Deep learning-based method and system for rapid generation of three-dimensional model of chest wall bone defect

Abstract

The application discloses a deep learning-based chest wall bone defect three-dimensional model rapid generation method and system, and relates to the technical field of graphic image processing and three-dimensional modeling. The deep learning-based chest wall bone defect three-dimensional model rapid generation method and system comprise the following steps: S1, constructing a chest wall image volume data set and preprocessing the chest wall image volume data set; S2, identifying the spatial distribution characteristics of chest wall bone related structures based on the chest wall image volume data set; S3, deducing a complete structure reference expression around a chest wall bone structure three-dimensional expression result; and S4, combining a complete chest wall bone three-dimensional structure model and the chest wall bone structure three-dimensional expression result to evaluate the spatial difference relationship. The method solves the problem that, in the three-dimensional modeling process of chest wall bone defects, there is a lack of a complete structure generation reference that can be calculated, and the completion result of the defect area has uncertainty in the spatial form.

Description

Technical Field

[0001] This invention relates to the fields of graphics and image processing and 3D modeling technology, specifically to a method and system for rapidly generating 3D models of chest wall bone defects based on deep learning. Background Technology

[0002] With the continuous development of medical imaging technology and computer-aided 3D modeling technology, chest tomographic imaging has become the primary technical means for acquiring spatial structural information of the chest wall bones. Computed tomography and magnetic resonance imaging (MRI) equipment can continuously acquire volumetric image data with spatial coordinate correlation, providing a data foundation for the 3D reconstruction of chest wall bone structures. Modeling and analyzing bony structures based on volumetric image data has become a common research and application direction in the fields of medical image processing and 3D structural representation.

[0003] For example, invention patent CN118762128B discloses a method and system for generating spinal models based on deep learning. It uses multiple CT images of the spine from various local perspectives, acquired from multiple angles, and incorporates image processing and analysis algorithms based on artificial intelligence and deep learning in the backend to analyze these images. This allows it to learn and capture key state features of the spine from these local perspectives, and then performs panoramic spinal state semantic fusion to reconstruct a more accurate spinal model. In this way, intelligent recognition and reconstruction of spinal models can be achieved based on CT images from different local perspectives, assisting doctors in diagnosing spinal diseases and formulating treatment plans.

[0004] For example, the invention patent with announcement number CN110084817B discloses a method for producing digital elevation models based on deep learning, which relates to the field of computer vision technology. By creating a set of target remote sensing images, creating a semantic segmentation network based on deep learning, training the semantic segmentation network, and creating a digital elevation model, it can effectively segment massive amounts of remote sensing images and improve segmentation efficiency.

[0005] The construction of existing 3D models of the chest wall bones is typically based on chest tomographic image data. Bone structure regions are extracted through manual or automated segmentation, and then interpolation, registration, and template derivation are used to generate the 3D structural model. Common implementation methods include grayscale feature-based segmentation, spatial connectivity-based structure extraction, and automated segmentation workflows based on learning models. These methods often employ a uniform processing path to generate the structure, resulting in limited adaptability to variations in the spatial continuity and temporal changes of image data.

[0006] To address the above issues, there is an urgent need for a method and system for rapidly generating 3D models of chest wall bone defects based on deep learning. Summary of the Invention

[0007] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method and system for rapid generation of 3D models of chest wall bone defects based on deep learning. This solves the problem that the lack of a computable complete structural reference in the 3D modeling process of chest wall bone defects leads to uncertainty in the spatial morphology of the completed defect area.

[0008] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a method for rapidly generating a three-dimensional model of chest wall bone defects based on deep learning, comprising the following steps: S1, acquiring basic image data generated during chest tomographic imaging, constructing a chest wall image volume dataset, and preprocessing the chest wall image volume dataset; S2, identifying the spatial distribution characteristics of chest wall bone-related structures based on the chest wall image volume dataset, forming a three-dimensional expression result of the chest wall bone structure consistent with the image coordinate system; S3, deriving a complete structural reference expression based on the three-dimensional expression result of the chest wall bone structure, generating a complete three-dimensional chest wall bone structure model aligned with the original structural space; S4, evaluating the spatial difference relationship by combining the complete three-dimensional chest wall bone structure model and the three-dimensional expression result of the chest wall bone structure, forming a three-dimensional structural model expression of the defect area, and outputting the three-dimensional model of the chest wall bone defect.

[0009] Furthermore, the specific steps for constructing the chest wall image volume dataset by acquiring the basic image data generated during the chest tomographic imaging process are as follows: Acquire the basic image data generated by the computed tomography (CT) scanner and magnetic resonance imaging (MRI) scanner during the chest tomographic imaging process. The basic image data includes: original image volume data, row direction spatial index number corresponding to the voxel, column direction spatial index number corresponding to the voxel, slice sequence information, inter-slice step size information, image volume size range information, effective imaging area identification information, and the image sampling timestamp sequence formed during continuous scanning; Perform preliminary structural organization and data merging of the basic image data according to the sampling time order to construct the chest wall image volume dataset.

[0010] Furthermore, the specific steps for preprocessing the chest wall image dataset are as follows: The arrangement of image volumes on the time and space axes is corrected through timestamp consistency verification and spatial index continuity checks; discrete noise voxels and discontinuous regions appearing in the image volumes are trimmed by combining three-dimensional connected component analysis and morphological structural element-based closing operations; histogram matching is introduced to align the overall grayscale distribution of the image volumes to address differences in image grayscale distribution under different scanning conditions; a sliding time window and exponentially weighted moving average algorithm are used in the time dimension to smooth intensity fluctuations between continuously sampled image volumes; and standardization and normalization are performed on the chest wall image dataset.

[0011] Further, the specific steps for identifying the spatial distribution features of chest wall bone-related structures based on the chest wall image volume dataset are as follows: Locate voxel positions along the storage order of the original image volume data, read the corresponding recorded grayscale values, and obtain the image voxel intensity values; determine the voxel's arrangement position in the row direction based on the row direction spatial index number stored in the image volume data, and obtain the voxel's lateral spatial index coordinates; determine the voxel's arrangement position in the column direction based on the column direction spatial index number stored in the image volume data, and obtain the voxel's longitudinal spatial index coordinates; combine the slice sequence information and inter-slice step size information recorded in the image volume data to calibrate the voxel's position in the slice scan sequence, and obtain the voxel's slice-oriented spatial index coordinates; within the spatial range defined by the image volume size range information and the effective imaging area identification information, collect the areas marked as effective regions in the image volume data. The set of voxels is used to determine the effective spatial region of the image. Within the effective spatial region, the image voxel intensity values ​​are read according to the combination relationship of the voxel horizontal spatial index coordinates, voxel vertical spatial index coordinates, and voxel layer directional spatial index coordinates. Based on the numerical differences between adjacent voxels, the variation amplitude of the image voxel intensity values ​​in the three spatial directions is calculated to form a spatial gradient expression of the image voxel intensity values. The magnitude of the spatial gradient expression is calculated to obtain the spatial gradient amplitude of the corresponding voxel position. The spatial gradient amplitude is squared to form the square value of the spatial structure intensity of a single voxel. Within the limited range of the effective spatial region of the image, the square values ​​of the spatial structure intensity corresponding to all voxels are accumulated according to the voxel volume unit to obtain the overall integral result of the square of the spatial structure intensity. The square root operation is performed on the overall integral result to obtain the effective value of the spatial structure.

[0012] Furthermore, the specific steps for forming a three-dimensional representation of the chest wall bone structure consistent with the image coordinate system are as follows: the effective value of the spatial structure is compared with the effective threshold of the structure in real time; when the effective value of the spatial structure is greater than or equal to the effective threshold of the structure, the chest wall bone structure extraction process continues and the three-dimensional representation of the chest wall bone structure is output; when the effective value of the spatial structure is less than the effective threshold of the structure, the output of the three-dimensional representation of the chest wall bone structure is stopped and the structure extraction failure status code is returned.

[0013] Furthermore, the specific steps for deriving the complete structural reference representation from the three-dimensional representation results of the chest wall bone structure are as follows: Based on the order of the image sampling timestamp sequence, trace back the corresponding slice sequence position or historical scan record along the scanning time sequence before the current image volume being analyzed, read the matching original image volume data, and obtain the historical sampled image value; locate the image sampling timestamp immediately preceding the historical sampled image value along the time sequence, read the corresponding original image volume data, and obtain the preceding sampled image value; combine the slice sequence or the distribution of multiple scan time points in the image sampling timestamp sequence, convert the continuously covered time range, determine the number of samples to participate in the analysis, and obtain the time sliding window. The time interval between adjacent image volume data is determined based on the time difference between adjacent time markers in the image sampling timestamp sequence, thus obtaining the adjacent sampling time interval. Within the image volume data range limited by the time window length, each historical sampled image value and its corresponding preceding sampled image value are selected sequentially in chronological order. The difference between the historical sampled image value and the preceding sampled image value is obtained, and the difference is divided by the adjacent sampling time interval to obtain the ratio result. The absolute value of the comparison result is taken to form the single-cycle rate of change amplitude. The amplitude accumulation result is obtained by summing all single-cycle rate of change amplitudes corresponding to the time window length. The amplitude accumulation result is divided by the time window length to obtain the time stability determination value.

[0014] Furthermore, the specific steps for generating a complete three-dimensional chest wall bone structure model aligned with the original structural space are as follows: The time stability determination value is compared in real time with a stability threshold, which includes a primary stability threshold and a secondary stability threshold; when the time stability determination value is less than the primary stability threshold, the complete structure generation path is entered and the complete three-dimensional chest wall bone structure model generation operation is performed; when the time stability determination value is greater than or equal to the primary stability threshold and less than the secondary stability threshold, the restricted structure generation path is entered, and the time stability determination value is written to the status record field; when the time stability determination value is greater than or equal to the secondary stability threshold, the structure generation path is blocked, and a prohibited status flag is generated.

[0015] Further, the specific steps for evaluating the spatial difference relationship by combining the complete three-dimensional structural model of the chest wall bone with the three-dimensional expression results of the chest wall bone structure are as follows: Based on the horizontal, vertical, and layer-wise neighborhood offsets, the spatial index coordinates of the target voxels are offset and located. Grayscale records from the original image volume data are read at the offset voxel positions to obtain the intensity values ​​of adjacent image voxels. The horizontal neighborhood offset is determined based on the index interval formed between adjacent voxels according to the row-wise spatial index number. The vertical neighborhood offset is determined based on the index interval formed between adjacent voxels according to the column-wise spatial index number. The positional difference between adjacent layer voxels is converted by combining the slice number information and the inter-layer step size information to determine the layer-wise neighborhood offset. The image voxel intensity value is read from the spatial position determined by the voxel's horizontal spatial index coordinates, vertical spatial index coordinates, and layer-direction spatial index coordinates. The voxel's horizontal, vertical, and layer-direction spatial index coordinates are then offset according to the horizontal, vertical, and layer-direction neighborhood offsets to obtain the coordinates of adjacent voxels. The adjacent image voxel intensity value is read at the adjacent voxel coordinates. The spatial difference amplitude is obtained by subtracting the adjacent image voxel intensity value from the image voxel intensity value and taking the absolute value. The absolute value of the second-order time variation of the image voxel intensity value at consecutive sampling time points is taken, and the absolute value is added to one to obtain the denominator. The spatial difference amplitude is divided by the denominator to obtain the differential confidence control value.

[0016] Furthermore, the specific steps for expressing and outputting the three-dimensional structural model of the defect area and the three-dimensional model of the chest wall bone defect are as follows: Differential confidence control values ​​are written into the differential result output control parameter set; based on the differential confidence control values, the voxel writing density parameter, the boundary smoothing intensity parameter, and the model output resolution parameter are simultaneously updated numerically; during the differential result writing process, the number of voxels involved in the writing is controlled according to the voxel writing density parameter, the smoothing processing amplitude of the differential structure boundary is controlled according to the boundary smoothing intensity parameter, and the mesh refinement degree of the three-dimensional structure of the defect area is controlled according to the model output resolution parameter; within each processing cycle, the differential confidence control values ​​continuously participate in the parameter update process, completing the differential structure writing and outputting the three-dimensional model of the chest wall bone defect.

[0017] The second aspect of this invention provides a rapid generation system for three-dimensional models of chest wall bone defects based on deep learning, comprising: an image data acquisition module for acquiring basic image data generated during chest tomographic image scanning, constructing a chest wall image volume dataset, and preprocessing the chest wall image volume dataset; a skeleton structure segmentation module for identifying the spatial distribution characteristics of chest wall bone-related structures based on the chest wall image volume dataset, forming a three-dimensional expression result of the chest wall bone structure consistent with the image coordinate system; a complete structure generation module for deriving a complete structure reference expression based on the three-dimensional expression result of the chest wall bone structure, generating a complete three-dimensional chest wall bone structure model aligned with the original structure space; and a defect difference output module for evaluating the spatial difference relationship by combining the complete three-dimensional chest wall bone structure model and the three-dimensional expression result of the chest wall bone structure, forming a three-dimensional structural model expression of the defect area and outputting the three-dimensional model of the chest wall bone defect.

[0018] Beneficial effects The present invention has the following beneficial effects: (1) This invention introduces a unified construction and preprocessing process for chest wall image volume dataset in the sovereign technical solution, which organizes the original image volume data, spatial index information and image sampling timestamp sequence in a structured manner, and constrains the image arrangement relationship through timestamp consistency check and spatial index continuity check, so that subsequent processing is always based on continuous and consistent image data, effectively reducing the impact of data dispersion under different equipment and different scanning conditions on spatial analysis results, and improving the stability and reproducibility of the overall data processing process.

[0019] (2) This invention uses a computational mechanism based on spatial gradient amplitude and effective spatial structure value to comprehensively evaluate the voxel structure intensity in the effective spatial area of ​​the image, and combines the effective structural threshold to control the output conditions of the three-dimensional expression results of the chest wall bone structure, so that the structure extraction process has clear data admission criteria. Compared with the existing methods that rely on manual experience or fixed rules, it can actively screen data areas with effective spatial characteristics at the software level, and improve the consistency and reliability of structure expression.

[0020] (3) This invention introduces a time stability judgment value in the process of complete structure reference derivation, and performs hierarchical control of the generation path based on the first-level stability threshold and the second-level stability threshold, so that the complete structure generation process is associated with the change state of the image sampling time series, avoiding direct entry into the complete structure generation process under data conditions with large time fluctuations, thereby forming a controlled generation path selection mechanism in software logic, and improving the consistency and controllability of the generation results in the continuous processing cycle.

[0021] (4) By directly participating the differential confidence control value in the continuous updating of voxel writing density parameter, boundary smoothing intensity parameter and output resolution parameter, the present invention transforms the output process of the three-dimensional structure of the defect region from static threshold control to dynamic parameter adjustment mode. Without introducing additional manual intervention, the output intensity is automatically adjusted according to the spatial difference and time change characteristics. Compared with the existing one-time differential output method, it enhances the adaptability of the defect structure expression under different data conditions.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method for rapidly generating a three-dimensional model of chest wall bone defects based on deep learning, as described in this invention. Figure 2 This is a structural diagram of the system for rapidly generating three-dimensional models of chest wall bone defects based on deep learning, as described in this invention. Figure 3 This is a distribution diagram of the time stability determination values ​​of the present invention; Figure 4 This is a schematic diagram of the shrinkage of the defect differential space according to the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1-4 This invention provides a technical solution: a method for rapidly generating a three-dimensional model of chest wall bone defects based on deep learning, comprising the following steps: S1, acquiring basic image data generated during chest tomographic imaging, constructing a chest wall image volume dataset, and preprocessing the chest wall image volume dataset; S2, identifying the spatial distribution characteristics of chest wall bone-related structures based on the chest wall image volume dataset, forming a three-dimensional expression result of the chest wall bone structure consistent with the image coordinate system; S3, deriving a complete structural reference expression based on the three-dimensional expression result of the chest wall bone structure, generating a complete three-dimensional chest wall bone structure model aligned with the original structural space; S4, evaluating the spatial difference relationship by combining the complete three-dimensional chest wall bone structure model and the three-dimensional expression result of the chest wall bone structure, forming a three-dimensional structural model expression of the defect area, and outputting the three-dimensional model of the chest wall bone defect.

[0026] Specifically, the steps for collecting basic image data generated during chest tomography and constructing a chest wall image volume dataset are as follows: Basic image data generated by the computed tomography (CT) scanner and magnetic resonance imaging (MRI) scanner during chest tomography are collected. This basic image data includes raw image volume data and spatial and temporal metadata corresponding to each image voxel. The raw image volume data carries the voxel grayscale distribution of the chest tomography image. Row and column spatial index numbers are generated by the imaging equipment during scanning according to pixel matrix arrangement rules and mapped to the two-dimensional arrangement of voxels in the image plane using image head information. This mapping is used to establish the ordered coordinate relationship of voxels in the horizontal and vertical directions during subsequent spatial structure analysis. Slice number information and inter-slice step size information are used to calibrate the voxels in the slice sequence. The layer position and physical spacing between adjacent slices provide a scale basis for 3D spatial reconstruction and voxel layer index calculation. The image volume size range information is used to limit the overall boundary of the image volume in 3D space, avoiding invalid areas from participating in subsequent calculations. The effective imaging area identification information is used to distinguish the imaging area from the non-imaging background area, providing a basis for determining the effective spatial area of ​​the image. The image sampling timestamp sequence formed during continuous scanning is used to record the sampling time of each image volume data, providing a time axis basis for the subsequent calculation of time stability judgment values. The basic image data is structurally organized and merged according to the order of the image sampling timestamp sequence, and the original image volume data is bound and stored with the corresponding spatial index information and time stamp information to form a chest wall image volume dataset with a unified coordinate system and time index relationship.

[0027] In this implementation plan, after the spatial index information and time stamp information in the basic image data are uniformly organized and bound, the chest wall image volume dataset forms a continuous and traceable data organization structure on the spatial coordinate system and time axis. This ensures that subsequent spatial structure analysis, temporal stability determination and differential operation are all based on a consistent data foundation, thereby reducing processing deviations caused by data misalignment or missing information and improving the stability and consistency of the overall image processing workflow.

[0028] Specifically, the preprocessing steps for the chest wall image dataset are as follows: The arrangement of image volumes along the time and space axes is corrected through timestamp consistency checks and spatial index continuity checks, ensuring that continuously sampled image volumes maintain a consistent correspondence in both temporal order and spatial location, avoiding the impact of sampling misorders and index jumps on subsequent analysis; Discrete noise voxels and discontinuous regions in the image volumes are trimmed using three-dimensional connected component analysis and morphological structural element-based closing operations, ensuring structural boundaries remain continuous at the voxel level and providing stable input for spatial structural representation; For different scanning... To address the differences in grayscale distribution of images under various conditions, a histogram matching method is introduced to align the overall grayscale distribution of the image volume, ensuring comparability in intensity and reducing interference from equipment differences and scanning parameter variations in subsequent processing. In the time dimension, a sliding time window and exponentially weighted moving average algorithm are employed to smooth intensity fluctuations between consecutively sampled image volumes, maintaining a continuous trend in the image volume sequence over time. Standardization and normalization are performed on the chest wall image volume dataset to ensure uniformity in numerical range and scale, meeting the consistency requirements of input data in subsequent computational processes.

[0029] In this implementation scheme, after preprocessing, the chest wall image dataset forms a consistent data foundation in terms of time sequence, spatial location, and grayscale distribution. This enables the images to have continuous and comparable structural features, reduces the impact of noise interference and discontinuous regions on spatial analysis, and ensures that subsequent spatial structure extraction, complete structure reference derivation, and differential evaluation processes are executed under stable input conditions, thereby improving the reliability and consistency of the overall processing flow.

[0030] Specifically, the steps for identifying the spatial distribution features of chest wall bone-related structures based on chest wall image volume datasets are as follows: Locate voxel positions along the storage order of the original image volume data, read the corresponding recorded grayscale values, and obtain the image voxel intensity values; determine the voxel's arrangement position in the row direction based on the row direction spatial index number stored in the image volume data, and obtain the voxel's lateral spatial index coordinates; determine the voxel's arrangement position in the column direction based on the column direction spatial index number stored in the image volume data, and obtain the voxel's longitudinal spatial index coordinates; combine the slice sequence information and inter-slice step size information recorded in the image volume data to calibrate the voxel's position in the slice scanning sequence, and obtain the voxel's slice-oriented spatial index coordinates; The effective imaging area identification information originates from the imaging field of view parameters and background area determination results generated by the scanning equipment during the image acquisition stage, and is used to distinguish the actual imaging area from the scanning edge or non-imaging area, within the image volume size range information and the effective imaging area identification information. Within the spatial range defined by the imaging region identification information, a set of voxels marked as effective regions in the image volume data is collected to determine the effective spatial region of the image. Within the effective spatial region of the image, the image voxel intensity values ​​are read according to the combination relationship of the voxel horizontal spatial index coordinates, voxel vertical spatial index coordinates, and voxel layer directional spatial index coordinates. Based on the numerical differences between adjacent voxels, the variation amplitude of the image voxel intensity values ​​in the three spatial directions is calculated to form a spatial gradient expression of the image voxel intensity values. The magnitude of the spatial gradient expression is calculated to obtain the spatial gradient amplitude of the corresponding voxel position. The spatial gradient amplitude is squared to form the square value of the spatial structure intensity of a single voxel. Within the defined effective spatial region of the image, the square values ​​of the spatial structure intensity corresponding to all voxels are accumulated according to the voxel volume unit to obtain the overall integral result of the square of the spatial structure intensity. The square root operation is performed on the overall integral result to obtain the effective value of the spatial structure.

[0031] The specific calculation method for the effective value of spatial structure is as follows: In the formula, Indicates the effective value of the spatial structure. Indicates the image voxel intensity value. Represents the horizontal spatial index coordinates of a voxel. Represents the vertical spatial index coordinates of a voxel. This represents the spatial index coordinates of the voxel layer. Indicates the effective spatial area of ​​the image.

[0032] In this implementation scheme, after the spatial distribution feature is identified, the chest wall image volume dataset forms a continuous and quantifiable structural intensity expression within the effective spatial area of ​​the image. This allows for a unified characterization of the boundary changes and structural concentration of chest wall bone-related structures in three-dimensional space, providing a stable and comparable spatial analysis basis for subsequent structural validity determination and complete structural reference derivation.

[0033] Specifically, the steps to generate a three-dimensional representation of the chest wall bone structure consistent with the image coordinate system are as follows: The effective value of the spatial structure is compared with the effective threshold in real time, providing a clear data entry point for the chest wall bone structure extraction process; when the effective value of the spatial structure is greater than or equal to the effective threshold, it is determined that the spatial structure intensity within the effective spatial area of ​​the image meets the requirements for continuous expression of the chest wall bone structure, and the chest wall bone structure extraction process continues, outputting the three-dimensional representation of the chest wall bone structure while maintaining the original image coordinate system; when the effective value of the spatial structure is less than the effective threshold, it is determined that the current image volume data is insufficient to support a stable chest wall bone structure representation, the output of the three-dimensional representation of the chest wall bone structure is stopped, and a structure extraction failure status code is returned to indicate the structure extraction status within the current processing cycle.

[0034] In this implementation plan, after the structural validity determination is completed, the output conditions of the three-dimensional expression results of the chest wall bone structure are clearly defined, so that the structural extraction process only enters the effective output path when the spatial structural strength meets the requirements, avoiding incomplete and unstable structural results from entering the subsequent processing flow, and ensuring the consistency and usability of the three-dimensional expression results of the chest wall bone structure in the spatial coordinate system.

[0035] Specifically, the steps for deriving the complete structural reference representation based on the three-dimensional representation of the chest wall bone structure are as follows: Based on the order of the image sampling timestamp sequence, trace back the corresponding slice sequence position or historical scan record along the scanning time sequence before the current image volume being analyzed, read the matching original image volume data, and obtain the historical sampled image values ​​to characterize the continuous state of the chest wall image in adjacent scan cycles; locate the image sampling timestamp immediately preceding the historical sampled image value along the time sequence, read the corresponding original image volume data, and obtain the preceding sampled image value to construct the change relationship between adjacent sampling cycles; combine the slice sequence or the distribution of multiple scan time points in the image sampling timestamp sequence to convert the continuously covered time range, determine the number of samples participating in the analysis, obtain the time sliding window length, and ensure that the time analysis range is consistent with the image sampling rhythm; based on the adjacent timestamps in the image sampling timestamp sequence... The time difference between adjacent image data is used to determine the time interval between adjacent sampling intervals, which reflects the actual span of image changes on the time axis. Within the image data range defined by the time window length, each historical sampled image value and its corresponding preceding sampled image value are selected sequentially in chronological order to construct a continuous time comparison sequence. The difference between the historical sampled image value and the preceding sampled image value is obtained, and the difference is normalized in combination with the adjacent sampling time interval to obtain a ratio result reflecting the amplitude of image changes per unit time. The absolute value of the comparison result is taken to form the single-cycle rate of change amplitude, so that changes in different directions are uniformly included in the same evaluation scale. The amplitude accumulation result is obtained by summing all single-cycle rate of change amplitudes corresponding to the time window length, and the amplitude accumulation result is divided by the time window length to obtain the time stability judgment value, which is used to characterize the overall stability of chest wall images during continuous sampling.

[0036] The specific calculation method for the time stability criterion value is as follows: In the formula, Indicates the time stability determination value. Represents historical sampled image values. This represents the value of the preceding sampled image. Indicates the length of the time sliding window. This indicates the time interval between adjacent sampling.

[0037] Table 1 shows the time stability determination value data table provided in the embodiments of this application. For determination 1, the historical sampled image value is set to 1200.10, the preceding sampled image value is set to 1199.80, the time window length is set to 1.00, and the adjacent sampling time interval is set to 0.50; for determination 2, the historical sampled image value is set to 1200.40, the preceding sampled image value is set to 1199.60, the time window length is set to 1.00, and the adjacent sampling time interval is set to 0.40; for determination 3, the historical sampled image value is set to 1200.30, the preceding sampled image value is set to 1199.80, and the adjacent sampling time interval is set to 0.40. 99.50, the time window length is set to 1.00, and the adjacent sampling time interval is set to 0.60; the historical sampled image value of judgment 4 is set to 1200.90, the previous sampled image value is set to 1199.40, the time window length is set to 1.00, and the adjacent sampling time interval is set to 0.50; the historical sampled image value of judgment 5 is set to 1200.60, the previous sampled image value is set to 1199.20, the time window length is set to 1.00, and the adjacent sampling time interval is set to 0.40.

[0038] Table 1. Time Stability Judgment Values ​​Data Table like Figure 3 The figure shows the distribution of time stability determination values ​​provided in the embodiments of this application. Combining the table and image data, it can be seen that the first-level stability threshold and the second-level stability threshold, serving as graded references for time stability, are 0.90 and 1.50, respectively. The time stability determination values ​​corresponding to the five sets of determinations are distributed between 0.60 and 3.50, showing an overall gradually increasing trend. The time stability determination value of determination 1 is 0.60, significantly lower than the first-level stability threshold, reflecting a relatively smooth change in the image sequence; the time stability determination value of determination 2 is 2.00, exceeding the first-level stability threshold, indicating that the amplitude of time change is beginning to increase; the time stability determination value of determination 3 is 1.33, near the first-level stability threshold; the time stability determination values ​​of determination 4 and determination 5 are 3.00 and 3.50, respectively, both exceeding the second-level stability threshold, indicating that the corresponding image sampling sequences have strong fluctuation characteristics in the time dimension.

[0039] In this implementation scheme, after the time stability determination value is formed, the change state of the chest wall image during the continuous sampling process is quantitatively characterized, so that the complete structure reference derivation process has a determination basis based on time characteristics. This ensures the continuity of the generation path under the condition that the image changes are smooth and the fluctuations are controlled, avoids the interference of time fluctuations on the consistency of the complete structure reference, and enhances the stability and controllability of the subsequent structure generation process.

[0040] Specifically, the steps for generating a complete 3D chest wall bone structure model aligned with the original structural space are as follows: The time stability determination value is compared in real time with a stability threshold, which includes a primary stability threshold and a secondary stability threshold, establishing a clear control relationship between the complete structure generation process and the temporal changes in the image. When the time stability determination value is less than the primary stability threshold, the chest wall image is determined to be stable in the time dimension, and the complete structure generation path is entered, executing the generation operation of the complete 3D chest wall bone structure model. The generation process maintains consistent alignment with the original structural space coordinate system. When the time stability determination value is greater than or equal to the primary stability threshold and less than the secondary stability threshold, the chest wall image is determined to be in an acceptable fluctuation state, entering the restricted structure generation path, and the time stability determination value is written into the status record field for subsequent consistency analysis and path control. When the time stability determination value is greater than or equal to the secondary stability threshold, the chest wall image is determined to be in an unstable state, blocking the structure generation path and generating a prohibited status marker to indicate that the current processing cycle does not meet the conditions for complete structure generation.

[0041] In this implementation scheme, after the generation path control is completed, a clear correspondence is established between the generation process of the complete three-dimensional chest wall bone structure model and the time change state. This ensures that the structure generation only enters the execution state under the condition that the image is stable or the fluctuation is controlled, and forms differentiated generation path selection in different stable intervals. This avoids unstable images from directly participating in the structure generation and ensures that the complete three-dimensional chest wall bone structure model remains consistent in terms of spatial alignment and continuity.

[0042] Specifically, the steps for evaluating spatial differences by combining the complete 3D model of the thoracic wall bone structure with the 3D representation results of the thoracic wall bone structure are as follows: Based on the horizontal, vertical, and layer-wise neighbor offsets, the spatial index coordinates of the target voxels are offset and located. Grayscale records from the original image volume data are read at the offset voxel positions to obtain the intensity values ​​of adjacent image voxels, used to characterize the intensity changes within the local spatial neighborhood. Based on the index interval formed between adjacent voxels using the row-wise spatial index number, the horizontal neighbor offset is determined, providing a clear index basis for the adjacency relationship of voxels in the row direction. Based on the index interval formed between adjacent voxels using the column-wise spatial index number, the vertical neighbor offset is determined, ensuring a consistent spatial step size for the adjacency relationship of voxels in the column direction. Combining slice number information and inter-layer step size information, the positional differences between adjacent layer voxels are converted to determine the layer-wise neighbor offset, used to reflect the positional differences of voxels in the layer direction. The physical spacing relationship is determined; image voxel intensity values ​​are read at spatial positions determined by voxel horizontal spatial index coordinates, voxel vertical spatial index coordinates, and voxel layer-direction spatial index coordinates; adjacent voxel coordinates are obtained by offsetting the voxel horizontal spatial index coordinates, voxel vertical spatial index coordinates, and voxel layer-direction spatial index coordinates according to the horizontal neighborhood offset, vertical neighborhood offset, and layer-direction neighborhood offset; adjacent image voxel intensity values ​​are read at adjacent voxel coordinates to construct voxel-level spatial contrast relationship; the spatial difference amplitude is obtained by subtracting the adjacent image voxel intensity values ​​from the image voxel intensity values ​​and taking the absolute value, which is used to characterize the intensity of local spatial differences; the absolute value of the second-order temporal change of image voxel intensity values ​​at continuous sampling time points is taken, and the absolute value result is added to one to obtain a denominator term, which is used to introduce a temporal change suppression term; the spatial difference amplitude is divided by the denominator term to obtain the differential confidence control value, which is used to comprehensively reflect the confidence level of spatial differences under temporal constraints.

[0043] The specific calculation method for the differential reliable control value is as follows: In the formula, Indicates the differential confidence control value. Indicates the image voxel intensity value. Indicates the voxel intensity values ​​of adjacent images. Represents the horizontal spatial index coordinates of a voxel. Represents the vertical spatial index coordinates of a voxel. This represents the spatial index coordinates of the voxel layer. This represents the horizontal neighborhood offset. This represents the vertical neighborhood offset. This indicates the offset of the layer to its neighborhood.

[0044] In this implementation plan, the spatial difference relationship assessment results form a unified judgment basis, and the difference intensity between the complete three-dimensional structural model of the chest wall bone and the three-dimensional expression result of the chest wall bone structure is stably characterized, so that the defect area maintains a consistent judgment scale under the combined effect of spatial changes and time constraints, and provides reliable data support for the output regulation of the three-dimensional structural expression of the defect area.

[0045] Specifically, the steps for forming a three-dimensional structural model of the defect area and outputting a three-dimensional model of the chest wall bone defect are as follows: Differential confidence control values ​​are written into the differential result output control parameter set to provide a unified control entry point for the differential result output process; based on the differential confidence control values, the voxel writing density parameter, boundary smoothing intensity parameter, and model output resolution parameter are simultaneously updated numerically. The voxel writing density parameter is used to limit the proportion of voxels participating in differential writing, controlling the spatial density of the defect structure representation; the boundary smoothing intensity parameter is used to adjust the continuous processing amplitude of the differential structure boundary at the voxel level, avoiding discrete or abrupt boundary states. The output resolution parameter is used to constrain the mesh refinement level of the 3D structure of the defect area during the output stage to match the subsequent structural requirements. During the differential result writing process, the number of voxels involved in the writing is controlled by the voxel writing density parameter, the smoothing intensity of the differential structure boundary is controlled by the boundary smoothing intensity parameter, and the mesh refinement degree of the 3D structure of the defect area is controlled by the model output resolution parameter. In each processing cycle, the completion of a single differential calculation and result writing is taken as the cycle boundary. The differential confidence control value continues to participate in the above parameter update process until the differential structure writing in the current cycle is completed and the 3D model of the chest wall bone defect is output.

[0046] In this implementation scheme, the output process of the three-dimensional structural model of the defect area is in a controlled state. A coordinated relationship is established between the degree of spatial refinement, boundary continuity and output density, so that the three-dimensional model of the chest wall bone defect can maintain structural integrity while having stable output quality. This ensures that the difference results exhibit consistent expression characteristics in the continuous processing process, and meets the requirements of model expression accuracy and reliability in the subsequent structural use stage.

[0047] like Figure 2The diagram shown is a structural schematic of the deep learning-based rapid generation system for 3D models of chest wall bone defects provided in this embodiment of the application. This system utilizes a deep learning-based method for rapidly generating 3D models of chest wall bone defects, including: an image data acquisition module, used to acquire basic image data generated during chest tomographic imaging and construct a chest wall image volume dataset according to unified data organization rules; and to perform temporal and spatial consistency correction, noise trimming, and numerical standardization on the chest wall image volume dataset to provide a stable data foundation for subsequent analysis; and a skeleton structure segmentation module, used for segmenting the chest wall image volume dataset... The system identifies the spatial distribution characteristics of chest wall bone-related structures under a unified coordinate system and generates a three-dimensional representation of the chest wall bone structure consistent with the image coordinate system, serving as the input basis for structural reference derivation. A complete structure generation module is used to construct a complete structural reference representation based on the three-dimensional representation of the chest wall bone structure, generating a complete three-dimensional chest wall bone structure model while maintaining spatial alignment, thus providing a calculable reference basis for the structural completion process. A defect difference output module is used to evaluate the spatial differences between the complete three-dimensional chest wall bone structure model and the three-dimensional representation of the chest wall bone structure under a unified spatial system, generating a three-dimensional structural model representation of the defect area and outputting a three-dimensional model of the chest wall bone defect for subsequent use.

[0048] In this implementation scheme, the modules work together to form a continuous processing chain. The image data is transformed from acquisition and processing to structural segmentation, reference generation and differential output under a unified spatial coordinate system, so that the chest wall image volume data is stably transformed into a three-dimensional structural model of the defect area. The overall process maintains coordination in terms of data consistency, structural alignment and output stability, providing a complete software processing foundation for the reliable generation of three-dimensional models of chest wall bone defects.

[0049] like Figure 4 The figure shows a schematic diagram of the shrinkage of the differential space in the defect area provided in an embodiment of this application. The figure displays the comparison results of the differential boundaries of the defect area against a white background, and uses lines to indicate two types of contours. The dashed contour represents the original differential boundary, and the solid contour represents the shrunken differential boundary obtained under the differential credibility control constraint. Combining the differential credibility control values ​​and corresponding area parameters marked in the figure, it can be seen that the original differential boundary exhibits outward expansion and irregular undulations in local areas, with a relatively wide boundary range. After introducing differential credibility control, the shrunken differential boundary converges inward overall, the boundary range is significantly reduced, and the contour shape is more compact and has higher continuity. The two types of boundaries maintain the same overall spatial position and orientation, and the comparison is intuitive and clear. This reflects the processing characteristics of the system performing spatial shrinkage and range control on the defect differential results under credibility constraints, thereby suppressing unstable differential regions and preserving reliable defect spatial representation.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for rapid generation of 3D models of chest wall bone defects based on deep learning, characterized in that, Includes the following steps: S1. Acquire basic image data generated during chest tomography scanning, construct chest wall image volume dataset, and preprocess the chest wall image volume dataset. S2, Based on the chest wall image volume dataset, identify the spatial distribution characteristics of chest wall bone-related structures to form a three-dimensional representation of chest wall bone structures consistent with the image coordinate system; S3, based on the three-dimensional expression of the chest wall bone structure, derive the complete structural reference expression and generate a complete three-dimensional chest wall bone structure model that is aligned with the original structural space; S4 combines the complete three-dimensional structural model of the chest wall bone with the three-dimensional expression results of the chest wall bone structure to evaluate the spatial difference relationship, form a three-dimensional structural model of the defect area, and output the three-dimensional model of the chest wall bone defect.

2. The method for rapid generation of three-dimensional models of chest wall bone defects based on deep learning according to claim 1, characterized in that: The specific steps for constructing a chest wall image volume dataset from the basic image data generated during chest tomographic imaging are as follows: Basic image data generated by computed tomography (CT) and magnetic resonance imaging (MRI) devices during chest CT imaging were acquired. The basic image data included: original image volume data, row-direction spatial index numbers corresponding to voxels, column-direction spatial index numbers corresponding to voxels, slice sequence information, inter-slice step size information, image volume size range information, effective imaging area identification information, and image sampling timestamp sequence formed during continuous scanning. The basic image data were preliminarily structured and merged according to the sampling time order to construct a chest wall image volume dataset.

3. The method for rapid generation of three-dimensional models of chest wall bone defects based on deep learning according to claim 1, characterized in that: The specific steps for preprocessing the chest wall image dataset are as follows: The arrangement of image volumes on the time and space axes is corrected by checking timestamp consistency and spatial index continuity. Discrete noise voxels and discontinuous regions in the image volumes are smoothed by combining three-dimensional connected component analysis and morphological structural element-based closing operations. Histogram matching is introduced to align the overall grayscale distribution of the image volumes to address differences in image grayscale distribution under different scanning conditions. In the time dimension, a sliding time window and exponentially weighted moving average algorithm are used to smooth the intensity fluctuations between continuously sampled image volumes. Standardization and normalization are performed on the chest wall image volume dataset.

4. The method for rapid generation of a three-dimensional model of chest wall bone defects based on deep learning according to claim 1, characterized in that: The specific steps for identifying the spatial distribution features of chest wall bone-related structures based on chest wall image data sets are as follows: The voxel positions are located along the storage order of the original image volume data, and the corresponding grayscale values ​​are read to obtain the image voxel intensity values. Based on the row direction spatial index number stored in the image volume data, the voxel arrangement position in the row direction is determined to obtain the voxel horizontal spatial index coordinates. Based on the column direction spatial index number stored in the image volume data, the voxel arrangement position in the column direction is determined to obtain the voxel vertical spatial index coordinates. Combining the slice sequence information and inter-slice step size information recorded in the image volume data, the position of the voxel in the slice scanning sequence is calibrated to obtain the voxel layer direction spatial index coordinates. Within the spatial range defined by the image volume size range information and the effective imaging area identification information, the set of voxels marked as effective areas in the image volume data is collected to determine the effective spatial area of ​​the image. Within the effective spatial region of the image, the image voxel intensity values ​​are read according to the combination relationship of the voxel horizontal spatial index coordinates, the voxel vertical spatial index coordinates, and the voxel layer directional spatial index coordinates. Based on the numerical differences between adjacent voxels, the variation amplitude of the image voxel intensity values ​​in the three spatial directions is calculated to form a spatial gradient expression of the image voxel intensity values. The magnitude of the spatial gradient expression is then calculated to obtain the spatial gradient amplitude at the corresponding voxel position. The spatial gradient magnitude is squared to form the squared value of the spatial structure intensity of a single voxel. Within the effective spatial region of the image, the squared values ​​of the spatial structure intensity corresponding to all voxels are accumulated according to the voxel volume unit to obtain the overall integral result of the squared spatial structure intensity. The square root operation is performed on the overall integral result to obtain the effective value of the spatial structure.

5. The method for rapid generation of three-dimensional models of chest wall bone defects based on deep learning according to claim 1, characterized in that: The specific steps for generating a three-dimensional representation of the chest wall bone structure consistent with the image coordinate system are as follows: The effective value of the spatial structure is compared with the effective threshold in real time. When the effective value of the spatial structure is greater than or equal to the effective threshold, the chest wall bone structure extraction process continues and the three-dimensional expression result of the chest wall bone structure is output. When the effective value of the spatial structure is less than the effective threshold, the output of the three-dimensional expression result of the chest wall bone structure is stopped and the structure extraction failure status code is returned.

6. The method for rapid generation of three-dimensional models of chest wall bone defects based on deep learning according to claim 1, characterized in that: The specific steps for deriving the complete structural reference representation from the three-dimensional representation of the chest wall bone structure are as follows: Based on the order of the image sampling timestamp sequence, the corresponding slice sequence position or historical scan record is traced back along the scanning time sequence before the current image volume being analyzed. The matching original image volume data is read to obtain the historical sampled image value. The image sampling timestamp immediately preceding the historical sampled image value is located along the time sequence, and the corresponding original image volume data is read to obtain the preceding sampled image value. Combining the slice sequence or the distribution of multiple scan time points in the image sampling timestamp sequence, the continuously covered time range is converted to determine the number of samples to be included in the analysis, and the time sliding window length is obtained. Based on the time difference between adjacent time stamps in the image sampling timestamp sequence, the time interval between adjacent image volume data is determined, thus obtaining the adjacent sampling time interval; Within the image volume data range defined by the time window length, each historical sampled image value and its corresponding preceding sampled image value are selected sequentially in chronological order. The difference between the historical sampled image value and the preceding sampled image value is obtained, and the difference is divided by the adjacent sampling time interval to obtain the ratio result. The absolute value of the comparison result is taken to form the single-cycle rate of change amplitude. The amplitude accumulation result is obtained by summing all single-cycle rate of change amplitudes corresponding to the time window length. The amplitude accumulation result is divided by the time window length to obtain the time stability determination value.

7. The method for rapid generation of a three-dimensional model of chest wall bone defects based on deep learning according to claim 1, characterized in that: The specific steps for generating a complete three-dimensional structural model of the chest wall bone aligned with the original structural space are as follows: The time stability determination value is compared with the stability threshold in real time. The stability threshold includes a primary stability threshold and a secondary stability threshold. When the time stability determination value is less than the primary stability threshold, the system enters the complete structure generation path and performs the generation of a complete three-dimensional chest wall bone structure model. When the time stability determination value is greater than or equal to the primary stability threshold and less than the secondary stability threshold, the system enters the restricted structure generation path and writes the time stability determination value into the status record field. When the time stability determination value is greater than or equal to the secondary stability threshold, the structure generation path is blocked and a prohibited status flag is generated.

8. The method for rapid generation of a three-dimensional model of chest wall bone defects based on deep learning according to claim 1, characterized in that: The specific steps for evaluating the spatial difference relationship by combining the complete three-dimensional structural model of the chest wall bones with the three-dimensional representation results of the chest wall bones are as follows: The spatial index coordinates of the target voxel are offset and located based on the horizontal neighborhood offset, vertical neighborhood offset and layer-to-neighbor offset. The grayscale records in the original image volume data are read at the offset voxel position to obtain the intensity values ​​of adjacent image voxels. Based on the index interval formed between adjacent voxels by the row direction spatial index number, the horizontal neighborhood offset is determined; based on the index interval formed between adjacent voxels by the column direction spatial index number, the vertical neighborhood offset is determined; the positional difference between adjacent layer voxels is converted by combining the slice sequence information and the inter-layer step size information to determine the layer-to-neighborhood offset. Image voxel intensity values ​​are read at spatial locations determined by voxel horizontal spatial index coordinates, voxel vertical spatial index coordinates, and voxel layer-direction spatial index coordinates. Adjacent voxel coordinates are obtained by offsetting the voxel horizontal spatial index coordinates, voxel vertical spatial index coordinates, and voxel layer-direction spatial index coordinates according to the horizontal neighborhood offset, vertical neighborhood offset, and layer-direction neighborhood offset. Adjacent image voxel intensity values ​​are read at adjacent voxel coordinates. The spatial difference amplitude is obtained by subtracting the adjacent image voxel intensity values ​​from the image voxel intensity values ​​and taking the absolute value. The absolute value of the second-order time change of the image voxel intensity value at consecutive sampling time points is taken, and the absolute value result is added to one to obtain the denominator term; the spatial difference amplitude is divided by the denominator term to obtain the differential confidence control value.

9. The method for rapid generation of a three-dimensional model of chest wall bone defects based on deep learning according to claim 1, characterized in that: The specific steps for expressing and outputting the three-dimensional structural model of the defect area and the three-dimensional model of the chest wall bone defect are as follows: The differential confidence control value is written into the differential result output control parameter set; based on the differential confidence control value, the voxel writing density parameter, the boundary smoothing intensity parameter, and the model output resolution parameter are simultaneously updated numerically; during the differential result writing process, the number of voxels involved in the writing is controlled according to the voxel writing density parameter, the smoothing processing amplitude of the differential structure boundary is controlled according to the boundary smoothing intensity parameter, and the mesh refinement degree of the three-dimensional structure of the defect area is controlled according to the model output resolution parameter; within each processing cycle, the differential confidence control value continuously participates in the parameter update process, completing the differential structure writing and outputting the three-dimensional model of the chest wall bone defect.

10. A system for rapidly generating three-dimensional models of chest wall bone defects based on deep learning, using the method for rapidly generating three-dimensional models of chest wall bone defects based on deep learning as described in any one of claims 1-9, characterized in that: The image data acquisition module is used to acquire basic image data generated during chest tomographic imaging, construct a chest wall image volume dataset, and preprocess the chest wall image volume dataset. The skeleton structure segmentation module is used to identify the spatial distribution characteristics of chest wall bone-related structures based on chest wall image volume datasets, and to form a three-dimensional representation of chest wall bone structures consistent with the image coordinate system. The complete structure generation module is used to derive the complete structural reference expression based on the three-dimensional expression result of the chest wall bone structure, and generate a complete three-dimensional chest wall bone structure model that is aligned with the original structural space. The defect differential output module is used to combine the complete three-dimensional structural model of the chest wall bone with the three-dimensional expression results of the chest wall bone structure to evaluate the spatial difference relationship, form a three-dimensional structural model expression of the defect area, and output the three-dimensional model of the chest wall bone defect.

Images