Machine learning based severe patient acute lung injury assessment system and method thereof

Abstract

The application discloses a severe patient acute lung injury evaluation system and method based on machine learning, relates to the field of acute lung injury evaluation, and comprises the following steps: acquiring chest multi-modal image data of a severe patient and performing standardized processing to generate a standard data set; dividing a lung region into structural blocks and extracting structural indexes, compensating for missing data through cross-modal space mapping or candidate block selection; identifying a modal conflict region and a structural mutation region to construct a scoring key region; after applying disturbance to screen an effective region, selecting a mean value or a weighted fusion strategy according to a consistency index; and combining dynamic weight to calculate a lung injury score and performing iterative optimization; the application can improve the identification accuracy of a key region of acute lung injury, and realizes accurate quantitative evaluation of the degree of lung injury.

Description

Technical Field

[0001] This invention relates to the field of acute lung injury assessment, specifically to a machine learning-based acute lung injury assessment system and method for critically ill patients. Background Technology

[0002] Acute lung injury (ALI) is a common respiratory complication in critically ill patients, characterized by rapid onset, severe condition, and rapid progression. In severe cases, it can develop into acute respiratory distress syndrome (ARDS), significantly increasing patient mortality. Early identification of ALI and objective assessment of its severity are crucial for guiding clinical intervention and improving prognosis. Imaging examinations, as non-invasive methods for assessing lung structure, play a key role in the diagnosis and evaluation of severe ALI.

[0003] However, traditional image assessment methods mainly rely on manual image reading or empirical rules for lesion identification and scoring, which suffers from problems such as high subjectivity, low efficiency, and inconsistent quantification standards, making it difficult to meet the needs of high-precision, structured assessment of lung injury in complex clinical scenarios. In recent years, machine learning methods have been widely used in the field of medical image processing. Leveraging the advantages of models in multidimensional feature extraction, complex pattern recognition, and data fusion, they can achieve automatic identification and assessment of structural indicators in multimodal data environments. However, the following shortcomings still exist: a lack of detailed subdivision of lung image structural regions and modeling of local structural mutations, making it impossible to accurately identify scoring-sensitive areas; and in actual multimodal image data, structural information is often missing due to imaging occlusion, resolution mismatch, etc., and existing solutions have failed to provide effective compensation mechanisms.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a machine learning-based assessment system and method for acute lung injury in critically ill patients.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention discloses a machine learning-based method for assessing acute lung injury in critically ill patients, comprising the following steps:

[0008] Acquire chest multimodal image data of critically ill patients, and perform standardization processing on the chest multimodal image data to generate a multimodal standard dataset;

[0009] The lung region is divided into several structural blocks, and structural indicators are extracted within the structural blocks based on the standard dataset. If structural indicators are missing in any modality, spatial mapping compensation is performed on the same structural blocks in other modalities first. If compensation is still not possible, the optimal candidate block is selected from the structural blocks in the current modality, and its structural indicators are extracted for compensation.

[0010] Based on the structural indicators, modal conflict areas and structural mutation areas in the structural blocks are identified, and the modal conflict areas and structural mutation areas are merged to construct the key scoring region;

[0011] A preset perturbation is applied to the structural indicators of the key scoring region and the perturbation impact value is calculated. It is then determined whether the perturbation impact value is greater than the preset impact threshold. If it is, the key scoring region is retained; otherwise, the key scoring region is removed.

[0012] A consistency index is calculated based on the structural indicators of the same key scoring region in all modalities. It is then determined whether the consistency index is greater than a preset consistency threshold. If it is, the structural indicators of the same key scoring region are fused by mean. Otherwise, a weighted fusion is performed based on the credibility of each modality.

[0013] The lung injury score of the critically ill patient was calculated based on the fused structural indicators and the dynamic weights of the structural indicators in the key scoring regions.

[0014] Determine whether the key scoring region meets the preset integrity condition. If yes, output the lung injury score; otherwise, reconstruct the key scoring region and calculate the lung injury score until the preset number of iterations is reached. If the integrity condition is still not met, output the average value of the lung injury score during all iterations.

[0015] Secondly, this invention discloses a machine learning-based acute lung injury assessment system for critically ill patients, comprising:

[0016] The data acquisition module is used to acquire multimodal chest imaging data of critically ill patients and to perform standardization processing on the multimodal chest imaging data to generate a multimodal standard dataset.

[0017] The data processing module is used to divide the lung region into several structural blocks, extract structural indicators within the structural blocks based on the standard dataset; if structural indicators are missing in any modality, spatial mapping compensation is preferentially performed on the same structural blocks in other modalities; if compensation is still not possible, the optimal candidate block is selected from the structural blocks in the current modality, and its structural indicators are extracted for compensation.

[0018] The region extraction module is used to identify modal conflict regions and structural abrupt change regions in the structural blocks based on the structural indicators, and to merge the modal conflict regions and structural abrupt change regions to construct the key scoring regions;

[0019] The region optimization module is used to apply a preset perturbation to the structural indicators of the key scoring region and calculate the perturbation impact value. It determines whether the perturbation impact value is greater than the preset impact threshold. If it is, the key scoring region is retained; otherwise, the key scoring region is removed.

[0020] The feature fusion module is used to calculate a consistency index based on the structural indicators of the same key scoring region in all modalities, and to determine whether the consistency index is greater than a preset consistency threshold. If it is, the structural indicators of the same key scoring region are fused by mean; otherwise, weighted fusion is performed based on the credibility of each modality.

[0021] The injury assessment module is used to calculate the lung injury score of the critically ill patient based on the fused structural indicators and the dynamic weights of the structural indicators in the key scoring regions.

[0022] The evaluation and optimization module is used to determine whether the key scoring region meets the preset integrity condition. If it does, the lung injury score is output; otherwise, the key scoring region is reconstructed and the lung injury score is calculated until the preset number of iterations is reached. If the integrity condition is still not met, the average value of the lung injury score during all iterations is output.

[0023] The beneficial effects of this invention are as follows:

[0024] 1. By introducing a modal consistency index and a credibility weighting mechanism, the data consistency is evaluated based on the correlation coefficient of the structural indicators between each modality. If the consistency is high, mean fusion is performed; if the consistency is poor, weighted fusion is performed based on modal credibility. This effectively avoids the evaluation error introduced by the traditional simple averaging method in the case of modal conflict, thereby improving the robustness and accuracy of the evaluation results.

[0025] 2. When structural indicators are missing in multimodal images, a cross-modal spatial mapping compensation strategy is prioritized. If cross-modal compensation fails, a block feature prediction model is constructed, and the optimal candidate block is selected based on feature similarity for intramodal compensation. This dual compensation mechanism effectively reduces the impact of missing data on lung injury assessment results while ensuring the rationality of structural block features.

[0026] 3. By integrating modal conflict detection and structural mutation identification, key scoring regions are constructed. Furthermore, perturbation sensitivity analysis is introduced to assess the impact of regions on the score by applying positive and negative perturbations to structural indicators, retaining only regions sensitive to score changes. This mechanism enhances the assessment model's ability to identify lesion-sensitive regions, reduces interference from redundant blocks, and improves the accuracy and interpretability of the assessment.

[0027] 4. By combining area proportion and key anatomical points as dual constraints, and adopting a dynamic expansion mechanism, the scoring area range can be adaptively adjusted to ensure the comprehensiveness and accuracy of the evaluation results. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is an overall block diagram of the method in Embodiment 1 of the present invention;

[0030] Figure 2 This is a flowchart illustrating the extraction of structural indicators in the method of Embodiment 1 of the present invention;

[0031] Figure 3 This is a flowchart of the lung injury assessment process in the method of Embodiment 1 of the present invention;

[0032] Figure 4 This is an overall block diagram of the system in Embodiment 2 of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Application Overview: In existing technologies, acute lung injury assessment primarily relies on manual interpretation of single-modal images, making it difficult to effectively integrate multimodal image data. Traditional methods employ simple averaging strategies during multimodal fusion, ignoring spatial and reliability differences between modalities, leading to unstable assessment results. When data is missing in a particular modality, existing technologies lack effective compensation mechanisms, affecting the completeness of the assessment. Furthermore, the weight allocation in scoring models is often fixed, failing to adapt to different lesion distributions and modal quality variations, resulting in inaccurate identification of sensitive areas.

[0035] To address the aforementioned issues, those skilled in the art have identified the need to establish a spatial mapping compensation mechanism for multimodal data to resolve the problem of missing structural indicators. By analyzing the relationship between intermodal feature conflicts and structural abrupt changes, a proposal is made to construct key scoring regions as the core of the evaluation. To address the shortcomings of traditional fusion strategies, a region selection mechanism based on perturbation sensitivity is designed, combined with dynamic weight adjustments to improve evaluation stability. The completeness of the scoring results is ensured through iterative optimization of the key region coverage.

[0036] Example 1:

[0037] like Figure 1-3 As shown, a machine learning-based method for assessing acute lung injury in critically ill patients includes the following steps: acquiring multimodal chest imaging data of critically ill patients; standardizing the multimodal chest imaging data to generate a standard multimodal dataset; dividing the lung region into several structural blocks; extracting structural indicators within the structural blocks based on the standard dataset; if structural indicators are missing in any modality, spatial mapping compensation is preferentially performed on the same structural blocks in other modalities; if compensation is still not possible, the optimal candidate block is selected from the structural blocks in the current modality, and its structural indicators are extracted for compensation; identifying modal conflict areas and structural mutation areas in the structural blocks based on the structural indicators; merging the modal conflict areas and structural mutation areas to construct a key scoring region; applying a preset perturbation to the structural indicators of the key scoring region and calculating the perturbation effect. The system calculates the impact value of the disturbance and determines whether it exceeds a preset impact threshold. If it does, the key scoring region is retained; otherwise, it is removed. A consistency index is calculated based on the structural indices of the same key scoring region across all modalities. The system then determines whether the consistency index exceeds a preset consistency threshold. If it does, the structural indices of the same key scoring region are averaged; otherwise, weighted fusion is performed based on the reliability of each modality. Based on the fused structural indices and their dynamic weights within the key scoring regions, the lung injury score of the critically ill patient is calculated. The system then determines whether the key scoring region meets a preset integrity condition. If it does, the lung injury score is output; otherwise, the key scoring region is reconstructed and the lung injury score is calculated again until a preset number of iterations is reached. If the integrity condition is still not met, the average lung injury score across all iterations is output.

[0038] Standardization processing refers to unifying the resolution, grayscale values, and spatial locations of images from different modalities. This can be achieved by adjusting image size, histogram equalization, and aligning anatomical landmarks, ensuring the comparability of multimodal data. Structural block partitioning refers to dividing the lung into grid cells based on bronchial tree bifurcation points. This can be achieved using rectangular grid segmentation, providing a spatial basis for local feature analysis. Spatial mapping compensation refers to finding structural indicators at the same location in other modalities to replace missing data. This can be achieved through coordinate system transformation and feature matching, addressing the problem of incomplete single-modal data. Optimal candidate block selection refers to searching for the nearest neighboring region with the highest feature similarity to the missing block in the current modality. This can be achieved using cosine similarity calculation and prediction models, ensuring the rationality of data compensation. Construction of key scoring regions refers to merging blocks with modal conflicts and structural abrupt changes to form the core evaluation region. This can be achieved through analysis of variance and Euclidean distance thresholding, focusing on sensitive evaluation areas. Dynamic weight calculation refers to adjusting indicator weights based on spatial variation and consistency deviation. This can be achieved through normalization and weighted summation, improving the adaptability of the scoring model.

[0039] Specifically, this method first standardizes multimodal data through image resolution adjustment, grayscale normalization, and anatomical landmark alignment. After dividing the lung into grid blocks aligned with bronchial bifurcations, the texture, density, and edge features of each block are extracted. When data for a block in a certain modality is missing, substitute data is preferentially obtained from corresponding locations in other modalities; if cross-modal compensation is not possible, the optimal candidate block is selected within the current modality through spatial distance filtering and feature prediction. Conflict regions are identified by calculating intermodal variance, and key assessment regions are determined by detecting feature mutations in adjacent blocks within a single modality. After applying positive and negative perturbations to the key regions, valid regions are selected based on score changes. In the multimodal data fusion stage, a mean or confidence-weighted strategy is adopted based on the correlation coefficient. Finally, the weights are dynamically adjusted by combining spatial changes and consistency deviations, and the coverage of key regions is iteratively optimized before outputting a lung injury score.

[0040] Traditional methods, which fuse multimodal data with fixed weights, cannot handle spatial feature differences and modal conflicts. This approach focuses on sensitive areas by constructing key scoring regions, and significantly improves assessment stability by combining perturbation analysis and dynamic weight adjustment. Existing techniques use simple interpolation to handle missing data; this approach compensates for this through cross-modal mapping and candidate block prediction, improving data integrity. Traditional scoring models rely on fixed anatomical partitions; this approach iteratively optimizes the coverage of key regions, enhancing adaptability to changes in lesion distribution.

[0041] Through the above technical solutions, this application effectively integrates the spatial and feature information of multimodal imaging data, resolving issues such as data missing compensation and modal conflicts. Perturbation impact analysis and dynamic weight adjustment improve the accuracy and robustness of lung injury scoring. An iterative optimization mechanism ensures that key regions cover all predefined anatomical landmarks, avoiding assessment omissions. This method provides reliable technical support for the early identification and quantitative assessment of acute lung injury in critically ill patients.

[0042] This application further proposes chest multimodal imaging data including CT scan images, X-ray images, and ultrasound images; standardization processing includes image resolution unification, grayscale value normalization, and spatial alignment; the specific process of standardization processing includes: unifying the resolution of different modal images to a preset reference resolution; achieving grayscale value normalization by stretching the grayscale histogram; and performing spatial alignment using the sternal angle and tracheal bifurcation point as anatomical reference points.

[0043] Chest multimodal imaging data refers to a collection of lung images acquired through different imaging devices, such as CT scanners, X-ray machines, and ultrasound equipment. Covering different tissue characteristics in the imaging modal compensates for the limitations of a single imaging method. Image resolution unification refers to adjusting the pixel size of different modalities to the same physical scale. This can be achieved using interpolation algorithms or downsampling methods to adjust the image size, eliminating the interference of resolution differences between modalities on subsequent structural indicator extraction. Gray-scale normalization eliminates differences in gray-scale response between different imaging devices. This can be achieved using histogram matching or linear stretching methods to map gray-scale values ​​to a unified range, enabling comparability of cross-modal data. Spatial alignment maps multimodal images to the same anatomical coordinate system. This can be achieved using rigid registration algorithms based on the sternal angle and tracheal bifurcation point, eliminating spatial offsets caused by patient positional differences through anatomical landmark alignment.

[0044] Specifically, after acquiring CT, X-ray, and ultrasound images, the resolution of the CT images is first adjusted to a preset baseline resolution (e.g., 0.5 mm / pixel). The X-ray images are then adjusted to the same resolution using a bicubic interpolation algorithm, while the ultrasound images undergo spatial calibration based on the physical parameters of the scanning probe. Subsequently, the Hounsfield units of the CT images are linearly transformed, the grayscale range of the X-ray images is mapped to the 0-255 range using histogram stretching, and the grayscale values ​​of the ultrasound images are normalized using a contrast enhancement algorithm. Finally, in the spatial registration stage, the three-dimensional coordinates of the sternal angle and tracheal bifurcation point in the CT images are detected and used as reference points. An affine transformation is then used to align the two-dimensional projection coordinate system of the X-ray images with the three-dimensional volumetric coordinate system of the ultrasound images to the anatomical space of the CT images, forming a multimodal dataset with a unified geometric baseline.

[0045] Traditional multimodal image processing often only performs simple scaling or grayscale adjustments without establishing spatial registration relationships based on anatomical structures, making it difficult to accurately correspond lesion locations between different modalities. This solution achieves rigid registration through anatomical landmarks, resolving the spatial misalignment problem in multimodal images. Furthermore, by combining resolution unification and grayscale normalization, it eliminates the impact of differences in device characteristics on structural index extraction.

[0046] Through the above technical solutions, this application achieves geometric and grayscale spatial consistency of multimodal image data, providing standardized input for subsequent structural block division and index extraction, avoiding feature matching errors caused by inconsistencies in resolution, grayscale distribution and spatial coordinates between modalities, significantly improving the accuracy of cross-modal data fusion, and supporting the reliability of subsequent lung injury scoring.

[0047] This application further proposes a structural block division based on lung anatomical features, dividing the lung region into multiple rectangular grid units, with the division boundaries of the rectangular grid units aligned with the bronchial tree bifurcation points; structural indicators include texture feature values, density distribution values, and edge sharpness; texture feature values ​​are obtained by calculating the contrast of pixel grayscale values ​​within the structural block; density distribution values ​​are obtained by statistically analyzing the mean of CT values ​​within the structural block; edge sharpness is quantified using the Canny operator response intensity; spatial mapping compensation includes spatially aligning missing structural blocks in one modality with known structural blocks in other modalities, using the latter's structural indicators as substitute data; the selection process for the optimal candidate block is as follows: calculating the physical spatial distance between all structural blocks in the current modality image and the structural block with missing structural indicators; setting... A distance threshold is used to filter out candidate blocks whose physical spatial distance is less than the threshold. If the candidate block set is empty, the distance threshold range is expanded and the filtering is repeated until at least one candidate block is found. The physical spatial distance is calculated based on the image coordinate system. The spatial coordinates and structural indicators of all complete structural blocks in the current modal image are extracted. Using the spatial coordinates as input and the vector of the structural indicator as output, a block feature prediction model is established through supervised learning. The spatial coordinates of structural blocks with missing structural indicators are input into the trained block feature prediction model, and the predicted structural indicators are output. The feature similarity between the candidate blocks and the predicted structural indicators is calculated. The structural block with the highest feature similarity is selected as the optimal candidate block. The feature similarity is obtained by calculating the cosine similarity between the vectors of the structural indicators.

[0048] The alignment of rectangular grid cell segmentation boundaries with bronchial tree bifurcation points refers to setting the grid division boundaries based on the anatomical landmarks of bronchial branches. Specifically, image registration techniques can be used to align grid division lines with the coordinates of bronchial bifurcation points, ensuring that structural block divisions match physiological structures and avoiding cross-anatomical region segmentation. Contrast calculation of texture feature values ​​involves analyzing the distribution differences of pixel grayscale values ​​in local areas. Specifically, a gray-level co-occurrence matrix algorithm can be used to extract texture contrast parameters to characterize the homogeneity or lesion degree of lung tissue. Spatial mapping compensation for positional alignment utilizes the spatial registration relationship of multimodal images to map the coordinates of missing regions to corresponding positions in other modalities. Specifically, an affine transformation algorithm can be used to achieve cross-modal coordinate transformation, thereby using complete data from other modalities to fill in the missing information of the current modality. Supervised learning of the block feature prediction model involves using the spatial coordinates of complete structural blocks as input features and structural index vectors as output labels. Specifically, random forest or neural network algorithms can be used to train the model to achieve spatial location-based structural index prediction, providing alternative data for missing blocks.

[0049] Specifically, based on standardized multimodal image data, the lungs are first divided into rectangular grid units according to anatomical landmarks of bronchial tree bifurcation points, with each unit corresponding to an independent structural block. For each structural block, three types of indicators—texture feature values, density distribution values, and edge sharpness—are extracted from images of different modalities. When an indicator is missing for a specific structural block in a particular modality, a spatial mapping compensation mechanism is used to map the spatial coordinates of that block to the corresponding location in other modalities to obtain alternative indicator data. If other modalities cannot provide compensation data, neighboring candidate blocks are selected based on physical spatial distance, and a pre-trained block feature prediction model is used to generate predicted indicators. Finally, the candidate block closest to the predicted result is selected as the compensation source based on cosine similarity. This process effectively solves the problem of data loss caused by imaging limitations, while ensuring the reliability of the compensated data through anatomical alignment and spatial correlation constraints.

[0050] Traditional methods typically employ interpolation or simple nearest-neighbor imputation when data is missing, failing to consider spatial mapping relationships between multiple modalities and anatomical constraints, thus easily introducing compensation errors. This approach achieves multi-level compensation for missing data through cross-modal spatial alignment compensation and machine learning-based block feature prediction, avoiding the impact of missing data from a single modality on the overall assessment. Furthermore, combining anatomical alignment at bronchial bifurcation points makes the segmentation of structural blocks more aligned with actual clinical needs.

[0051] Through the above technical solutions, this application can effectively address the problem of missing local information in multimodal image data. By combining spatial mapping compensation and machine learning prediction, it improves the accuracy of missing data compensation and avoids assessment bias caused by incomplete data. Simultaneously, the block division and structural index extraction methods based on anatomical structures enhance the sensitivity detection capability of local lesion areas, providing more reliable basic data for subsequent lung injury scoring.

[0052] This application further proposes a method for identifying modal conflict zones and structural abrupt change zones in structural blocks based on structural indices. The identification process of modal conflict zones includes: calculating the variance of structural indices of the same structural block in different modes; if the variance is greater than the conflict determination threshold, it is determined to be a modal conflict zone. The identification process of structural abrupt change zones includes: calculating the Euclidean distance difference between structural indices of a structural block and adjacent structural blocks within a single mode; if the difference exceeds a preset abrupt change threshold, it is marked as a structural abrupt change zone.

[0053] Modal conflict zones are areas marked due to significant differences in structural indicators between different modalities of imagery. Specifically, variance calculation can be used to quantify these differences, reflecting the degree of fluctuation in structural indicators of the same structural block across different modalities. The conflict determination threshold can be an empirical value pre-set based on the distribution of multimodal data, for example, set to 0.5, to filter out modally inconsistent areas of interest. Structural aberration zones are areas in the same modal of imagery where structural indicators abruptly change. Specifically, Euclidean distance difference calculation can be used to capture local anomalies, achieved by comparing the differences in structural indicator vectors between the current block and its neighboring blocks. The preset aberration threshold can be set based on statistical results from normal lung imaging data, for example, set to twice the average difference in structural indicators between adjacent blocks, to identify abnormal structural changes.

[0054] Specifically, in the identification of modal conflict zones, for the same structural block, structural indices such as texture feature values ​​and density distribution values ​​are extracted under CT, X-ray, and ultrasound modalities, and the variances of these indices across different modalities are calculated. If the variance exceeds the conflict threshold, it indicates that the region has significant inconsistencies in multimodal data and needs to be marked as a modal conflict zone. In the identification of structural aberration zones, for image data of a certain modality, each structural block is traversed, and the Euclidean distance difference between its structural index vector and that of adjacent blocks is calculated. If the difference exceeds a preset aberration threshold, the region is determined to have a local structural aberration. By merging modal conflict zones and structural aberration zones, a key scoring region covering multimodal inconsistencies and local anomalies can be constructed, providing a data foundation for subsequent refined scoring.

[0055] Traditional methods often rely solely on a single modality or fail to clearly distinguish between intermodal differences and local structural anomalies, leading to inaccurate selection of scoring regions. This method, by quantifying intermodal variance and local abrupt changes within a single modality, can dynamically identify conflicting regions in multimodal data and anomalous regions within images. The resulting critical scoring region reflects both cross-modal inconsistencies and captures local structural changes within a single modality, effectively improving the coverage and accuracy of scoring-sensitive regions.

[0056] Through the above technical solution, this application solves the problem of scoring region selection bias caused by the ineffective quantification of intermodal differences and the inaccurate identification of local structural abnormalities in the prior art. By jointly determining the modal conflict zone and the structural mutation zone, key areas that significantly affect the lung injury score can be dynamically screened, avoiding scoring errors introduced by the failure to handle intermodal conflicts or local mutations, thereby improving the reliability and sensitivity of lung injury assessment results.

[0057] This application further proposes to calculate the average value of each type of structural indicator in all key scoring regions, and to obtain the benchmark score by weighting and summing the average values ​​of different types of structural indicators according to preset weights; to apply a preset fixed proportion of positive and negative adjustments to each type of structural indicator in the key scoring regions, and to calculate the benchmark score based on the adjusted structural indicators, thereby obtaining the positive score and the negative score respectively; the formula for calculating the disturbance impact value is: the disturbance impact value is equal to the sum of the absolute value of the difference between the positive score value and the benchmark score value plus the absolute value of the difference between the negative score value and the benchmark score value, divided by two.

[0058] The baseline score is a reference score formed by weighting the average of different types of structural indicators. This can be achieved by setting different importance weights for the structural indicators; for example, setting the weight of texture features to 0.4, density distribution to 0.3, and edge sharpness to 0.3. This baseline score quantifies the basic state of the key scoring region when undisturbed. Preset fixed-proportion positive and negative adjustments refer to the controlled increase or decrease of structural indicators, which can be achieved using percentage adjustments, such as a 5% fluctuation. By applying bidirectional perturbation, the sensitivity of the key scoring region to indicator changes can be tested. The perturbation impact value quantifies the stability of the key region by calculating the average deviation of the score before and after the perturbation, which can be achieved using the absolute value averaging method. This indicator is used to screen regions that significantly affect the overall scoring result, thereby optimizing the selection of key scoring regions.

[0059] Specifically, after the initial construction of the key scoring regions is completed, its effectiveness needs to be verified through perturbation testing. First, various structural indicators are extracted from all key scoring regions, and their average values ​​are calculated. These average values ​​are then weighted and summed according to preset weights to form a baseline score. Subsequently, each type of structural indicator is adjusted positively and negatively. For example, increasing the density distribution value by 5% and recalculating the score yields a positive score; then decreasing the density distribution value by 5% yields a negative score. Finally, the perturbation impact value is determined by calculating the average deviation of these two perturbation score values ​​from the baseline value. If this value is higher than a preset threshold, it indicates that the region is sensitive to indicator changes and should be retained in the key scoring regions; otherwise, it is removed. This dynamic screening mechanism eliminates redundant regions with low contribution to the overall score, improving the reliability of the evaluation results.

[0060] Traditional methods often rely on fixed rules or static thresholds when constructing scoring regions, failing to consider the potential impact of indicator perturbations on scoring results. For example, some existing schemes directly use manually defined anatomical regions as scoring regions, ignoring differences in lesion distribution and indicator sensitivity among different patients. This scheme, however, achieves dynamic optimization and selection of key scoring regions through bidirectional perturbation testing and influence value calculation, effectively avoiding scoring bias caused by improper region selection.

[0061] Through the above technical solution, this application solves the problem of the lack of dynamic adaptability in the selection of scoring regions in existing lung injury assessment methods. By quantifying the impact of perturbations on the scoring results, it can automatically identify and retain key regions sensitive to lung injury scoring, while eliminating regions with poor stability. This mechanism significantly improves the anti-interference ability and accuracy of the scoring results, especially when the patient's lung lesions are complexly distributed or when there is local noise in the imaging data, effectively reducing assessment errors.

[0062] This application further proposes a method for calculating the consistency index by combining all modes in pairs and calculating the correlation coefficient. Specifically, this involves combining all modes in pairs, calculating the correlation coefficient between structural indicators for each combination, averaging the correlation coefficients to obtain the consistency index, and determining the correlation coefficient by the ratio of the product of covariance and standard deviation.

[0063] The pairwise combination of all modalities refers to pairing different imaging modalities, such as CT, X-ray, and ultrasound, for example, CT and X-ray, CT and ultrasound, and X-ray and ultrasound. This step is used to comprehensively cover all possible correlations between multimodal data, ensuring the comprehensiveness of the consistency assessment. The correlation coefficient is a numerical value reflecting the degree of linear correlation between structural indicators of two modalities within the same structural block. Specifically, it can be calculated using the Pearson correlation coefficient formula, obtained by dividing the covariance by the product of the standard deviations of the structural indicators of each modality. This value can quantify the consistency of the changing trends of structural indicators between modalities. The ratio of the product of covariance and standard deviation is calculated by multiplying the standard deviations of the two modal structural indicators after calculating their covariances, then dividing the covariance by this product value. This is used to eliminate the influence of dimensional differences on the correlation assessment, achieving a standardized correlation measure. The consistency index is a comprehensive consistency measure obtained by averaging the correlation coefficients of all pairwise modal combinations. For example, by adding the three correlation coefficients of the three modal combinations and averaging them, this index can comprehensively reflect the overall consistency level of multimodal data and provide a quantitative basis for the selection of subsequent fusion strategies.

[0064] Specifically, after obtaining the structural indices for each modality, the modalities are first paired to form combinations, such as CT and X-ray, CT and ultrasound, and X-ray and ultrasound. For each combination, corresponding structural indices, such as texture feature values ​​and density distribution values, are extracted from the same key scoring region. The covariance between each pair of modal structural indices is calculated, and the standard deviation of each modal structural indices is also calculated separately. The covariance is divided by the product of the two standard deviations to obtain the correlation coefficient of the combination. This process is repeated until all pairwise combinations are calculated, and the arithmetic mean of all correlation coefficients is taken to obtain the final consistency index. If the consistency index exceeds a preset threshold, it indicates that the structural indices among the multimodalities are highly consistent, and the mean fusion method is used; if it is below the threshold, it indicates that there are significant differences between the modalities, and weighted fusion is performed according to the credibility of each modality. For example, when the correlation coefficient between CT and X-ray is 0.85, that between CT and ultrasound is 0.72, and that between X-ray and ultrasound is 0.65, the consistency index is (0.85+0.72+0.65) / 3=0.74. Assuming the preset consistency threshold is 0.7, mean fusion is then performed.

[0065] Traditional multimodal fusion methods typically calculate the correlation coefficient of single-modal combinations or simply average all modal data, failing to systematically evaluate the interrelationships between all modalities. For example, existing techniques may only calculate the correlation between CT and X-ray, ignoring the combination of CT and ultrasound, resulting in the failure to effectively identify data biases between some modalities. This method, by exhaustively enumerating all pairwise modal combinations and calculating the average correlation coefficient, can more comprehensively capture the consistency and differences of multimodal data, avoiding evaluation biases caused by single-combination calculations, thereby improving the scientific rigor of fusion strategy selection.

[0066] Through the above technical solution, this application can effectively identify local inconsistencies in multimodal data. For example, when there are significant differences between the structural indicators of ultrasound and CT due to imaging quality, the correlation coefficient calculation can accurately reflect this difference, thereby triggering a confidence-based weighted fusion strategy. This approach ensures the stability of the evaluation results while improving adaptability to modal quality fluctuations, solving the error accumulation problem caused by simple averaging fusion in existing technologies. It is particularly suitable for complex clinical scenarios where some modal imaging quality is poor.

[0067] This application further proposes to calculate the absolute difference between the structural index and the manually labeled reference structural index for each modality, average all absolute differences under the same modality to obtain the average deviation value, and normalize the average deviation value by taking the reciprocal of the average deviation value as the modality confidence weight.

[0068] Among them, manually labeled reference structural indicators refer to clinically significant structural feature values ​​manually marked by professional physicians based on imaging data. This can be achieved by averaging the values ​​of two labeled indicators, providing a benchmark for the true value of structural indicators. Absolute difference refers to the absolute value of the difference between the structural indicators extracted by the algorithm and the manually labeled values ​​within the same structural block. This can be achieved by calculating point-by-point, taking the absolute values, and then summing them, used to measure the degree of deviation between modal data and the true value. Average deviation value is the arithmetic mean of the absolute differences of all structural blocks within the same modality. This can be achieved by summing the values ​​and dividing by the total number of blocks, reflecting the overall level of deviation of the modality from the true value. Reciprocal operation refers to taking the reciprocal of the average deviation value mathematically. This can be achieved through a numerical conversion formula, converting the deviation value into a positive confidence index. Normalization processing involves adjusting the reciprocal results of each modality to a weight ratio that sums to 1. This can be achieved through linear scaling, ensuring the comparability of confidence weights across different modalities.

[0069] Specifically, after obtaining the multimodal structural indices, the algorithm-extracted structural indices are first compared block-by-block with manually labeled reference values. For example, for a structural block in the CT modality, the algorithm-calculated density distribution value is 85 HU, while the manually labeled value is 80 HU, resulting in an absolute difference of 5. By averaging the absolute differences across all blocks in this modality, assuming an average deviation of 3.2, its reciprocal is approximately 0.3125. If the average deviation for the ultrasound modality is 4.0, its reciprocal is 0.25. Adding the reciprocals of 0.3125 and 0.25 for both modalities yields a sum of 0.5625. After normalization, the confidence weight for the CT modality is approximately 0.3125 / 0.5625 ≈ 0.556, and for the ultrasound modality, it is approximately 0.25 / 0.5625 ≈ 0.444. Therefore, during weighted fusion, the CT modality data with higher confidence will have a higher weight, thus reducing the impact of modalities with larger deviations on the final score.

[0070] Traditional methods typically perform multimodal data fusion based on preset fixed weights or simple averaging strategies, such as assigning fixed weights of 0.6 and 0.4 to CT and X-ray modalities, respectively. However, in real-world scenarios, the imaging quality of different modalities may fluctuate due to differences in equipment, patient positioning, and other factors, making it impossible for fixed weights to dynamically adapt to data deviations. This proposed solution quantifies the deviation between modal data and manually labeled true values, transforming credibility calculation into a mathematical derivation process. This allows weight allocation to be dynamically adjusted according to actual data quality, preventing the spread of scoring errors caused by distortion in a single modality.

[0071] Through the above technical solution, this application solves the problem of unreasonable weight allocation caused by quality differences in multimodal data fusion. For example, when the structural indicators of some regions in ultrasound images deviate significantly due to lung gas interference, their reliability weight is automatically reduced, while the CT modality receives a higher weight due to its higher data stability, thereby significantly improving the clinical consistency of the fusion results. At the same time, the objective calculation process based on manually labeled reference values ​​reduces the reliance on empirical parameter settings and enhances the robustness of the evaluation system.

[0072] This application further proposes a dynamic weight calculation process including: based on each structural block in the scoring key region, calculating the absolute value of the difference between the structural indicators after fusion between adjacent structural blocks to obtain the spatial variation; normalizing all spatial variations in the scoring key region to obtain the variation amplitude factor; in the same scoring key region, calculating the standard deviation of the structural indicators before fusion under different modalities to obtain the consistency deviation value; obtaining the consistency adjustment factor after normalization of the consistency deviation; and weighting the variation amplitude factor and the consistency adjustment factor according to a set ratio to obtain the dynamic weight of each scoring key region.

[0073] Spatial variation refers to the degree of difference in structural indicators between adjacent structural blocks after fusion. This can be achieved by calculating the absolute difference in structural indicators between adjacent blocks, reflecting the magnitude of local structural changes within the key scoring region. Normalization involves converting values ​​of different dimensions or ranges into a unified proportion. Specifically, the min-max normalization method can be used to map spatial variation to the 0-1 interval, eliminating the impact of dimensional differences on weight calculation. Consistency deviation refers to the dispersion of structural indicators in different modalities for the same key scoring region. This can be achieved by calculating the standard deviation, measuring the reliability differences of multimodal data in this region. Dynamic weights refer to the contribution of the key scoring region to the final lung injury score. This can be obtained by linearly weighting the variation amplitude factor and the consistency adjustment factor, for example, setting a ratio of 6:4 or 7:3 to achieve a balance between spatial heterogeneity and modal consistency.

[0074] Specifically, when calculating the dynamic weights, the process first iterates through all structural blocks within the critical scoring region, calculating the absolute value of the difference between the fused structural indices of each block and its adjacent blocks, forming a set of spatial variations. This set is then normalized so that the normalized value for the largest variation is 1, and the smallest is 0. Next, for the same critical scoring region, the original structural indices before fusion are extracted from each modality, their standard deviations are calculated and normalized, resulting in a consistency adjustment factor reflecting the degree of deviation between modalities. Finally, the normalized variation amplitude factor and the consistency adjustment factor are weighted and summed according to a preset ratio, for example, the variation amplitude factor accounts for 60% and the consistency adjustment factor accounts for 40%, generating a dynamic weight value. This weight value directly participates in the weighted calculation of the lung injury score.

[0075] Traditional methods employ fixed weight allocation strategies, which cannot adapt to the structural heterogeneity and modal reliability fluctuations in different key scoring regions. This proposed solution, however, introduces a dual adjustment mechanism of spatial variation and consistency bias. This allows for dynamic weight adjustment based on the degree of structural abrupt changes and the level of intermodal consistency within a region. For example, higher weights are assigned to regions with significant structural differences and high modal consistency, thereby improving the objectivity and adaptability of the scoring.

[0076] Through the above technical solution, this application solves the problem that static weights cannot reflect changes in local structural features and differences in modal reliability, and realizes dynamic optimization of the weights of key scoring regions. In cases of uneven lesion distribution or fluctuations in the quality of some modal data, this solution can automatically adjust the contribution of each region to the score. For example, it can reduce the weight of regions with blurred edges but high modal consistency, while increasing the weight of regions with significant structural changes and consistent multimodal data, thereby significantly improving the accuracy of lung injury scoring and the reliability of clinical assessment results.

[0077] This application further proposes preset integrity conditions, including that the total area of ​​the scoring key region accounts for a proportion of the lung region that is not less than a set lower limit, and that it covers all predefined anatomical key points; the process of reconstructing the scoring key region includes selecting several structural blocks whose spatial location is continuous with the edge of the scoring key region, adding them to the current scoring key region, and forming an expanded scoring key region.

[0078] The preset integrity conditions refer to the minimum coverage area and key anatomical location coverage requirements that the scoring key region must meet. This can be achieved by setting area percentage thresholds and a predefined list of anatomical coordinates to ensure the scoring region has sufficient representativeness and coverage of key lesion areas. The process of reconstructing the scoring key region involves supplementing the region through neighborhood expansion when the integrity conditions are not met. Specifically, this can be achieved by using an edge continuity detection algorithm to filter adjacent structural blocks and iteratively adding them to expand the region, thus resolving the problem of insufficient coverage in the initial scoring region.

[0079] Specifically, after selecting the key scoring regions, the system automatically calculates the proportion of their total area to the lung region and compares it with a preset lower limit. Simultaneously, it checks whether all predefined anatomical key points are included within the key scoring regions. If the above conditions are not simultaneously met, it selects adjacent structural blocks that are physically continuous with the current region from its edge and merges them. The expanded region undergoes integrity verification again until the requirements are met or the preset iteration limit is reached. If the conditions are still not met when the iteration terminates, the average score generated in the previous iterations is used as the final output.

[0080] In some specific implementations, predefined anatomical key points may include landmark locations such as bronchial bifurcation points and lobular junctions, with a lower limit set, for example, 30% of the total lung area. When expanding the scoring key area, a spatial adjacency algorithm can be used to calculate the distance between candidate blocks and the edge of the current area, prioritizing the merging of blocks with the smallest distance.

[0081] Traditional methods rely solely on a single area threshold when assessing the integrity of the scoring region, neglecting the coverage requirements of key anatomical locations and easily leading to the omission of important lesion areas. This proposed solution, however, combines area proportion and key anatomical points as dual constraints, employing a dynamic expansion mechanism to adaptively adjust the scoring region range, ensuring the comprehensiveness and accuracy of the assessment results.

[0082] Through the above technical solution, this application can effectively avoid assessment bias caused by incomplete coverage of key scoring areas. Especially when there are lesions in key anatomical areas such as the edge of the lung lobe or the bifurcation of the bronchus, the clinical applicability of the lung injury score and the reliability of the assessment results are significantly improved by forcibly covering predefined key points.

[0083] Example 2:

[0084] like Figure 4 As shown, the machine learning-based acute lung injury assessment system for critically ill patients includes:

[0085] The data acquisition module is used to acquire multimodal chest imaging data of critically ill patients and to perform standardization processing on the multimodal chest imaging data to generate a multimodal standard dataset.

[0086] The data processing module is used to divide the lung region into several structural blocks, extract structural indicators within the structural blocks based on the standard dataset; if structural indicators are missing in any modality, spatial mapping compensation is preferentially performed on the same structural blocks in other modalities; if compensation is still not possible, the optimal candidate block is selected from the structural blocks in the current modality, and its structural indicators are extracted for compensation.

[0087] The region extraction module is used to identify modal conflict regions and structural abrupt change regions in the structural blocks based on the structural indicators, and to merge the modal conflict regions and structural abrupt change regions to construct the key scoring regions;

[0088] The region optimization module is used to apply a preset perturbation to the structural indicators of the key scoring region and calculate the perturbation impact value. It determines whether the perturbation impact value is greater than the preset impact threshold. If it is, the key scoring region is retained; otherwise, the key scoring region is removed.

[0089] The feature fusion module is used to calculate a consistency index based on the structural indicators of the same key scoring region in all modalities, and to determine whether the consistency index is greater than a preset consistency threshold. If it is, the structural indicators of the same key scoring region are fused by mean; otherwise, weighted fusion is performed based on the credibility of each modality.

[0090] The injury assessment module is used to calculate the lung injury score of the critically ill patient based on the fused structural indicators and the dynamic weights of the structural indicators in the key scoring regions.

[0091] The evaluation and optimization module is used to determine whether the key scoring region meets the preset integrity condition. If it does, the lung injury score is output; otherwise, the key scoring region is reconstructed and the lung injury score is calculated until the preset number of iterations is reached. If the integrity condition is still not met, the average value of the lung injury score during all iterations is output.

[0092] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0093] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0094] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A machine learning-based method for assessing acute lung injury in critically ill patients, characterized in that, Includes the following steps: Acquire chest multimodal image data of critically ill patients, and perform standardization processing on the chest multimodal image data to generate a multimodal standard dataset; The lung region is divided into several structural blocks, and structural indicators are extracted within the structural blocks based on the standard dataset. If structural indicators are missing in any mode, spatial mapping compensation is first performed on the same structural blocks in other modes; if compensation is still not possible, the optimal candidate block is selected from the structural blocks in the current mode, and its structural indicators are extracted for compensation. Based on the structural indicators, modal conflict areas and structural mutation areas in the structural blocks are identified, and the modal conflict areas and structural mutation areas are merged to construct the key scoring region; A preset perturbation is applied to the structural indicators of the key scoring region and the perturbation impact value is calculated. It is then determined whether the perturbation impact value is greater than the preset impact threshold. If it is, the key scoring region is retained; otherwise, the key scoring region is removed. A consistency index is calculated based on the structural indicators of the same key scoring region in all modalities. It is then determined whether the consistency index is greater than a preset consistency threshold. If it is, the structural indicators of the same key scoring region are fused by mean. Otherwise, a weighted fusion is performed based on the credibility of each modality. The lung injury score of the critically ill patient was calculated based on the fused structural indicators and the dynamic weights of the structural indicators in the key scoring regions. Determine whether the key scoring area meets the preset integrity condition; if so, output the lung injury score. Otherwise, the key scoring region is reconstructed and the lung injury score is calculated until the preset number of iterations is reached. If the integrity condition is still not met, the average value of the lung injury score during all iterations is output.

2. The machine learning-based method for assessing acute lung injury in critically ill patients according to claim 1, characterized in that: The chest multimodal imaging data includes CT scan images, X-ray images, and ultrasound images; The standardization process includes image resolution unification, grayscale value normalization, and spatial alignment. The specific process of the standardization process includes: unifying the resolution of images of different modalities to a preset reference resolution; normalizing gray values ​​by stretching the gray histogram; and performing spatial alignment using the sternal angle and tracheal bifurcation point as anatomical reference points.

3. The machine learning-based method for assessing acute lung injury in critically ill patients according to claim 2, characterized in that: The division of the structural blocks is based on the anatomical features of the lungs, dividing the lung region into multiple rectangular grid units, with the dividing boundaries of the rectangular grid units aligned with the bifurcation points of the bronchial tree. The structural indicators include texture feature values, density distribution values, and edge sharpness; The texture feature values ​​are obtained by calculating the contrast of pixel grayscale values ​​within the structural block; The density distribution value is obtained by taking the mean of the CT values ​​within the statistical structural block; the edge sharpness is quantized by the response intensity of the Canny operator. The spatial mapping compensation includes spatially aligning the missing structural blocks in a mode with the known structural blocks in other modes, and using the latter's structural indices as substitute data. The process of selecting the optimal candidate block is as follows: calculate the physical spatial distance between all structural blocks in the current modal image and structural blocks with missing structural indicators; Set a distance threshold and filter out a set of candidate blocks whose physical spatial distance is less than the distance threshold; if the set of candidate blocks is empty, expand the range of the distance threshold and filter again until at least one candidate block is found; wherein the physical spatial distance is calculated based on the image coordinate system; Extract the spatial coordinates and structural indices of all complete structural blocks in the current modal image; use the spatial coordinates as input and the vector of structural indices as output to establish a block feature prediction model through supervised learning; input the spatial coordinates of structural blocks with missing structural indices into the trained block feature prediction model and output the predicted structural indices. Calculate the feature similarity between candidate blocks and predicted structural indicators; The structural block with the highest feature similarity is selected as the optimal candidate block; the feature similarity is obtained by calculating the cosine similarity between the vectors of structural indices.

4. The machine learning-based method for assessing acute lung injury in critically ill patients according to claim 3, characterized in that: The process of identifying modal conflict zones includes: calculating the variance of structural indices of the same structural block in different modes; if the variance is greater than the conflict determination threshold, it is determined to be a modal conflict zone. The identification process of the structural mutation region includes: calculating the Euclidean distance difference between the structural index of the structural block and the adjacent structural block within a single mode; if the difference exceeds the preset mutation threshold, it is marked as a structural mutation region.

5. The machine learning-based method for assessing acute lung injury in critically ill patients according to claim 4, characterized in that: The calculation process for the disturbance impact value includes: For each type of structural indicator, the average value is calculated in all key scoring areas. The average values ​​of different types of structural indicators are weighted and summed according to preset weights to obtain the benchmark score. Within the key scoring area, a preset fixed proportion of positive and negative adjustments are applied to each type of structural indicator. The scoring value is calculated based on the adjusted structural indicator, resulting in positive and negative scoring values ​​respectively. The formula for calculating the disturbance impact value is: Disturbance impact value = ((|positive score value - baseline score value|) + (|negative score value - baseline score value|)) ÷ 2.

6. The machine learning-based method for assessing acute lung injury in critically ill patients according to claim 5, characterized in that: The calculation process of the consistency index includes: All modalities are paired, and the correlation coefficient between the structural indices of each pair is calculated. The consistency index is obtained by averaging all the correlation coefficients. The correlation coefficient is the ratio of the product of the covariance among modal structure indices to the standard deviation of each modal structure indices.

7. The machine learning-based method for assessing acute lung injury in critically ill patients according to claim 6, characterized in that: The calculation process for the reliability of each mode includes: For each modality, the absolute difference between the structural index and the manually labeled reference structural index is calculated. The average deviation value is obtained by averaging all the absolute differences in the same modality. The reciprocal of the average deviation value is then normalized and used as the modality confidence weight.

8. The machine learning-based method for assessing acute lung injury in critically ill patients according to claim 7, characterized in that: The calculation process of the dynamic weights includes: Based on each structural block in the key scoring region, the absolute value of the difference between the structural indicators after fusion between adjacent structural blocks is calculated to obtain the spatial variation; all spatial variations in the key scoring region are normalized to obtain the variation amplitude factor. Within the same critical scoring region, for structural indices before fusion under different modalities, their standard deviations are calculated to obtain consistency deviation values; these consistency deviations are then normalized to obtain consistency adjustment factors. The variation amplitude factor and consistency adjustment factor are weighted according to a set ratio to obtain the dynamic weight of each key scoring region.

9. The machine learning-based method for assessing acute lung injury in critically ill patients according to claim 8, characterized in that: The preset integrity conditions include: the total area of ​​the scoring key region accounts for a proportion of the lung region that is not less than a set lower limit, and it covers all predefined anatomical key points; The process of reconstructing the key scoring region includes: selecting several structural blocks whose spatial location is continuous with the edge of the key scoring region, and adding them to the current key scoring region to form an extended key scoring region.

10. A machine learning-based assessment system for acute lung injury in critically ill patients, characterized in that: Using a machine learning-based assessment method for acute lung injury in critically ill patients as described in any one of claims 1 to 9, comprising: The data acquisition module is used to acquire chest multimodal image data of critically ill patients and to perform standardization processing on the chest multimodal image data to generate a multimodal standard dataset. The data processing module is used to divide the lung region into several structural blocks, extract structural indicators within the structural blocks based on the standard dataset; if structural indicators are missing in any modality, spatial mapping compensation is preferentially performed on the same structural blocks in other modalities; if compensation is still not possible, the optimal candidate block is selected from the structural blocks in the current modality, and its structural indicators are extracted for compensation. The region extraction module is used to identify modal conflict regions and structural abrupt change regions in the structural blocks based on the structural indicators, and to merge the modal conflict regions and structural abrupt change regions to construct key scoring regions; The region optimization module is used to apply a preset perturbation to the structural indicators of the key scoring region and calculate the perturbation impact value. It determines whether the perturbation impact value is greater than the preset impact threshold. If it is, the key scoring region is retained; otherwise, the key scoring region is removed. The feature fusion module is used to calculate a consistency index based on the structural indicators of the same key scoring region in all modalities, and to determine whether the consistency index is greater than a preset consistency threshold. If it is, the structural indicators of the same key scoring region are fused by mean; otherwise, weighted fusion is performed based on the credibility of each modality. The injury assessment module is used to calculate the lung injury score of the critically ill patient based on the fused structural indicators and the dynamic weights of the structural indicators in the key scoring regions. The evaluation and optimization module is used to determine whether the key scoring region meets the preset integrity condition. If it does, the lung injury score is output; otherwise, the key scoring region is reconstructed and the lung injury score is calculated until the preset number of iterations is reached. If the integrity condition is still not met, the average value of the lung injury score during all iterations is output.

Images