A deep learning-based lesion positioning method and system

Abstract

This invention discloses a lesion localization method and system based on deep learning, comprising: acquiring medical image data; preprocessing the images to obtain standardized data; inputting the standardized data into an improved nnFormer to construct a potential lesion morphological field; fusing the morphological field and image data to generate a multi-source lesion mechanism parameter field; segmenting the lesion occurrence mechanism parameter field, establishing evolution paths, and inversely reasoning about boundary regions; calculating topological and physiological characteristics, and jointly assessing the final lesion region using energy; and outputting the lesion localization result, including spatial location and regional extent. This invention achieves stable and accurate localization of lesion regions in medical images by constructing a potential lesion morphological field from standardized medical images using an improved nnFormer, and by fusing multi-source mechanism parameter fields, morphological evolution inverse reasoning, and topological-physiological joint energy assessment.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence technology, and in particular to a lesion localization method and system based on deep learning. Background Technology

[0002] In the field of medical imaging-assisted diagnosis, lesion localization is a crucial foundation for clinical diagnosis and treatment planning. With the widespread application of medical imaging equipment such as computed tomography (CT) and magnetic resonance imaging (MRI), computer-based automatic lesion localization technology has gradually become a research hotspot. Current lesion localization methods mainly rely on manual image interpretation, rule-based image processing methods, and deep learning-based detection or segmentation methods. Among these, manual methods are highly dependent on physician experience, involve high workload, and are difficult to guarantee consistency; traditional image processing methods based on thresholding, edge analysis, or region growing are sensitive to noise and changes in imaging conditions, exhibiting poor robustness.

[0003] In recent years, deep learning methods have been widely applied to the field of lesion localization in medical imaging. A common approach is to use convolutional neural networks or Transformer structures for end-to-end lesion detection or segmentation in medical images. While these methods improve localization accuracy to some extent, they typically rely solely on pixel-level features or high-level semantic features of the image, neglecting the intrinsic connections between lesion occurrence and tissue structure, physiological mechanisms, and morphological evolution. Furthermore, existing deep learning methods often use a single model to directly output lesion results, lacking systematic modeling of the lesion's morphological evolution and spatial structural stability. This makes them prone to inaccurate localization and unstable boundaries in scenarios involving early-stage lesions, small-volume lesions, or lesions with complex morphologies.

[0004] The post-processing methods for lesion localization results in the current technology are relatively simple, usually only performing morphological operations or simple connected component screening. They fail to comprehensively constrain the localization results from the perspectives of topological structure and physiological rationality, resulting in deficiencies in spatial continuity, structural integrity and medical rationality of the localization results.

[0005] Therefore, how to provide a lesion localization method and system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a lesion localization method and system based on deep learning. This invention introduces an improved nnFormer model to obtain the morphological distribution of target organs and local structures, and on this basis, constructs a potential lesion morphological field. Furthermore, it integrates multi-source information such as gray-level gradient changes, tissue density shifts, directional consistency, and deformation tension to construct a lesion occurrence mechanism parameter field. Combined with inverse reasoning of lesion morphological evolution and topological-physiological joint energy assessment, it achieves precise localization of lesion regions in medical images. This invention fully utilizes deep learning, medical image processing, and morphological analysis technologies to systematically characterize the occurrence mechanism and spatial structural features of lesions, effectively improving the localization accuracy of early-stage, small-volume, and morphologically complex lesions. It possesses advantages such as high stability of localization results, strong rationality of boundary structures, and good medical consistency.

[0007] A lesion localization method based on deep learning according to an embodiment of the present invention includes:

[0008] Acquire medical image data of the target object, preprocess the medical image data to obtain standardized medical image data;

[0009] Standardized medical imaging data is input into the improved nnFormer to obtain the morphological distribution results of target organs and local structures, and a potential lesion morphological field is constructed based on the morphological distribution results.

[0010] Based on standardized medical imaging data and potential lesion morphology field, the gray-level gradient change rate field, tissue density offset field, orientation consistency tensor field and deformation tension field are calculated to construct the lesion occurrence mechanism parameter field.

[0011] Threshold segmentation and connectivity analysis are performed based on the parameter field of lesion occurrence mechanism to determine the potential lesion occurrence area. Multiple lesion morphology evolution paths are constructed for the potential lesion occurrence area, and reverse reasoning of morphological evolution is performed to obtain the initial estimated area of ​​lesion boundary.

[0012] Topological and physiological structural features are calculated for the initially estimated lesion boundary region. Based on the topological and physiological structural features, a joint energy assessment is performed to determine the final lesion region that satisfies both topological stability and physiological rationality.

[0013] The lesion localization result is output based on the final lesion area, and the lesion localization result includes the spatial location of the lesion and the corresponding area range.

[0014] Optionally, the medical imaging data includes computed tomography (CT) images, magnetic resonance imaging (MRI) images, positron emission tomography (PET) images, ultrasound images, and X-ray images.

[0015] Optionally, the preprocessing of the medical image data includes sequentially performing noise suppression processing, grayscale intensity normalization processing, and spatial resampling processing on the medical image data.

[0016] Optionally, the construction of the potential lesion morphology field based on the morphological distribution results includes:

[0017] Standardized medical image data is input into the input end of the improved nnFormer model. The standardized medical image data is three-dimensional volume data. The improved nnFormer model divides the three-dimensional volume data into voxel blocks at the input end to obtain multiple voxel block sequences with fixed voxel sizes. Each voxel block sequence is converted into a feature sequence.

[0018] An improved nnFormer model is constructed, which includes an organ structure anchoring coding layer, a multi-scale morphological consistency fusion layer, and a boundary topology stable decoding layer.

[0019] In the organ structure anchoring coding layer, organ anchoring feature generation processing is performed on the feature sequence. The organ anchoring feature generation processing includes generating organ anchoring maps for candidate structural regions of the target organ, and using the organ anchoring maps to spatially constrain the attention weights during the self-attention calculation process to obtain organ anchoring structural features.

[0020] In the multi-scale morphological consistency fusion layer, the organ anchoring structural features are resampled at multiple voxel scales, local self-attention encoding is performed at each scale, and the encoding results at multiple scales are fused with consistency alignment. The consistency alignment fusion includes fusion with consistency constraints on features at the same spatial location at different scales to generate fused morphological features that maintain the continuity of the overall morphology of the organ.

[0021] In the boundary topologically stable decoding layer, the fused morphological features are decoded to obtain the morphological distribution results of the target organ and local structures. In the decoding process, a topologically stable boundary constraint operation is introduced. The topologically stable boundary constraint operation includes connectivity preservation and hole suppression processing of the boundary candidate regions of the decoding output to obtain the topologically stable organ boundary morphological distribution results.

[0022] A potential lesion morphology field is constructed based on the topologically stable organ boundary morphology distribution results. The potential lesion morphology field represents the potential distribution area of ​​lesions within the target organ that is consistent with the local structural morphology continuity.

[0023] Optionally, the construction of the lesion occurrence mechanism parameter field includes:

[0024] Based on standardized medical image data, the gray-level difference value of each voxel is calculated by central difference along three spatial axes. The combined amplitude of the difference values ​​in the three directions is used as the gray-level gradient change intensity of the voxel. The gray-level gradient change intensity of each voxel constitutes the gray-level gradient change rate field.

[0025] A fixed-size three-dimensional neighborhood window is constructed with each voxel as the center. The mean and standard deviation of the neighborhood gray value are calculated. The difference between the voxel gray value and the neighborhood mean is normalized according to the neighborhood standard deviation to obtain the density shift intensity. The density shift intensity of each voxel constitutes the tissue density shift field.

[0026] Based on the gray-level difference value, the outer product of the gray-level difference value is accumulated in the fixed-size three-dimensional neighborhood window of each voxel to obtain the structure tensor. The structure tensor is decomposed into features to obtain the energy proportion of the main direction and the energy proportion of the secondary direction. The difference between the two is used as the directional consistency intensity. The directional consistency intensity of each voxel constitutes the directional consistency tensor field.

[0027] Based on the morphological field of potential lesions, a morphological gating weight distribution is generated. The gray-level gradient rate of change field, tissue density offset field and orientation consistency tensor field are weighted and accumulated within a neighborhood window using the morphological gating weight distribution to obtain the morphologically gated gradient field, density offset field and orientation consistency field. Under the constraint of the morphological gating weight distribution, three-dimensional differentiable registration is performed on standardized medical images to obtain the displacement field. The local spatial change amplitude of the displacement field is calculated as the deformation tension intensity. The deformation tension intensity of each voxel is used to construct the deformation tension field.

[0028] The gradient field, density shift field, orientation consistency field, deformation tension field, and potential lesion morphology field after morphological gating are normalized by interval, and corresponding weight coefficients are set and linearly weighted and summed to obtain the lesion occurrence mechanism parameter field. The parameter values ​​are weighted and averaged according to the morphological gating weight distribution within a fixed-size neighborhood window to obtain a spatially continuous lesion occurrence mechanism parameter field.

[0029] Optionally, the reverse reasoning of the morphological evolution to obtain the initial estimated region of the lesion boundary includes:

[0030] Based on the parameter field of lesion occurrence mechanism, a candidate voxel set is obtained by threshold segmentation. Three-dimensional connected region analysis is performed on the candidate voxel set to remove connected regions with a voxel number less than a preset threshold, thereby obtaining the potential lesion occurrence region.

[0031] For each connected region of the potential lesion occurrence area, the center position, principal direction and boundary curvature distribution are calculated, and a seed point set is generated. The seed point set includes the center seed point, the boundary high curvature seed point and the local extremum seed point of the parameter field.

[0032] Multiple lesion morphological evolution paths are constructed based on a set of seed points. The morphological evolution paths consist of multiple morphological states arranged in sequence. The multiple lesion morphological evolution paths include: a central expansion path obtained by performing isotropic expansion based on the central seed point and alternately performing boundary smoothing; a directional guidance path obtained by performing anisotropic expansion along the main direction based on the boundary high curvature seed point; and a mechanism guidance path obtained by extrapolating layer by layer along the gradient ascent direction of the parameter field based on the local extreme value seed point of the parameter field and performing curvature constraints.

[0033] For each connected domain, reverse reasoning of morphological evolution is performed on multiple lesion morphological evolution paths. The reverse reasoning of morphological evolution includes traversing backward from the end of the path to the beginning and extracting the set of newly added boundary voxels at each level. For the set of newly added boundary voxels, image consistency score and mechanism consistency score are calculated.

[0034] The newly added boundary voxel set is screened based on the image consistency score and the mechanism consistency score. Consistency fusion processing is performed on the boundary voxels obtained from multiple paths. The consistency fusion processing includes retaining boundary voxels that appear repeatedly in multiple paths and removing boundary voxels that appear in a single path and whose neighborhood support is lower than a preset threshold, so as to obtain the initial estimated region of the lesion boundary.

[0035] Optionally, determining the final lesion region that satisfies topological stability and physiological rationality includes:

[0036] Based on the initial estimated region of the lesion boundary, topological structure analysis is performed on the voxels to obtain the topological feature information corresponding to each voxel. The topological feature information includes the connectivity status of the region where the voxel is located, the existence status of local voids, and the boundary curvature change status.

[0037] The topological stability assessment results are constructed based on topological feature information. The topological stability assessment results are obtained by comprehensively quantifying the degree of connectivity disruption, the degree of void formation, and the degree of abrupt change in boundary curvature.

[0038] Based on standardized medical imaging data and lesion occurrence mechanism parameter field, physiological structural feature information of voxels in the initial estimation area of ​​lesion boundary is calculated. The physiological structural feature information includes the gray level heterogeneity in the neighborhood of voxels, the range of tissue density change and the intensity of local deformation tension. Physiological rationality assessment results are generated based on the physiological structural feature information.

[0039] The topological stability assessment results and the physiological rationality assessment results are jointly evaluated. The joint evaluation adopts a hierarchical judgment method. First, voxels that do not meet the requirements of connectivity and boundary integrity are screened out based on the topological stability assessment results. Then, voxels that do not conform to the characteristics of lesion tissue are screened out based on the physiological rationality assessment results, thereby obtaining the joint energy assessment results.

[0040] Based on the joint energy assessment results, regional consistency optimization processing is performed within the initial lesion boundary area. The regional consistency optimization processing includes preserving spatially continuous voxel regions with stable joint energy assessment results, suppressing isolated voxels and regions with discontinuous boundaries, and finally determining the final lesion region that satisfies topological stability and physiological rationality.

[0041] Optionally, the step of outputting lesion localization results based on the final lesion region includes:

[0042] Based on the final region of the lesion, the spatial coordinate information of all voxels constituting the final region of the lesion is extracted to form a set of lesion voxels;

[0043] Spatial location statistics of the lesion voxel set are performed. By summarizing and calculating the spatial coordinates of each voxel in the lesion voxel set, the spatial center position of the lesion in the medical image is determined.

[0044] Based on the coordinate distribution range of each voxel in the lesion voxel set in three spatial dimensions, the smallest spatial enclosing region that can completely cover the lesion voxel set is determined.

[0045] The lesion voxel set, spatial center position, and spatial surrounding area are mapped to the original medical image coordinate system, and the lesion localization result is output in the form of structured data. The lesion localization result includes the spatial location identifier of the lesion and the corresponding regional range identifier.

[0046] A lesion localization system based on deep learning according to an embodiment of the present invention includes the following modules:

[0047] The image preprocessing module is used to acquire medical image data and perform preprocessing to obtain standardized medical image data;

[0048] The morphological field construction module is used to input standardized medical image data into the improved nnFormer model to obtain organ morphological distribution results and construct the morphological field of potential lesions.

[0049] The mechanism parameter field construction module is used to calculate the gray-level gradient change rate field, tissue density offset field, orientation consistency tensor field and deformation tension field based on standardized medical image data and potential lesion morphology field, and to construct the lesion occurrence mechanism parameter field.

[0050] The boundary estimation module is used to determine the potential lesion area based on the lesion occurrence mechanism parameter field, construct multiple morphological evolution paths and perform morphological evolution reverse reasoning to obtain the preliminary lesion boundary area.

[0051] The final region determination module is used to calculate the topological and physiological structural features of the initially estimated lesion boundary region, and to perform joint energy assessment to determine the final lesion region.

[0052] The positioning output module is used to output the spatial location and range of the lesion based on the final lesion area.

[0053] The beneficial effects of this invention are:

[0054] This invention effectively overcomes the shortcomings of existing technologies that rely solely on pixel-level features for lesion identification by introducing morphological distribution modeling of organs and local structures into the deep learning lesion localization process. By improving nnFormer to obtain overall morphological and local structural information of the target organ and constructing a potential lesion morphological field, the lesion localization process is conducted under the constraints of the global organ structure. This helps reduce background tissue interference and improves the accuracy of lesion region identification, maintaining good localization stability, especially in images with complex organ boundaries or low contrast.

[0055] This invention constructs a parameter field for lesion development by fusing multi-source information such as grayscale gradient changes, tissue density shifts, directional consistency, and deformation tension. Based on this parameter field, a reverse reasoning mechanism for lesion morphology evolution is introduced, enabling the lesion localization process to simultaneously consider the intrinsic mechanism of lesion formation and its spatial evolution. This technical solution effectively reduces the impact of misjudgments caused by a single model, exhibits stronger adaptability to early lesions, small-volume lesions, and irregularly shaped lesions, and improves the continuity of lesion boundaries and the stability of localization results.

[0056] This invention introduces a joint energy assessment mechanism combining topological and physiological structural features during the lesion localization stage. This mechanism constrains the structural rationality of the lesion boundary, ensuring that the final localization result conforms to medical imaging characteristics in terms of spatial connectivity, boundary integrity, and physiological rationality. This comprehensive constraint method effectively suppresses the generation of isolated false detection areas and unreasonable boundaries, improving the reliability and consistency of lesion localization results and providing more accurate and stable auxiliary information for clinical diagnosis and treatment. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart of a deep learning-based lesion localization method proposed in this invention;

[0059] Figure 2 This is a schematic diagram of the structure of a lesion localization system based on deep learning proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figure 1 A deep learning-based lesion localization method includes:

[0062] Acquire medical image data of the target object, preprocess the medical image data to obtain standardized medical image data;

[0063] Standardized medical imaging data is input into the improved nnFormer to obtain the morphological distribution results of target organs and local structures, and a potential lesion morphological field is constructed based on the morphological distribution results.

[0064] Based on standardized medical imaging data and potential lesion morphology field, the gray-level gradient change rate field, tissue density offset field, orientation consistency tensor field and deformation tension field are calculated to construct the lesion occurrence mechanism parameter field.

[0065] Threshold segmentation and connectivity analysis are performed based on the parameter field of lesion occurrence mechanism to determine the potential lesion occurrence area. Multiple lesion morphology evolution paths are constructed for the potential lesion occurrence area, and reverse reasoning of morphological evolution is performed to obtain the initial estimated area of ​​lesion boundary.

[0066] Topological and physiological structural features are calculated for the initially estimated lesion boundary region. Based on the topological and physiological structural features, a joint energy assessment is performed to determine the final lesion region that satisfies both topological stability and physiological rationality.

[0067] The lesion localization result is output based on the final lesion area, and the lesion localization result includes the spatial location of the lesion and the corresponding area range.

[0068] In this embodiment, the medical imaging data includes computed tomography (CT) images, magnetic resonance imaging (MRI) images, positron emission tomography (PET) images, ultrasound images, and X-ray images.

[0069] In this embodiment, the preprocessing of medical image data includes sequentially performing noise suppression processing, grayscale intensity normalization processing, and spatial resampling processing on the medical image data.

[0070] In this embodiment, the construction of the potential lesion morphology field based on the morphological distribution results includes:

[0071] Standardized medical image data is input into the input end of the improved nnFormer model. The standardized medical image data is three-dimensional volume data. The improved nnFormer model divides the three-dimensional volume data into voxel blocks at the input end to obtain multiple voxel block sequences with fixed voxel sizes. Each voxel block sequence is converted into a feature sequence.

[0072] An improved nnFormer model is constructed, which includes an organ structure anchoring coding layer, a multi-scale morphological consistency fusion layer, and a boundary topology stable decoding layer.

[0073] In the organ structure anchoring coding layer, organ anchoring feature generation processing is performed on the feature sequence. This process includes generating organ anchoring maps for candidate structural regions of the target organ, and spatially constraining the attention weights using these organ anchoring maps during self-attention computation to obtain organ anchoring structural features. Specifically, generating organ anchoring maps for candidate structural regions of the target organ involves:

[0074] Based on the organ morphology distribution results of standardized medical imaging data, the target organ is coarsely located in three-dimensional voxel space to obtain the initial candidate region of the target organ. Morphological closure operation and hole filling are performed on the initial candidate region to form a continuous candidate structural region.

[0075] The organ boundary zone and the organ interior zone are extracted within the candidate structure region. The organ boundary zone is obtained by extending the boundary of the candidate structure region inward and outward according to the preset voxel distance. The organ interior zone is the remaining region of the candidate structure region after removing the organ boundary zone. Different anchoring weights are assigned to the organ boundary zone and the organ interior zone respectively.

[0076] The anchoring weights are written into a spatial weight matrix that corresponds one-to-one with the input feature sequence according to the voxel coordinates. The spatial weight matrix takes the first preset weight at the position inside the organ, the second preset weight at the position at the organ boundary, and the third preset weight at the position outside the candidate structure region, thereby forming an organ anchoring mapping.

[0077] In the multi-scale morphological consistency fusion layer, the organ anchoring structural features are resampled at multiple voxel scales, local self-attention encoding is performed at each scale, and the encoding results at multiple scales are fused with consistency alignment. The consistency alignment fusion includes fusion with consistency constraints on features at the same spatial location at different scales to generate fused morphological features that maintain the continuity of the overall morphology of the organ.

[0078] In the boundary topologically stable decoding layer, the fused morphological features are decoded to obtain the morphological distribution results of the target organ and local structures. A topologically stable boundary constraint operation is introduced during the decoding process. This operation includes connectivity preservation and hole suppression processing on the boundary candidate regions output by the decoding, resulting in a topologically stable organ boundary morphological distribution result. Specifically, the connectivity preservation and hole suppression processing on the boundary candidate regions output by the decoding is as follows:

[0079] The morphological distribution results of the decoded output are thresholded to obtain the boundary candidate binary region. Three-dimensional connected component labeling is performed on the boundary candidate binary region. Connected components with a voxel count greater than the preset threshold and connected to the main region of the target organ are retained, while isolated connected components that are not connected to the main region of the target organ are removed.

[0080] Three-dimensional hole detection is performed on the retained boundary candidate binary regions. Hole detection is achieved by inverting the boundary candidate binary regions, marking connected components, and filtering closed connected components located inside the main body region of the target organ to obtain the hole regions. A filling operation is performed on the hole regions to achieve hole suppression.

[0081] Boundary consistency smoothing is performed on the candidate binary regions of the boundary that have achieved connectivity preservation and hole suppression. Boundary consistency smoothing includes performing three-dimensional morphological opening and closing operations on the boundary voxels and locally correcting the boundary curvature mutant voxels to obtain the topologically stable organ boundary morphological distribution results.

[0082] A potential lesion morphology field is constructed based on the topologically stable organ boundary morphology distribution results. The potential lesion morphology field represents the potential distribution area of ​​lesions within the target organ that is consistent with the local structural morphology continuity.

[0083] In this embodiment, the construction of the lesion occurrence mechanism parameter field includes:

[0084] Based on standardized medical image data, the gray-level difference value of each voxel is calculated using central difference along three spatial axes. The combined amplitude of the difference values ​​in the three directions is used as the gray-level gradient change intensity of the voxel. The gray-level gradient change intensity of each voxel constitutes the gray-level gradient change rate field. Specifically, the gray-level difference value of each voxel is calculated using central difference along three spatial axes as follows:

[0085] Select the target voxel in the standardized medical image data, and obtain the gray values ​​of the previous and next adjacent voxels of the target voxel in the three spatial axes.

[0086] The center difference is calculated for each of the three spatial axes. The gray difference value is obtained by taking the difference between the gray value of the next adjacent voxel and the gray value of the previous adjacent voxel, and combining it with the voxel spacing after spatial resampling.

[0087] When the target voxel is located at the boundary of the volume data, resulting in the absence of the previous or next adjacent voxel in the direction, the center difference result of the direction is replaced with the one-sided difference result. The gray difference value of the direction is obtained by using the difference between the gray value of the target voxel and the gray value of the only adjacent voxel, combined with the voxel spacing. Thus, the gray difference values ​​of the target voxel in the three spatial axis directions are obtained.

[0088] A fixed-size 3D neighborhood window is constructed centered on each voxel. The mean and standard deviation of the neighborhood grayscale are calculated. The difference between the voxel grayscale value and the neighborhood mean is normalized according to the neighborhood standard deviation to obtain the density shift intensity. The density shift intensity of each voxel constitutes the tissue density shift field. The calculation of the neighborhood grayscale mean and standard deviation is as follows:

[0089] A fixed-size three-dimensional neighborhood window is extracted from the standardized medical image data with the target voxel as the center, and the set of all voxels within the three-dimensional neighborhood window is determined.

[0090] The summation of all voxel gray values ​​within the 3D neighborhood window is divided by the number of voxels in the window to obtain the mean neighborhood gray value corresponding to the target voxel.

[0091] After obtaining the mean gray value of the neighborhood, the difference between the gray value of each voxel in the three-dimensional neighborhood window and the mean gray value of the neighborhood is squared and summed. The summation result is divided by the number of voxels in the window and square rooted to obtain the standard deviation of the neighborhood gray value corresponding to the target voxel.

[0092] Based on the gray-level difference value, the outer product of the gray-level difference value is accumulated in the fixed-size three-dimensional neighborhood window of each voxel to obtain the structure tensor. The structure tensor is decomposed into features to obtain the energy proportion of the main direction and the energy proportion of the secondary direction. The difference between the two is used as the directional consistency intensity. The directional consistency intensity of each voxel constitutes the directional consistency tensor field.

[0093] Based on the morphological field of potential lesions, a morphological gating weight distribution is generated. The gray-level gradient rate of change field, tissue density offset field and orientation consistency tensor field are weighted and accumulated within a neighborhood window using the morphological gating weight distribution to obtain the morphologically gated gradient field, density offset field and orientation consistency field. Under the constraint of the morphological gating weight distribution, three-dimensional differentiable registration is performed on standardized medical images to obtain the displacement field. The local spatial change amplitude of the displacement field is calculated as the deformation tension intensity. The deformation tension intensity of each voxel is used to construct the deformation tension field.

[0094] The gradient field, density shift field, orientation consistency field, deformation tension field, and potential lesion morphology field after morphological gating are normalized by interval, and corresponding weight coefficients are set and linearly weighted and summed to obtain the lesion occurrence mechanism parameter field. The parameter values ​​are weighted and averaged according to the morphological gating weight distribution within a fixed-size neighborhood window to obtain a spatially continuous lesion occurrence mechanism parameter field.

[0095] In this embodiment, the step of performing reverse reasoning of morphological evolution to obtain the initial estimated region of the lesion boundary includes:

[0096] Based on the parameter field of lesion occurrence mechanism, a candidate voxel set is obtained by threshold segmentation. Three-dimensional connected region analysis is performed on the candidate voxel set to remove connected regions with a voxel number less than a preset threshold, thereby obtaining the potential lesion occurrence region.

[0097] For each connected region of the potential lesion area, the center location, principal direction, and boundary curvature distribution are calculated, and a seed point set is generated. This seed point set includes a center seed point, boundary high curvature seed points, and local extremum seed points of the parameter field. Specifically, the calculation of the center location, principal direction, and boundary curvature distribution for each connected region of the potential lesion area is as follows:

[0098] The spatial coordinates of all voxels within the connected domain are summarized, and the average of the coordinate values ​​in the three spatial directions is calculated to obtain the center position of the connected domain.

[0099] Using the center position as a reference, the spatial coordinates of each voxel in the connected domain are centered. Based on the centered coordinate set, the covariance relationship of the three-dimensional coordinate distribution is calculated, and the covariance relationship is eigenvalued. The eigenvector corresponding to the direction of maximum variance is taken as the main direction of the connected domain.

[0100] Boundary extraction is performed on the connected components to obtain a boundary set composed of boundary voxels. A local neighborhood boundary point set of a fixed size is extracted from each boundary voxel in the boundary set. A local surface is fitted to the local neighborhood boundary point set, and the curvature value at the boundary voxel is obtained from the fitted surface. The curvature values ​​of all boundary voxels are statistically summarized and binned according to curvature intervals to form the boundary curvature distribution of the connected components.

[0101] Multiple lesion morphological evolution paths are constructed based on a set of seed points. The morphological evolution paths consist of multiple morphological states arranged in sequence. The multiple lesion morphological evolution paths include: a central expansion path obtained by performing isotropic expansion based on the central seed point and alternately performing boundary smoothing; a directional guidance path obtained by performing anisotropic expansion along the main direction based on the boundary high curvature seed point; and a mechanism guidance path obtained by extrapolating layer by layer along the gradient ascent direction of the parameter field based on the local extreme value seed point of the parameter field and performing curvature constraints.

[0102] For each connected component, reverse morphological evolution reasoning is performed along multiple lesion morphological evolution paths. This reverse morphological evolution reasoning includes traversing backwards from the end of the path to the beginning and extracting the newly added boundary voxel set at each level. For the newly added boundary voxel set, image consistency score and mechanism consistency score are calculated, where:

[0103] The image consistency score is determined by the intensity of gray-level gradient change and tissue density shift at the newly added boundary voxel, while the mechanism consistency score is determined by the lesion occurrence mechanism parameter value and parameter field spatial continuity at the newly added boundary voxel.

[0104] The newly added boundary voxel set is screened based on the image consistency score and the mechanism consistency score. Consistency fusion processing is performed on the boundary voxels obtained from multiple paths. The consistency fusion processing includes retaining boundary voxels that appear repeatedly in multiple paths and removing boundary voxels that appear in a single path and whose neighborhood support is lower than a preset threshold, so as to obtain the initial estimated region of the lesion boundary.

[0105] In this embodiment, determining the final lesion region that satisfies topological stability and physiological rationality includes:

[0106] Based on the initial estimated region of the lesion boundary, topological structure analysis is performed on the voxels to obtain the topological feature information corresponding to each voxel. The topological feature information includes the connectivity status of the region where the voxel is located, the existence status of local voids, and the boundary curvature change status.

[0107] The topological stability assessment results are constructed based on topological feature information. The topological stability assessment results are obtained by comprehensively quantifying the degree of connectivity disruption, the degree of void formation, and the degree of abrupt change in boundary curvature.

[0108] Based on standardized medical imaging data and lesion pathogenesis parameter fields, physiological structural feature information is calculated for voxels within the initial estimation region of the lesion boundary. This physiological structural feature information includes the level of gray-level heterogeneity in the voxel neighborhood, the magnitude of tissue density variation, and the intensity of local deformation tension. A physiological rationality assessment result is generated based on this physiological structural feature information. Specifically, the calculation of the physiological structural feature information involves:

[0109] A fixed-size three-dimensional neighborhood window is extracted centered on the target voxel within the initially estimated boundary of the lesion. The gray value set of all voxels within the three-dimensional neighborhood window is obtained, the variance of the gray value set is calculated, and the variance is used as the gray heterogeneity level.

[0110] Within the same three-dimensional neighborhood window, obtain the set of lesion occurrence mechanism parameter values ​​corresponding to each voxel, calculate the maximum and minimum values ​​of the parameter value set respectively, and take the difference between the maximum and minimum values ​​as the tissue density change range;

[0111] Within the same three-dimensional neighborhood window, the values ​​of the deformation tension field at each voxel are obtained, the mean of the set of values ​​is calculated, and the mean is used as the local deformation tension intensity to obtain the physiological structural feature information of the target voxel.

[0112] The topological stability assessment results and the physiological rationality assessment results are jointly evaluated. The joint evaluation adopts a hierarchical judgment method. First, voxels that do not meet the requirements of connectivity and boundary integrity are screened out based on the topological stability assessment results. Then, voxels that do not conform to the characteristics of lesion tissue are screened out based on the physiological rationality assessment results, thereby obtaining the joint energy assessment results.

[0113] Based on the joint energy assessment results, regional consistency optimization processing is performed within the initial lesion boundary area. The regional consistency optimization processing includes preserving spatially continuous voxel regions with stable joint energy assessment results, suppressing isolated voxels and regions with discontinuous boundaries, and finally determining the final lesion region that satisfies topological stability and physiological rationality.

[0114] In this embodiment, the step of outputting lesion localization results based on the final lesion region includes:

[0115] Based on the final region of the lesion, the spatial coordinate information of all voxels constituting the final region of the lesion is extracted to form a set of lesion voxels;

[0116] Spatial location statistics of the lesion voxel set are performed. By summarizing and calculating the spatial coordinates of each voxel in the lesion voxel set, the spatial center position of the lesion in the medical image is determined.

[0117] Based on the coordinate distribution range of each voxel in the lesion voxel set in three spatial dimensions, the smallest spatial enclosing region that can completely cover the lesion voxel set is determined.

[0118] The lesion voxel set, spatial center position, and spatial surrounding area are mapped to the original medical image coordinate system, and the lesion localization result is output in the form of structured data. The lesion localization result includes the spatial location identifier of the lesion and the corresponding regional range identifier.

[0119] refer to Figure 2 A lesion localization system based on deep learning includes the following modules:

[0120] The image preprocessing module is used to acquire medical image data and perform preprocessing to obtain standardized medical image data;

[0121] The morphological field construction module is used to input standardized medical image data into the improved nnFormer model to obtain organ morphological distribution results and construct the morphological field of potential lesions.

[0122] The mechanism parameter field construction module is used to calculate the gray-level gradient change rate field, tissue density offset field, orientation consistency tensor field and deformation tension field based on standardized medical image data and potential lesion morphology field, and to construct the lesion occurrence mechanism parameter field.

[0123] The boundary estimation module is used to determine the potential lesion area based on the lesion occurrence mechanism parameter field, construct multiple morphological evolution paths and perform morphological evolution reverse reasoning to obtain the preliminary lesion boundary area.

[0124] The final region determination module is used to calculate the topological and physiological structural features of the initially estimated lesion boundary region, and to perform joint energy assessment to determine the final lesion region.

[0125] The positioning output module is used to output the spatial location and range of the lesion based on the final lesion area.

[0126] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to the imaging center of a tertiary-level Class A hospital. In the early stage, liver lesions are usually small in size, have low gray-scale contrast, and the boundary transition with the surrounding normal liver parenchyma is not obvious. Traditional methods based on pixel features or a single deep learning model are prone to problems such as missed detection of lesions or unstable boundary localization, especially when the lesion is irregular in shape or adjacent to vascular structures.

[0127] In this embodiment, data from 180 patients who underwent enhanced abdominal CT scans were selected, including 120 clinically diagnosed cases of liver lesions and 60 non-lesion control cases. All image data were in standard DICOM format with an image resolution of 512×512 and a slice thickness of 1.25mm. The lesion areas were annotated by two experienced radiologists as an evaluation reference.

[0128] In practical applications, the system first preprocesses the acquired medical image data, including noise suppression, grayscale normalization, and spatial resampling, to eliminate differences caused by different scanning equipment and parameters. Subsequently, the standardized medical image data is input into the improved nnFormer model to obtain the morphological distribution results of the overall and local structures of the liver, and a potential lesion morphological field is constructed based on this morphological distribution, so that the lesion localization process is constrained by the overall structure of the organ.

[0129] The system combines standardized medical imaging data with a potential lesion morphology field to calculate a gray-level gradient change rate field, a tissue density offset field, a orientation consistency tensor field, and a deformation tension field, and integrates these to construct a lesion occurrence mechanism parameter field. Subsequently, threshold segmentation and connectivity analysis are performed on the lesion occurrence mechanism parameter field to determine potential lesion occurrence regions. Multiple lesion morphological evolution paths are constructed for each potential occurrence region, and the initial estimated lesion boundary region is obtained through inverse reasoning of morphological evolution. Then, topological and physiological structural feature calculations are performed on the initially estimated lesion boundary region. Joint energy assessment determines the final lesion region that satisfies topological stability and physiological rationality. Finally, the system outputs the spatial location of the lesion and its corresponding regional extent to assist doctors in diagnosis.

[0130] To verify the actual effectiveness of the method of the present invention, the present invention was compared with three common methods, including the traditional morphology rule-based method, the three-dimensional U-Net lesion segmentation method, and the general Transformer medical segmentation method.

[0131] Table 1. Performance comparison of different lesion localization methods in liver CT images.

[0132]

[0133] As shown in Table 1, the average localization error of the morphological rule method is 6.2 mm, which is significantly affected by noise, contrast, and individual differences. The 3D U-Net and general Transformer have average errors of 4.5 mm and 4.1 mm respectively, demonstrating a significant improvement in localization accuracy. The method of this invention further reduces the error to 3.0 mm, indicating that under the constraint of "improving nnFormer to obtain organ morphological distribution and constructing a potential lesion morphological field, fusing multi-source fields to construct a lesion occurrence mechanism parameter field, and inverse reasoning of morphological evolution," the lesion center and range are closer to those of manual annotation. The errors for small lesion subsets and complex morphological subsets are 3.4 mm and 3.2 mm respectively, which are slightly higher than the overall average but still better than the comparison methods, consistent with the objective situation of weak signals in small lesions and irregular boundaries in complex morphological subsets.

[0134] In terms of lesion boundary integrity rate, the morphological rule method achieved 73.4%, which is prone to boundary loss or breakage; the 3D U-Net and general Transformer achieved 81.6% and 84.2%, respectively; the method of this invention achieved 89.5%, and 87.2% and 88.6% in the subset of small lesions and the subset of complex morphology, respectively. This indicates that after obtaining the initial boundary estimation region, this invention can effectively suppress isolated false detections and maintain boundary continuity and connectivity by performing consistency screening and optimization of the results through joint energy assessment of topological features and physiological structural features.

[0135] In terms of the combined performance of small lesion detection rate and processing time, the morphological rule method achieved a detection rate of 60.8%, while 3D U-Net and the general Transformer achieved 72.3% and 75.6%, respectively. The method of this invention improved to 83.7%, maintaining rates of 81.9% and 82.4% in the small lesion and complex morphology subsets, respectively, demonstrating stronger adaptability to difficult lesions. Regarding processing time, this invention achieved 12.6 seconds per case, slightly higher than 3D U-Net and the general Transformer, but lower than the 17.9 seconds per case of the morphological rule method. This indicates that even after introducing parameter field construction, morphological evolution inverse reasoning, and joint energy assessment, it still possesses clinically acceptable operational efficiency.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A lesion localization method based on deep learning, characterized in that, include: Acquire medical image data of the target object, preprocess the medical image data to obtain standardized medical image data; Standardized medical imaging data is input into the improved nnFormer to obtain the morphological distribution results of target organs and local structures, and a potential lesion morphological field is constructed based on the morphological distribution results. Based on standardized medical imaging data and potential lesion morphology field, the gray-level gradient change rate field, tissue density offset field, orientation consistency tensor field and deformation tension field are calculated to construct the lesion occurrence mechanism parameter field. Threshold segmentation and connectivity analysis are performed based on the parameter field of lesion occurrence mechanism to determine the potential lesion occurrence area. Multiple lesion morphology evolution paths are constructed for the potential lesion occurrence area, and reverse reasoning of morphological evolution is performed to obtain the initial estimated area of ​​lesion boundary. Topological and physiological structural features are calculated for the initially estimated lesion boundary region. Based on the topological and physiological structural features, a joint energy assessment is performed to determine the final lesion region that satisfies both topological stability and physiological rationality. The lesion localization result is output based on the final lesion area, and the lesion localization result includes the spatial location of the lesion and the corresponding area range.

2. The lesion localization method based on deep learning according to claim 1, characterized in that, The medical imaging data includes computed tomography (CT) images, magnetic resonance imaging (MRI) images, positron emission tomography (PET) images, ultrasound images, and X-ray images.

3. The lesion localization method based on deep learning according to claim 1, characterized in that, The preprocessing of medical image data includes sequentially performing noise suppression processing, grayscale intensity normalization processing, and spatial resampling processing on the medical image data.

4. The lesion localization method based on deep learning according to claim 1, characterized in that, The construction of the potential lesion morphology field based on the morphological distribution results includes: Standardized medical image data is input into the input end of the improved nnFormer model. The standardized medical image data is three-dimensional volume data. The improved nnFormer model divides the three-dimensional volume data into voxel blocks at the input end to obtain multiple voxel block sequences with fixed voxel sizes. Each voxel block sequence is converted into a feature sequence. An improved nnFormer model is constructed, which includes an organ structure anchoring coding layer, a multi-scale morphological consistency fusion layer, and a boundary topology stable decoding layer. In the organ structure anchoring coding layer, organ anchoring feature generation processing is performed on the feature sequence. The organ anchoring feature generation processing includes generating organ anchoring maps for candidate structural regions of the target organ, and using the organ anchoring maps to spatially constrain the attention weights during the self-attention calculation process to obtain organ anchoring structural features. In the multi-scale morphological consistency fusion layer, the organ anchoring structural features are resampled at multiple voxel scales, local self-attention encoding is performed at each scale, and the encoding results at multiple scales are fused with consistency alignment. The consistency alignment fusion includes fusion with consistency constraints on features at the same spatial location at different scales to generate fused morphological features that maintain the continuity of the overall morphology of the organ. In the boundary topologically stable decoding layer, the fused morphological features are decoded to obtain the morphological distribution results of the target organ and local structures. In the decoding process, a topologically stable boundary constraint operation is introduced. The topologically stable boundary constraint operation includes connectivity preservation and hole suppression processing of the boundary candidate regions of the decoding output to obtain the topologically stable organ boundary morphological distribution results. A potential lesion morphology field is constructed based on the topologically stable organ boundary morphology distribution results. The potential lesion morphology field represents the potential distribution area of ​​lesions within the target organ that is consistent with the local structural morphology continuity.

5. The lesion localization method based on deep learning according to claim 1, characterized in that, The constructed lesion occurrence mechanism parameter field includes: Based on standardized medical image data, the gray-level difference value of each voxel is calculated by central difference along three spatial axes. The combined amplitude of the difference values ​​in the three directions is used as the gray-level gradient change intensity of the voxel. The gray-level gradient change intensity of each voxel constitutes the gray-level gradient change rate field. A fixed-size three-dimensional neighborhood window is constructed with each voxel as the center. The mean and standard deviation of the neighborhood gray value are calculated. The difference between the voxel gray value and the neighborhood mean is normalized according to the neighborhood standard deviation to obtain the density shift intensity. The density shift intensity of each voxel constitutes the tissue density shift field. Based on the gray-level difference value, the outer product of the gray-level difference value is accumulated in the fixed-size three-dimensional neighborhood window of each voxel to obtain the structure tensor. The structure tensor is decomposed into features to obtain the energy proportion of the main direction and the energy proportion of the secondary direction. The difference between the two is used as the directional consistency intensity. The directional consistency intensity of each voxel constitutes the directional consistency tensor field. Based on the morphological field of potential lesions, a morphological gating weight distribution is generated. The gray-level gradient rate of change field, tissue density offset field and orientation consistency tensor field are weighted and accumulated within a neighborhood window using the morphological gating weight distribution to obtain the morphologically gated gradient field, density offset field and orientation consistency field. Under the constraint of the morphological gating weight distribution, three-dimensional differentiable registration is performed on standardized medical images to obtain the displacement field. The local spatial change amplitude of the displacement field is calculated as the deformation tension intensity. The deformation tension intensity of each voxel is used to construct the deformation tension field. The gradient field, density shift field, orientation consistency field, deformation tension field, and potential lesion morphology field after morphological gating are normalized by interval, and corresponding weight coefficients are set and linearly weighted and summed to obtain the lesion occurrence mechanism parameter field. The parameter values ​​are weighted and averaged according to the morphological gating weight distribution within a fixed-size neighborhood window to obtain a spatially continuous lesion occurrence mechanism parameter field.

6. The lesion localization method based on deep learning according to claim 1, characterized in that, The reverse reasoning of the execution morphology evolution obtains the initial estimated region of the lesion boundary, including: Based on the parameter field of lesion occurrence mechanism, a candidate voxel set is obtained by threshold segmentation. Three-dimensional connected region analysis is performed on the candidate voxel set to remove connected regions with a voxel number less than a preset threshold, thereby obtaining the potential lesion occurrence region. For each connected region of the potential lesion occurrence area, the center position, principal direction and boundary curvature distribution are calculated, and a seed point set is generated. The seed point set includes the center seed point, the boundary high curvature seed point and the local extremum seed point of the parameter field. Multiple lesion morphological evolution paths are constructed based on a set of seed points. The morphological evolution paths consist of multiple morphological states arranged in sequence. The multiple lesion morphological evolution paths include: a central expansion path obtained by performing isotropic expansion based on the central seed point and alternately performing boundary smoothing; a directional guidance path obtained by performing anisotropic expansion along the main direction based on the boundary high curvature seed point; and a mechanism guidance path obtained by extrapolating layer by layer along the gradient ascent direction of the parameter field based on the local extreme value seed point of the parameter field and performing curvature constraints. For each connected domain, reverse reasoning of morphological evolution is performed on multiple lesion morphological evolution paths. The reverse reasoning of morphological evolution includes traversing backward from the end of the path to the beginning and extracting the set of newly added boundary voxels at each level. For the set of newly added boundary voxels, image consistency score and mechanism consistency score are calculated. The newly added boundary voxel set is screened based on the image consistency score and the mechanism consistency score. Consistency fusion processing is performed on the boundary voxels obtained from multiple paths. The consistency fusion processing includes retaining boundary voxels that appear repeatedly in multiple paths and removing boundary voxels that appear in a single path and whose neighborhood support is lower than a preset threshold, so as to obtain the initial estimated region of the lesion boundary.

7. The lesion localization method based on deep learning according to claim 1, characterized in that, The determination of the final lesion region that satisfies topological stability and physiological rationality includes: Based on the initial estimated region of the lesion boundary, topological structure analysis is performed on the voxels to obtain the topological feature information corresponding to each voxel. The topological feature information includes the connectivity status of the region where the voxel is located, the existence status of local voids, and the boundary curvature change status. The topological stability assessment results are constructed based on topological feature information. The topological stability assessment results are obtained by comprehensively quantifying the degree of connectivity disruption, the degree of void formation, and the degree of abrupt change in boundary curvature. Based on standardized medical imaging data and lesion occurrence mechanism parameter field, physiological structural feature information of voxels in the initial estimation area of ​​lesion boundary is calculated. The physiological structural feature information includes the gray level heterogeneity in the neighborhood of voxels, the range of tissue density change and the intensity of local deformation tension. Physiological rationality assessment results are generated based on the physiological structural feature information. The topological stability assessment results and the physiological rationality assessment results are jointly evaluated. The joint evaluation adopts a hierarchical judgment method. First, voxels that do not meet the requirements of connectivity and boundary integrity are screened out based on the topological stability assessment results. Then, voxels that do not conform to the characteristics of lesion tissue are screened out based on the physiological rationality assessment results, thereby obtaining the joint energy assessment results. Based on the joint energy assessment results, regional consistency optimization processing is performed within the initial lesion boundary area. The regional consistency optimization processing includes preserving spatially continuous voxel regions with stable joint energy assessment results, suppressing isolated voxels and regions with discontinuous boundaries, and finally determining the final lesion region that satisfies topological stability and physiological rationality.

8. The lesion localization method based on deep learning according to claim 1, characterized in that, The lesion localization result output based on the final lesion region includes: Based on the final region of the lesion, the spatial coordinate information of all voxels constituting the final region of the lesion is extracted to form a set of lesion voxels; Spatial location statistics of the lesion voxel set are performed. By summarizing and calculating the spatial coordinates of each voxel in the lesion voxel set, the spatial center position of the lesion in the medical image is determined. Based on the coordinate distribution range of each voxel in the lesion voxel set in three spatial dimensions, the smallest spatial enclosing region that can completely cover the lesion voxel set is determined. The lesion voxel set, spatial center position, and spatial surrounding area are mapped to the original medical image coordinate system, and the lesion localization result is output in the form of structured data. The lesion localization result includes the spatial location identifier of the lesion and the corresponding regional range identifier.

9. A lesion localization system based on deep learning, comprising executing the lesion localization method based on deep learning as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The image preprocessing module is used to acquire medical image data and perform preprocessing to obtain standardized medical image data; The morphological field construction module is used to input standardized medical image data into the improved nnFormer model to obtain organ morphological distribution results and construct the morphological field of potential lesions. The mechanism parameter field construction module is used to calculate the gray-level gradient change rate field, tissue density offset field, orientation consistency tensor field and deformation tension field based on standardized medical image data and potential lesion morphology field, and to construct the lesion occurrence mechanism parameter field. The boundary estimation module is used to determine the potential lesion area based on the lesion occurrence mechanism parameter field, construct multiple morphological evolution paths and perform morphological evolution reverse reasoning to obtain the preliminary lesion boundary area. The final region determination module is used to calculate the topological and physiological structural features of the initially estimated lesion boundary region, and to perform joint energy assessment to determine the final lesion region. The positioning output module is used to output the spatial location and range of the lesion based on the final area of ​​the lesion.

Images