Deep learning based risk prediction system for lung volume reduction surgery for emphysema-induced bullae

Abstract

This invention discloses a deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion, belonging to the field of image processing technology. The system includes: a surgical data acquisition module, a preoperative image enhancement module, a surgical risk prediction model construction module, and a surgical risk prediction module. Specifically, this invention extracts the effective region of lung tissue, obtains multi-scale denoised and enhanced lung tissue images based on dynamic noise suppression factors and noise region masks, and obtains adaptively enhanced lung tissue images based on adaptive contrast stretching, providing high-quality image input for multimodal risk feature fusion. It generates cross-modal fusion features based on a gating mechanism, generates multi-scale fusion risk features through convolutional layers, and splits them into two branches. Interactively enhanced branch features are obtained based on interactive attention weights, and deep risk features of each branch are extracted, improving the clinical applicability and accuracy of risk prediction for lung bullae reduction surgery under lung expansion.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically referring to a deep learning-based risk prediction system for lung bullae reduction surgery. Background Technology

[0002] The Lung Bulla Volume Reduction Surgery Risk Prediction System under Lung Expansion is a system that uses deep learning technology to analyze multidimensional patient data and predict the risk of complications after lung bulla volume reduction surgery under lung expansion. It enables early and accurate quantitative assessment of complication risk, assists clinicians in developing individualized surgical plans and intervention strategies, thereby improving surgical safety and reducing the incidence of postoperative complications.

[0003] However, existing risk prediction systems for pulmonary bullae reduction surgery under lung expansion suffer from uneven grayscale distribution, blurred bullae boundaries, and redundant background interference in preoperative CT images. This leads to incomplete extraction of risk-related image features, low recognition accuracy, and strong noise interference, affecting the accuracy of subsequent surgical risk prediction. Furthermore, existing risk prediction systems for pulmonary bullae reduction surgery under lung expansion are affected by multiple modal factors, making it difficult to effectively integrate heterogeneous data. Moreover, risk features exhibit multi-scale and channel redundancy, resulting in insufficient ability to characterize individualized surgical risks and limited prediction accuracy. Summary of the Invention

[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based risk prediction system for pulmonary bullae reduction surgery under lung expansion. Addressing the problems in existing pulmonary bullae reduction surgery risk prediction systems, such as uneven grayscale distribution, blurred bullae boundaries, and background redundancy in preoperative CT images, which lead to incomplete risk-related image feature extraction, low discrimination, and strong noise interference, thus affecting the accuracy of subsequent surgical risk prediction, this solution extracts the effective lung tissue region based on lung tissue masking, generates a basic lung tissue image labeled with bullae, obtains a globally balanced correction image based on a globally normalized image and a locally enhanced image, generates a scale fusion lower bound factor and a scale fusion upper bound factor, obtains a multi-scale denoised and enhanced lung tissue image based on a dynamic noise suppression factor and a noise region mask, and obtains an adaptively enhanced lung tissue image based on adaptive contrast stretching, thus enhancing the distinguishability of risk lesion region features and providing a basis for multi-modal risk prediction. This approach provides high-quality image input, reducing the interference of image quality defects on risk prediction and improving the accuracy of risk prediction for sub-bulbar volume reduction surgery. Addressing the limitations of existing sub-bulbar volume reduction surgery risk prediction systems, which suffer from the combined influence of multimodal factors, difficulty in effectively fusing heterogeneous data, and the multi-scale and channel redundancy of risk features leading to insufficient representation of individualized surgical risks and limited prediction accuracy, this solution generates cross-modal fusion features based on a gating mechanism. Multi-scale fusion risk features are generated through convolutional layers and split into two branches. Interactive attention weights are used to obtain interactively enhanced branch features. Deep risk features of each branch are extracted, weighted, and then an adaptive fusion feature is obtained. The predicted output complication risk level accurately matches the risk feature distribution of different patients, improving the representation and prediction ability of individual patient complication risks and enhancing the clinical applicability and accuracy of sub-bulbar volume reduction surgery risk prediction.

[0005] The present invention provides a deep learning-based risk prediction system for lung bullae volume reduction surgery, which includes a surgical data acquisition module, a preoperative image enhancement module, a surgical risk prediction model construction module, and a surgical risk prediction module.

[0006] The surgical data acquisition module collects historical data on patients undergoing lung bullae volume reduction surgery under lung expansion.

[0007] The preoperative image enhancement module extracts the effective area of ​​lung tissue based on lung tissue mask, generates a basic lung tissue image after bullae marking, obtains a global equalization correction image based on global normalized image and local enhancement image, generates scale fusion lower bound factor and scale fusion upper bound factor, obtains a multi-scale denoising enhanced lung tissue image based on dynamic noise suppression factor and noise region mask, and obtains an adaptively enhanced lung tissue image based on adaptive contrast stretching.

[0008] The surgical risk prediction model construction module generates cross-modal fusion features based on a gating mechanism, generates multi-scale fusion risk features through convolutional layers, and splits them into two branch features. Based on interactive attention weights, it obtains interactively enhanced branch features, extracts the deep risk features of each branch, and obtains adaptive fusion features after weighting, predicting and outputting the risk level of complications.

[0009] The surgical risk prediction module collects real-time data of patients undergoing lung bullae volume reduction surgery under lung expansion, performs preoperative image enhancement, and obtains the risk of postoperative complications based on the surgical risk prediction model.

[0010] Furthermore, the surgical data acquisition module collects historical data on patients undergoing lung bullae reduction surgery under lung expansion, including preoperative CT images, physiological data, clinical data, and complication risk levels, with the complication risk level used as a data label.

[0011] Furthermore, the preoperative image enhancement module includes a pulmonary bulla labeling unit, an image fusion correction unit, a multi-scale feature extraction unit, a multi-scale noise reduction enhancement unit, and a partitioned adaptive enhancement unit; specifically, it includes the following:

[0012] The lung bullae labeling unit performs grayscale standardization on preoperative CT images, obtains lung tissue masks through grayscale thresholding, extracts the effective lung tissue region, calculates the mean and standard deviation of grayscale values ​​of the effective lung tissue region, generates an adaptive threshold, obtains the initial lung bullae mask, performs morphological opening and closing operations sequentially to obtain the final lung bullae mask, crops the image background region based on the lung tissue mask, and marks the lung bullae region in the image to obtain the basic lung tissue image after lung bullae labeling.

[0013] Image fusion and correction unit: Performs global gamma correction and normalization on the basic lung tissue image to obtain a global normalized image; performs local CLAHE enhancement on the bullous region to obtain a local enhanced image; performs weighted fusion on the global normalized image and the local enhanced image based on the final bullous mask to obtain an enhanced fused image; and then performs global CLAHE and inverse normalization on the enhanced fused image to obtain a global balanced correction image.

[0014] Multi-scale feature extraction unit: Performs dual-scale Gaussian blur processing on the globally equalized correction image to generate fine-scale feature images and coarse-scale feature images respectively. Normalizes the feature images at each scale and calculates the gray mean and standard deviation of the feature images at each scale to generate scale fusion lower bound factor and scale fusion upper bound factor.

[0015] Multi-scale noise reduction and enhancement unit: Based on the lower bound factor and upper bound factor of scale fusion, nonlinear fusion operation is performed on the feature image of each scale, and weighted fusion is performed to obtain a multi-scale fused image. The gray mean and standard deviation of the background region are calculated to generate a dynamic noise suppression factor and a noise region mask. Noise suppression is performed on the multi-scale fused image, and then gamma correction and inverse normalization are performed to obtain a lung tissue image after multi-scale noise reduction and enhancement.

[0016] The partitioned adaptive enhancement unit divides the multi-scale denoised and enhanced lung tissue image into bullous candidate regions and normal lung tissue regions based on the final bullous mask. The brightness variance of the two regions is calculated separately, and an adaptive brightness offset coefficient is obtained by fusing them according to the variance weight. A partitioned adaptive contrast scaling coefficient is generated according to the region type. The multi-scale denoised and enhanced lung tissue image is subjected to adaptive contrast stretching, and the gray value of the bullous marker is retained by gray-scale cropping to obtain the adaptively enhanced lung tissue image.

[0017] Furthermore, the surgical risk prediction model construction module, based on a deep neural network, constructs a surgical risk prediction model to predict the risk of complications after a patient undergoes bullous lung reduction surgery with lung expansion. It includes a cross-modal fusion unit, a multi-scale risk feature extraction unit, a branch interaction enhancement unit, an adaptive fusion unit, and a risk prediction unit; specifically, it includes the following:

[0018] Cross-modal fusion unit: Extracts image features from adaptively enhanced lung tissue images through three layers of 3×3 convolution, performs data encoding and normalization on physiological and clinical data, extracts joint features through fully connected layers, dynamically generates weights for image features and joint features using a gating mechanism, maps the joint features to dimensions that match the image features, and then performs weighted fusion to obtain cross-modal fusion features.

[0019] Multi-scale risk feature extraction unit: Features are extracted stepwise from cross-modal fusion features through 5 convolutional layers and the resolution is reduced. Risk enhancement, upsampling activation and feature fusion are performed on the feature maps of layers 2 to 5 to generate multi-scale fused risk features.

[0020] Branch interaction enhancement unit: The multi-scale fused risk features are evenly divided into two branch features according to the channel dimension. For each branch feature, it is passed through a fully connected layer and a Sigmoid activation function to generate its corresponding interaction attention weight. The branch features are modulated based on the interaction attention weight to obtain the two branch features after interaction enhancement.

[0021] The adaptive fusion unit maps the features of the two branches after the interaction enhancement through a fully connected layer, then performs layer normalization and GELU activation, followed by dimensional reshaping. The deep risk features of each branch are extracted using a multilayer perceptron and concatenated along the channel dimension. The dimension is compressed by global average pooling. The fusion weights are generated by a fully connected layer, batch normalization, and sigmoid and softmax activation. The deep risk features of the two branches are weighted based on the fusion weights and concatenated along the channel to obtain the adaptive fusion features.

[0022] Risk prediction unit: After reshaping the adaptive fusion features into a one-dimensional vector, the probability distribution of complication risk levels is obtained through two fully connected layers and the Softmax function. The complication risk level with the largest value in the probability distribution is selected as the prediction label output, thus completing the construction of the surgical risk prediction model.

[0023] Furthermore, the surgical risk prediction module collects real-time data on patients undergoing bullous volume reduction surgery under lung expansion, including preoperative CT images, physiological data, and clinical data. After preoperative image enhancement of the preoperative CT images, they are input together with the physiological and clinical data into the surgical risk prediction model for processing. Based on the output prediction labels, the risk of complications after the patient undergoes bullous volume reduction surgery under lung expansion is obtained in real time.

[0024] The beneficial effects achieved by the present invention using the above solution are as follows:

[0025] (1) In response to the problems of uneven gray-level distribution, blurred bullae boundaries and background redundancy in the existing CT preoperative images of the lung bullae volume reduction surgery risk prediction system, which lead to incomplete extraction of risk-related image features, low recognition and strong noise interference, and affect the accuracy of subsequent surgical risk prediction, this solution extracts the effective area of ​​lung tissue based on lung tissue mask, generates a basic lung tissue image after bullae labeling, accurately segments the lung tissue and bullae regions, and removes background interference; obtains a globally balanced correction image based on the globally normalized image and the locally enhanced image, taking into account both global gray-level balance and local bullae detail enhancement; generates a scale fusion lower bound factor and a scale fusion upper bound factor to realize the quantitative association and adaptive constraint of multi-scale features; obtains a multi-scale denoised and enhanced lung tissue image based on the dynamic noise suppression factor and the noise region mask to reduce the interference of noise on risk features; obtains an adaptively enhanced lung tissue image based on adaptive contrast stretching to enhance the feature discrimination of risk lesion regions, provide high-quality image input for multi-modal risk feature fusion, reduce the interference of image quality defects on risk prediction, and improve the accuracy of lung bullae volume reduction surgery risk prediction.

[0026] (2) To address the problems in existing risk prediction systems for lung expansion surgery with bullous volume reduction, which are affected by multiple modal factors, making it difficult to effectively integrate heterogeneous data, and the risk features have multi-scale and channel redundancy, resulting in insufficient ability to represent individualized surgical risks and limited prediction accuracy, this solution generates cross-modal fusion features based on a gating mechanism to achieve adaptive weighted fusion of multimodal data; generates multi-scale fusion risk features through convolutional layers to comprehensively cover multi-dimensional risk influencing factors of lung expansion surgery and improve the completeness of feature representation; and splits the features into two branches, obtaining interactive enhanced branch features based on interactive attention weights to improve the identification and representation effectiveness of multi-scale risk features for lung expansion surgery risks; extracts the deep risk features of each branch, and obtains adaptive fusion features after weighting, predicts and outputs the risk level of complications, accurately matches the risk feature distribution of different patients, improves the representation and prediction ability of individual patients' complication risks, and improves the clinical applicability and accuracy of risk prediction for lung expansion surgery with bullous volume reduction. Attached Figure Description

[0027] Figure 1 A schematic diagram of the deep learning-based risk prediction system for lung bullae reduction surgery provided by the present invention;

[0028] Figure 2 This is a schematic diagram of the preoperative image enhancement module;

[0029] Figure 3 A schematic diagram of the module for building a surgical risk prediction model.

[0030] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0033] Example 1, see Figure 1 The present invention provides a deep learning-based system for predicting the risk of lung bullae reduction surgery, which includes a surgical data acquisition module, a preoperative image enhancement module, a surgical risk prediction model construction module, and a surgical risk prediction module.

[0034] The surgical data acquisition module collects historical data on lung bullae volume reduction surgery performed on patients with lung expansion and sends the data to the preoperative image enhancement module.

[0035] The preoperative image enhancement module receives data sent by the surgical data acquisition module, extracts the effective area of ​​lung tissue based on the lung tissue mask, generates a basic lung tissue image after bullae marking, obtains a global equalization correction image based on the global normalized image and the local enhancement image, generates a scale fusion lower bound factor and a scale fusion upper bound factor, obtains a multi-scale denoising enhanced lung tissue image based on the dynamic noise suppression factor and the noise region mask, obtains an adaptively enhanced lung tissue image based on adaptive contrast stretching, and sends the data to the surgical risk prediction model construction module.

[0036] The surgical risk prediction model construction module receives data sent by the preoperative image enhancement module, generates cross-modal fusion features based on the gating mechanism, generates multi-scale fusion risk features through convolutional layers, and splits them into two branches. Based on interactive attention weights, it obtains interactively enhanced branch features, extracts the deep risk features of each branch, and obtains adaptive fusion features after weighting. It predicts and outputs the complication risk level and sends the data to the surgical risk prediction module.

[0037] The surgical risk prediction module receives data sent by the surgical risk prediction model construction module, collects real-time data of patients undergoing lung bullae volume reduction surgery under lung expansion, performs preoperative image enhancement, and obtains the risk of postoperative complications based on the surgical risk prediction model.

[0038] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the surgical data acquisition module, historical patient data on lung bullae volume reduction surgery under lung expansion is collected, including preoperative CT images, physiological data, clinical data and complication risk level, and the complication risk level is used as a data label.

[0039] The physiological data include body temperature, blood pressure, heart rate, respiratory rate, complete blood count, coagulation function indicators, blood biochemistry indicators, vital capacity, body mass index, and body fat percentage.

[0040] The clinical data includes past medical history, allergy history, family medical history, surgical history, blood transfusion history, smoking history, alcohol consumption history, and medication history;

[0041] The risk levels for the complications are categorized as low, medium, and high.

[0042] Example 3, see Figure 1 and Figure 2 This embodiment, based on the above embodiment, includes a pulmonary bulla labeling unit, an image fusion correction unit, a multi-scale feature extraction unit, a multi-scale noise reduction enhancement unit, and a partitioned adaptive enhancement unit in the preoperative image enhancement module; specifically, it includes the following:

[0043] Pulmonary bullae labeling unit: Preoperative CT images of pulmonary bullae under lung expansion suffer from uneven gray-level distribution, blurred boundaries between lung tissue and bullae, and interference from redundant background areas, making it impossible to accurately locate the core region of the pulmonary bullae. Gray-level standardization was performed on the preoperative CT images, mapping pixel values ​​to the range of 0-255. A lung tissue mask was obtained through gray-level thresholding, and the effective region of lung tissue was extracted. The mean and standard deviation of the gray-level of the effective region of lung tissue were calculated to generate an adaptive threshold, resulting in an initial pulmonary bullae mask. Morphological opening and closing operations were performed sequentially to obtain the final pulmonary bullae mask. The background region of the image was cropped based on the lung tissue mask, and the pulmonary bullae region was marked in the image, resulting in a basic image of lung tissue with bullae labeling. The adaptive threshold fits the gray-level distribution characteristics of lung tissue after lung expansion, and morphological operations optimize the integrity of the bullae mask, providing precise regional constraints for subsequent enhancement. The formulas used are as follows:

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] In the formula, This is the original grayscale value at pixel location x in the pre-CT image, where x is the pixel coordinate. and These are the lower and upper limits of grayscale in lung tissue, respectively. , , It is a logical "OR". , and These are the lung tissue mask at position x, the initial bulla mask, and the final bulla mask, respectively. It is the gray value at position x in the image after gray-level normalization. and These are the mean and standard deviation of gray values ​​in the effective area of ​​lung tissue, respectively. The effective area of ​​lung tissue is composed of... and The pixels constitute the input, and k is the adaptive threshold coefficient. , It is a morphological dilation operation. It is a morphological erosion operation, K small and K large These are 3×3 small-size and 5×5 large-size structural elements, respectively. λ is the grayscale value of the bullous region marker, λ=50. It is the gray value at position x in the basic image of lung tissue after bullae labeling;

[0049] Image fusion and correction unit: Single global enhancement tends to over-enhance noise in the bullous region and weaken details, while single local enhancement cannot take into account the overall contrast of lung tissue. After lung expansion, the density difference of lung tissue is large, and conventional enhancement can easily lead to regional grayscale distortion. Global gamma correction and normalization are performed on the basic lung tissue image to map grayscale values ​​to the 0-1 range, resulting in a globally normalized image. Local CLAHE enhancement is performed on the bullous region to obtain a locally enhanced image. Based on the final bullous mask, the globally normalized image and the locally enhanced image are weighted and fused to obtain an enhanced fused image. Global CLAHE and inverse normalization are then performed on the enhanced fused image to restore grayscale values ​​to the 0-255 range, resulting in a globally balanced corrected image. This effectively improves image readability without over-enhancement, providing a high-quality image foundation for feature extraction. The formulas used are as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] In the formula, , , and These represent the grayscale values ​​at position x of the globally normalized image, the locally enhanced image, the enhanced fused image, and the globally equalized corrected image, respectively. It is the normalization function, γ is the gamma correction exponent, γ=0.5. This is a contrast-limited adaptive histogram equalization function. C1 and C2 are the contrast limits for local and global CLAHE, respectively, with C1=3.0 and C2=2.0. S1 and S2 are the block sizes for local and global CLAHE, respectively, with S1 being 4×4 and S2 being 8×8. It is an inverse normalization function;

[0055] Multi-scale feature extraction unit: Under lung expansion, bullae exhibit differences in size, shape, and density. Single-scale features can only capture single-granularity information, easily missing multi-scale risk features such as microbulbs and edge textures, resulting in incomplete feature representation. The globally balanced corrected image undergoes dual-scale Gaussian blur processing to generate fine-scale and coarse-scale feature images. Each scale's feature image is normalized, mapping gray values ​​to the 0-1 range. The mean and standard deviation of gray values ​​for each scale's feature image are calculated to generate a lower and upper bound factor for scale fusion. The fine-scale feature captures microbulbs and texture details, while the coarse-scale feature captures the overall lung tissue and bullae outlines. The fusion factor achieves adaptive association of multi-scale features, conforming to the multi-scale distribution characteristics of bullae after lung expansion, thus improving the completeness of feature representation. The formulas used are as follows:

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] In the formula, and σ1 and σ2 are the gray values ​​at position x of the fine-scale and coarse-scale feature images, respectively, and the standard deviations of the Gaussian blur at the fine and coarse scales are σ1=1.0 and σ2=3.0, respectively. It is a Gaussian blur function. It is the gray value at position x after the feature image is normalized, and α is an adjustment parameter. μ σ and σ σ These are the mean gray level and standard deviation of the feature image, respectively. and These are the lower bound factor and the upper bound factor of scale fusion at position x, respectively.

[0061] Multi-scale noise reduction and enhancement unit: The inherent noise of preoperative CT images is amplified after multi-scale fusion. Background noise easily interferes with the extraction of pulmonary bullae features. Low-density areas of lung tissue after lung expansion are more sensitive to noise, and conventional noise reduction easily loses details. Based on the lower and upper bound factors of scale fusion, nonlinear fusion operations are performed on the feature images of each scale. Weighted fusion is used to obtain a multi-scale fused image. The mean and standard deviation of grayscale in the background area are calculated to generate a dynamic noise suppression factor and a noise region mask. Noise suppression is applied to the multi-scale fused image, followed by gamma correction and inverse normalization to restore the grayscale values ​​to the range of 0-255, resulting in a multi-scale noise-reduced and enhanced lung tissue image. Dynamic noise suppression conforms to the noise distribution characteristics of the background area, and multi-scale fusion improves feature robustness and reduces the interference of noise on risk features. The formulas used are as follows:

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] In the formula, It is the grayscale value calculated at position x. and These are the grayscale values ​​at position x after calculation at the fine and coarse scales, respectively. ω1 is the gray value at position x in the multi-scale fused image, and ω2 are the fusion weights for the fine and coarse scales, respectively, with ω1 + ω2 = 1. and These are the mean and standard deviation of the grayscale values ​​of the background area, respectively. The background area consists of... The pixels constitute It is a smoothing term. β is the noise suppression intensity adjustment coefficient. , It is the dynamic noise suppression factor at position x. It is a global equalization correction of the local variance within the x-neighborhood of a pixel in the image. It is the threshold for determining noise regions. , It is the noise region mask at position x. It is the grayscale value at position x after noise suppression. It is the maximum gray value of the image after noise suppression. It is the gray value at position x in the lung tissue image after multi-scale noise reduction and enhancement;

[0069] A partitioned adaptive enhancement unit is used. Significant differences in brightness and contrast exist between the bullous region and normal lung tissue under lung expansion surgery. Uniform enhancement can lead to over-brightness in the bullous region or loss of detail in normal tissue, failing to highlight risk-related regional feature differences. Based on the final bullous mask, the multi-scale denoising and enhanced lung tissue image is divided into bullous candidate regions and normal lung tissue regions. The brightness variance of the two regions is calculated separately, and an adaptive brightness offset coefficient is obtained by fusing them according to variance weights. A partitioned adaptive contrast scaling coefficient is generated based on the region type. Adaptive contrast stretching is applied to the multi-scale denoising and enhanced lung tissue image, and grayscale cropping preserves the bullous marker grayscale values, resulting in an adaptively enhanced lung tissue image. Partitioned enhancement aligns with the pathological feature differences between bullous and normal tissue during lung expansion surgery, enhancing the risk feature identification of the bullous region while preserving baseline information of normal tissue, providing highly discriminative image features for cross-modal fusion. The formulas used are as follows:

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] In the formula, and These represent the grayscale values ​​of the candidate bullae region and the normal lung tissue region at location x, respectively. It is the adaptive brightness offset coefficient, and δ is the contrast enhancement intensity control parameter. ω bulla and ω normal These are the variance weights of candidate bullae regions and normal lung tissue regions, respectively. ω bulla +ω normal =1, and These are the brightness variances of candidate bullae regions and normal lung tissue regions, respectively. It is the adaptive contrast scaling factor for the partition at position x. It is a regularization parameter. It is a grayscale cropping function. It is the gray value at position x in the adaptively enhanced lung tissue image.

[0076] By performing the above operations, this solution addresses the problems in existing risk prediction systems for pulmonary bullae reduction surgery under lung expansion. These problems include uneven grayscale distribution, blurred bullae boundaries, and redundant background interference in preoperative CT images, leading to incomplete extraction of risk-related image features, low discrimination, and strong noise interference, which affect the accuracy of subsequent surgical risk prediction. This solution extracts the effective lung tissue region based on lung tissue masking, generates a basic lung tissue image labeled with bullae, accurately segments the lung tissue and bullae regions, and removes background interference. A globally balanced corrected image is obtained based on a globally normalized image and a locally enhanced image, balancing global grayscale balance with local bullae detail enhancement. A scale fusion lower bound factor and a scale fusion upper bound factor are generated to achieve quantitative association and adaptive constraints of multi-scale features. A multi-scale denoised and enhanced lung tissue image is obtained based on a dynamic noise suppression factor and a noise region mask, reducing noise interference with risk features. An adaptively enhanced lung tissue image is obtained based on adaptive contrast stretching, enhancing the distinguishability of risk lesion region features. This provides high-quality image input for multi-modal risk feature fusion, reduces the interference of image quality defects on risk prediction, and improves the accuracy of risk prediction for pulmonary bullae reduction surgery under lung expansion.

[0077] Example 4, see Figure 1 and Figure 3 This embodiment, based on the above embodiment, constructs a surgical risk prediction model based on a deep neural network in the surgical risk prediction model construction module. This model predicts the risk of complications for patients undergoing bullous lung reduction surgery with lung expansion. It includes a cross-modal fusion unit, a multi-scale risk feature extraction unit, a branch interaction enhancement unit, an adaptive fusion unit, and a risk prediction unit. Specifically, it includes the following:

[0078] Cross-modal fusion unit: Imaging, physiological, and clinical data have different modalities and significant dimensional differences. Direct splicing can easily lead to feature conflicts. Lung expansion surgery risk requires the fusion of multimodal information; a single modality cannot comprehensively represent the risk. Image features are extracted from adaptively enhanced lung tissue images using three 3×3 convolutional layers. Physiological and clinical data are encoded and normalized. Joint features are extracted through fully connected layers. A gating mechanism dynamically generates weights for image features and joint features. After mapping the joint features to dimensions matching the image features, weighted fusion is performed to obtain cross-modal fused features. Data encoding uses One-Hot encoding to convert categorical data into numerical data. Data normalization uses a max-min scaling method to unify numerical data to the range [0, 1]. The gating mechanism adaptively learns modal importance, aligning with the differentiated contribution of imaging morphology, physiological state, and clinical history to risk during lung expansion surgery, avoiding modal redundancy and feature conflicts, and generating a unified cross-modal risk representation. The formulas used are as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] In the formula, I image and I unite These are image features and joint features, It is a 3×3 convolution operation. This is a batch normalization operation. L1 is a linear rectified activation function, and L2 is an adaptively enhanced lung tissue image. It is a fully connected layer operation, X unite These are physiological and clinical data after data encoding and normalization. ω image and ω unite These are the initial weights for image features and joint features, respectively. , , It is the Sigmoid activation function. It is a global average pooling operation. It is a dimension reshaping operation, and F is a cross-modal fusion feature;

[0083] A multi-scale risk feature extraction unit is used. Risk-related information in cross-modal fusion features is distributed across different resolution levels. Single-scale convolution easily loses fine-grained risk details. The risk of lung expansion surgery is strongly correlated with multi-scale features such as bullae size, lung function, and medical history. Five convolutional layers progressively extract features from the cross-modal fusion features and reduce the resolution. Each layer includes 3×3 convolution, batch normalization, ReLU activation, and max pooling with a stride of 2. Risk enhancement, upsampling activation, and feature fusion are performed on the feature maps from layers 2 to 5 to generate multi-scale fused risk features. Progressive downsampling preserves global risk correlations, and mid-layer feature enhancement strengthens key risk signals, improving the model's ability to capture multi-dimensional risks of lung expansion surgery. The formulas used are as follows:

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] In the formula, A i A i-1 A1 and A2 are the feature maps output by the i-th, (i-1)-th, and 1st layers, respectively. It is a max pooling operation with a step size of 2. and These are the feature maps of the (i-1)th layer and the first layer after activation enhancement, respectively. It is a 2x upsampling operation. It is a 1×1 convolution operation, r i It is the risk enhancement factor of the i-th layer. , It is element-wise multiplication. It is a channel-by-channel splicing, A cat It is the total feature map after multi-scale feature splicing and fusion, A risk It is a multi-scale fusion risk characteristic;

[0089] Branch interaction enhancement unit: Information redundancy exists between multi-scale risk feature channels. A single feature stream cannot distinguish between core and secondary risk channels. The channel importance of lung expansion surgery risk features varies, easily leading to feature representation redundancy. The multi-scale fused risk features are uniformly split into two branch features according to the channel dimension. For each branch feature, a fully connected layer and a sigmoid activation function are passed to generate its corresponding interaction attention weight. Based on the interaction attention weight, the branch features are modulated to obtain two branch features with enhanced interaction. Channel splitting and attention modulation achieve feature decoupling, dynamically strengthening high-contribution risk channel features, suppressing redundant information, improving the discriminability and effectiveness of multi-scale risk features, and adapting to the channel distribution characteristics of lung expansion surgery risk features. The formulas used are as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, A G and A H These are the first branch features and the second branch features, respectively. It is a channel splitting operation. and These are the features enhanced by the first branch interaction and the features enhanced by the second branch interaction, respectively.

[0094] Adaptive fusion unit; the dual-branch enhanced features need to be reasonably fused. Fixed weights cannot adapt to the feature differences in lung expansion surgery risk among different patients, which may lead to the weakening of risk information in one branch. The two branches of interactively enhanced features are respectively mapped through fully connected layers, then processed by layer normalization and GELU activation, followed by dimensional reshaping. The deep risk features of each branch are extracted using a multilayer perceptron and concatenated along the channel dimension. The dimension is compressed by global average pooling. After passing through fully connected layers, batch normalization, and sigmoid and softmax activation, adaptively adjustable fusion weights are generated. The deep risk features of the two branches are weighted based on the fusion weights and concatenated along the channel to obtain adaptive fusion features. The adaptive weights fit the risk feature distribution of individual patients, avoiding the generality defects of fixed weights, making full use of the complementary information of the dual branches, and improving the feature representation accuracy of individual lung expansion surgery risk. The formula used is as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] In the formula, It is the feature of the j-th branch after feature transformation, layer normalization, and GELU activation processing by the fully connected layer, where j is the branch index. It is a layer normalization operation. It is the GELU activation function. It is the feature enhanced by the interaction of the j-th branch. It is the deep risk characteristic of the j-th branch. It is a multilayer perceptron, where η is the fusion weight of the first branch. It is the Softmax activation function. and These are the deep risk characteristics of the first and second branches, respectively. AM It is an adaptive fusion feature;

[0100] Risk prediction unit: After reshaping the adaptive fusion features into a one-dimensional vector, the probability distribution of complication risk levels is obtained through two fully connected layers and the Softmax function. The complication risk level with the largest value in the probability distribution is selected as the prediction label output, thus completing the construction of the surgical risk prediction model.

[0101] By performing the above operations, this solution addresses the problems in existing risk prediction systems for lung bullae reduction surgery under lung expansion. These systems are affected by multiple modal factors, making it difficult to effectively integrate heterogeneous data. Furthermore, risk features exhibit multi-scale characteristics and channel redundancy, resulting in insufficient ability to represent individualized surgical risks and limited prediction accuracy. This solution generates cross-modal fusion features based on a gating mechanism, achieving adaptive weighted fusion of multimodal data. Multi-scale fusion risk features are generated through convolutional layers to comprehensively cover multi-dimensional risk influencing factors in lung expansion surgery, improving the completeness of feature representation. The features are then split into two branches, and interactive attention weights are used to obtain interactively enhanced branch features, improving the identification and representation effectiveness of multi-scale risk features for lung expansion surgery risks. Deep risk features from each branch are extracted, weighted, and then adaptively fused to predict complication risk levels. This accurately matches the risk feature distribution of different patients, improving the representation and prediction ability of individual patient complication risks, and enhancing the clinical practicality and accuracy of risk prediction for lung bullae reduction surgery under lung expansion.

[0102] Example 5, see Figure 1 This embodiment is based on the above embodiment. In the surgical risk prediction module, real-time data of patients undergoing lung bullae reduction surgery under lung expansion is collected, including preoperative CT images, physiological data and clinical data. After preoperative image enhancement of the preoperative CT images, they are input together with the physiological data and clinical data into the surgical risk prediction model for processing. Based on the output prediction label, the risk of complications after lung bullae reduction surgery under lung expansion is obtained in real time.

[0103] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0105] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion, characterized by: It includes a surgical data acquisition module, a preoperative image enhancement module, a surgical risk prediction model construction module, and a surgical risk prediction module; The surgical data acquisition module collects historical data on patients undergoing lung bullae volume reduction surgery under lung expansion. The preoperative image enhancement module extracts the effective area of ​​lung tissue based on lung tissue mask, generates a basic lung tissue image after bullae marking, obtains a global equalization correction image based on global normalized image and local enhancement image, generates scale fusion lower bound factor and scale fusion upper bound factor, obtains a multi-scale denoising enhanced lung tissue image based on dynamic noise suppression factor and noise region mask, and obtains an adaptively enhanced lung tissue image based on adaptive contrast stretching. The surgical risk prediction model construction module generates cross-modal fusion features based on a gating mechanism, generates multi-scale fusion risk features through convolutional layers, and splits them into two branch features. Based on interactive attention weights, it obtains interactively enhanced branch features, extracts the deep risk features of each branch, and obtains adaptive fusion features after weighting, predicting and outputting the risk level of complications. The surgical risk prediction module collects real-time data of patients undergoing lung bullae volume reduction surgery under lung expansion, performs preoperative image enhancement, and obtains the risk of postoperative complications based on the surgical risk prediction model.

2. The deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion as described in claim 1, characterized in that: The preoperative image enhancement module includes a pulmonary bullae labeling unit, an image fusion correction unit, a multi-scale feature extraction unit, a multi-scale noise reduction enhancement unit, and a region-adaptive enhancement unit; specifically, it includes the following: Pulmonary bullae marker unit; Image fusion and correction unit: Performs global gamma correction and normalization on the basic lung tissue image to obtain a global normalized image; performs local CLAHE enhancement on the bullous region to obtain a local enhanced image; performs weighted fusion on the global normalized image and the local enhanced image based on the final bullous mask to obtain an enhanced fused image; and then performs global CLAHE and inverse normalization on the enhanced fused image to obtain a global balanced correction image. Multi-scale feature extraction unit; The globally equalized correction image is subjected to dual-scale Gaussian blurring to generate fine-scale feature images and coarse-scale feature images respectively. The feature images at each scale are normalized, and the gray mean and standard deviation of the feature images at each scale are calculated to generate the scale fusion lower bound factor and scale fusion upper bound factor. Multi-scale noise reduction and enhancement unit; Partition Adaptive Enhancement Unit.

3. The deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion as described in claim 2, characterized in that: The pulmonary bullae labeling unit performs grayscale normalization on the preoperative CT image, obtains a lung tissue mask through grayscale thresholding, extracts the effective lung tissue region, calculates the mean and standard deviation of the grayscale of the effective lung tissue region, generates an adaptive threshold, obtains the initial pulmonary bullae mask, performs morphological opening and closing operations sequentially to obtain the final pulmonary bullae mask, crops the image background region based on the lung tissue mask, and marks the pulmonary bullae region in the image to obtain the basic lung tissue image after pulmonary bullae labeling.

4. The deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion as described in claim 2, characterized in that: The multi-scale noise reduction and enhancement unit performs nonlinear fusion operations on the feature images of each scale based on the scale fusion lower bound factor and the scale fusion upper bound factor, and obtains a multi-scale fused image by weighted fusion. It calculates the gray mean and standard deviation of the background region, generates a dynamic noise suppression factor and a noise region mask, performs noise suppression on the multi-scale fused image, and then performs gamma correction and inverse normalization processing to obtain a lung tissue image after multi-scale noise reduction and enhancement.

5. The deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion as described in claim 2, characterized in that: The partition adaptive enhancement unit divides the multi-scale denoised and enhanced lung tissue image into bullous candidate regions and normal lung tissue regions based on the final bullous mask. It calculates the brightness variance of the two regions respectively, and obtains an adaptive brightness offset coefficient by fusing them according to the variance weight. It generates a partition adaptive contrast scaling coefficient according to the region type, performs adaptive contrast stretching on the multi-scale denoised and enhanced lung tissue image, and retains the bullous marker gray value by gray-scale cropping to obtain the adaptively enhanced lung tissue image.

6. The deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion as described in claim 1, characterized in that: The surgical risk prediction model construction module, based on a deep neural network, predicts the risk of complications after a patient undergoes bullous lung reduction surgery with lung expansion. It includes a cross-modal fusion unit, a multi-scale risk feature extraction unit, a branch interaction enhancement unit, an adaptive fusion unit, and a risk prediction unit; specifically, it includes the following: Cross-modal fusion unit: Extracts image features from adaptively enhanced lung tissue images through three layers of 3×3 convolution, performs data encoding and normalization on physiological and clinical data, extracts joint features through fully connected layers, dynamically generates weights for image features and joint features using a gating mechanism, maps the joint features to dimensions that match the image features, and then performs weighted fusion to obtain cross-modal fusion features. Multi-scale risk feature extraction unit: Features are extracted stepwise from cross-modal fusion features through 5 convolutional layers and the resolution is reduced. Risk enhancement, upsampling activation and feature fusion are performed on the feature maps of layers 2 to 5 to generate multi-scale fused risk features. Branch interaction enhancement unit: The multi-scale fused risk features are evenly divided into two branch features according to the channel dimension. For each branch feature, it is passed through a fully connected layer and a Sigmoid activation function to generate its corresponding interaction attention weight. The branch features are modulated based on the interaction attention weight to obtain the two branch features after interaction enhancement. Adaptive fusion unit; Risk prediction unit: After reshaping the adaptive fusion features into a one-dimensional vector, the probability distribution of complication risk levels is obtained through two fully connected layers and the Softmax function. The complication risk level with the largest value in the probability distribution is selected as the prediction label output, thus completing the construction of the surgical risk prediction model.

7. The deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion as described in claim 6, characterized in that: The adaptive fusion unit maps the features of the two branches after interaction enhancement through a fully connected layer, then performs layer normalization and GELU activation, followed by dimensional reshaping. The deep risk features of each branch are extracted using a multilayer perceptron and concatenated along the channel dimension. The dimension is compressed by global average pooling. The fusion weights are generated by a fully connected layer, batch normalization, and sigmoid and softmax activation. The deep risk features of the two branches are weighted based on the fusion weights and concatenated along the channel to obtain the adaptive fusion feature.

8. The deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion as described in claim 1, characterized in that: The surgical data acquisition module collects historical data on patients undergoing lung bullae reduction surgery with lung expansion, including preoperative CT images, physiological data, clinical data, and complication risk levels, with the complication risk level used as a data label.

9. The deep learning-based risk prediction system for lung bullae reduction surgery under lung expansion as described in claim 1, characterized in that: The surgical risk prediction module collects real-time data on patients undergoing bullous volume reduction surgery under lung expansion, including preoperative CT images, physiological data, and clinical data. After preoperative image enhancement of the preoperative CT images, they are input together with the physiological and clinical data into the surgical risk prediction model for processing. Based on the output prediction labels, the risk of complications after bullous volume reduction surgery under lung expansion is obtained in real time.

Images