Dynamic DR lung field area diagnosis system based on deep learning

Abstract

The invention discloses a dynamic DR lung field area diagnosis system based on deep learning, and relates to the technical field of medical imaging, and the system comprises an image collection optimization module which is used for capturing a low-dose dynamic DR sequence of a complete period of calm respiration and best effort respiration of a patient, generating a dynamic DR image, and optimizing the dynamic DR image through employing a multi-mode transfer learning strategy, outputting a DR time sequence tensor; and the intelligent diagnosis module is used for carrying out disease cause classification through a gating chart convolutional network in combination with the lung field dilation function ratio and the space structure complexity index, outputting a classification result, and generating a structured report in combination with the lung field dilation function ratio, the space structure complexity index and the continuous lung field area change curve. The differential homeomorphic transformation field is generated through the Lie group exponential mapping algorithm, and the lung lobe boundary connectivity and the bronchial tree topological structure are strictly kept in the three-dimensional deformation process.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and in particular to a dynamic DR lung field area diagnostic system based on deep learning. Background Technology

[0002] In recent years, dynamic digital radiography (DR) has been widely used in the field of functional chest imaging, combining deep learning algorithms to achieve dynamic monitoring of lung field morphology. This method captures images of the respiratory cycle through low-dose continuous exposures, segments lung field contours using a temporal convolutional network, and estimates respiratory motion displacement based on optical flow. A multimodal transfer learning strategy further enhances the robustness of cross-device image analysis, providing a data foundation for quantifying lung function.

[0003] Existing technologies have shortcomings in compensating for respiratory motion artifacts. Traditional deformation models cannot forcibly satisfy anatomical topological constraints, leading to tearing of lung lobe boundaries and distortion of microstructures during deep breathing. Geometric distortions are passed on to subsequent functional quantification processes, causing large errors in area calculation and drift in structural features, affecting the accurate identification of early COPD lesions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based dynamic DR lung field area diagnostic system to solve the problem of functional quantification error propagation caused by topological distortion of lung field anatomy during dynamic breathing.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a deep learning-based dynamic DR lung field area diagnostic system, comprising: an image acquisition and optimization module, which captures low-dose dynamic DR sequences of a patient's complete cycle of calm and expeditious breathing, generates dynamic DR images, optimizes the dynamic DR images using a multimodal transfer learning strategy, and outputs a DR temporal tensor; a displacement analysis module, which uses a convolutional segmentation network to segment the DR temporal tensor frame by frame, generates an initial binarized lung field mask, extracts the initial lung field contour of each frame, inputs the initial lung field contours of all frames into a spatiotemporal convolutional neural network, analyzes the lung field contour displacement vector between adjacent frames frame by frame, and outputs a velocity field mapping map; and a deformation compensation module, which, based on the velocity field mapping map, generates a reversible deformation field using a Lie group exponential mapping algorithm, enforces the topological invariance constraint of the anatomical structure, constructs a differential homeomorphic transformation field, and performs dynamic deformation compensation on the lung field area ... displacement analysis module, which uses a convolutional segmentation network to capture the lung field area of ​​a patient's complete cycle of calm and expedited breathing, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates a low-dose dynamic DR sequence, generates The DR images undergo frame-by-frame geometric compensation to output a temporal sequence of lung fields. The functional quantification module uses a convolutional segmentation network to re-segment the lung field temporal sequence, generating a binary lung field mask, calculating the lung field area value per frame, forming a continuous lung field area change curve, extracting the mean area during quiet breathing and the peak area during deep breathing, and calculating the lung field expansion function ratio. The structural analysis module spatially partitions the lung field temporal sequence, constructing an anatomical set covering local to global areas, extracting the grayscale distribution characteristics of the anatomical set, and generating a spatial structural complexity index. The intelligent diagnosis module combines the lung field expansion function ratio and the spatial structural complexity index, performs etiological classification using a gated graph convolutional network, outputs the classification results, and generates a structured report by combining the lung field expansion function ratio, spatial structural complexity index, and continuous lung field area change curve.

[0008] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for outputting the DR time-series tensor are as follows:

[0009] Low-dose dynamic DR sequences are captured to generate dynamic DR images by capturing the complete cycle of a patient's calm breathing and expedited breathing.

[0010] Dynamic DR images are input into a multi-scale self-supervised augmentation network, and optimized dynamic DR sequences are generated through cascaded wavelet decomposition and spatial residual learning.

[0011] Neural radiation field reconstruction is performed on the optimized dynamic DR sequence to generate the DR temporal tensor.

[0012] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for extracting the initial lung field contour of each frame are as follows:

[0013] The DR temporal tensor is decomposed into frame-by-frame three-dimensional volume data along the time dimension, and input into a convolutional segmentation network of deformable convolutional layers and Transformer encoders. The position of the convolutional sampling points is dynamically adjusted to adapt to respiratory deformation, and cross-frame attention is used to constrain segmentation continuity, outputting an initial binarized lung field mask for each frame.

[0014] A morphological refinement operation is performed on the initial binarized lung field mask for each frame to extract the initial lung field contour.

[0015] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for outputting the velocity field mapping are as follows:

[0016] Register the initial lung field contours of all frames to a pre-stored standard lobar partition template to generate a lung field contour sequence with anatomical markers;

[0017] The lung field contour sequence with anatomical markers is input into a two-branch spatiotemporal convolutional network. Geometric features are extracted through the spatial branch, while anatomical constraint features are extracted through the anatomical branch.

[0018] Geometric features and anatomical constraint features are fused through a mutual attention mechanism to output the lung field contour displacement vector between adjacent frames;

[0019] The lung field contour displacement vector is encoded by time differentiation and spatial normalization to generate a velocity field mapping.

[0020] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for constructing the differential homeomorphic transformation field are as follows:

[0021] Multiscale continuous homology analysis was performed on the velocity field mapping to extract the generation and extinction interval parameters of the lung field connectivity components and the annular structure.

[0022] The parameters of the birth and death intervals are combined with the velocity field mapping diagram to perform Lie group exponential mapping, generating the initial deformation field, and the Jacobian matrix is ​​calculated.

[0023] The initial deformation field and Jacobian matrix are input into the topology constraint optimizer. The topology durability loss value is calculated based on the birth and death interval parameters. The deformation field is iteratively optimized by combining the determinant value of the Jacobian matrix with constraints.

[0024] When the determinant of the Jacobian matrix of the optimized deformation field is greater than the preset invertibility threshold and the topological durability loss is less than the preset convergence threshold, the differential homeomorphic transformation field is output.

[0025] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for outputting the temporal image sequence of the lung field are as follows:

[0026] Based on dynamic DR imaging, the X-ray attenuation coefficient of major tissue regions in the lung field was calculated.

[0027] By combining the X-ray attenuation coefficient with the differential homeomorphic transformation field, geometrically corrected images are generated through ray tracing compensation.

[0028] Anisotropic diffusion is performed on the geometrically corrected images to output a temporal sequence of lung fields.

[0029] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for forming a continuous lung field area change curve are as follows:

[0030] The lung field temporal image sequence is input into a dynamic deformation adaptive convolutional network, and the lung field is segmented frame by frame through the convolution kernel deformation mechanism to generate a binary lung field mask.

[0031] Morphological refinement and subpixel gradient correction are performed on each frame of the binarized lung field mask to generate subpixel precision lung field contours.

[0032] Based on the sub-pixel precision lung field contour, the lung field area value of a single frame is calculated by Green's theorem area integration.

[0033] Adaptive temporal filtering is applied to the lung field area values ​​of a single frame to generate continuous lung field area change curves.

[0034] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for calculating the lung field expansion function ratio are as follows:

[0035] Identify the time windows between quiet breathing and deep breathing from continuous lung field area change curves;

[0036] The mean area was obtained within the time window of the quiet breathing period as the mean area of ​​the quiet breathing period;

[0037] The maximum area value within the deep breathing period time window is extracted as the peak area during the deep breathing period.

[0038] The lung field expansion function ratio is calculated by dynamically coupling the mean area during quiet breathing and the peak area during deep breathing.

[0039] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for generating the spatial structure complexity index are as follows:

[0040] Spatial regional subdivision of the temporal imaging sequence of the lung field was performed to form a three-layer anatomical set of alveolar, lobular, and lobar levels;

[0041] Extract the grayscale distribution of pixels within the three-layer anatomical set, generate a grayscale feature histogram, and extract the grayscale distribution features;

[0042] By integrating grayscale distribution features through multi-scale feature fusion rules, a spatial structure complexity index is output.

[0043] As a preferred embodiment of the deep learning-based dynamic DR lung field area diagnostic system of the present invention, the specific steps for generating the structured report are as follows:

[0044] A dynamic lung field function map is constructed based on the lung field expansion function ratio, spatial structure complexity index, and continuous lung field area change curve.

[0045] The dynamic lung field functional map is input into a gated graph convolutional network, and etiological classification is performed through node attribute fusion and dynamic edge connection, and the classification results are output.

[0046] The system combines lung field expansion function ratio, spatial structure complexity index, classification results, and continuous lung field area change curves to generate a structured report using a pre-stored clinical guideline template.

[0047] The beneficial effects of this invention are as follows: By generating a differential homeomorphic transformation field through the Lie group exponential mapping algorithm, the connectivity of lung lobe boundaries and the topological structure of the bronchial tree are strictly maintained during three-dimensional deformation. After performing geometric compensation frame by frame on dynamic DR images, the anatomical fidelity of the output lung field temporal image sequence is significantly improved, enhancing topological conservation, structural integrity, and functional quantification, thereby increasing the visualization accuracy of alveolar expansion trajectories during deep breathing and providing reliable anatomical support for the diagnosis of prodromal COPD. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0049] Figure 1 This is a block diagram of a deep learning-based dynamic DR lung field area diagnostic system.

[0050] Figure 2 The flowchart for generating DR time series tensors.

[0051] Figure 3 A flowchart for calculating the expansion function ratio for dynamic coupling.

[0052] Figure 4 A flowchart for generating a velocity field mapping diagram. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0056] Reference Figures 1-4 This is one embodiment of the present invention, which provides a deep learning-based dynamic DR lung field area diagnostic system, comprising the following steps:

[0057] The image acquisition optimization module captures low-dose dynamic DR sequences of the patient's complete cycle of calm breathing and expedited breathing, generates dynamic DR images, and optimizes the dynamic DR images using a multimodal transfer learning strategy, outputting DR temporal tensors.

[0058] Furthermore, low-dose dynamic DR sequences are captured during the patient's complete cycle of calm breathing and expedited breathing to generate dynamic DR images.

[0059] Specifically, a low-dose pulsed X-ray tube is operated to continuously output a beam of radiation. Simultaneously, a flat panel detector array is activated to receive attenuated radiation signals that penetrate the patient's chest. The patient completes a full respiratory cycle consisting of calm inhalation, calm exhalation, maximal inhalation, and maximal exhalation according to instructions. The flat panel detector array converts the received attenuated radiation signals that penetrate the patient's chest into an electrical signal array. The electrical signal array generates digital projection data through an analog-to-digital converter circuit. The digital projection data is then reconstructed using a filtered back-projection algorithm to generate dynamic DR images.

[0060] Dynamic DR images are input into a multi-scale self-supervised augmentation network, and optimized dynamic DR sequences are generated through cascaded wavelet decomposition and spatial residual learning.

[0061] It should be noted that the construction process of the multi-scale self-supervised augmentation network is as follows: The Haar wavelet decomposition layer serves as the input interface to receive dynamic DR images and outputs low-frequency approximate components and horizontal / vertical / diagonal high-frequency detail components. The residual dense block structure connects the low-frequency processing channel, and integrates dense connection layers and cross-layer skip connection units. The directional gradient convolutional layer connects the high-frequency processing channel and configures multi-directional separable convolutional kernel groups. The output of the residual dense block structure generates enhanced low-frequency features, and the output of the directional gradient convolutional layer generates enhanced high-frequency features. The inverse wavelet reconstruction layer receives the enhanced low-frequency features and the enhanced high-frequency features and performs fusion reconstruction to finally obtain the multi-scale self-supervised augmentation network.

[0062] Specifically, dynamic DR images are input into a multi-scale self-supervised enhancement network, and Haar wavelet decomposition is performed to generate low-frequency approximation components and high-frequency detail components. The low-frequency approximation components are input into residual dense blocks to enhance tissue contrast, and the high-frequency detail components are input into directional gradient convolution kernels to enhance edge texture. The enhanced low-frequency approximation components and high-frequency detail components are then reconstructed by inverse wavelet transform to generate an optimized dynamic DR sequence.

[0063] Neural radiation field reconstruction is performed on the optimized dynamic DR sequence to generate the DR temporal tensor.

[0064] Specifically, the framework for reconstructing a neural radiation field by inputting a dynamic DR sequence is optimized. X-ray direction parameters are loaded onto the neural radiation field, which generate a sequence of ray path sampling points. This sequence is then input into a radiation density prediction engine, which outputs spatial radiation density predictions and grayscale predictions. These spatial radiation density predictions and grayscale predictions are then combined using differentiable X-ray projection accumulation to generate reconstructed frame data. Finally, the reconstructed frame data is combined along the time dimension to form a DR temporal tensor.

[0065] The displacement analysis module uses a convolutional segmentation network to segment the DR temporal tensor frame by frame, generate an initial binarized lung field mask, and extract the initial lung field contour of each frame. The initial lung field contours of all frames are input into a spatiotemporal convolutional neural network, which analyzes the lung field contour displacement vector between adjacent frames frame by frame and outputs a velocity field mapping map.

[0066] Furthermore, the DR temporal tensor is decomposed into frame-by-frame three-dimensional volume data along the time dimension, which is then input into a convolutional segmentation network consisting of deformable convolutional layers and a Transformer encoder. The position of the convolutional sampling points is dynamically adjusted to adapt to respiratory deformation, and cross-frame attention is used to constrain segmentation continuity, outputting an initial binarized lung field mask for each frame.

[0067] It should be noted that the pre-training process of the convolutional segmentation network is as follows: A labeled lung image dataset is used for training. The image is standardized and pre-processed and then resampled to an isotropic resolution. The processed image is input into a deformable convolutional layer to extract deformation features. The deformation features are input into a Transformer encoder to capture long-range spatial dependencies. The encoded features are fused with deep semantics and shallow details through a multi-scale feature pyramid. The fusion result is upsampled to the original resolution. The backpropagation algorithm updates the network weight parameters. The network is iteratively optimized until the segmentation accuracy on the validation set converges, and finally, the convolutional segmentation network is obtained.

[0068] Specifically, the DR temporal tensor is decomposed into independent three-dimensional volume data according to time frames. The three-dimensional volume data is input into the deformable convolutional layer of the convolutional segmentation network. The deformable convolutional layer samples the grid coordinates according to the respiratory motion offset. The offset sampled coordinates are used to extract deformation compensation features, which are input into the Transformer encoder to integrate the temporal correlation features of adjacent frames. The temporal correlation features are mapped to pixel classification results through the decoding layer. The pixel classification results are converted into the initial binarized lung field mask for each frame.

[0069] A morphological refinement operation is performed on the initial binarized lung field mask for each frame to extract the initial lung field contour.

[0070] Specifically, an iterative erosion algorithm is used to erode the initial binarized lung field mask. The skeletonization algorithm uses 3×3 structuring elements to scan the initial binarized lung field mask. During the scanning process, isolated noise points are eliminated and the main structure of the lung field is preserved. The skeletonization algorithm outputs a thinned binarized lung field mask with a single pixel width. The thinned binarized lung field mask is input into a boundary tracking algorithm to identify the contour inflection points of connected regions in the thinned binarized lung field mask. The contour inflection points are connected in spatial order to form a closed curve, which is the initial lung field contour.

[0071] Register the initial lung field contours of all frames to a pre-stored standard lobar partition template to generate a lung field contour sequence with anatomical markers;

[0072] It should be noted that the pre-stored standard lung lobe partition template refers to a dataset that includes spatial reference coordinates of the lung lobe anatomical structure of healthy individuals and corresponding anatomical identifiers.

[0073] Specifically, the initial lung field contours of all frames are input into the point set registration process. The point set registration process loads the pre-stored standard lung lobe partition template, extracts the feature point coordinates of the initial lung field contours of each frame, performs minimum distance matching between the feature point coordinates and the anatomical marker point coordinates of the standard lung lobe partition template, calculates the rotation and translation transformation matrix for minimum distance matching, and applies the rotation and translation transformation matrix to align the initial lung field contours to the standard lung lobe partition template space. The aligned contours inherit the anatomical markers of the standard lung lobe partition template, generating a lung field contour sequence with anatomical markers.

[0074] The lung field contour sequence with anatomical markers is input into a two-branch spatiotemporal convolutional network. Geometric features are extracted through the spatial branch, while anatomical constraint features are extracted through the anatomical branch.

[0075] Specifically, the lung field contour sequence with anatomical markers is input into a bi-branch spatiotemporal convolutional network. The spatial branch extracts the contour motion geometric features of the lung field contour sequence with anatomical markers, the anatomical branch extracts the lung lobe boundary constraint features of the lung field contour sequence with anatomical markers, the spatial branch outputs geometric features, and the anatomical branch outputs anatomical constraint features.

[0076] Geometric features and anatomical constraint features are fused through a mutual attention mechanism to output the lung field contour displacement vector between adjacent frames;

[0077] Specifically, the geometric features and anatomical constraint features are input into a mutual attention mechanism, which maps the geometric features to generate a query vector, maps the anatomical constraint features to generate a key vector and a value vector, the query vector and key vector determine the association weights, the association weights are applied to the value vectors to generate weighted features, the weighted features and geometric features are subjected to residual connections, the connection results are normalized by layers to output a fused feature vector, the fused feature vector is input into a fully connected regression layer to generate displacement parameters, and the displacement parameters are converted into lung field contour displacement vectors between adjacent frames.

[0078] The lung field contour displacement vector is encoded by time differentiation and spatial normalization to generate a velocity field mapping.

[0079] Specifically, the time differentiation operation process uses the central difference algorithm to obtain the displacement change rate of adjacent time points. The displacement change rate reflects the motion velocity of the lung field contour points. The motion velocity of the lung field contour points is spatially normalized and encoded. The spatial normalization encoding is loaded with the maximum anatomical size parameter of the lung field to generate a normalized scaling factor for each contour point. The normalized scaling factor is used to adjust the motion velocity. The adjusted motion velocity is converted into a three-channel tensor format, and the output is the velocity field mapping map.

[0080] The deformation compensation module, based on the velocity field mapping map, generates a reversible deformation field through the Lie group exponential mapping algorithm, forces the anatomical structure topological invariance constraint to satisfy, constructs a differential homeomorphic transformation field, performs geometric compensation on dynamic DR images frame by frame, and outputs a lung field temporal image sequence.

[0081] Furthermore, multi-scale continuous cohomology analysis was performed on the velocity field mapping to extract the generation and extinction interval parameters of the lung field connectivity components and the ring structure;

[0082] Specifically, the velocity field map is input into a multi-scale filtering process to generate a multi-scale filtered velocity field map. The multi-scale filtered velocity field map is then subjected to a continuous homology analysis process. This process constructs a simple complex structure, which obtains the homology groups of each connected component. The evolution of the homology groups records the birth and death times of the connected components. Simultaneously, the formation and disappearance times of the ring structure are detected, and the birth and death times are extracted as birth and death interval parameters. These parameters include lung field connected component parameters and ring structure parameters.

[0083] Specifically, the birth-death interval parameters and the velocity field mapping are combined using a Lie group exponential mapping to generate the initial deformation field, and the Jacobian matrix is ​​calculated, with the expression as follows:

[0084]

[0085] In the formula, J(r,t) represents the Jacobian matrix, r represents the three-dimensional spatial coordinate vector, and t represents the current time. Represents the gradient operator, Represents the gradient of a three-dimensional spatial coordinate vector. This represents the composite operator for differential operators, exp represents matrix exponentiation, τ = t0 represents the start time of integration, and τ represents the time integration variable. denoted as the time dimension integral from the initial time to the current time, v represents the velocity field mapping, φ(r,τ) represents the value of the initial deformation field at the three-dimensional spatial coordinate vector r and the time integration variable τ, dτ represents the time integration infinitesimal element, and D(r) represents the spatial constraint matrix at the three-dimensional spatial coordinate vector r based on the birth and death interval parameters.

[0086] The initial deformation field and Jacobian matrix are input into the topology constraint optimizer. The topology durability loss value is calculated based on the birth and death interval parameters. The deformation field is iteratively optimized by combining the determinant value of the Jacobian matrix with constraints.

[0087] Specifically, based on the topological durability loss value, the topological stability of the anatomical structure is quantified. At the same time, the determinant value of the Jacobian matrix is ​​extracted to evaluate the local deformation invertibility. The gradient descent algorithm is used to iteratively update the initial deformation field parameters and output the optimized deformation field.

[0088] Specifically, the initial deformation field and Jacobian matrix are input into the topology constraint optimizer, and the topology durability loss value is calculated based on the birth-death interval parameters. The expression is as follows:

[0089]

[0090] In the formula, L represents the topological persistence loss value, and N represents the number of connected components in all lung fields. The value represents the reciprocal of the number of connected components in all lung fields, k represents the anatomical scale index, and K represents the total number of anatomical scales. This represents summation over all anatomical scales, ω k This represents the exponentially decaying weighted coefficient of the anatomical scale k. The square operation represents the L2 norm. M represents the boundary differential operator. k An anatomical structure entity representing the anatomical scale k. The boundary curve representing the anatomical structure at anatomical scale k. Let λ represent the closed path integral along the boundary curve, n represent the unit normal vector of the boundary curve, dl represent the arc length infinitesimal element of the boundary curve, and λ represent the arc length infinitesimal element of the boundary curve. k b represents the adjustment coefficient for the topological constraint strength at anatomical scale k. k c represents the birth time parameter representing the anatomical scale k. k The time to death parameter represents the anatomical structural scale k, (b k -c k ) indicates the length of the duration of the birth and death time interval.

[0091] It should be noted that the exponentially decaying weighted coefficient of the anatomical scale k strengthens the protection weight of the microscale through exponential decay. The example values ​​are k = 1, ω1 = 0.368, k = 2, ω2 = 0.135, k = 3, ω3 = 0.05. The topological constraint strength adjustment coefficient of the anatomical scale k is inversely proportional to the characteristic size of the anatomical structure based on the stability of the birth and death intervals of the structure at each scale in the continuous homology analysis. The example values ​​are k = 1, λ1 = 0.9, k = 2, λ2 = 0.45, k = 3, λ3 = 0.3.

[0092] When the determinant of the Jacobian matrix of the optimized deformation field is greater than the preset invertibility threshold and the topological durability loss is less than the preset convergence threshold, the differential homeomorphic transformation field is output.

[0093] It should be noted that the preset reversibility threshold is based on the mathematical constraint of deformation reversibility of the Lie group exponential mapping, and the safety boundary is determined by statistical analysis of the deformation amplitude of respiratory movements in healthy individuals. The example value is 0.12. The preset convergence threshold is based on the error range of topological stability in continuous homology analysis, combined with calibration in multi-center clinical trials. The example value is 0.08.

[0094] Specifically, during the topology constraint optimization process, the determinant value of the Jacobian matrix of the optimized deformation field is continuously monitored, and the current topology persistence loss value is recorded. The determinant value of the Jacobian matrix is ​​compared with a preset invertibility threshold, and the topology persistence loss value is compared with a preset convergence threshold. When the determinant value of the Jacobian matrix is ​​continuously greater than the preset invertibility threshold and the topology persistence loss value is stably less than the preset convergence threshold, it is determined that the optimized deformation field satisfies the differential homeomorphism condition, the iterative optimization process is immediately terminated, the final optimized deformation field parameters are extracted, and a differential homeomorphic transformation field is generated.

[0095] It should be noted that when the determinant of the Jacobian matrix is ​​consistently less than the preset invertibility threshold and the topological persistence loss is consistently less than the preset convergence threshold, the invertibility constraint mechanism is strengthened; when the determinant is consistently greater than the preset invertibility threshold and the topological loss is consistently greater than the preset convergence threshold, the topological constraint mechanism is strengthened; if neither condition is met, the dual constraints are strengthened simultaneously and the learning rate is reduced, and the deformation field parameters are iteratively updated through the gradient descent algorithm until both thresholds are met simultaneously.

[0096] Specifically, based on dynamic DR imaging, the X-ray attenuation coefficient of the main tissue regions of the lung field is calculated, and the expression is as follows:

[0097]

[0098] In the formula, μ(r) represents the X-ray attenuation coefficient at the three-dimensional spatial coordinate vector r, and E max E represents the upper limit of the X-ray energy spectrum. min E represents the lower limit of the X-ray energy spectrum. max -E min Indicates the total width of the X-ray energy spectrum coverage. The expression represents the integral operation within the X-ray energy spectrum, ρ(r) represents the tissue density mapping value at the three-dimensional spatial coordinate vector r, ln represents the natural logarithm operation, I0 represents the incident X-ray intensity reference value, I(r) represents the emitted X-ray intensity value at the three-dimensional spatial coordinate vector r, (I0 / I(r)) represents the ratio of incident to emitted intensity, erf represents the Gaussian error integral operation, α represents the gradient sensitivity adjustment coefficient, and ||·|| represents the Euclidean norm operation of the vector. represents the Gaussian gradient operator, * represents the spatial convolution operator, and dE represents the energy integral infinitesimal.

[0099] It should be noted that the gradient sensitivity adjustment coefficient is derived from the statistical characteristics of the gradient amplitude distribution of dynamic DR images, and the example value is 0.02.

[0100] By combining the X-ray attenuation coefficient with the differential homeomorphic transformation field, geometrically corrected images are generated through ray tracing compensation.

[0101] Specifically, the X-ray attenuation coefficient distribution of the dynamic DR image is loaded into the ray tracing engine. The differential homeomorphic transformation field provides a three-dimensional spatial deformation correction mapping. The ray tracing engine emits a virtual ray beam from the X-ray source. The virtual ray beam penetrates the spatial coordinates corrected by the differential homeomorphic transformation field. The X-ray attenuation coefficient value is sampled at the corrected spatial coordinates. The X-ray attenuation coefficient value is accumulated along the ray path. The accumulated result is converted into a logarithmic spatial intensity value. The logarithmic spatial intensity value generates the pixel grayscale of the geometrically corrected image. The geometrically corrected image retains the timestamp and spatial resolution of the dynamic DR image and is output as a geometrically corrected image.

[0102] Anisotropic diffusion is performed on the geometrically corrected images to output a temporal sequence of lung fields.

[0103] Specifically, the geometrically corrected image is input into the anisotropic diffusion process to obtain the local gray-level gradient amplitude of the image. The diffusion intensity weight is controlled according to the local gray-level gradient amplitude. High-intensity diffusion is performed in the lung parenchyma in the low-gradient region, while weak diffusion is maintained at the lung lobe boundary in the high-gradient region. The pixel gray levels of the geometrically corrected image are iteratively updated. The updated geometrically corrected image re-obtains the local gray-level gradient amplitude and diffusion intensity weight, and outputs the diffused image. The diffused images are combined in the original time sequence to generate a lung field time-series image sequence.

[0104] The functional quantization module uses a convolutional segmentation network to re-segment the temporal image sequence of the lung field, generate a binary lung field mask, calculate the lung field area value of a single frame, form a continuous lung field area change curve, extract the mean area during quiet breathing and the peak area during deep breathing, and calculate the lung field expansion function ratio.

[0105] Furthermore, the lung field temporal image sequence is input into a dynamic deformation adaptive convolutional network, and the lung field is segmented frame by frame through the convolution kernel deformation mechanism to generate a binary lung field mask;

[0106] It should be noted that a dynamic digital X-ray imaging simulation dataset is used as the input to the basic 3D segmentation architecture. A physically simulated respiratory motion deformation field is applied to generate a temporal sequence containing artifacts. This artifact-containing temporal sequence is then used to predict the sampling point offset of the convolutional kernel under the supervision of a displacement-aware learning mechanism. The prediction results enable the convolutional kernel deformation to adapt to the dynamic changes in respiration. Annotated real dynamic digital X-ray imaging data is then input to fine-tune the top-level parameters of the network, ultimately resulting in a dynamically deformable adaptive convolutional network.

[0107] Specifically, the lung field temporal image sequence is input into a dynamic deformation adaptive convolutional network. The dynamic deformation adaptive convolutional network loads the lung field temporal image sequence of consecutive frames. The convolution kernel deformation mechanism dynamically adjusts the coordinates of the sampling points according to the feature differences between adjacent frames. The adjusted sampling point coordinates are used to extract a deformation compensation feature map. The deformation compensation feature map is enhanced with spatial context information through multi-layer convolution operations. A probability distribution map is generated by sigmoid activation. The probability distribution map is binarized to generate a binarized lung field mask.

[0108] Morphological refinement and subpixel gradient correction are performed on each frame of the binarized lung field mask to generate subpixel precision lung field contours.

[0109] Specifically, the morphological refinement operation employs a skeletonization algorithm to iteratively refine the binarized lung field mask. The skeletonization algorithm eliminates redundant pixels while preserving the main structure, outputting a refined binarized lung field mask. The refined binarized lung field mask is then input into a subpixel gradient correction process to obtain the grayscale gradient field of the corresponding frame in the binarized lung field temporal image sequence. The grayscale gradient field drives the subpixel-level displacement of the edge points of the refined binarized lung field mask, optimizing the spatial coordinate accuracy of the contour points. The optimized contour points are then connected sequentially according to their adjacency relationship to generate a subpixel-precision lung field contour.

[0110] Specifically, based on the sub-pixel precision lung field contour, the area value of the lung field in a single frame is calculated using Green's theorem area integral, expressed as:

[0111]

[0112] In the formula, A f This represents the lung field area value in frame f, where f represents the frame index, and Ω. f This represents the subpixel precision lung field contour closure curve of frame f. Indicates along the closed curve Ω f The path integral, where u represents the x-coordinate of the contour point, d y Let Γ represent the differential of the ordinate of the profile curve, y represent the ordinate of the profile point, du represent the differential of the abscissa of the profile curve, Γ represent the deformation correction factor, and p represent the coordinate vector of the profile point. This represents the gradient vector of the contour point coordinates, and |·| represents the absolute value operation.

[0113] The deformation correction factor is derived from sub-pixel gradient correction. The deformation error is determined by analyzing the gradient amplitude distribution of the lung field contour during dynamic breathing. The example value is 0.92.

[0114] Adaptive temporal filtering is applied to the lung field area values ​​of a single frame to generate continuous lung field area change curves.

[0115] Specifically, the lung field area value of a single frame is input into the adaptive temporal filtering process, and the fluctuation characteristics of the area values ​​of consecutive frames are input into the adaptive temporal filtering process. The size of the filtering window is dynamically adjusted according to the fluctuation characteristics. Narrow window Gaussian filtering is applied in the small fluctuation stage, and wide window mid-range filtering is enabled in the large fluctuation stage. The filtering process smooths high-frequency noise and preserves the breathing trend. The filtered lung field area value is updated frame by frame to generate a continuous lung field area change curve.

[0116] Identify the time windows between quiet breathing and deep breathing from continuous lung field area change curves;

[0117] Specifically, the continuous lung field area change curve is input into the respiratory cycle analysis process. The respiratory cycle analysis process detects local maxima points of the continuous lung field area change curve and marks the time of deep inspiration. It also detects local minima points of the continuous lung field area change curve and marks the time of deep expiration. The average slope between adjacent maxima and minima points is calculated. If the absolute value of the average slope is less than the quiet breathing threshold, it is determined to be a quiet breathing segment. If the absolute value of the average slope is greater than the deep breathing threshold, it is determined to be a deep breathing segment. The start and end times of the quiet breathing segment are marked to form a quiet breathing period time window. The start and end times of the deep breathing segment are marked to form a deep breathing period time window.

[0118] It should be noted that the quiet breathing threshold is based on the statistical distribution of the average slope of the lung field area change curve during the quiet breathing period in healthy individuals, with an example value of 0.05; the deep breathing threshold is based on the slope of expedited breathing in the pulmonary function test standards, with an example value of 0.15.

[0119] The mean area was obtained within the time window of the quiet breathing period as the mean area of ​​the quiet breathing period;

[0120] Specifically, extract the data points of the continuous lung field area change curve corresponding to the time window of the quiet breathing period, obtain the arithmetic mean of the continuous lung field area change curve data points, and use the arithmetic mean as the mean area during the quiet breathing period.

[0121] The maximum area value within the deep breathing period time window is extracted as the peak area during the deep breathing period.

[0122] Specifically, the continuous lung field area change curve data segment within the deep breathing period time window is extracted. The continuous lung field area change curve data segment contains a timestamp and a binary sequence. All area values ​​in the binary sequence are scanned to generate an area value set. The maximum value in the area value set is identified as the area peak during the deep breathing period.

[0123] Specifically, the lung field expansion function ratio is calculated by dynamically coupling the mean area during quiet breathing and the peak area during deep breathing. The expression is as follows:

[0124]

[0125] In the formula, R represents the lung field expansion function ratio, and A peak The area under the respiratory tract is the peak area during deep breathing, and T represents the length of the quiet breathing time window. It represents the reciprocal of the length of the quiet breathing period time window. The integral operation is performed within the time window of the quiet breathing period. T1 represents the set of time windows of the quiet breathing period, dt represents the time element, β represents the respiratory cycle variability compensation coefficient, δ represents the dynamic time warping distance operation, and T2 represents the set of time windows of the deep breathing period.

[0126] It should be noted that the respiratory cycle variability compensation coefficient is derived from the statistical difference in respiratory phase drift between healthy individuals and COPD patients in multicenter clinical trials, with an example value of 1.2.

[0127] The structural analysis module performs spatial region segmentation on the temporal image sequence of the lung field, constructs an anatomical set covering the local to the global, extracts the gray-level distribution features of the anatomical set, and generates a spatial structural complexity index.

[0128] Furthermore, the temporal imaging sequence of the lung field is spatially segmented to form a three-layer anatomical set of alveolar, lobular, and lobar levels.

[0129] Specifically, the lung field temporal image sequence is input into a multi-scale space for subdivision. A predefined anatomical reference atlas is used to divide and generate a lobe-level anatomical set based on the lobe-level boundary coordinates. A lobule-level hexagonal grid is loaded into the lobe-level anatomical set to divide and generate a lobule-level anatomical set. An alveolar-level statistical shape template is applied into the lobule-level anatomical set to divide and generate an alveolar-level anatomical set. The three-layer anatomical set is then output.

[0130] It should be noted that the predefined anatomical reference atlas refers to standardized spatial partitioning data of lung anatomy generated based on annotations of healthy individuals, which includes three levels of spatial labeling information: lobe-level boundary coordinates, lobule-level hexagonal grids, and alveolar-level statistical shape templates.

[0131] Extract the grayscale distribution of pixels within the three-layer anatomical set, generate a grayscale feature histogram, and extract the grayscale distribution features;

[0132] Specifically, a three-layer anatomical set is loaded onto the lung field temporal image sequence. Gray value sets are extracted from pixels within the alveolar-level anatomical set to generate an alveolar-level gray-level feature histogram. Gray value sets are extracted from pixels within the lobular-level anatomical set to generate a lobular-level gray-level feature histogram. Gray value sets are extracted from pixels within the lobar-level anatomical set to generate a lobar-level gray-level feature histogram. The mean, variance, skewness, and kurtosis are calculated from the alveolar-level gray-level feature histogram as alveolar-level gray-level distribution features. The mean, variance, skewness, and kurtosis are calculated from the lobular-level gray-level feature histogram as lobular-level gray-level distribution features. The mean, variance, skewness, and kurtosis are calculated from the lobar-level gray-level feature histogram as lobar-level gray-level distribution features. The gray-level distribution features are then output.

[0133] By integrating grayscale distribution features through multi-scale feature fusion rules, a spatial structure complexity index is output.

[0134] Specifically, the grayscale distribution features are input into the multi-scale feature fusion process. The alveolar grayscale distribution features are assigned the maximum weight coefficient, the lobular grayscale distribution features are assigned the medium weight coefficient, and the lobar grayscale features are assigned the minimum weight coefficient. The weighted grayscale distribution features are then subjected to principal component analysis for dimensionality reduction. The dimensionality reduction results are then standardized and scaled to obtain the entropy measure of each component in the scaled results. The entropy measure characterizes the degree of spatial structure disorder and outputs a spatial structure complexity index.

[0135] The intelligent diagnostic module combines the lung field expansion function ratio and spatial structure complexity index, uses a gated graph convolutional network to classify etiologies, outputs classification results, and generates a structured report by combining the lung field expansion function ratio, spatial structure complexity index and continuous lung field area change curve.

[0136] Furthermore, a dynamic lung field function map is constructed based on the lung field expansion function ratio, spatial structure complexity index, and continuous lung field area change curve.

[0137] Specifically, the continuous lung field area change curve is converted into a time-area value coordinate point sequence. The time-area value coordinate point sequence is input into the two-dimensional matrix generation process to form a base matrix. The lung field expansion function ratio is written as a global parameter into the header information of the base matrix. The spatial structure complexity index is mapped to the color depth value of each time point in the base matrix. The color depth value reflects the degree of local structural disorder. The integrated data is encapsulated in the DICOM-SR structured report format and outputs a dynamic lung field function map.

[0138] The dynamic lung field functional map is input into a gated graph convolutional network, and etiological classification is performed through node attribute fusion and dynamic edge connection, and the classification results are output.

[0139] It should be noted that the pre-training process of the gated graph convolutional network is performed using a multi-center lung disease dataset. The multi-center lung disease dataset includes dynamic lung field functional map samples and gold standard etiology annotations. The graph structure nodes are configured to correspond to the time points of the dynamic lung field functional maps. The node attributes are loaded with the area value at the time point, the spatial structure complexity index, and the lung field expansion function ratio. Dynamic edge connections are set based on respiratory phase similarity weights. The graph convolutional layer aggregates the features of neighboring nodes. The gated recurrent unit controls the update of node feature states. The node feature input is updated and mapped to the etiology category dimension. The cross-entropy loss between the predicted category and the gold standard etiology annotation is obtained. The backpropagation algorithm optimizes the network weight parameters. The network is iteratively trained until the classification accuracy on the validation set converges. The network parameters are saved to obtain the gated graph convolutional network.

[0140] Specifically, the dynamic lung field functional map is input into a gated graph convolutional network, which parses the node attribute data of the dynamic lung field functional map. The node attribute data includes the area value at time points, the spatial structure complexity index, and the lung field expansion function ratio parameter. The dynamic edge connection mechanism establishes the connection weights of temporally adjacent nodes based on the similarity of respiratory phases. The node attribute data is aggregated into the features of neighboring nodes through graph convolutional layers. The gating unit controls the intensity of node feature state update. The updated node features are passed through a fully connected classification layer. The fully connected classification layer outputs the probability distribution of etiology categories. The probability distribution of etiology categories is used to determine the final classification result through ArgMax operation.

[0141] The system combines lung field expansion function ratio, spatial structure complexity index, classification results, and continuous lung field area change curves to generate a structured report using a pre-stored clinical guideline template.

[0142] It should be noted that the pre-stored clinical guideline template refers to a structured text framework compiled based on medical diagnostic standards, which includes disease definition description paragraphs, quantitative parameter matching tables, and diagnostic conclusion generation rules.

[0143] Specifically, the lung field expansion function ratio, spatial structure complexity index, classification results, and continuous lung field area change curve are input into the clinical report generation engine. The clinical report generation engine loads a pre-stored clinical guideline template, which contains diagnostic criteria paragraphs for diseases such as COPD and pulmonary fibrosis. The lung field expansion function ratio is filled into the functional quantification field of the clinical guideline template, the spatial structure complexity index is filled into the structural analysis field of the clinical guideline template, the classification results are matched with the etiology conclusion field of the clinical guideline template, the continuous lung field area change curve is converted into a waveform graph, and inserted into the image attachment area of ​​the clinical guideline template. The clinical report generation engine automatically combines and fills the results to generate a complete report draft, which is then converted into a structured report by the DICOM-SR encoder.

[0144] In summary, this invention generates a differential homeomorphic transformation field using a Lie group exponential mapping algorithm, strictly preserving the connectivity of lung lobe boundaries and the bronchial tree topology during three-dimensional deformation. After performing geometric compensation frame-by-frame on dynamic DR images, the anatomical fidelity of the output lung field temporal image sequence is significantly improved, enhancing topological conservation, structural integrity, and functional quantification. This improves the visualization accuracy of alveolar expansion trajectories during deep breathing, providing reliable anatomical support for the diagnosis of prodromal COPD.

[0145] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based dynamic DR lung field area diagnostic system, characterized in that: include, The image acquisition optimization module captures low-dose dynamic DR sequences of the patient's complete cycle of calm breathing and expedited breathing, generates dynamic DR images, and optimizes the dynamic DR images using a multimodal transfer learning strategy, outputting DR temporal tensors. The displacement analysis module uses a convolutional segmentation network to segment the DR temporal tensor frame by frame, generate an initial binarized lung field mask, and extract the initial lung field contour of each frame. The initial lung field contours of all frames are input into a spatiotemporal convolutional neural network, which analyzes the lung field contour displacement vector between adjacent frames frame by frame and outputs a velocity field mapping map. The deformation compensation module, based on the velocity field mapping map, generates a reversible deformation field through the Lie group exponential mapping algorithm, forces the anatomical structure topological invariance constraint to satisfy the differential homeomorphic transformation field, performs geometric compensation on dynamic DR images frame by frame, and outputs lung field temporal image sequence. The functional quantization module uses a convolutional segmentation network to re-segment the temporal image sequence of the lung field, generating a binary lung field mask, calculating the lung field area value of a single frame, forming a continuous lung field area change curve, and extracting the mean, maximum, minimum, and area change rate ((maximum area value - minimum area value) / minimum area value) during quiet breathing; and the mean, maximum, minimum, and area change rate ((maximum area value - minimum area value) / minimum area value) and peak area during deep breathing, calculating the lung field expansion function ratio; obtaining relevant area estimation values ​​based on the lung field area estimation model of human physiological indicators of normal lung function populations, and calculating the inspiratory index ((actual maximum area value - estimated mean area value) / estimated mean area value) and expiratory index ((actual minimum area value - estimated mean area value) / estimated mean area value) during quiet breathing, and the inspiratory index ((actual maximum area value - estimated mean area value) / estimated mean area value) and expiratory index ((actual minimum area value - estimated mean area value) / estimated mean area value) during deep breathing. The structural analysis module performs spatial region segmentation on the lung field temporal image sequence, constructs an anatomical set covering the local to the global, extracts the gray-level distribution features of the anatomical set, and generates a spatial structural complexity index. The intelligent diagnostic module combines the lung field expansion function ratio and spatial structure complexity index, uses a gated graph convolutional network to classify etiologies, outputs classification results, and generates a structured report by combining the lung field expansion function ratio, spatial structure complexity index and continuous lung field area change curve.

2. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 1, characterized in that: The specific steps for outputting the DR temporal tensor are as follows. Low-dose dynamic DR sequences are captured to generate dynamic DR images by capturing the complete cycle of a patient's calm breathing and expedited breathing. Dynamic DR images are input into a multi-scale self-supervised augmentation network, and optimized dynamic DR sequences are generated through cascaded wavelet decomposition and spatial residual learning. Neural radiation field reconstruction is performed on the optimized dynamic DR sequence to generate the DR temporal tensor.

3. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 2, characterized in that: The specific steps for extracting the initial lung field contour for each frame are as follows: The DR temporal tensor is decomposed into frame-by-frame three-dimensional volume data along the time dimension, and input into a convolutional segmentation network of deformable convolutional layers and Transformer encoders. The position of the convolutional sampling points is dynamically adjusted to adapt to respiratory deformation, and cross-frame attention is used to constrain segmentation continuity, outputting an initial binarized lung field mask for each frame. A morphological refinement operation is performed on the initial binarized lung field mask for each frame to extract the initial lung field contour.

4. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 3, characterized in that: The specific steps for generating the output velocity field mapping are as follows. Register the initial lung field contours of all frames to a pre-stored standard lobar partition template to generate a lung field contour sequence with anatomical markers; The lung field contour sequence with anatomical markers is input into a two-branch spatiotemporal convolutional network. Geometric features are extracted through the spatial branch, while anatomical constraint features are extracted through the anatomical branch. Geometric features and anatomical constraint features are fused through a mutual attention mechanism to output the lung field contour displacement vector between adjacent frames; The lung field contour displacement vector is encoded by time differentiation and spatial normalization to generate a velocity field mapping.

5. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 4, characterized in that: The specific steps for constructing the differential homeomorphic transformation field are as follows. Multiscale continuous homology analysis was performed on the velocity field mapping to extract the generation and extinction interval parameters of the lung field connectivity components and the annular structure. The parameters of the birth and death intervals are combined with the velocity field mapping diagram to perform Lie group exponential mapping, generating the initial deformation field, and the Jacobian matrix is ​​calculated. The initial deformation field and Jacobian matrix are input into the topology constraint optimizer. The topology durability loss value is calculated based on the birth and death interval parameters. The deformation field is iteratively optimized by combining the determinant value of the Jacobian matrix with constraints. When the determinant of the Jacobian matrix of the optimized deformation field is greater than the preset invertibility threshold and the topological durability loss is less than the preset convergence threshold, the differential homeomorphic transformation field is output.

6. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 5, characterized in that: The specific steps for outputting the temporal image sequence of the lung fields are as follows. Based on dynamic DR imaging, the X-ray attenuation coefficient of major tissue regions in the lung field was calculated. By combining the X-ray attenuation coefficient with the differential homeomorphic transformation field, geometrically corrected images are generated through ray tracing compensation. Anisotropic diffusion is performed on the geometrically corrected images to output a temporal sequence of lung fields.

7. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 6, characterized in that: The specific steps for forming a continuous lung field area change curve are as follows. The lung field temporal image sequence is input into a dynamic deformation adaptive convolutional network, and the lung field is segmented frame by frame through the convolution kernel deformation mechanism to generate a binary lung field mask. Morphological refinement and subpixel gradient correction are performed on each frame of the binarized lung field mask to generate subpixel precision lung field contours. Based on the sub-pixel precision lung field contour, the lung field area value of a single frame is calculated by Green's theorem area integration. Adaptive temporal filtering is applied to the lung field area values ​​of a single frame to generate continuous lung field area change curves.

8. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 7, characterized in that: The specific steps for calculating the lung field expansion function ratio are as follows: Identify the time windows between quiet breathing and deep breathing from continuous lung field area change curves; The mean area was obtained within the time window of the quiet breathing period as the mean area of ​​the quiet breathing period; The maximum area value within the deep breathing period time window is extracted as the peak area during the deep breathing period. The lung field expansion function ratio is calculated by dynamically coupling the mean area during quiet breathing and the peak area during deep breathing.

9. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 8, characterized in that: The specific steps for generating the spatial structure complexity index are as follows: Spatial regional subdivision of the temporal imaging sequence of the lung field was performed to form a three-layer anatomical set of alveolar, lobular, and lobar levels; Extract the grayscale distribution of pixels within the three-layer anatomical set, generate a grayscale feature histogram, and extract the grayscale distribution features; By integrating grayscale distribution features through multi-scale feature fusion rules, a spatial structure complexity index is output.

10. The deep learning-based dynamic DR lung field area diagnostic system as described in claim 9, characterized in that: The specific steps for generating the structured report are as follows: A dynamic lung field function map is constructed based on the lung field expansion function ratio, spatial structure complexity index, and continuous lung field area change curve. The dynamic lung field functional map is input into a gated graph convolutional network, and etiological classification is performed through node attribute fusion and dynamic edge connection, and the classification results are output. The system combines lung field expansion function ratio, spatial structure complexity index, classification results, and continuous lung field area change curves to generate a structured report using a pre-stored clinical guideline template.

Images