresponding electrode needle detection prediction signal sequence.

[0047] Specifically, the exploration path correction module includes an error acquisition unit, a correction model training unit, and an optimal path unit;

[0048] The error acquisition unit uses the electrode needle signal in the same time dimension as the real-time scanning image and the corresponding electrode needle exploration prediction signal sequence to obtain the exploration signal error in the same time dimension;

[0049] The correction model training unit is used to input the suspected detection area, safe detection boundary, and detection signal error of the scanned image into the detection path correction model for training, thereby obtaining a trained detection path correction model;

[0050] The optimal path unit is used to correct the model based on the trained exploration path. It simulates and generates multiple exploration paths by scanning the suspected exploration area and safe exploration boundary of the image, and performs exploration risk assessment on the simulated multiple exploration paths to obtain the optimal exploration path.

[0051] A computer-readable storage medium stores computer instructions, which, when executed, execute an AI-assisted method for diagnosing pleural invasion of lung cancer.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] To address the shortcomings of the existing technology, the present invention constructs a regional prediction model by combining YOLOv8 with an attention matching network, inputs an enhanced training set into the model for training, and optimizes model performance using a built-in regional electrical signal conversion error function. This enables the model to more accurately identify suspected exploration areas and reduce misjudgments. Secondly, by configuring a safe exploration boundary and combining real-time electrode needle signals with predicted signal sequences to obtain exploration signal errors, accurate correction and risk assessment of the exploration path are achieved, effectively solving the diagnostic error problem caused by path deviation in the existing technology, generating an optimal exploration path, and greatly improving the safety and accuracy of the operation. Finally, the exploration path is dynamically adjusted based on the real-time scan image and electrode needle signals, providing the exploration results of the suspected exploration area and the corresponding invasion exploration score. This real-time feedback mechanism not only helps doctors make more accurate clinical decisions, but also reduces the risks that may arise during the operation. In addition, in the process of generating the exploration path, the present invention also obtains corresponding dynamic safe exploration boundaries based on the physiological state of different subjects, further ensuring the safety of the explored subjects during the exploration process. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of an AI-assisted method for diagnosing pleural invasion of lung cancer according to Example 1 of the present invention;

[0055] Figure 2 This is a structural diagram of the regional prediction model in Example 1 of the present invention;

[0056] Figure 3 This is a module diagram of an AI-assisted lung cancer pleural invasion diagnosis assistance system according to Example 2 of the present invention. DETAILED DESCRIPTION

[0057] Example 1

[0058] See also Figure 1 The present invention provides an embodiment of an AI-assisted method for diagnosing pleural invasion of lung cancer, comprising the following steps:

[0059] S1. Obtaining scanned images and electrode needle signals for preprocessing, synchronizing the preprocessed data in time, and performing enhanced annotation on the scanned images and electrode needle signals to obtain an enhanced training set. Furthermore, in this embodiment, the scanned images are CT images or MRI images obtained during the diagnosis of pleural invasion of lung cancer.

[0060] Furthermore, in this embodiment, the step of acquiring the enhanced training set includes:

[0061] S101, preprocessing the acquired scanned image using a grayscale algorithm and a filtering denoising algorithm, while removing low-frequency baseline drift components in the electrode needle signal using a high-pass filter to obtain a preprocessed scanned image and electrode needle signal;

[0062] S102: Based on the pre-processed scanned image, mark the violated areas and potential violated areas using an expert experience method, and mark the levels of the violated areas and potential violated areas to obtain a marked scanned image;

[0063] In this embodiment, the location, size, and morphology of the lung cancer tumor in the scanned image are first annotated. This process is achieved by manually outlining the tumor boundary or using existing medical image segmentation algorithms for automatic or semi-automatic segmentation, followed by manual correction. Secondly, the location and morphology of the pleura are annotated, especially the part of the pleura adjacent to the tumor or potentially invaded. Thirdly, based on medical knowledge and clinical experience, potential areas of lung cancer pleural invasion are marked, such as the contact interface between the tumor and the pleura, and areas of pleural thickening or nodular changes.

[0064] S103, based on the corresponding marked area in the marked scanned image, using an association algorithm to screen and obtain the electrode signal sequence and corresponding correlation degree corresponding to the same position area from the pre-processed electrode needle signals;

[0065] Exemplarily, during the acquisition process of the scanning image and the electrode needle signal, the acquisition timestamp of each data point is recorded, and the scanning image and the electrode needle signal within the same time point or time interval are matched by timestamp matching; for example, assuming that the scanning image is acquired at a speed of 10 frames per second and the electrode needle signal is acquired at a speed of 1000 data points per second, the image frame and electrode needle signal data segment corresponding to a certain moment are found according to the timestamp; if the acquisition frequencies of the scanning image and the electrode needle signal are inconsistent, the scanning image and the electrode needle signal are matched by interpolation.

[0066] S104 , pairing the electrode signal sequences corresponding to the same position area with the same position area of the scanned image through the correlation degree to obtain an enhanced training set.

[0067] In this embodiment, the labels also include information such as whether the tumor has invaded the pleura (yes / no), the degree of invasion (mild, moderate, severe, etc.), and the diagnosis-related feature information in the scan image and electrode needle signal is used as auxiliary labels.

[0068] The process first preprocesses the scanned image using a grayscale algorithm and a filtering denoising algorithm to ensure image clarity and consistency; at the same time, a high-pass filter is used to remove the low-frequency baseline drift component in the electrode needle signal, which improves the signal purity while effectively reducing noise interference and enhancing spatiotemporal synchronization; secondly, the expert experience method is used to annotate the invaded areas and potential invaded areas, and their levels are marked, generating a detailed annotated scan image, ensuring the accuracy and authority of the annotation, making the model training closer to the actual clinical situation; thirdly, the association algorithm is used to screen out the electrode signal sequence corresponding to the annotated area and its correlation degree from the preprocessed electrode needle signal. This precise matching of multimodal data not only enriches the information content of the training samples, but also ensures the consistency and correlation between different modal data, providing more comprehensive data support for model learning; finally, the electrode signal sequence in the same position area is associated and paired with the scanned image through the correlation degree to construct an enhanced training set, achieving a high degree of fusion of multimodal data, and significantly improving the accuracy and generalization ability of the model in identifying lesion areas.

[0069] S2. Build a regional prediction model based on the YOLOv8 algorithm and the attention matching network. Input the enhanced training set into the regional prediction model and train it using the built-in regional electrical signal conversion error function to obtain a trained regional prediction model.

[0070] Furthermore, in this embodiment, the regional prediction model includes an improved backbone layer and an improved bottleneck layer and detection layer; the improved backbone layer includes the original first low-dimensional feature extraction sublayer, the second low-dimensional feature extraction sublayer and the improved third to ninth high-dimensional feature extraction sublayers of the MobileViT network, and the improved bottleneck layer includes the original bottleneck sublayer and the ECA channel attention sublayer; further, in this embodiment, the MobileViT network enhances the network's ability to combine local and global information by integrating the advantages of CNN and ViT.

[0071] Furthermore, in this embodiment, the first low-dimensional feature extraction sublayer and the second low-dimensional feature extraction sublayer are original sublayers in the backbone layer of the original YOLOv8 algorithm, and are constructed through the Darknet-53 network.

[0072] See also Figure 2 Furthermore, in this embodiment, the steps of constructing and training the regional prediction model include:

[0073] S201, inputting the enhanced training set into the improved backbone layer, performing feature extraction on the scanned image in the enhanced training set through the first low-dimensional feature extraction sublayer and the second low-dimensional feature extraction sublayer to obtain initial image extraction features, and simultaneously inputting the electrode signal sequence corresponding to the current input scanned image in the enhanced training set into the one-dimensional convolution sublayer and the dimensionality adjustment sublayer to obtain initial electrode signal features of the same dimension as the initial image extraction features;

[0074] S202, cascade-fusing the initial electrode signal features and the initial image extraction features and inputting them into the third to ninth high-dimensional feature extraction sublayers connected in series to obtain a first high-dimensional image-signal extraction feature;

[0075] S203: Input the first high-dimensional image-signal extraction feature into the original bottleneck sublayer to perform upsampling feature extraction, and fuse the features extracted from the corresponding sub-blocks in the original bottleneck sublayer with the high-dimensional image-signal extraction features obtained from the corresponding sub-layer in the improved backbone layer through the improved fusion path between the backbone layer and the bottleneck layer in the original YOLOv8 to obtain a first fused image-signal extraction feature.

[0076] Furthermore, the improved fusion path in this embodiment is achieved by adding a convolution kernel size of 1 in each fusion path between the backbone layer and the bottleneck layer in the original YOLOv8 network. The improved fusion path includes the first fusion path, the second fusion path and the third fusion path, and the corresponding structure is the same as the fusion path structure between the backbone layer and the bottleneck layer in the original YOLOv8 network.

[0077] S204: Input the first fused image-signal extracted features into the ECA channel attention sublayer, and perform enhanced correlation fusion based on the correlation between the electrode signals and the corresponding regions in the corresponding scanned images to obtain enhanced fusion features;

[0078] S205. Input the enhanced fusion features into the regional prediction sublayer, regional assessment sublayer, and regional electrical signal conversion sublayer in the detection layer, respectively, to obtain the predicted infringed region anchor frame and anchor frame error, regional infringed assessment level and level error assessment probability, and the corresponding predicted anchor frame region electrode signal and regional electrode signal prediction error.

[0079] Furthermore, in this embodiment, the regional prediction sublayer is constructed by the YOLOv8 original prediction head; the regional evaluation sublayer is constructed by the comprehensive fuzzy evaluation algorithm; and the regional electrical signal conversion sublayer is constructed by the fully connected network.

[0080] S206, constructing a regional electrical signal conversion error function using the regional electrode signal prediction error, the anchor frame error, and the level error evaluation probability, and using the regional electrical signal conversion error function for training, while obtaining the loss of each round of training;

[0081] S207: Set a loss threshold and an early stopping value. When the loss of each round of training is less than the loss threshold within the continuous early stopping value range, a trained regional prediction model is obtained. Furthermore, in this embodiment, the early stopping value is set to 10.

[0082] During the training process, the improved backbone layer performs preliminary feature extraction on the scanned image through the first low-dimensional feature extraction sublayer and the second low-dimensional feature extraction sublayer, and combines the one-dimensional convolution sublayer and the dimension adjustment sublayer to process the electrode signal sequence, and obtains the initial electrode signal features with the same dimension as the scanned image features, ensuring the effective fusion of multimodal data in the early stage; secondly, the initial electrode signal features and the initial image extraction features are cascaded and input into the high-dimensional feature extraction sublayer, and the ECA channel attention mechanism is introduced to optimize the bottleneck layer, realizing the efficient extraction and fusion of deep-level features. The improved fusion path is achieved by adding a convolution kernel size of The two-dimensional convolutional network not only enhances the correlation between features, but also reduces the number of model parameters, realizes cross-channel information integration and increases nonlinear characteristics, effectively integrates multi-scale information, and further improves the model's ability to identify complex lesion areas. Third, the ECA channel attention sublayer enhances association fusion, allowing the model to more accurately capture the relationship between electrode signals and corresponding areas in the scanned image, generating enhanced fusion features. The ECA channel attention mechanism in the bottleneck layer also enhances the capture of long-range dependencies, improving the network's feature expression and fusion capabilities. Subsequently, the regional prediction sublayer, regional evaluation sublayer, and regional electrical signal conversion sublayer in the detection layer are responsible for intrusion region anchor box prediction, evaluation level prediction, and electrode signal prediction, respectively, providing comprehensive diagnostic information. Finally, the regional electrical signal conversion error function is constructed using the regional electrical signal prediction error, anchor box error, and level error evaluation probability, and is trained by setting a loss threshold and early stopping value. This not only ensures that the model converges to the optimal solution, but also improves the model's generalization ability, effectively avoids overfitting problems, and ensures the stability and reliability of the model in practical applications.

[0083] S3, inputting the acquired scan image into the regional prediction model to obtain a suspected detection area and a corresponding electrode needle detection prediction signal sequence;

[0084] S4. Configure a dynamic safe exploration boundary, use the electrode needle signal in the same time dimension as the scan image in S3 and the corresponding electrode needle exploration prediction signal sequence to obtain the exploration signal error, and input the suspected exploration area, safe exploration boundary, and exploration signal error of the scan image into the configured exploration path correction model for training, and obtain simulated multiple exploration paths, and perform exploration risk assessment on the simulated multiple exploration paths to obtain the optimal exploration path;

[0085] Furthermore, in this embodiment, the steps of constructing the dynamic security exploration boundary include:

[0086] S401, determining an initial safe exploration boundary interval based on the exploration area, and obtaining structural parameters and respiratory parameters corresponding to each exploration sub-area in the exploration area corresponding to different ages;

[0087] Furthermore, the initial safety exploration boundary interval in this embodiment is obtained by combining the safety exploration boundary and exploration area of historical cases through an expert experience algorithm.

[0088] Furthermore, different exploration areas (such as the area of the lung used to diagnose pleural invasion of lung cancer) have specific structural and functional characteristics. For example, the lungs have specific blood vessels, trachea, and nerve distributions. The location and function of these structures determine that the electrode needle cannot enter certain areas at will during exploration, otherwise serious damage may be caused. By understanding this knowledge, a preliminary and generally applicable safe exploration range can be determined, that is, the initial safe exploration boundary interval. At the same time, the physiological state of the human body at different age stages is different. The structural parameters (such as tissue elasticity, thickness, etc.) and respiratory parameters (such as respiratory frequency, amplitude, etc.) corresponding to each exploration sub-area in the corresponding exploration area of different ages are obtained. These parameters reflect the physiological differences of human bodies of different age groups in this exploration area, providing data support for the subsequent construction of regulatory factors.

[0089] S402. Different age groups are divided by expert experience, and a safe exploration boundary adjustment factor is constructed using the structural parameters and respiratory parameters corresponding to each exploration sub-region under the exploration region corresponding to the divided age groups;

[0090] Furthermore, there are significant differences in physiological functions, tissue structures, etc. between people of different age groups. These differences will affect the setting of the safe exploration boundary. By using the structural parameters and respiratory parameters of each exploration sub-area under the exploration area corresponding to the divided age groups, a safe exploration boundary adjustment factor is constructed through a certain mathematical model or empirical formula; this adjustment factor can quantify the degree of influence of physiological differences in different age groups on the safe exploration boundary. For example, aging may lead to a decrease in tissue elasticity and changes in respiratory function, and the adjustment factor will adjust the safe exploration boundary accordingly based on these changes.

[0091] S403: Using the safety exploration boundary adjustment factor and the initial safety exploration boundary interval, obtain dynamic safety exploration boundaries corresponding to different age groups.

[0092] Furthermore, the constructed safety exploration boundary adjustment factor is applied to the initial safety exploration boundary interval to obtain the dynamic safety exploration boundary corresponding to different age groups; this dynamic adjustment takes into account the physiological characteristics of different age groups, making the safety exploration boundary more in line with the individual's actual situation and improving the safety and accuracy of the exploration.

[0093] For example, based on knowledge of lung anatomy, the locations of important structures such as the large blood vessels and trachea of the lungs are known. To prevent the electrode needle from damaging these structures, the initial safe exploration boundary interval is set to an area at least 5 mm away from the large blood vessels and trachea of the lungs. At the same time, the structural parameters of each exploration sub-area corresponding to this exploration area are obtained for different ages (such as children, adults, and the elderly) (such as children's lung tissue has better elasticity, with an elasticity coefficient of 0.8; the elderly's lung tissue has poor elasticity, with an elasticity coefficient of 0.4) and respiratory parameters (such as children's faster breathing rate, 25 times per minute; adults 15 times per minute; the elderly 18 times per minute).

[0094] Based on experience, medical experts divide age groups into children (0-12 years old), adults (13-60 years old), and the elderly (61 years old and above). For children, because their lung tissue is relatively tender and their respiratory rate is fast, a safe exploration boundary adjustment factor of 1.2 is constructed based on the structural parameters and respiratory parameters of their exploration sub-area (which means that the safety range is appropriately expanded based on the initial boundary). For the elderly, taking into account factors such as decreased elasticity of lung tissue and weakened respiratory function, an adjustment factor of 0.8 is constructed (which means that the safety range is appropriately narrowed, because the elderly's lungs have poor tolerance to damage and need to be operated with more caution).

[0095] For children, the initial safe exploration boundary interval is an area at least 5 mm away from the major pulmonary vessels and trachea. After applying an adjustment factor of 1.2, the dynamic safe exploration boundary becomes an area at least 6 mm (5×1.2 = 6) away from the major pulmonary vessels and trachea; for the elderly, after applying an adjustment factor of 0.8, the dynamic safe exploration boundary becomes an area at least 4 mm (5×0.8 = 4) away from the major pulmonary vessels and trachea.

[0096] Furthermore, in this embodiment, the training steps of the exploration path correction model include:

[0097] S411, build an exploration path correction model based on the path algorithm, input the dynamic safety exploration boundary and the infringing area anchor frame predicted by the scanned image into the exploration path correction model, and generate the dynamic safety exploration boundary. Nexploration paths and their corresponding electrode signals and the safety assessment score corresponding to each exploration path;

[0098] S412: Set a safety assessment score threshold and a path error threshold, and construct a path error for the exploration area using the electrode signals corresponding to the exploration path and the electrode signals of the anchor frame area predicted by the regional prediction model;

[0099] S413. The exploration path correction model is trained using the safety assessment score threshold, the path error threshold, the safety assessment score corresponding to each exploration path, and the exploration area path error. When all simulated exploration paths satisfy the condition that the exploration area path error is less than the path error threshold and the safety assessment score corresponding to each exploration path is greater than the safety assessment score threshold, a trained exploration path correction model is obtained.

[0100] Furthermore, in this embodiment, the step of obtaining the optimal exploration path includes:

[0101] S414: Obtain a dynamic safety exploration boundary of the exploration object based on the basic information and physiological status information corresponding to the exploration object;

[0102] S415: Input the dynamic safe exploration boundary of the exploration object, the predicted suspected exploration area of the exploration object, and the corresponding electrode needle exploration prediction signal sequence into the trained exploration path correction model, and obtain the exploration path that meets the threshold conditions of S413 and has the maximum safety assessment score and the minimum exploration area path error as the optimal exploration path;

[0103] S416: Feedback the parameters corresponding to the optimal exploration path to the electrode needle, and correct the real-time exploration path of the electrode needle at the current moment.

[0104] This process determines the initial safe exploration margin interval based on anatomical and physiological knowledge and obtains structural and respiratory parameters corresponding to different age groups, ensuring the scientific basis for setting the exploration margin. Secondly, expert experience is used to divide different age groups, and a safe exploration margin adjustment factor is constructed based on the structural and respiratory parameters of each age group. This process quantifies the impact of physiological differences across age groups on the exploration margin, making the safe margin more tailored to individual circumstances. For example, for children, whose tissues are more delicate and their respiratory rates are faster, the adjustment factor is set to 1.2, appropriately widening the safe margin. For the elderly, whose tissue elasticity is reduced and respiratory function is weakened, the adjustment factor is set to 0.8, narrowing the safe margin to increase operational caution. Thirdly, the adjustment factor is applied to the initial safe exploration margin interval to obtain a dynamic safe exploration margin. This dynamic adjustment takes into account individual physiological characteristics and improves exploration safety and accuracy. In this example, the dynamic safe exploration margin for children expands from 5mm to 6mm, while that for the elderly decreases from 5mm to 4mm, reflecting personalized safety considerations.

[0105] During the training process of the exploration path correction model, the exploration path can be optimized through path algorithms, threshold settings and error evaluation to ensure the performance of the training model and meet clinical safety and accuracy requirements. The optimal exploration path is obtained through the exploration path correction model and fed back to correct the real-time exploration path of the electrode needle, which can ensure that the exploration operation is more accurate and safe, and realize system optimization from data processing, model training to actual operation, which helps to improve the reliability and effectiveness of diagnosis and treatment. In addition, the process's systematic process and parameter adjustment mechanism fully consider individual differences, and can provide more targeted exploration services for different patients, thereby improving the quality and effect of the entire diagnostic process.

[0106] S5. Integrate the trained regional prediction model and exploration path correction model into the electrode needle, and perform auxiliary correction on the electrode needle exploration path based on the real-time acquired scanning image and electrode needle signal to obtain the exploration results of the suspected exploration area and the corresponding invasion exploration score.

[0107] Furthermore, in this embodiment, the detection result includes whether it is an invasion area, and an invasion detection score obtained through an evaluation algorithm according to specific parameters of the corresponding invasion area.

[0108] Example 2

[0109] See also Figure 3 Another embodiment provided by the present invention is an AI-assisted lung cancer pleural invasion diagnosis auxiliary system, comprising: a data processing module, a region prediction module, a detection path correction module, and an integrated output module;

[0110] Data processing module, used for acquisition and enhancement preprocessing of lung cancer pleural invasion scan images;

[0111] The data processing module includes a data synchronization acquisition unit and a labeling unit;

[0112] A data synchronization acquisition unit is used to synchronously acquire a scan image of lung cancer pleural invasion and an electrode needle signal that matches the scan image area. A labeling unit is used to synchronously label the acquired scan image and electrode needle signal using expert experience to obtain an enhanced training set. Furthermore, in this embodiment, the area corresponding to the scan image is paired with the electrode signal obtained by exploring the same area in the electrode needle signal.

[0113] The regional prediction module is used for model training and obtaining the suspected exploration area and the corresponding electrode needle exploration prediction signal sequence;

[0114] The regional prediction module includes a model training unit and a prediction unit;

[0115] The model training unit is used to build a regional prediction model based on the YOLOv8 algorithm and the attention matching network. The enhanced training set is input into the regional prediction model and trained using the built-in regional electrical signal conversion error function to obtain a trained regional prediction model.

[0116] A prediction unit, configured to input the acquired scanned image into a region prediction model to obtain a sequence of predicted detection signals of a suspected detection region and a corresponding electrode needle detection region;

[0117] The exploration path correction module is used to train the exploration path correction model and obtain the optimal exploration path;

[0118] The exploration path correction module includes an error acquisition unit, a correction model training unit and an optimal path unit;

[0119] The error acquisition unit uses the electrode needle signal in the same time dimension as the real-time scanning image and the corresponding electrode needle exploration prediction signal sequence to obtain the exploration signal error in the same time dimension;

[0120] The correction model training unit is used to input the suspected detection area, safe detection boundary, and detection signal error of the scanned image into the detection path correction model for training, thereby obtaining a trained detection path correction model;

[0121] The optimal path unit is used to modify the model based on the trained exploration path. It simulates and generates multiple exploration paths by scanning the suspected exploration area and safe exploration boundary of the image, and performs exploration risk assessment on the simulated multiple exploration paths to obtain the optimal exploration path.

[0122] The integrated output module is used to integrate the trained regional prediction model and exploration path correction model into the electrode needle, and to assist in correcting the electrode needle exploration path based on the real-time scan image and electrode needle signal to obtain the exploration results of the suspected exploration area and the corresponding invasion exploration score.

[0123] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the scope of protection of the purpose of the present invention and the claims, and all of these are protected by the present invention.

[0124] If the technical solution disclosed herein involves personal information, the product using the technical solution disclosed herein has clearly informed the individual of the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using the technical solution disclosed herein has obtained the individual's separate consent before processing the sensitive personal information and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, a clear and prominent sign is set to inform that the individual has entered the personal information collection scope and that personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that they agree to the collection of their personal information; or on the personal information processing device, when the personal information processing rules are notified by obvious signs / information, the individual's authorization is obtained through pop-up information or by asking the individual to upload their personal information. The personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. An AI-assisted method for diagnosing pleural invasion of lung cancer, characterized by: include: S1. Obtain scanned images and electrode needle signals for preprocessing, synchronize the preprocessed data in time, and enhance and annotate the scanned images and electrode needle signals to obtain an enhanced training set; S2. Build a regional prediction model based on the YOLOv8 algorithm and the attention matching network. Input the enhanced training set into the regional prediction model and train it using the built-in regional electrical signal conversion error function to obtain a trained regional prediction model. S3, inputting the acquired scan image into the regional prediction model to obtain a suspected detection area and a corresponding electrode needle detection prediction signal sequence; S4. Configure a dynamic safe exploration boundary, use the electrode needle signal in the same time dimension as the scan image in S3 and the corresponding electrode needle exploration prediction signal sequence to obtain the exploration signal error, and input the suspected exploration area, safe exploration boundary, and exploration signal error of the scan image into the configured exploration path correction model for training, and obtain simulated multiple exploration paths, and perform exploration risk assessment on the simulated multiple exploration paths to obtain the optimal exploration path; S5. Integrate the trained regional prediction model and exploration path correction model into the electrode needle, and perform auxiliary correction on the electrode needle exploration path based on the real-time acquired scanning image and electrode needle signal to obtain the exploration result of the suspected exploration area and the corresponding invasion exploration score; The training steps of the exploration path correction model include: S411, build an exploration path correction model based on the path algorithm, input the dynamic safety exploration boundary and the infringing area anchor frame predicted by the scanned image into the exploration path correction model, and generate the dynamic safety exploration boundary. N exploration paths and their corresponding electrode signals and the safety assessment score corresponding to each exploration path; S412: Set a safety assessment score threshold and a path error threshold, and construct a path error for the exploration area using the electrode signals corresponding to the exploration path and the electrode signals of the anchor frame area predicted by the regional prediction model; S413. Train the exploration path correction model using the safety assessment score threshold, the path error threshold, the safety assessment score corresponding to each exploration path, and the exploration area path error. When all simulated exploration paths satisfy the requirement that the exploration area path error is less than the path error threshold and the safety assessment score corresponding to each exploration path is greater than the safety assessment score threshold, obtain a trained exploration path correction model. The step of obtaining the optimal exploration path includes: S414: Obtain a dynamic safety exploration boundary of the exploration object based on the basic information and physiological status information corresponding to the exploration object; S415: Input the dynamic safe exploration boundary of the exploration object, the predicted suspected exploration area of the exploration object, and the corresponding electrode needle exploration prediction signal sequence into the trained exploration path correction model, and obtain the exploration path that meets the threshold conditions of S413 and has the maximum safety assessment score and the minimum exploration area path error as the optimal exploration path; S416: Feedback the parameters corresponding to the optimal exploration path to the electrode needle, and correct the real-time exploration path of the electrode needle at the current moment.

2. The AI-assisted method for diagnosing pleural invasion of lung cancer according to claim 1, characterized in that: The step of acquiring the enhanced training set includes: S101, preprocessing the acquired scanned image using a grayscale algorithm and a filtering denoising algorithm, while removing low-frequency baseline drift components in the electrode needle signal using a high-pass filter to obtain a preprocessed scanned image and electrode needle signal; S102: Based on the pre-processed scanned image, mark the violated areas and potential violated areas using an expert experience method, and mark the levels of the violated areas and potential violated areas to obtain a marked scanned image; S103, based on the corresponding marked area in the marked scanned image, using an association algorithm to screen and obtain the electrode signal sequence and corresponding correlation degree corresponding to the same position area from the pre-processed electrode needle signals; S104 , pairing the electrode signal sequences corresponding to the same position area with the same position area of the scanned image through the correlation degree to obtain an enhanced training set.

3. The AI-assisted method for diagnosing pleural invasion of lung cancer according to claim 2, wherein: The region prediction model includes an improved backbone layer, an improved bottleneck layer and a detection layer; the improved backbone layer includes the original first low-dimensional feature extraction sublayer, the second low-dimensional feature extraction sublayer and the third to ninth high-dimensional feature extraction sublayers improved from the MobileViT network, and the improved bottleneck layer includes the original bottleneck sublayer and the ECA channel attention sublayer; The steps of constructing and training the regional prediction model include: S201, inputting the enhanced training set into the improved backbone layer, performing feature extraction on the scanned image in the enhanced training set through the first low-dimensional feature extraction sublayer and the second low-dimensional feature extraction sublayer to obtain initial image extraction features, and simultaneously inputting the electrode signal sequence corresponding to the current input scanned image in the enhanced training set into the one-dimensional convolution sublayer and the dimensionality adjustment sublayer to obtain initial electrode signal features of the same dimension as the initial image extraction features; S202, cascade-fusing the initial electrode signal features and the initial image extraction features and inputting them into the third to ninth high-dimensional feature extraction sublayers connected in series to obtain a first high-dimensional image-signal extraction feature; S203: Input the first high-dimensional image-signal extraction feature into the original bottleneck sublayer to perform upsampling feature extraction, and fuse the features extracted from the corresponding sub-blocks in the original bottleneck sublayer with the high-dimensional image-signal extraction features obtained from the corresponding sub-layer in the improved backbone layer through the improved fusion path between the backbone layer and the bottleneck layer in the original YOLOv8 to obtain a first fused image-signal extraction feature. The improved fusion path is achieved by adding a convolution kernel of size 1 in each fusion path between the backbone layer and the bottleneck layer in the original YOLOv8 network. The two-dimensional convolutional network is constructed.

4. The AI-assisted method for diagnosing pleural invasion of lung cancer according to claim 3, wherein: The steps of constructing and training the regional prediction model also include: S204, inputting the first fused image-signal extracted feature into the ECA channel attention sublayer, performing enhanced correlation fusion based on the correlation between the electrode signal and the corresponding area in the corresponding scanned image, to obtain an enhanced fusion feature; S205, inputting the enhanced fusion features into the regional prediction sublayer, regional assessment sublayer, and regional electrical signal conversion sublayer of the detection layer, respectively, to obtain the predicted infringed region anchor frame and anchor frame error, regional infringed assessment level and level error assessment probability, and the corresponding predicted anchor frame region electrode signal and regional electrode signal prediction error; S206, constructing a regional electrical signal conversion error function using the regional electrode signal prediction error, the anchor frame error, and the level error evaluation probability, and using the regional electrical signal conversion error function for training, while obtaining the loss of each round of training; S207: Set a loss threshold and an early stopping value. When the loss of each round of training is less than the loss threshold within a continuous early stopping value range, a trained regional prediction model is obtained.

5. The AI-assisted method for diagnosing pleural invasion of lung cancer according to claim 4, characterized in that: The steps of constructing the dynamic security exploration boundary include: S401, determining an initial safe exploration boundary interval based on the exploration area, and obtaining structural parameters and respiratory parameters corresponding to each exploration sub-area in the exploration area corresponding to different ages; S402. Different age groups are divided by expert experience, and a safe exploration boundary adjustment factor is constructed using the structural parameters and respiratory parameters corresponding to each exploration sub-region under the exploration region corresponding to the divided age groups; S403: Using the safety exploration boundary adjustment factor and the initial safety exploration boundary interval, obtain dynamic safety exploration boundaries corresponding to different age groups.

6. An AI-assisted lung cancer pleural invasion diagnosis assistance system, which is used to implement the AI-assisted lung cancer pleural invasion diagnosis assistance method according to any one of claims 1 to 5, characterized in that: include: Data processing module, area prediction module, exploration path correction module and integrated output module; The data processing module includes a data synchronization acquisition unit and a marking unit; The data synchronous acquisition unit is used to synchronously acquire a scan image of lung cancer pleural invasion and an electrode needle signal that matches the scan image area; the labeling unit is used to synchronously label the acquired scan image and electrode needle signal accordingly using an expert experience method to obtain an enhanced training set; The regional prediction module includes a model training unit and a prediction unit; The model training unit is used to build a regional prediction model based on the YOLOv8 algorithm and the attention matching network, input the enhanced training set into the regional prediction model and use the built-in regional electrical signal conversion error function for training to obtain a trained regional prediction model; The prediction unit is used to input the acquired scan image into the region prediction model to obtain a suspected detection region and a corresponding electrode needle detection prediction signal sequence.

7. The AI-assisted lung cancer pleural invasion diagnosis auxiliary system according to claim 6, characterized in that: The exploration path correction module includes an error acquisition unit, a correction model training unit and an optimal path unit; The error acquisition unit uses the electrode needle signal in the same time dimension as the real-time scanning image and the corresponding electrode needle exploration prediction signal sequence to obtain the exploration signal error in the same time dimension; The correction model training unit is used to input the suspected detection area, safe detection boundary, and detection signal error of the scanned image into the detection path correction model for training, thereby obtaining a trained detection path correction model; The optimal path unit is used to correct the model of the trained exploration path, simulate and generate multiple exploration paths by scanning the suspected exploration area and safe exploration boundary of the image, and perform exploration risk assessment on the simulated multiple exploration paths to obtain the optimal exploration path.

8. A computer-readable storage medium, characterized in that Computer instructions are stored thereon, and when the computer instructions are executed, the AI-assisted lung cancer pleural invasion diagnosis auxiliary method according to any one of claims 1 to 5 is executed.

Images