A medical X-ray image analysis method and system based on deep learning

Abstract

The application provides a medical X-ray image analysis method and system based on deep learning, comprising: acquiring chest X-ray images of the same patient at different time points; identifying a lung nodule region from the chest X-ray images and acquiring two-dimensional image performance information of the nodule; inferring morphological information of the nodule in a three-dimensional space based on the two-dimensional image performance information of the nodule, combining imaging physical principles and anatomical prior knowledge; acquiring three-dimensional morphological features of the nodule from the inferred three-dimensional morphological information of the nodule; comparing the three-dimensional morphological features of the nodule acquired at different time points; judging pathological changes of the nodule according to the comparison result; and outputting the pathological change judgment result of the nodule. The application can effectively distinguish the real pathological changes of the lung nodule from the imaging false changes caused by non-pathological factors.

Description

Technical Field

[0001] This application relates to the field of image analysis technology, and more specifically, to a medical X-ray image analysis method and system based on deep learning. Background Technology

[0002] Medical X-ray image analysis systems play a crucial role in assisting radiologists in identifying pulmonary nodules. However, during long-term follow-up of pulmonary nodules, due to the inherent two-dimensional projection characteristics of X-ray images and subtle changes in non-pathological factors such as patient position, respiratory status, and X-ray projection angle, the system struggles to accurately distinguish between true pathological changes and pseudo-changes on the images. This complicates the physician's diagnosis and affects the system's practical application value. Summary of the Invention

[0003] This application provides a deep learning-based medical X-ray image analysis method and system to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application discloses a deep learning-based medical X-ray image analysis method for distinguishing between true pathological changes in pulmonary nodules and pseudo-pathological changes in imaging caused by non-pathological factors, including: Obtain chest X-ray images of the same patient taken at different time points; Identify lung nodule regions from the chest X-ray image and obtain two-dimensional imaging information of the nodules; Based on the two-dimensional image representation information of the nodule, combined with imaging physics principles and prior anatomical knowledge, the morphological information of the nodule in three-dimensional space is inferred. From the inferred three-dimensional morphological information of the nodule, the three-dimensional morphological features of the nodule are obtained; Compare the three-dimensional morphological features of the nodule obtained at different time points; Based on the comparison results, the pathological changes of the nodule can be determined; Output the pathological changes of the nodule.

[0005] Secondly, this application also discloses a deep learning-based medical X-ray image analysis system for distinguishing between true pathological changes in pulmonary nodules and false imaging changes caused by non-pathological factors. The system includes: The image acquisition module is used to acquire chest X-ray images of the same patient taken at different time points; The nodule recognition module is used to identify lung nodule regions from the chest X-ray image and obtain two-dimensional image representation information of the nodule; The three-dimensional morphology inference module is used to infer the morphological information of the nodule in three-dimensional space based on the two-dimensional image representation information of the nodule, combined with imaging physics principles and prior anatomical knowledge. The three-dimensional feature acquisition module is used to acquire the three-dimensional morphological features of the nodule from the inferred three-dimensional morphological information of the nodule; The feature comparison module is used to compare the three-dimensional morphological features of the nodule obtained at different time points; The pathological assessment module is used to determine the pathological changes of the nodule based on the comparison results. The results output module is used to output the pathological changes of the nodule.

[0006] Compared with the prior art, this application has at least the following beneficial effects: This application directly quantifies and compares the true three-dimensional attributes of nodules by comparing the three-dimensional morphological features acquired at different time points. This method is more robust and accurate than existing technologies that rely solely on two-dimensional image features. For example, a nodule whose two-dimensional projected shape may change under different projection angles may maintain a stable volume and major and minor axis dimensions in three-dimensional space. This application can capture this three-dimensional stability, thereby more reliably determining the pathological changes of nodules.

[0007] This application effectively overcomes the shortcomings of existing technologies in handling the two-dimensional projection characteristics of X-ray images and interference from non-pathological factors by introducing three-dimensional morphological inference and three-dimensional feature comparison. Therefore, it can more robustly and accurately distinguish between true pathological changes and radiographic pseudo-changes in pulmonary nodules, significantly improving the application value of medical X-ray image analysis systems in long-term follow-up of pulmonary nodules and providing radiologists with a more precise auxiliary diagnostic tool. Attached Figure Description

[0008] Figure 1 A flowchart illustrating a deep learning-based medical X-ray image analysis method provided in this application; Figure 2 This is a schematic diagram of the structure of a deep learning-based medical X-ray image analysis system provided in this application. Detailed Implementation

[0009] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0010] like Figure 1 As shown, this application proposes a deep learning-based medical X-ray image analysis method to distinguish between true pathological changes in pulmonary nodules and pseudo-pathological changes in imaging caused by non-pathological factors, including: Obtain chest X-ray images of the same patient taken at different time points; Identify pulmonary nodule regions from chest X-ray images and obtain two-dimensional imaging information of the nodules; Based on the two-dimensional image representation information of nodules, combined with imaging physics principles and prior anatomical knowledge, the morphological information of nodules in three-dimensional space is inferred. The three-dimensional morphological features of the nodules are obtained from the inferred three-dimensional morphological information of the nodules. Compare the three-dimensional morphological features of nodules obtained at different time points; Based on the comparison results, determine the pathological changes of the nodules; Output the pathological changes of the nodules.

[0011] This application, by introducing three-dimensional morphological inference and uncertainty quantification, can more accurately assess the true pathological changes of pulmonary nodules, effectively avoiding the problem of traditional two-dimensional analysis methods being susceptible to interference from non-pathological factors, thereby improving the robustness and reliability of diagnosis.

[0012] Chest X-ray images refer to two-dimensional images obtained by imaging a patient's chest using X-ray technology. They are typically used for preliminary screening and follow-up of lung diseases.

[0013] Pulmonary nodules are dense shadows that appear as focal, round or irregular shapes on chest X-ray images, and are usually less than 3 centimeters in size.

[0014] Two-dimensional image representation information refers to the visual features of nodules extracted from X-ray images, such as their projected area, shape, grayscale distribution, or edge features.

[0015] Imaging physics refers to the physical process by which X-rays attenuate to different degrees due to differences in density and atomic number when they penetrate human tissue, thus creating grayscale differences on the detector.

[0016] Prior anatomical knowledge refers to normal anatomical information about the lungs and surrounding structures, such as the location, shape, and interrelationships of the lobes, bronchi, blood vessels, ribs, etc.

[0017] Morphological information in three-dimensional space refers to the geometric shape, size, and position of nodules in real three-dimensional space, rather than their projection onto two-dimensional images.

[0018] Three-dimensional morphological characteristics refer to quantitative indicators used to describe the three-dimensional morphology of nodules, such as volume, major and minor axis dimensions, surface smoothness, and internal density uniformity.

[0019] Pathological changes refer to changes in the true morphology or tissue structure of nodules caused by disease progression (such as tumor growth or inflammation resolution).

[0020] Non-pathological imaging pseudo-changes refer to changes in the shape or size of nodules on images caused by non-disease factors such as imaging conditions (e.g., patient position, breathing, or projection angle) or image processing.

[0021] The following is a further detailed description of the deep learning-based medical X-ray image analysis method proposed in this application: First, the method includes acquiring chest X-ray images of the same patient taken at different time points. In practice, image acquisition can be achieved in several ways. For example, radiology technicians can manually retrieve and import historical X-ray image data of patients from the hospital's PACS (Picture Archiving and Communication System). Alternatively, an automated interface can be established with the PACS system, allowing the system to automatically download the required chest X-ray images in batches from the PACS based on the patient ID and examination date. Furthermore, traditional X-ray films can be digitized using a scanner, converted into digital image files, and then imported into the system for analysis.

[0022] Secondly, the method includes the steps of identifying lung nodule regions from chest X-ray images and obtaining two-dimensional image representation information of the nodules. In implementation, methods based on traditional image processing techniques can be used to identify nodule regions. For example, algorithms such as threshold segmentation, region growing, or edge detection can be used, combined with morphological operations (such as opening and closing operations) to initially delineate the regions suspected of having nodules. For obtaining the two-dimensional image representation information, geometric features such as the projected area, perimeter, circularity, or rectangularity of these initially identified regions can be calculated, as well as gray-level statistical features such as the average gray value, standard deviation, skewness, and kurtosis.

[0023] Furthermore, this method includes a step of inferring the morphological information of the nodule in three-dimensional space based on the two-dimensional image representation information of the nodule, combined with imaging physics principles and prior anatomical knowledge. In specific implementation, a rule-based reasoning method can be adopted. For example, a series of rules can be preset: if the nodule is elliptical and has clear edges in the two-dimensional image, its three-dimensional shape may be spherical or ellipsoidal; if the nodule has blurred edges and overlaps with the ribs, the uncertainty of its three-dimensional shape inference is high. These rules can be constructed by combining the projection geometry of X-ray imaging and lung anatomy (such as the orientation of blood vessels or bronchi). Through these rules, the possible morphology of the nodule in three-dimensional space can be inferred from the two-dimensional image features.

[0024] Next, the method includes the step of obtaining the three-dimensional morphological features of the nodule from the inferred three-dimensional morphological information. In practice, geometric parameters can be directly extracted from the inferred three-dimensional morphological information. For example, if the nodule is inferred to be approximately spherical, its diameter can be calculated; if it is inferred to be ellipsoidal, its major axis, minor axis, and median axis lengths can be calculated. Furthermore, features such as its three-dimensional volume, surface area, or density can also be calculated. These features can be calculated based on simple geometric formulas or obtained through voxel counting or surface mesh analysis of the three-dimensional morphological model.

[0025] The method then includes a step of comparing the three-dimensional morphological features of nodules acquired at different time points. In practice, direct numerical comparison can be used. For example, the three-dimensional volumes of nodules acquired at different time points can be subtracted to obtain the volume change; the major and minor axis dimensions of nodules acquired at different time points can be compared to determine whether they have increased or decreased. Alternatively, the percentage changes in these features can be calculated to standardize the assessment of the degree of change.

[0026] Finally, the method includes steps for judging the pathological changes of the nodules based on the comparison results and outputting the judgment results. In implementation, a series of judgment thresholds can be preset. For example, if the three-dimensional volume of the nodule increases by more than 20% or the absolute volume increases by more than a certain threshold (e.g., 50 cubic millimeters), it is judged as pathological progression; if the volume shrinks by more than 20%, it is judged as pathological regression. If the changes are within the preset threshold range, it is judged as stable. The judgment results can be output to the user interface in text form (e.g., nodule stable, nodule enlarged, or nodule shrank), or integrated into the PACS system in the form of a structured report.

[0027] The deep learning-based medical X-ray image analysis method proposed in this application works by integrating and analyzing multi-dimensional information to overcome the limitations of traditional two-dimensional X-ray image analysis, thereby achieving accurate differentiation of the true pathological changes in pulmonary nodules.

[0028] Specifically, this method first acquires chest X-ray images of the same patient taken at different time points, providing foundational data for subsequent dynamic analysis. These images contain two-dimensional projection information of the nodules at different time points.

[0029] The subsequent steps of identifying pulmonary nodule regions from chest X-ray images and obtaining two-dimensional imaging information of the nodules aim to accurately locate the nodules from the complex chest background and extract their visible features in a two-dimensional plane. These two-dimensional features serve as the starting point for further inference of their three-dimensional morphology.

[0030] By combining two-dimensional image features with the physical processes of X-ray imaging and the anatomical structure of the lungs, the system can infer the possible morphology of a nodule in real three-dimensional space from a two-dimensional projection. For example, a nodule that appears flat on a two-dimensional image may be inferred as the projection of a spherical nodule at a specific angle, considering its overlap with the ribs and the X-ray projection angle. This inference of three-dimensional morphology greatly reduces the ambiguity introduced by two-dimensional projection.

[0031] From the inferred three-dimensional morphological information of the nodule, the three-dimensional morphological features of the nodule are obtained. This step transforms the abstract three-dimensional morphology into quantifiable indicators, such as volume or major and minor axis dimensions. These three-dimensional features more accurately reflect the actual size and shape of the nodule than two-dimensional features, and are less affected by non-pathological factors (such as changes in the projection angle).

[0032] Subsequently, the step of comparing the three-dimensional morphological features of the nodules acquired at different time points directly compares the true three-dimensional attributes of the nodules, rather than merely comparing two-dimensional projections. This comparison can more accurately reflect the actual growth or regression of the nodules and effectively filter out false changes in imaging caused by non-pathological factors such as patient position and respiration.

[0033] Ultimately, based on the comparison results, the steps for judging the pathological changes of nodules and outputting the judgment results of pathological changes are based on robust three-dimensional morphological feature comparison, which can provide more reliable judgments of pathological changes. For example, only when the three-dimensional volume or major and minor axis dimensions of the nodule have indeed changed significantly after considering the uncertainty will it be judged as a pathological change, thereby avoiding misjudgment of false changes.

[0034] Overall, this application constructs a more robust and accurate framework for analyzing pulmonary nodules by introducing an inference mechanism from two-dimensional images to three-dimensional morphology and comparing three-dimensional features on this basis. This method enables the system to gain a deeper understanding of the true changes in nodules, thereby providing a more reliable basis for clinical diagnosis and effectively solving the problem of misjudgment of false changes caused by non-pathological factors in existing technologies.

[0035] This application fundamentally solves the problems of information loss and ambiguity caused by two-dimensional projection by inferring the three-dimensional morphology of nodules. For example, in the prior art, a nodule whose projected area on a two-dimensional image is slightly increased due to changes in the patient's respiratory depth may be misjudged as a real growth. However, in this application, by combining imaging physics principles and prior anatomical knowledge, the system can identify that such two-dimensional changes are caused by non-pathological factors and infer that the true three-dimensional morphology of the nodule has not changed significantly, thus avoiding misjudgment.

[0036] Furthermore, this application directly quantifies and compares the true three-dimensional attributes of nodules by comparing the three-dimensional morphological features of nodules acquired at different time points. This method is more robust and accurate than the comparison based solely on two-dimensional image features in existing technologies. For example, a nodule whose two-dimensional projected shape may change under different projection angles may maintain a stable volume and major and minor axis dimensions in three-dimensional space. This application can capture this three-dimensional stability, thereby more reliably determining the pathological changes of nodules.

[0037] In summary, this application effectively overcomes the shortcomings of existing technologies in handling the two-dimensional projection characteristics and non-pathological interference of X-ray images by introducing three-dimensional morphological inference and three-dimensional feature comparison. This method can more robustly and accurately distinguish between true pathological changes and radiographic pseudo-changes in pulmonary nodules, significantly improving the application value of medical X-ray image analysis systems in long-term follow-up of pulmonary nodules and providing radiologists with a more precise auxiliary diagnostic tool.

[0038] In some embodiments, the step of inferring the morphological information of the nodule in three-dimensional space based on the two-dimensional image representation information of the nodule, combined with imaging physics principles and prior anatomical knowledge, includes: Two-dimensional projection regions of suspected pulmonary nodules in chest X-ray images are identified, and two-dimensional image representation information of the nodules is extracted. The two-dimensional image representation information includes nodule coarse region identification information, local texture feature information, gray-level gradient and contrast pattern information, and boundary feature information with surrounding dense structures. Based on the two-dimensional image representation information of the nodules, combined with the pre-set lung anatomy knowledge and the physical principles of X-ray imaging, the probabilistic morphology of the nodules in three-dimensional space is inferred from the two-dimensional image representation information. The inference process includes a pre-set three-dimensional morphology template library, candidate morphology generation and projection simulation, morphology probability assignment and expected three-dimensional feature calculation. The uncertainty of the inferred three-dimensional probabilistic morphology of nodules is quantified. Uncertainty quantification takes into account the fuzziness of two-dimensional projection, the degree of overlap with dense structures, and the inherent limitations of inference from a single image. Based on the three-dimensional probabilistic morphology and degree of uncertainty of the nodules, the three-dimensional morphological characteristics of the nodules are inferred.

[0039] The identification of two-dimensional projection regions of suspected pulmonary nodules in chest X-ray images refers to the preliminary localization of potentially nodule-containing two-dimensional regions in X-ray images using image processing and pattern recognition techniques. Extracting two-dimensional image representation information of the nodules involves acquiring a series of data describing the characteristics of the nodules in the two-dimensional images from these identified regions. Specifically, the coarse nodule region identification information can be understood as the approximate location and extent of the nodule in the image; local texture feature information refers to the gray-level variation patterns and structural details within the nodule region, such as smoothness or roughness; gray-level gradient and contrast pattern information reflects the clarity of the nodule edges and the gray-level differences with surrounding tissues; and boundary feature information with surrounding dense structures refers to the image representation when the nodule is adjacent to high-density structures such as blood vessels or bones, such as the degree of overlap or boundary fusion. This information collectively constitutes a comprehensive description of the two-dimensional image of the nodule, providing a foundation for subsequent three-dimensional inference.

[0040] Based on the two-dimensional image representation of the nodule, combined with pre-defined lung anatomy and the physical principles of X-ray imaging, the probabilistic morphology of the nodule in three-dimensional space is inferred. Here, probabilistic morphology refers to the multiple possible shapes of the nodule in three-dimensional space and their corresponding probability distributions, rather than a single, definitive shape. This inference process aims to address the inherent information loss in two-dimensional projection. Specifically, a pre-defined three-dimensional morphology template library contains various common three-dimensional morphological models of lung nodules, such as spherical, ellipsoidal, lobulated, or spiky shapes. Candidate morphology generation and projection simulation involve selecting or generating a series of possible three-dimensional nodule morphologies from the template library and simulating the two-dimensional projection images produced by these morphologies under X-ray irradiation. Morphological probability assignment involves comparing the simulated projection with the actual two-dimensional image and assigning a probability value to each candidate three-dimensional morphology based on the degree of matching, representing its likelihood of being a true nodule morphology. Expected three-dimensional feature calculation involves calculating the expected three-dimensional volume and expected three-dimensional major and minor axis dimensions of the nodule based on these probability values ​​and the corresponding three-dimensional morphology.

[0041] Quantifying the uncertainty of the inferred three-dimensional probabilistic morphology of the nodule is a key aspect of this scheme. Uncertainty quantification aims to assess the reliability of the three-dimensional morphology inference results. Specifically, the ambiguity of the two-dimensional projection refers to the possibility that the nodule edges in the X-ray image may be unclear, making it difficult to accurately define its two-dimensional contour; the degree of overlap with dense structures refers to the possibility that the nodule may overlap with high-density structures such as blood vessels or ribs in the two-dimensional image, making the nodule boundary difficult to distinguish; the inherent limitation of single-image inference means that the depth information of the nodule cannot be fully recovered from a single two-dimensional X-ray image, and there is a possibility that multiple three-dimensional morphologies correspond to the same two-dimensional projection. By comprehensively considering these factors, a quantified uncertainty score can be obtained, reflecting the confidence level of the three-dimensional morphology inference results.

[0042] Therefore, based on the three-dimensional probabilistic morphology and degree of uncertainty of the nodules, the three-dimensional morphological features of the nodules are obtained. This means that the final obtained three-dimensional morphological features not only contain the most likely morphological information of the nodules, but also include their reliability assessment, making the description of the nodule morphology more comprehensive and robust.

[0043] This application first identifies and extracts detailed two-dimensional image information of nodules, providing rich and multi-dimensional input for subsequent three-dimensional inference. Furthermore, by combining pre-defined lung anatomy knowledge and the physical principles of X-ray imaging, a pre-defined three-dimensional morphology template library is used to generate candidate morphologies and simulate projections. These morphologies are then assigned probabilistic values, enabling the inference of the probabilistic morphology of the nodule in three-dimensional space from the two-dimensional image. This probabilistic inference method effectively addresses the inherent ambiguity and information gaps in two-dimensional images. More importantly, this application quantifies the uncertainty of the inferred three-dimensional probabilistic morphology of the nodule, considering the ambiguity of two-dimensional projection, the degree of overlap with dense structures, and the inherent limitations of single-image inference, thus objectively assessing the reliability of the inference results. Therefore, when acquiring the three-dimensional morphological features of the nodule, not only its most probable morphology is considered, but its uncertainty range is also included, making the final three-dimensional morphological features more robust and reliable.

[0044] In some embodiments, the step of quantifying the degree of uncertainty in the inferred three-dimensional probabilistic morphology of the nodule includes: To obtain imaging parameter information of chest X-ray images and information on the patient's respiratory status; Identify stable anatomical landmarks in chest X-ray images and obtain information on the relative positional changes of anatomical landmarks in images at different time points; The calculation results of morphological distribution entropy are calibrated based on imaging parameter information, patient respiratory status information, and relative positional changes of anatomical landmarks in images at different time points. The calculation results of the contribution of two-dimensional feature ambiguity are calibrated based on imaging parameter information, patient respiratory status information, and relative position changes of anatomical landmarks in images at different time points. Based on the calibrated morphological distribution entropy and the calibrated two-dimensional feature ambiguity contribution, the comprehensive uncertainty score of the three-dimensional probabilistic morphology of the nodule is inferred.

[0045] Specifically, acquiring imaging parameter information for chest X-ray images refers to obtaining data such as tube voltage, tube current, exposure time, focal spot size, detector type, and acquisition angle used during image acquisition. These parameters directly affect image contrast, spatial resolution, and noise level, thus impacting the accuracy of extracting two-dimensional image representation information of nodules. Patient respiratory status information can be understood as the patient's inspiratory depth, expiratory depth, or breath-holding state during image acquisition. Respiratory movements cause displacement and deformation of lung structures and nodules in the image, introducing additional ambiguity and uncertainty.

[0046] Identifying stable anatomical landmarks in chest X-ray images and obtaining information on their relative positional changes at different time points refers to automatically or semi-automatically identifying relatively stable anatomical structures in the lung region, such as ribs, vertebrae, tracheal bifurcation, or the edge of the cardiac silhouette, using image processing technology. The relative positional changes of these anatomical landmarks at different time points can serve as an objective basis for assessing changes in patient position, respiratory movement amplitude, and the degree of geometric distortion that may exist during image acquisition.

[0047] Based on imaging parameters, patient respiratory status, and relative positional changes of anatomical landmarks, the calculation results of morphological distribution entropy are calibrated. The purpose is to incorporate external uncertainties during image acquisition into the assessment of the inherent uncertainty in the three-dimensional probabilistic morphology of nodules. For example, when imaging parameters result in high image noise or the patient's respiratory movements are large, the dispersion of the inferred three-dimensional morphology of the nodule (i.e., morphological distribution entropy) should be increased accordingly to more accurately reflect its inherent uncertainty. Similarly, the calculation results of two-dimensional feature ambiguity contribution are calibrated to quantify the impact of image quality on the reliability of two-dimensional feature extraction. For example, if the image is significantly blurred due to respiratory movements, the contribution of two-dimensional feature ambiguity should be amplified to reflect its negative impact on three-dimensional morphological inference.

[0048] This application calibrates the uncertainty of the inferred three-dimensional probabilistic morphology of nodules by acquiring and utilizing imaging parameter information, patient respiratory status information, and information on the relative positional changes of stable anatomical landmarks. It is precisely because these external factors are explicitly incorporated into the uncertainty quantification model that the calculation results of morphological distribution entropy and the contribution of two-dimensional feature ambiguity can more comprehensively and accurately reflect the true reliability of the inferred three-dimensional morphology of nodules. Therefore, misjudgments caused by fluctuations in image quality or changes in the patient's physiological state can be effectively avoided, thereby improving the robustness of the entire analysis method.

[0049] The following is a specific example to illustrate this.

[0050] Suppose we are analyzing two chest X-ray images of the same patient taken at different time points. The first image was acquired under standard imaging parameters while the patient was holding their breath well; the second image was acquired under slightly different exposure parameters while the patient was exhibiting slight respiratory movements. Without the calibration mechanism described in this application, the system might directly calculate the morphological distribution entropy and the contribution of two-dimensional feature ambiguity based on the two-dimensional image representation information. It might mistakenly assume that the uncertainty of the second image is similar to that of the first image due to slight ambiguity, or even misjudge that the nodules have changed due to minor changes in certain features.

[0051] However, when using the scheme of this application, the system first acquires the imaging parameter information of each of the two images and the patient's respiratory status information. For example, the exposure time of the second image may be slightly longer, and the respiratory status shows slight respiratory movements. Simultaneously, the system identifies stable anatomical landmarks (such as ribs) in both images and finds that the relative positional changes of these landmarks in the second image are slightly greater than in the first image, further confirming the presence of respiratory movements. Based on this information, the system calibrates the morphological distribution entropy and two-dimensional feature ambiguity contribution of the second image. Specifically, due to the difference in imaging parameters and the influence of respiratory movements, the morphological distribution entropy and two-dimensional feature ambiguity contribution of the second image are appropriately increased to reflect the higher uncertainty of the inferred three-dimensional morphological information. Finally, by combining the calibrated morphological distribution entropy and two-dimensional feature ambiguity contribution, a more accurate comprehensive uncertainty score can be obtained. This score will clearly indicate that the reliability of 3D morphological inference from the second image is lower than that from the first image, thus enabling a more cautious assessment of changes when comparing the three-dimensional morphological features of nodules at different time points, avoiding misjudging changes caused by differences in image acquisition conditions as true pathological changes.

[0052] In some embodiments, the step of inferring the three-dimensional morphological features of the nodule based on the three-dimensional probabilistic morphology and degree of uncertainty includes: Calculate the expected three-dimensional volume and expected three-dimensional major and minor axis dimensions of the nodule; Based on the degree of uncertainty of the nodules, a fluctuation range is set for the expected three-dimensional volume and the expected three-dimensional major and minor axis dimensions, respectively, to obtain the range of the expected three-dimensional volume and the range of the expected three-dimensional major and minor axis dimensions; Compare the expected three-dimensional volume range and the expected three-dimensional major and minor axis size range obtained at different time points; Based on the comparison results of the expected three-dimensional volume range and the expected three-dimensional major and minor axis size range, and combined with the degree of uncertainty, the three-dimensional morphological characteristics of the nodule are inferred.

[0053] Specifically, when acquiring the three-dimensional morphological features of a nodule, it is first necessary to calculate the expected three-dimensional volume and expected three-dimensional major and minor axis dimensions of the nodule. The expected three-dimensional volume refers to the most probable three-dimensional volume value calculated using methods such as weighted averaging based on the nodule's three-dimensional probabilistic morphological distribution. The expected three-dimensional major and minor axis dimensions refer to the length and width of the nodule in three-dimensional space along its main axis directions, also calculated using a weighted approach based on the probabilistic morphological distribution. These expected values ​​represent the most probable three-dimensional morphological parameters of the nodule under the current image information.

[0054] Considering the ambiguity of two-dimensional projection, the degree of overlap with dense structures, and the inherent limitations of inference from a single image, which lead to uncertainty in the three-dimensional probabilistic morphology of nodules, it is necessary to set a fluctuation range for the aforementioned expected three-dimensional volume and expected three-dimensional major and minor axis dimensions. This fluctuation range is set based on the quantified degree of uncertainty; for example, the higher the degree of uncertainty, the larger the fluctuation range. By setting the fluctuation range, the intervals for the expected three-dimensional volume and the expected three-dimensional major and minor axis dimensions can be obtained. These intervals reflect the possible value ranges of the nodule's three-dimensional morphological parameters, thus more comprehensively characterizing the nodule's morphological features, rather than a single precise value.

[0055] To determine the pathological changes in a nodule, it is necessary to compare the expected three-dimensional volume range and the expected three-dimensional major and minor axis size range obtained at different time points. This comparison can be achieved using methods such as range overlap analysis or range distance measurement. For example, if the ranges at two time points do not overlap at all, it indicates a significant change in nodule morphology; if the ranges have a high degree of overlap, it may mean that the morphological change is not significant or falls within the range of uncertainty.

[0056] Based on the comparison results of the expected three-dimensional volume range and the expected three-dimensional major and minor axis size range, and combined with the degree of uncertainty, the three-dimensional morphological characteristics of the nodule are inferred. For example, when the range comparison results show a significant increase in the nodule volume or size range and a low degree of uncertainty, it can be determined that the nodule has undergone genuine growth changes. Conversely, if the range change is within the uncertainty range, it may be considered a false change in imaging.

[0057] This application introduces expected three-dimensional volume and expected three-dimensional major and minor axis dimensions, and sets fluctuation ranges to form intervals. This effectively addresses the uncertainties arising from two-dimensional projection ambiguity, overlap with dense structures, and the inherent limitations of inference from a single image when acquiring three-dimensional morphological features of nodules. The aforementioned basic scheme may rely solely on comparisons of a single expected value, making it susceptible to interference from uncertainties and leading to misjudgments of the true pathological changes in nodules. By quantifying and transforming uncertainty into fluctuation ranges of morphological parameters, this scheme can more comprehensively reflect the possible range of nodule morphology. When comparing these intervals at different time points, not only the central trend of the morphology is considered, but also its variability. Only when the intervals of morphological parameters undergo significant and non-overlapping changes are they considered true pathological changes, thus avoiding false changes caused by image noise or inference errors. This interval-based comparison method makes the judgment of three-dimensional morphological features of nodules more robust and reliable.

[0058] In some embodiments, the steps of identifying pulmonary nodule regions from chest X-ray images and obtaining two-dimensional imaging information of the nodules include: Chest X-ray images are processed to optimize image quality; Multiscale analysis was performed on the processed images to identify the two-dimensional projection regions of suspected pulmonary nodules; For the two-dimensional projection region of the identified suspected lung nodules, the uniformity of its local gray-level distribution is calculated, and the gradient change at its edge is analyzed. When the edge of the two-dimensional projection area of ​​a suspected pulmonary nodule overlaps with the edge of a dense anatomical structure, the nodule and the dense structure are distinguished based on the gray value and texture features of the overlapping area. The segmentation threshold is adjusted based on the local gray-scale distribution of the nodules, edge gradient changes, and differentiation results. Based on the adjusted segmentation threshold, the two-dimensional projection contour of the nodule is drawn; Based on the drawn two-dimensional projection outline, the projected area and shape irregularity of the nodule are calculated as the two-dimensional image representation information of the nodule.

[0059] Acquiring chest X-ray images refers to receiving raw X-ray image data from imaging equipment or storage systems. Processing chest X-ray images to optimize image quality includes operations such as noise reduction, contrast enhancement, and artifact removal. The aim is to improve image clarity and analyzability, providing high-quality input for subsequent nodule identification. Multi-scale analysis of the processed images involves scanning and extracting features using different spatial resolutions or filter sizes. The goal is to effectively identify the two-dimensional projection regions of suspected lung nodules of different sizes and shapes.

[0060] For the two-dimensional projection region of a suspected lung nodule, the uniformity of its local gray-level distribution is calculated, which can be understood as assessing the degree of fluctuation in pixel gray-level values ​​within the region. Analyzing the gradient changes at its edges refers to detecting the steepness and direction of gray-level value changes at the boundary of the region, aiming to preliminarily characterize the internal density features and boundary clarity of the nodule. When the edge of the two-dimensional projection region of a suspected lung nodule overlaps with the edge of a dense anatomical structure, such as a blood vessel or rib, the gray-level values ​​and texture features of the overlapping area are used to distinguish the nodule from the dense structure. This aims to avoid misidentifying non-nodular structures as nodules or misclassifying a portion of a nodule as other structures.

[0061] Based on the local grayscale distribution, edge gradient changes, and differentiation results of the nodules, the segmentation threshold is adjusted. This involves dynamically setting a grayscale threshold to separate the nodules from the background, taking into account both the nodule's inherent image characteristics and its differentiation from surrounding structures. The aim is to achieve more accurate nodule boundary segmentation. Based on the adjusted segmentation threshold, a two-dimensional projection contour of the nodule is drawn. This involves binarizing the image using this threshold to delineate the nodule's precise boundaries. Finally, based on the drawn two-dimensional projection contour, the nodule's projected area and shape irregularity are calculated as two-dimensional image representation information. This provides quantified two-dimensional morphological features for subsequent analysis and comparison.

[0062] This application utilizes systematic preprocessing and multi-scale analysis of chest X-ray images to effectively identify the two-dimensional projection regions of suspected pulmonary nodules. Especially in complex cases where nodule edges overlap with dense anatomical structures, meticulous analysis of the grayscale values ​​and texture features of the overlapping areas accurately distinguishes nodules from non-nodule structures, thus avoiding missegmentation. Furthermore, by combining the local grayscale distribution and edge gradient changes of the nodule, the segmentation threshold is adaptively adjusted to ensure accurate depiction of the nodule's two-dimensional projection contour, ultimately obtaining reliable two-dimensional image representation information.

[0063] In some embodiments, when the edge of the two-dimensional projection region of a suspected pulmonary nodule overlaps with the edge of a dense anatomical structure, the step of distinguishing the nodule from the dense structure based on the grayscale value and texture features of the overlapping region includes: Obtain the grayscale distribution information of the overlapping region; Analyze the edge gradient direction of the overlapping region; Calculate the local texture periodicity of overlapping regions; Assess the connectivity between the overlapping region and the main region of the nodule; Assess the connectivity between the overlapping regions and the main dense structural regions; and Based on the gray value distribution information of the overlapping region, the edge gradient direction of the overlapping region, the local texture periodicity of the overlapping region, the connectivity between the overlapping region and the main region of the nodule, and the connectivity between the overlapping region and the main region of the dense structure, nodules and dense structures can be distinguished.

[0064] Obtaining grayscale distribution information in overlapping regions involves statistically analyzing the pixel grayscale values ​​within these regions. This includes calculating grayscale histograms, average grayscale values, or grayscale variances to characterize the density and uniformity of the region. The purpose is to preliminarily identify whether tissue components of varying densities exist within the overlapping regions.

[0065] Analyzing the edge gradient direction of overlapping regions involves determining the direction and intensity of edges by calculating the gradient vectors of pixels within the overlapping region. For example, operators like Sobel, Prewitt, or Canny can be used to detect edges and analyze their directional distribution. The aim is to reveal the boundary features between different structures within the overlapping region; for instance, nodule edges are typically smoother, while blood vessel or bone edges may be sharper and more directional.

[0066] Calculating the local texture periodicity of overlapping regions refers to quantifying the repeatability and regularity of texture within these regions using methods such as Fourier transform, wavelet analysis, or gray-level co-occurrence matrix. For example, vascular structures may exhibit certain periodic or directional textures, while the texture within nodules is usually more random or uniform. The aim is to further distinguish nodules from surrounding dense structures using texture features.

[0067] Assessing the connectivity between overlapping regions and the main nodule region involves analyzing the spatial continuity and similarity in grayscale and texture between the overlapping regions and the identified main nodule portion. For example, the degree of connectivity can be determined through morphological operations (such as dilation or erosion) or connected component analysis. The aim is to determine whether the overlapping region is an extension of the nodule.

[0068] Assessing the connectivity between overlapping regions and the main dense anatomical regions involves analyzing the spatial continuity and characteristic similarity between the overlapping regions and surrounding known dense anatomical structures (such as blood vessels or ribs). The aim is to determine whether the overlapping regions are part of these dense structures.

[0069] This application, by comprehensively considering the grayscale distribution information, edge gradient direction, local texture periodicity, and connectivity with the main nodule and dense structure regions of the overlapping area, can more precisely analyze the complex structure of the overlapping area. Methods relying solely on grayscale values ​​and texture features often struggle to accurately distinguish between nodules and dense structures such as blood vessels and ribs commonly found in X-ray images, leading to inaccurate nodule segmentation or false positives. By introducing edge gradient direction and local texture periodicity, the system can capture subtle morphological and textural differences between different structures. Furthermore, by evaluating the connectivity of the overlapping area with the main nodule and dense structure regions, it further utilizes the spatial continuity information of the structure, thereby effectively distinguishing nodules from non-nodular dense structures.

[0070] In some embodiments, the steps of distinguishing nodules from dense structures based on the gray value distribution information of the overlapping region, the edge gradient direction of the overlapping region, the local texture periodicity of the overlapping region, the connectivity between the overlapping region and the main body region of the nodule, and the connectivity between the overlapping region and the main body region of the dense structure include: The overlapping regions are further subdivided into multiple sub-regions; For each sub-region, the gray value distribution characteristics, edge gradient direction characteristics, and texture periodicity characteristics within it are calculated to identify tiny dense structures in the sub-region that correspond to the characteristics of blood vessel branches or fine bone structures. Based on the spatial distribution and connectivity of the tiny dense structures, they are separated from the overlapping regions; Assess the connectivity between the separated overlapping region and the main nodule region; Assess the connectivity between the separated overlapping region and the main dense structural region; and Based on the gray value distribution information of the separated overlapping regions, the edge gradient direction, the local texture periodicity, the connectivity between the separated overlapping regions and the main region of the nodules, and the connectivity between the separated overlapping regions and the main region of the dense structure, nodules and dense structures can be distinguished.

[0071] Specifically, subdividing the overlapping region into multiple sub-regions refers to dividing the original nodule and dense structure overlap area into smaller, independently analyzable local units. For example, methods such as mesh generation, superpixel segmentation, or adaptive segmentation based on local features can be used to decompose the overlapping region into several sub-regions with relatively uniform characteristics. The aim is to capture and process the heterogeneity within the overlapping region more precisely.

[0072] Specifically, for each sub-region, the gray-level distribution characteristics, edge gradient direction characteristics, and texture periodicity characteristics are calculated. This can be understood as extracting local image features for each subdivided sub-region. Specifically, gray-level distribution characteristics can include average gray-level values, gray-level histograms, or variance; edge gradient direction characteristics can include the edge direction distribution detected by Sobel, Prewitt, or Canny operators; and texture periodicity characteristics can be calculated using methods such as Gabor filters, Local Binary Patterns (LBP), or Gray-Level Co-occurrence Matrix (GLCM). The aim is to provide detailed local imaging data for subsequent identification of small, dense structures.

[0073] Identifying minute, dense structures within a subregion that correspond to the features of vascular branches or fine skeletal structures specifically involves using pre-defined anatomical knowledge and image pattern recognition techniques to locate structures with specific morphology, density, and connectivity patterns within that subregion. For example, vascular branches typically exhibit linear, branching, or tree-like structures with relatively high and continuous grayscale values; fine skeletal structures may present as denser, more sharply defined linear or dot-like structures. The aim is to accurately identify and locate these non-nodular, dense structures.

[0074] Separating small, dense structures from overlapping regions based on their spatial distribution and connectivity involves removing them from the original overlapping region after identification using image processing techniques. For example, morphological operations, connected component analysis, or graph-cut-based methods can be employed to peel these identified small, dense structures from the overlapping region, resulting in a cleaned-up overlapping region. The aim is to eliminate the interference of these non-nodular dense structures on nodule morphology assessment.

[0075] Assessing the connectivity between the separated overlapping regions and the main nodule region, as well as the connectivity between the separated overlapping regions and the main dense structure region, involves re-analyzing the spatial proximity and connectivity between the remaining overlapping regions and the main nodule and dense structure regions after removing the small dense structures. For example, Euclidean distance, the number of connected components, or graph-theory-based connectivity indices can be calculated. The aim is to more accurately determine the affiliation of the remaining overlapping regions after eliminating interfering factors.

[0076] Distinguishing between nodules and dense structures based on the grayscale distribution, edge gradient direction, local texture periodicity, connectivity between the separated overlapping regions and the main nodule region, and connectivity between the separated overlapping regions and the main dense structure region, involves using updated image features and connectivity assessment results after the aforementioned fine processing to make a final classification of the remaining overlapping regions. The aim is to achieve accurate delineation of the boundaries between nodules and dense structures.

[0077] This application achieves more refined feature analysis of local areas by subdividing the overlapping region, enabling the capture of minute structural features that may be overlooked by traditional methods. Secondly, by explicitly identifying and separating minute dense structures such as vascular branches or fine skeletal structures, it effectively eliminates the interference of these non-nodular factors in nodule boundary determination. It is precisely because these minute dense structures are accurately identified and removed that the subsequent assessment of the connectivity between the separated overlapping region and the main body of the nodule and dense structure is more accurate, thus avoiding misjudgments caused by the interweaving of blood vessels or bones with the nodule. This step-by-step and refined processing fundamentally improves the robustness and accuracy of distinguishing nodules from dense structures in complex overlapping regions.

[0078] The following is a specific example to illustrate this.

[0079] Suppose that in an acquired chest X-ray image, the two-dimensional projection region of a suspected pulmonary nodule partially overlaps with a pulmonary blood vessel. First, this overlapping region is subdivided into multiple sub-regions, for example, by segmenting it into several small blocks with similar characteristics using a watershed algorithm based on gray-level gradients. Next, for each sub-region, its average gray value, edge orientation histogram, and local texture periodicity are calculated. For example, if a sub-region exhibits linear, high-gray-level characteristics with strong texture periodicity, and its edge gradient direction aligns with the direction of the blood vessel, this sub-region is likely to be identified as part of a blood vessel branch. Once these blood vessel branches are identified, the system separates them from the original overlapping region based on their spatial distribution and connectivity with surrounding pixels, for example, by setting a connectivity threshold to mark sub-regions connected to the main body of the blood vessel as blood vessels. After the blood vessel structure is separated, the remaining overlapping region (i.e., the purified overlapping region) is re-evaluated for its connectivity with the main nodule region and the main dense structure region (now excluding blood vessels). For example, if the separated overlapping region has high connectivity with the main nodule region but low connectivity with the dense structure region, the remaining region will ultimately be identified as part of the nodule. In this way, even if the nodule is closely adjacent to or partially intertwined with a blood vessel, the true boundary of the nodule can be more accurately depicted, avoiding misidentification of blood vessels as nodule edges, thereby improving the accuracy of nodule segmentation.

[0080] In some embodiments, the step of identifying minute, dense structures in a subregion that correspond to features of vascular branches or fine bone structures includes: Obtain grayscale value distribution information, edge gradient direction information, and local texture periodicity information of the sub-region; Based on the grayscale value distribution information, analyze the density change trend within the sub-region; Analyze the edge orientation and connectivity of sub-regions based on edge gradient direction information; Based on the periodicity information of local textures, analyze the texture patterns within sub-regions; Based on density variation trends, edge orientation and connectivity, as well as texture patterns, we can identify tiny dense structures in sub-regions that are arranged in a linear, branching or mesh-like pattern. Based on the degree of matching between the micro-dense structure and the spatial morphology and connection pattern of the preset vascular branch or fine skeletal structure, it is determined whether the micro-dense structure is a vascular branch or fine skeletal structure. When tiny, dense structures are identified as vascular branches or fine skeletal structures, they are separated from the sub-regions.

[0081] Specifically, acquiring grayscale distribution information, edge gradient direction information, and local texture periodicity information of a sub-region refers to quantifying the image features of the sub-region using image processing algorithms, such as histogram analysis, Sobel or Canny operators for edge detection, and Gabor filters or wavelet transforms for texture analysis. Grayscale distribution information reflects the statistical characteristics of pixel brightness within the region and can be used to assess structure density; edge gradient direction information reveals the orientation and intensity of structural boundaries, helping to identify linear or branching structures; and local texture periodicity information describes the repetitiveness or regularity of pixel arrangement within the region, which is crucial for distinguishing the smooth texture of blood vessels from the coarse texture of bones.

[0082] Analyzing the density variation trend within a sub-region based on grayscale value distribution information can be understood as determining whether a density gradient or density non-uniformity exists within the sub-region by calculating statistical measures such as the mean, variance, skewness, or kurtosis of grayscale values, or by analyzing the shape of the grayscale histogram. For example, blood vessels typically exhibit a trend of high density at the center and gradually decreasing density at the edges. Analyzing the edge orientation and connectivity of a sub-region based on edge gradient direction information involves statistically analyzing the gradient direction of edge pixels to identify edges with specific directions (such as linear or branching) and assessing whether these edges are interconnected to form a continuous structure. Analyzing the texture patterns within a sub-region based on local texture periodicity information can be understood as identifying whether periodic or non-periodic texture patterns exist within the sub-region by calculating texture features (such as energy, contrast, correlation, or entropy). For example, the texture of blood vessels may exhibit a certain directionality and smoothness, while the texture of bones may be coarser and more irregular.

[0083] Based on density variation trends, edge orientation and connectivity, and texture patterns, the system identifies minute, dense structures arranged linearly, branchingly, or in a network within sub-regions. Specifically, pattern recognition or machine learning algorithms can be used, taking the extracted features as input, to train the model to identify patterns that conform to the typical morphology of vascular branches or fine skeletal structures. For example, linear structures may correspond to blood vessels, branching structures may correspond to vascular branches or bronchi, and network structures may correspond to fine bones.

[0084] The system determines whether a micro-dense structure is a vascular branch or a fine skeletal structure based on the degree of matching between its spatial morphology and connection patterns and those of predefined vascular branches or fine skeletal structures. This can be achieved by establishing a database containing the spatial morphology (e.g., length, width, curvature, or branching angle) and connection patterns (e.g., number of branch points or connection topology) of typical vascular branches and fine skeletal structures. The features of the identified micro-dense structures are then matched against predefined patterns in the database using similarity analysis methods, such as shape descriptors or topology matching algorithms, to quantify the degree of matching. When the degree of matching reaches a predefined threshold, the micro-dense structure can be determined to belong to either a vascular branch or a fine skeletal structure.

[0085] When tiny dense structures are identified as vascular branches or fine skeletal structures, they are separated from the subregion. Separation can be achieved through methods such as pixel marking, region growing, or morphological operations (e.g., erosion or opening) to ensure that these non-nodular dense structures are accurately removed from the nodular region for subsequent analysis of the nodules themselves.

[0086] This application comprehensively analyzes the grayscale distribution, edge gradient direction, and local texture periodicity information of sub-regions to more fully and meticulously characterize the imaging features of tiny dense structures. Specifically, density variation trends help distinguish solid nodules from hollow or semi-hollow structures such as blood vessels; edge orientation and connectivity effectively capture the linear, branching, or network topological structures of blood vessels and bones; and texture regularity further provides a basis for distinguishing different tissue types. By matching these multi-dimensional features with the spatial morphology and connection patterns of preset vascular branches or fine bone structures, this application can achieve accurate identification and judgment of tiny dense structures. This method based on multi-feature fusion and pattern matching significantly improves the accuracy and robustness of distinguishing nodules from non-nodular dense structures in complex imaging backgrounds, effectively avoiding misjudgments or omissions caused by single features or simple recognition rules in traditional methods.

[0087] Suppose that in a chest X-ray image, a two-dimensional projection of a suspected pulmonary nodule partially overlaps with a blood vessel. First, this overlapping area is subdivided into multiple sub-regions. For one sub-region, the system acquires its grayscale distribution information, finding that the grayscale value is higher at the center and gradually decreases towards the edges, exhibiting the typical density variation trend of a blood vessel. Simultaneously, analyzing its edge gradient direction information reveals a clear linear orientation, forming a continuous branching structure with the edges of adjacent sub-regions. Furthermore, calculating its local texture periodicity information shows that its texture is relatively smooth and has a certain directionality, consistent with the texture characteristics of a blood vessel.

[0088] Based on these features, the system identified a linear, minute, dense structure within the sub-region. Subsequently, the spatial morphological features (such as length or curvature) and connectivity features (such as the number of branch points) of this minute, dense structure were extracted and matched against a pre-defined database of vascular branching structures. Similarity calculations revealed a high match between this minute, dense structure and a typical pulmonary vascular branching pattern in the database. Therefore, the system identified this minute, dense structure as a vascular branch and separated it from the sub-region. This refined identification and separation process ensured accurate depiction of the nodule's true boundaries, avoiding interference from vascular structures in the nodule morphology analysis.

[0089] In some embodiments, the step of determining whether a micro-dense structure is a vascular branch or a fine skeletal structure based on the degree of matching between the micro-dense structure and the spatial morphology and connection pattern of a preset vascular branch or fine skeletal structure includes: To obtain the spatial morphological characteristics and connection pattern characteristics of tiny dense structures; Obtain a pre-defined library of spatial morphology and connection patterns of vascular branches or fine skeletal structures; For each structural pattern in the preset library, calculate its similarity to the spatial morphological features and connection pattern features of tiny dense structures; Identify structural patterns in a pre-defined library that have the highest similarity to tiny, dense structures; When the highest similarity is lower than the preset matching threshold, the tiny dense structure is marked as an unknown variant structure; When the highest similarity is higher than or equal to the preset matching threshold, the tiny dense structure is judged as the blood vessel branch or small bone structure corresponding to the highest similarity. When tiny dense structures are labeled as unknown variant structures, assess the connectivity between the tiny dense structures and the main nodule region; When tiny dense structures are labeled as unknown variant structures, assess the connectivity between the tiny dense structures and the main region of the dense structure. Based on the connectivity assessment results between the unknown variant structure and the main region of the nodule, and the connectivity assessment results between the unknown variant structure and the main region of the dense structure, it is determined whether the unknown variant structure is part of the nodule.

[0090] Specifically, spatial morphological features can be understood as quantitative indicators describing the geometric shape, size, or orientation of minute, dense structures, such as their aspect ratio, curvature, branching angle, volume, or projected area. Connectivity pattern features refer to the spatial connections between this structure and other adjacent structures (such as blood vessels, bronchi, or pleura), the strength of these connections, and the distribution of connection points. These features can be extracted from image data using image processing and pattern recognition algorithms.

[0091] The pre-defined spatial morphology and connectivity pattern library of vascular branches or small skeletal structures can be understood as a pre-established database containing various typical vascular branch and small skeletal structure feature templates. This library can be constructed through analysis of a large number of normal lung X-ray images, expert annotation, or based on anatomical models. For each structural pattern in the pre-defined library, calculating its similarity to the spatial morphological and connectivity features of tiny, dense structures involves using an appropriate similarity measurement algorithm (such as Euclidean distance, cosine similarity, or machine learning-based feature matching algorithms) to quantify the degree of matching between the structure to be analyzed and the templates in the library.

[0092] The matching threshold is a pre-defined value used to distinguish whether a structure highly matches a known pattern. When the calculated highest similarity is below this threshold, it indicates that the tiny, dense structure does not perfectly conform to any known typical vascular branching or fine skeletal structure pattern, and is thus marked as an unknown variant structure. This marking mechanism avoids simply classifying it as a known structure or ignoring it directly, but rather treats it as a special case requiring further analysis.

[0093] When a small, dense structure is labeled as an unknown variant, its connectivity with the main nodule region and the main dense structure region is evaluated. Connectivity evaluation may include calculating spatial distance, analyzing grayscale continuity, texture similarity, or analyzing path connectivity using graph theory methods. The aim is to determine whether the unknown variant structure is more spatially connected to the nodule or to surrounding dense structures (such as blood vessels or bronchial walls).

[0094] Finally, based on the assessment of the connectivity between the unknown variant structure and the main regions of the nodule and the dense structure, it can be determined whether the unknown variant structure is part of the nodule. For example, if the structure has a stronger connectivity with the main region of the nodule and a weaker connectivity with the main region of the dense structure, it may be judged as part of the nodule; otherwise, it may be judged as a non-nodule structure.

[0095] This application effectively addresses the potential classification inaccuracies when dealing with atypical or ambiguous small dense structures by introducing similarity calculation, matching thresholds, and further connectivity assessment of unknown variant structures. Specifically, firstly, by quantifying the similarity between small dense structures and preset patterns, structures that highly match typical vascular branches or fine skeletal structures can be identified more precisely. Secondly, when the similarity does not reach the preset threshold, it is not directly classified but rather labeled as an unknown variant structure, avoiding arbitrary judgments when information is insufficient. Finally, by conducting in-depth assessment of the connectivity between these unknown variant structures and the main body of nodules and dense structures, their interaction with the surrounding environment can be used to assist in determining their classification. This allows for more accurate differentiation of nodules from overlapping vascular or skeletal structures in complex imaging backgrounds, improving the robustness of the distinction.

[0096] The following is a specific example to illustrate this.

[0097] Suppose a tiny, dense structure is identified in a subregion of a chest X-ray image. First, the spatial morphological features of this tiny, dense structure are extracted (e.g., it appears as a slightly curved line, approximately 3 mm long and 1 mm wide) and connectivity features (e.g., one end appears to be connected to a larger vascular structure, while the other end vaguely points to a suspected nodule region). Next, these features are compared to templates in a pre-defined library of vascular branches and small skeletal structures for similarity calculation.

[0098] If the calculation shows that the similarity between the tiny dense structure and a typical small vascular branching pattern in the database is 0.85, while the preset matching threshold is 0.90, then because 0.85 is lower than 0.90, the tiny dense structure is marked as an unknown variant structure. At this point, the system will not immediately classify it as a vascular branch, but will further evaluate its connectivity with the main nodule region and the main dense structure region.

[0099] Connectivity assessment revealed that the unknown variant structure exhibited good grayscale continuity and was spatially close to the main region of the suspected nodule, resulting in a connectivity score of 0.7. In contrast, its connectivity score with the main region of known large blood vessel structures was 0.4. Based on these assessment results, the system determined that the unknown variant structure was more likely connected to the main region of the nodule and therefore classified it as part of the nodule. This approach avoids misclassifying a structure as non-nodular due to subtle differences in morphology from typical vascular branches, thereby improving the accuracy of nodule identification.

[0100] like Figure 2 As shown in the embodiments of this application, a medical X-ray image analysis system based on deep learning is also disclosed to distinguish between true pathological changes in pulmonary nodules and false imaging changes caused by non-pathological factors, including: Image acquisition module 1 is used to acquire chest X-ray images of the same patient taken at different time points; Nodule recognition module 2 is used to identify lung nodule regions from chest X-ray images and obtain two-dimensional image representation information of the nodules; The 3D morphology inference module 3 is used to infer the morphological information of the nodule in three-dimensional space based on the two-dimensional image representation information of the nodule, combined with imaging physics principles and prior anatomical knowledge. The three-dimensional feature acquisition module 4 is used to acquire the three-dimensional morphological features of the nodules from the inferred three-dimensional morphological information of the nodules; Feature comparison module 5 is used to compare the three-dimensional morphological features of nodules acquired at different time points; The pathological judgment module 6 is used to judge the pathological changes of the nodules based on the comparison results. Result output module 7 is used to output the pathological change judgment results of the nodules.

[0101] The deep learning-based medical X-ray image analysis system proposed in this application can accurately distinguish between the true pathological changes of pulmonary nodules and the false changes in imaging caused by non-pathological factors.

[0102] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A medical X-ray image analysis method based on deep learning, characterized by, This method is used to distinguish between true pathological changes in pulmonary nodules and pseudo-pathological changes in imaging caused by non-pathological factors, including: Obtain chest X-ray images of the same patient taken at different time points; Identify pulmonary nodule regions from chest X-ray images and obtain two-dimensional imaging information of the nodules; Based on the two-dimensional image representation information of nodules, combined with imaging physics principles and prior anatomical knowledge, the morphological information of nodules in three-dimensional space is inferred. The three-dimensional morphological features of the nodules are obtained from the inferred three-dimensional morphological information of the nodules. Compare the three-dimensional morphological features of nodules obtained at different time points; Based on the comparison results, determine the pathological changes of the nodules; Output the pathological changes of the nodules. 2.The medical X-ray image analysis method based on deep learning according to claim 1, characterized in that, The steps for inferring the morphological information of nodules in three-dimensional space based on the two-dimensional image representation information of nodules, combined with imaging physics principles and prior anatomical knowledge, include: Two-dimensional projection regions of suspected pulmonary nodules in chest X-ray images are identified, and two-dimensional image representation information of the nodules is extracted. The two-dimensional image representation information includes nodule coarse region identification information, local texture feature information, gray-level gradient and contrast pattern information, and boundary feature information with surrounding dense structures. Based on the two-dimensional imaging information of the nodules, combined with the pre-set knowledge of lung anatomy and the physical principles of X-ray imaging, the probabilistic morphology of the nodules in three-dimensional space is inferred from the two-dimensional imaging information. The degree of uncertainty in the quantified inferred three-dimensional probabilistic morphology of nodules; Based on the three-dimensional probabilistic morphology and degree of uncertainty of the nodules, the three-dimensional morphological characteristics of the nodules are inferred. 3.The medical X-ray image analysis method based on deep learning according to claim 2, characterized in that, The steps for quantifying the uncertainty of the three-dimensional probability morphology of the nodules include: To obtain imaging parameter information of chest X-ray images and information on the patient's respiratory status; Identify stable anatomical landmarks in chest X-ray images and obtain information on the relative positional changes of anatomical landmarks in images at different time points; The calculation results of morphological distribution entropy are calibrated based on imaging parameter information, patient respiratory status information, and relative positional changes of anatomical landmarks in images at different time points. The calculation results of the contribution of two-dimensional feature ambiguity are calibrated based on imaging parameter information, patient respiratory status information, and relative position changes of anatomical landmarks in images at different time points. Based on the calibrated morphological distribution entropy and the calibrated two-dimensional feature ambiguity contribution, the comprehensive uncertainty score of the three-dimensional probabilistic morphology of the nodule is inferred. 4.The medical X-ray image analysis method based on deep learning according to claim 2, characterized in that, The step of inferring the three-dimensional morphological features of a nodule based on its three-dimensional probabilistic morphology and degree of uncertainty includes: Calculate the expected three-dimensional volume and expected three-dimensional major and minor axis dimensions of the nodule; Based on the degree of uncertainty of the nodules, a fluctuation range is set for the expected three-dimensional volume and the expected three-dimensional major and minor axis dimensions, respectively, to obtain the range of the expected three-dimensional volume and the range of the expected three-dimensional major and minor axis dimensions; Compare the expected three-dimensional volume range and the expected three-dimensional major and minor axis size range obtained at different time points; Based on the comparison results of the expected three-dimensional volume range and the expected three-dimensional major and minor axis size range, and combined with the degree of uncertainty, the three-dimensional morphological characteristics of the nodule are inferred. 5.The medical X-ray image analysis method based on deep learning according to claim 1, characterized in that, The steps of identifying lung nodule regions from chest X-ray images and obtaining two-dimensional imaging information of the nodules include: Chest X-ray images are processed to optimize image quality; Multiscale analysis was performed on the processed images to identify the two-dimensional projection regions of suspected pulmonary nodules; For the two-dimensional projection region of the identified suspected lung nodules, the uniformity of its local gray-level distribution is calculated, and the gradient change at its edge is analyzed. When the edge of the two-dimensional projection area of ​​a suspected pulmonary nodule overlaps with the edge of a dense anatomical structure, the nodule and the dense structure are distinguished based on the gray value and texture features of the overlapping area. The segmentation threshold is adjusted based on the local gray-scale distribution of the nodules, edge gradient changes, and differentiation results. Based on the adjusted segmentation threshold, the two-dimensional projection contour of the nodule is drawn; Based on the drawn two-dimensional projection outline, the projected area and shape irregularity of the nodule are calculated as the two-dimensional image representation information of the nodule.

6. The medical X-ray image analysis method based on deep learning according to claim 5, characterized in that, When the edge of the two-dimensional projection region of a suspected pulmonary nodule overlaps with the edge of a dense anatomical structure, the step of distinguishing the nodule from the dense structure based on the grayscale value and texture features of the overlapping region includes: Obtain the grayscale distribution information of the overlapping region; Analyze the edge gradient direction of the overlapping region; Calculate the local texture periodicity of overlapping regions; Assess the connectivity between the overlapping region and the main region of the nodule; Assess the connectivity between the overlapping regions and the main dense structural regions; and Based on the gray value distribution information of the overlapping region, the edge gradient direction of the overlapping region, the local texture periodicity of the overlapping region, the connectivity between the overlapping region and the main region of the nodule, and the connectivity between the overlapping region and the main region of the dense structure, nodules and dense structures can be distinguished.

7. The medical X-ray image analysis method based on deep learning according to claim 6, characterized in that, The steps for distinguishing nodules from dense structures based on the grayscale distribution information of the overlapping region, the edge gradient direction of the overlapping region, the local texture periodicity of the overlapping region, the connectivity between the overlapping region and the main region of the nodule, and the connectivity between the overlapping region and the main region of the dense structure include: The overlapping regions are further subdivided into multiple sub-regions; For each sub-region, the gray value distribution characteristics, edge gradient direction characteristics, and texture periodicity characteristics within it are calculated to identify tiny dense structures in the sub-region that correspond to the characteristics of blood vessel branches or fine bone structures. Based on the spatial distribution and connectivity of the tiny dense structures, they are separated from the overlapping regions; Assess the connectivity between the separated overlapping region and the main nodule region; Evaluate the connectivity between the separated overlapping region and the main dense structural region; Based on the gray value distribution information of the separated overlapping regions, the edge gradient direction, the local texture periodicity, the connectivity between the separated overlapping regions and the main region of the nodules, and the connectivity between the separated overlapping regions and the main region of the dense structure, nodules and dense structures can be distinguished. 8.The medical X-ray image analysis method based on deep learning according to claim 7, characterized in that, The step of identifying tiny, dense structures in a subregion that correspond to features of vascular branches or fine bone structures includes: Obtain grayscale value distribution information, edge gradient direction information, and local texture periodicity information of the sub-region; Based on the grayscale value distribution information, analyze the density change trend within the sub-region; Analyze the edge orientation and connectivity of sub-regions based on edge gradient direction information; Based on the periodicity information of local textures, analyze the texture patterns within sub-regions; Based on density variation trends, edge orientation and connectivity, as well as texture patterns, we can identify tiny dense structures in sub-regions that are arranged in a linear, branching or mesh-like pattern. Based on the degree of matching between the micro-dense structure and the spatial morphology and connection pattern of the preset vascular branch or fine skeletal structure, it is determined whether the micro-dense structure is a vascular branch or fine skeletal structure. When tiny, dense structures are identified as vascular branches or fine skeletal structures, they are separated from the sub-regions. 9.The medical X-ray image analysis method based on deep learning according to claim 8, characterized in that, The step of determining whether a micro-dense structure is a vascular branch or a fine skeletal structure based on the degree of matching between the micro-dense structure and the spatial morphology and connection pattern of a preset vascular branch or fine skeletal structure includes: To obtain the spatial morphological characteristics and connection pattern characteristics of tiny dense structures; Obtain a pre-defined library of spatial morphology and connection patterns of vascular branches or fine skeletal structures; For each structural pattern in the preset library, calculate its similarity to the spatial morphological features and connection pattern features of tiny dense structures; Identify structural patterns in a pre-defined library that have the highest similarity to tiny, dense structures; When the highest similarity is lower than the preset matching threshold, the tiny dense structure is marked as an unknown variant structure; When the highest similarity is higher than or equal to the preset matching threshold, the tiny dense structure is judged as the blood vessel branch or small bone structure corresponding to the highest similarity. When micro-dense structures are labeled as unknown variant structures, assess the connectivity between micro-dense structures and the main nodule region, and assess the connectivity between micro-dense structures and the main dense structure region. Based on the connectivity assessment results between the unknown variant structure and the main region of the nodule, and the connectivity assessment results between the unknown variant structure and the main region of the dense structure, it is determined whether the unknown variant structure is part of the nodule.

10. A medical X-ray image analysis system based on deep learning, characterized by, This system is used to distinguish between true pathological changes in pulmonary nodules and pseudo-pathological changes in imaging caused by non-pathological factors, including: The image acquisition module is used to acquire chest X-ray images of the same patient taken at different time points; The nodule recognition module is used to identify lung nodule regions from chest X-ray images and obtain two-dimensional imaging information of the nodules; The three-dimensional morphology inference module is used to infer the morphological information of nodules in three-dimensional space based on the two-dimensional image representation information of nodules, combined with imaging physics principles and prior anatomical knowledge. The three-dimensional feature acquisition module is used to obtain the three-dimensional morphological features of the nodules from the inferred three-dimensional morphological information of the nodules. The feature comparison module is used to compare the three-dimensional morphological features of nodules acquired at different time points; The pathological assessment module is used to determine the pathological changes of nodules based on the comparison results. The results output module is used to output the pathological changes of the nodules.

Images