An intelligent chest CT image processing system
By combining multi-scale filtering and frequency domain edge enhancement modules with edge fusion correction and segmentation models, the problem of insufficient response of existing chest CT image processing systems to small blood vessels and multi-scale lesions is solved, and efficient processing and accurate segmentation of blood vessels and lesions are achieved.
Patent Information
- Application Number
- CN202510816852.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-18
- Publication Date
- 2025-09-30
- Estimated Expiration
- 2045-06-18
AI Technical Summary
Existing chest CT image processing systems lack sensitivity to small blood vessels, improperly handle frequency domain noise and artifacts, and lack specialized guidance based on vascular and bronchial anatomical structures, resulting in poor processing effects. They also lack response to multi-scale lesions and lack a mechanism to specifically focus on boundaries, which makes segmentation prone to over-smoothing or fracture, resulting in poor accuracy of processing results.
A multi-scale filtering processing module, a frequency domain edge enhancement module, an edge fusion correction module, and a chest CT image segmentation model are used to calculate gradients at 1mm and 2mm voxel intervals, respectively. Differential processing is performed in different spectral bands based on noise characteristics to guide vascular enhancement. Branch A refines the nodule edge and branch B enhances the macroscopic consistency of the lung lobe. Combined with five-layer parallel refinement and global downsampling, the decoding layer focuses on the interface to optimize the segmentation contour.
It improves the effect and accuracy of chest CT image processing, enhances the sensitive response to small blood vessels and large-scale lesions, reduces false positive noise, and improves the accuracy and completeness of segmentation.
Smart Images
Figure CN120355700B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing, and in particular to an intelligent chest CT image processing system. Background Art
[0002] Chest CT image processing systems perform preprocessing, such as noise reduction, enhancement, and correction, as well as segmentation and identification of different tissues and organs, three-dimensional reconstruction, measurement and analysis, and lesion detection on chest CT scan images. However, typical chest CT image processing systems suffer from insufficient sensitivity to small blood vessels, improper handling of frequency-domain noise and artifacts, and a lack of specialized guidance based on vascular and bronchial anatomy, resulting in poor processing results. They also suffer from insufficient response to multi-scale lesions and lack a mechanism for specialized boundary focus, leading to oversmoothing or fragmentation of the segmentation of blood vessels, lung parenchyma interfaces, and slender branches, resulting in poor processing accuracy. Summary of the Invention
[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an intelligent chest CT image processing system. In view of the problems that the general chest CT image processing system has insufficient sensitivity to small blood vessels, improper processing of frequency domain noise and artifacts, and lack of special guidance based on the anatomical structure of blood vessels and bronchial tubes, resulting in poor final processing effect, the present invention calculates the gradient with 1mm and 2mm voxel intervals respectively, and can sensitively respond to small nodules and large-scale pleural edges; performs differentiated processing in different spectral bands based on noise characteristics, guides the focus on retaining and enhancing the course of blood vessels, while suppressing high-frequency noise and restoring low-frequency large structures; through edge fusion correction, takes into account the adaptive enhancement of small-scale nodules and large-scale lesions; thereby improving Chest CT image processing effect; In view of the fact that general chest CT image processing systems have insufficient response to multi-scale lesions and lack a mechanism to specifically focus on boundaries, which leads to excessive smoothing or breakage of the segmentation of blood vessels, lung parenchyma interfaces and slender branches, and poor accuracy of processing results, this scheme uses branch A to refine the weak contrast between nodule edges and artifacts, and branch B to enhance the macroscopic consistency of the lung lobes; through five-layer parallel refinement + global joint downsampling, both small-scale ground-glass nodules and large-scale lung lobe structures are taken into account; the decoding layer suppresses the learned area, focuses on blood vessels, and lung parenchyma interfaces through interface focusing; the loss introduces boundary smoothing, optimizes the overall segmentation contour, and reduces false positive burr noise; thereby improving the accuracy of chest CT image processing.
[0004] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent chest CT image processing system, which includes an image acquisition module, a multi-scale filtering processing module, a frequency domain edge enhancement module, an edge fusion correction module, an enhancement loss design module, a chest CT image segmentation model establishment module and a chest CT image processing module;
[0005] The image acquisition module acquires historical chest CT images and constructs a chest CT sample set;
[0006] The multi-scale filtering processing module calculates the change intensity map of the chest CT image at 1mm and 2mm voxel intervals respectively, and generates the multi-scale change intensity map by upsampling alignment and scale-weighted fusion;
[0007] The frequency domain edge enhancement module performs 2DFFT separation on each layer of the multi-scale variation intensity map into three spectrum segments: low, medium, and high frequency. Multi-scale residual features are extracted using 5×5 and 1×1 convolutions. The vascular response is integrated into the medium frequency residual to reconstruct the restored chest CT gradient map.
[0008] The edge fusion correction module concatenates the original input features of the chest CT image with the repaired gradient map, and then performs weighted fusion of features through channel attention and attention map, and then outputs the chest CT correction feature map through three layers of 3×3 convolution and residual connection;
[0009] The enhancement loss design module constructs a total loss function of the enhancement network based on high-quality chest CT reference images;
[0010] The chest CT image segmentation model building module constructs an encoding-bottleneck-decoding network and trains the chest CT image segmentation model using normalized active contour and boundary length smoothing loss;
[0011] The chest CT image processing module processes real-time chest CT images based on a multi-scale filtering processing module, a frequency domain edge enhancement module, an edge fusion correction module and a chest CT image segmentation model.
[0012] Furthermore, the image acquisition module acquires historical chest CT images; standardizes intensity values; labels high-quality chest CT reference images; labels all historical chest CT images; and obtains a chest CT sample set.
[0013] Furthermore, the multi-scale filtering processing module uses multi-scale Sobel filtering to estimate the gradient map for the chest CT image. , expressed as: ; The Sobel outputs with voxel spacing of 1mm and 2mm are upsampled, aligned, and weighted fused, expressed as: ;in, 、 and is the gradient component of the filter output; is the fused multi-scale change intensity map; is the scale weighting coefficient; and These are the intensity change maps calculated at 1mm and 2mm voxel intervals, respectively.
[0014] Furthermore, the frequency domain edge enhancement module adds multi-spectral band analysis and Frangi-vascular enhancement guidance based on the frequency domain modified amplitude spectrum; 2DFFT is performed on each layer of CT gradient map, and the spectrum is separated into three bands: low frequency, medium frequency, and high frequency, which are expressed as: ; ;in, is the Fourier transform of the 2D slice; A is the frequency domain separation result, 、 and Represent the low-frequency, medium-frequency and high-frequency spectrum components respectively; the spatial domain extracts multi-scale features through the convolutional network, which is expressed as: ;in, is the multi-scale feature residual, k is the spectrum segment index; It uses a 5×5 convolution kernel on the 2D plane; For channel integration; is the activation function; for the vascular structure, the Frangi filter is used to generate the vascular response map V and Fusion, expressed as: ;in, is the intermediate frequency spectrum component after vascular response fusion; γ is the vascular enhancement coefficient; reconstruction fusion is expressed as: ;in, is the final enhanced and repaired gradient map; IFFT is the inverse Fourier transform reconstruction; P is the phase spectrum.
[0015] Furthermore, the edge fusion correction module adds space-slice attention in addition to channel attention to dynamically adjust gradient and chest CT image features; by combining features and dual attention, we obtain: ; Channel attention is expressed as: ; ;in, is the original input feature map; U is the joint feature after splicing; is the channel attention vector; and are the two channel attention sub-vectors after splitting; is a linear transformation; LayerNorm is layer normalization; is segmentation; Sigmoid is the Sigmoid activation function; 3D convolution + Sigmoid processing is performed on U to generate a spatial attention map , expressed as: ; Feature fusion is expressed as: ; ; Finally, after three layers of 3×3 convolution + residual output, it is expressed as: ;in, is 3D convolution; S is the feature fusion result; is a weighted feature map; 、 and It is a three-convolution operation; is the final enhanced output feature map.
[0016] Furthermore, the enhancement loss design module enhances the chest CT sample set based on the multi-scale filtering processing module, the frequency domain edge enhancement module and the edge fusion correction module, compares the sample set based on the high-quality chest CT reference image, and obtains the chest CT segmentation sample set based on loss training; adds edge loss; pixel consistency loss Expressed as: ;Edge consistency loss , expressed as: ; Final total loss function Expressed as: ; Where H and W are the height and width of the chest CT image; is the number of channels and rows and columns; is the enhanced pixel value; is the pixel value of the high-quality chest CT reference image; is a smoothing term; is the edge detection operator; is the loss weight; is the gradient constraint term.
[0017] Furthermore, the chest CT image segmentation model building module builds a segmentation model based on the chest CT segmentation sample set. The overall process includes three stages: encoding, bottleneck, and decoding. It introduces dual-branch refinement feature extraction in the encoding layer, applies multi-scale hybrid convolutional attention in the core compression layer, and combines the reverse attention boundary module in the decoding layer. Finally, boundary accuracy loss is used to enhance boundary learning. Specifically, it includes the following contents:
[0018] Architecture Overview: Input is an enhanced chest CT image; after the initial convolution, it enters a five-layer encoder, a bottleneck, and then a five-layer decoder; five lossless transmission channels are used between the encoding and decoding layers; and the deepest layer focuses on enhancing the blood vessel-lung parenchyma interface.
[0019] Encoder design; introduce dual-branch parallel extraction; branch A is expressed as: ;in, is the output of branch A; is a depth-wise separable convolution, 3×3 is the convolution kernel; a is the balanced residual coefficient; is the i-th layer of the encoder; is the example normalization; branch B is expressed as: ;in, It is the output of branch B; It is an axial depth-wise separable convolution with a convolution kernel of 7×7; concatenation is expressed as: ;in, 、 and They are branch A, branch B and the encoder layer i output after splicing; It is a concatenation operation; the number of channels is doubled and then downsampled;
[0020] Core compression layer design; hybrid convolution + attention processing is used for chest CT, expressed as: ; ; The attention weight is expressed as: ; ; ; Split output Expressed as: ;in, is the mixed convolution output; is the encoder output; It is point-wise convolution; is the point-by-point summation output; d is the quantitative response intensity, highlighting the key lesion area; yes mean; v is the response mean; is the attention weight; It is a balancing item;
[0021] Boundary enhancement; a dual mechanism of skip connection feature fusion and reverse attention boundary is introduced in the decoding stage; the first three levels of decoding refine the lung tissue microstructure, which can be expressed as: ;in, and They are the feature maps after upsampling of the i+1th layer and the i-th layer decoder respectively; BAM is boundary attention; for the deepest two layers of decoding, the prior prediction → reverse attention mechanism is introduced; the prior prediction generation is expressed as: ; ; Reverse attention boundary fusion is expressed as: ; ;in, is the boundary prior; is the reverse attention weight; is the feature after inverse attention weighting; is the skip connection feature; is the fusion coefficient;
[0022] Loss function design; add boundary length smoothing term to boundary accuracy loss; segmentation loss function is expressed as: ; ; ;in, is the total segmentation loss function; is the loss of boundary accuracy; is the boundary length smoothing term; is the loss weight; g and u are pixel coordinates; is the predicted segmentation result; is the true segmentation label; is the predicted segmentation gradient.
[0023] Furthermore, the chest CT image processing module collects chest CT images in real time, enhances them based on the multi-scale filtering processing module, the frequency domain edge enhancement module and the edge fusion correction module, and inputs them into the chest CT image segmentation model, and uses the segmentation results output by the model as the chest CT image processing results.
[0024] The beneficial effects achieved by the present invention using the above scheme are as follows:
[0025] (1) In view of the fact that general chest CT image processing systems are not sensitive enough to small blood vessels, do not handle frequency domain noise and artifacts properly, and lack specialized guidance based on the anatomical structure of blood vessels and bronchus, resulting in poor final processing effects, this scheme calculates gradients at 1mm and 2mm voxel intervals, respectively, and can respond sensitively to small nodules and large-scale pleural edges; it performs differentiated processing in different spectral bands based on noise characteristics, guiding the focus on retaining and enhancing the course of blood vessels, while suppressing high-frequency noise and restoring low-frequency large structures; through edge fusion correction, it takes into account the adaptive enhancement of small-scale nodules and large-scale lesions; thereby improving the processing effect of chest CT images.
[0026] (2) In view of the problem that general chest CT image processing systems have insufficient response to multi-scale lesions and lack a mechanism to focus on boundaries, which leads to excessive smoothing or breakage of the segmentation of blood vessels, lung parenchyma interfaces and slender branches, and poor accuracy of processing results, this scheme refines the nodule edge and weak contrast of artifacts through branch A, and strengthens the macroscopic consistency of the lung lobe through branch B; through five-layer parallel refinement + global joint downsampling, both small-scale ground-glass nodules and large-scale lung lobe structures are taken into account; the decoding layer suppresses the learned area, focuses on blood vessels, and lung parenchyma interfaces through interface focusing; the loss introduces boundary smoothing to optimize the overall segmentation contour and reduce false positive burr noise; thereby improving the accuracy of chest CT image processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A schematic diagram of the process of an intelligent chest CT image processing system provided by the present invention;
[0028] Figure 2 Schematic diagram of the process of building a module for the chest CT image segmentation model.
[0029] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION
[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0031] In the description of the present invention, it should be understood that terms such as "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they should not be understood as limiting the present invention.
[0032] Example 1, see Figure 1 The present invention provides an intelligent chest CT image processing system, comprising an image acquisition module, a multi-scale filtering processing module, a frequency domain edge enhancement module, an edge fusion correction module, an enhancement loss design module, a chest CT image segmentation model establishment module and a chest CT image processing module;
[0033] The image acquisition module acquires historical chest CT images and constructs a chest CT sample set;
[0034] The multi-scale filtering processing module calculates the change intensity map of the chest CT image at 1mm and 2mm voxel intervals respectively, and generates the multi-scale change intensity map by upsampling alignment and scale-weighted fusion;
[0035] The frequency domain edge enhancement module performs 2DFFT separation on each layer of the multi-scale variation intensity map into three spectrum segments: low, medium, and high frequency. Multi-scale residual features are extracted using 5×5 and 1×1 convolutions. The vascular response is integrated into the medium frequency residual to reconstruct the restored chest CT gradient map.
[0036] The edge fusion correction module concatenates the original input features of the chest CT image with the repaired gradient map, and then performs weighted fusion of features through channel attention and attention map, and then outputs the chest CT correction feature map through three layers of 3×3 convolution and residual connection;
[0037] The enhancement loss design module constructs a total loss function of the enhancement network based on high-quality chest CT reference images;
[0038] The chest CT image segmentation model building module constructs an encoding-bottleneck-decoding network and trains the chest CT image segmentation model using normalized active contour and boundary length smoothing loss;
[0039] The chest CT image processing module processes real-time chest CT images based on a multi-scale filtering processing module, a frequency domain edge enhancement module, an edge fusion correction module and a chest CT image segmentation model.
[0040] Example 2, see Figure 1 This embodiment is based on the above embodiment. The image acquisition module acquires historical chest CT images; standardizes the intensity value; truncates the HU value range to [−1000, 400], and linearly normalizes it to [0, 1]; labels high-quality chest CT reference images; labels all historical chest CT images with lung lobe and blood vessel segmentation labels; and obtains a chest CT sample set.
[0041] Example 3, see Figure 1 This embodiment is based on the above embodiment. The multi-scale filtering processing module uses multi-scale Sobel filtering to estimate the gradient map in view of the large differences in the sizes of blood vessels, trachea and lesions in chest CT images. , expressed as: The Sobel outputs with voxel spacing of 1mm and 2mm are upsampled, aligned, and weighted fused to improve the edge sensitivity of lesions of different sizes, expressed as: ;in, 、 and is the gradient component of the filter output; is the fused multi-scale change intensity map; is the scale weighting coefficient; and The change intensity maps are calculated at 1mm and 2mm voxel intervals respectively; taking the continuity between slices into consideration, the processing effect of blood vessels and bronchial branches is better; multi-scale fusion takes into account the gradient response of large-scale pleural edges and tiny nodules.
[0042] Example 4, see Figure 1 This embodiment is based on the above embodiment. The frequency domain edge enhancement module adds multi-spectral band analysis and Frangi-vascular enhancement guidance based on the frequency domain correction amplitude spectrum, targeting the characteristics of CT noise and artifacts. A 2DFFT is performed on each layer of CT gradient map, and the spectrum is separated into three bands: low frequency, medium frequency, and high frequency. The result is expressed as follows: ; ;in, is the Fourier transform of the 2D slice; A is the frequency domain separation result, 、 and Represent the low-frequency, medium-frequency and high-frequency spectrum components respectively; the spatial domain extracts multi-scale features through the convolutional network, which is expressed as: ;in, is the multi-scale feature residual, k is the spectrum segment index; It uses a 5×5 convolution kernel on the 2D plane; For channel integration; is the activation function; for the vascular structure, the Frangi filter is used to generate the vascular response map V and Fusion, expressed as: ;in, is the intermediate frequency spectrum component after vascular response fusion; γ is the vascular enhancement coefficient; reconstruction fusion is expressed as: ;in, It is the final enhanced and repaired gradient map; IFFT is the inverse Fourier transform reconstruction; P is the phase spectrum; through multi-spectral band processing, the weak gradients of small blood vessels and bronchioles are finely restored, and Frangi guidance focuses on preserving the vascular structure, suppressing noise while taking into account anatomical details.
[0043] Example 5, see Figure 1 Based on the above embodiment, this embodiment further adds space-slice attention to the edge fusion correction module in addition to channel attention, dynamically adjusting the gradient and chest CT image features; by combining features and dual attention, we obtain: ; Channel attention is expressed as: ; ;in, is the original input feature map; U is the joint feature after splicing; is the channel attention vector; and are the two channel attention sub-vectors after splitting; is a linear transformation; LayerNorm is layer normalization; is segmentation; Sigmoid is the Sigmoid activation function; 3D convolution + Sigmoid processing is performed on U to generate a spatial attention map , expressed as: ; Feature fusion is expressed as: ; ; Finally, after three layers of 3×3 convolution + residual output, it is expressed as: ;in, is 3D convolution; S is the feature fusion result; is a weighted feature map; 、 and It is a three-convolution operation; It is the final enhanced output feature map; slice dimension attention ensures that gradient information can also be dynamically emphasized or suppressed along the axis, further improving the integrity of 3D continuous lesions; channel + spatial dual attention enables both small-scale nodules and large-scale lesions to be adaptively enhanced.
[0044] Example 6, see Figure 1 This embodiment is based on the above embodiment. The enhancement loss design module enhances the chest CT sample set based on the multi-scale filtering processing module, the frequency domain edge enhancement module and the edge fusion correction module. The chest CT sample set is compared with the high-quality chest CT reference image and the chest CT segmentation sample set is obtained based on loss training. The edge loss is added to ensure that the enhanced chest CT image has high-quality semantics and edge information. The pixel consistency loss is used to determine the image quality. Expressed as: ;Edge consistency loss Helps restore the clear boundaries of the lesion area, expressed as: ; Final total loss function Expressed as: ; Where H and W are the height and width of the chest CT image; is the number of channels and rows and columns; is the enhanced pixel value; is the pixel value of the high-quality chest CT reference image; is a smoothing term; is the edge detection operator; is the loss weight; is the gradient constraint term.
[0045] By performing the above operations, the general chest CT image processing system has insufficient sensitivity to small blood vessels, improper processing of frequency domain noise and artifacts, and lack of specialized guidance based on vascular and bronchial anatomical structures, resulting in poor final processing effects. This scheme calculates gradients at 1mm and 2mm voxel intervals, respectively, and can sensitively respond to small nodules and large-scale pleural edges; differentiates noise characteristics in different spectral bands to guide the focus on retaining and enhancing blood vessel courses, while suppressing high-frequency noise and restoring low-frequency large structures; through edge fusion correction, adaptive enhancement of small-scale nodules and large-scale lesions is taken into account, thereby improving the chest CT image processing effect.
[0046] Example 7, see Figure 1 and Figure 2This embodiment is based on the above embodiment. The chest CT image segmentation model establishment module establishes a segmentation model based on the chest CT segmentation sample set. The overall process includes three stages: encoding, bottleneck, and decoding. In the encoding layer, dual-branch refinement feature extraction is introduced, multi-scale hybrid convolutional attention is applied in the core compression layer, and the reverse attention boundary module is combined in the decoding layer. Finally, boundary accuracy loss is used to enhance boundary learning. Specifically, it includes the following contents:
[0047] Architecture Overview: Input is an enhanced chest CT scan; after the initial convolution, it enters a five-layer encoder, a bottleneck, and then a five-layer decoder. Five lossless transmission channels are used between the encoding and decoding layers to ensure that both lung lobe details and global structure are restored simultaneously. The deepest layers are specifically enhanced at the vascular-parenchymal interface.
[0048] Encoder design: To take into account both the small nodules of ground-glass shadows and the overall anatomical structure of the lung lobes in chest CT, a dual-branch parallel extraction is introduced. Branch A captures the edges of small nodules, subtle textures, and weak contrast of artifacts. Branch A is expressed as: ;in, is the output of branch A; is a depth-wise separable convolution, 3×3 is the convolution kernel; a is the balanced residual coefficient; is the i-th layer of the encoder; is sample normalization; branch B strengthens the consistency of the macroscopic structure of the lung lobe and suppresses artifacts and noise; branch B is expressed as: ;in, It is the output of branch B; It is an axial depth-wise separable convolution with a convolution kernel of 7×7; concatenation is expressed as: ;in, 、 and They are branch A, branch B and the encoder layer i output after splicing; It is a splicing operation; the number of channels is doubled and then downsampled; through dual branches, the lung nodule edges are refined and the lung lobe contours are robustly modeled, improving the segmentation quality of both early and late lesions;
[0049] Core compression layer design: To meet the multi-scale segmentation requirements of the coexistence of small nodules and large-scale lesions in chest CT, a hybrid convolution + attention processing is used, which can be expressed as: ; ; The attention weight is expressed as: ; ; ; Split output Expressed as: ;in, is the mixed convolution output; is the encoder output; It is point-wise convolution; is the point-by-point summation output; d is the quantitative response intensity, highlighting the key lesion area; yes mean; v is the response mean; is the attention weight; It is a balancing term; it adaptively highlights key lesion responses, suppresses background artifacts, and outputs features containing rich multi-scale and spatial information in the core compression layer;
[0050] Boundary enhancement; a dual mechanism of skip connection feature fusion and reverse attention boundary is introduced in the decoding stage; the first three levels of decoding refine the lung tissue microstructure, which can be expressed as: ;in, and They are the feature maps after upsampling of the i+1th layer and the ith layer decoder respectively; BAM is boundary attention; it highlights the cellular structure of lung tissue; for the deepest two layers of decoding, the prior prediction → reverse attention mechanism is introduced, with the boundary as the focus. By suppressing the learned area and focusing on the blood vessel-lung parenchyma junction, the accuracy of segmentation of slender structures and interfaces can be significantly improved, and the boundary blur caused by over-smoothing can be reduced; the prior prediction generation is represented as: ; ; Reverse attention boundary fusion is expressed as: ; ;in, is the boundary prior; is the reverse attention weight; is the feature after inverse attention weighting; is the skip connection feature; is the fusion coefficient;
[0051] Loss function design; add boundary length smoothing term to boundary accuracy loss; segmentation loss function is expressed as: ; ; ;in, is the total segmentation loss function; is the loss of boundary accuracy; is the boundary length smoothing term; is the loss weight; g and u are pixel coordinates; is the predicted segmentation result; is the true segmentation label; is the predicted segmentation gradient.
[0052] By performing the above operations, the general chest CT image processing system has insufficient response to multi-scale lesions and lacks a mechanism to specifically focus on boundaries, which leads to excessive smoothing or breakage of the segmentation of blood vessels, lung parenchyma interfaces and slender branches, and poor processing accuracy. This scheme uses branch A to refine the weak contrast between nodule edges and artifacts, and branch B to enhance the macroscopic consistency of the lung lobes; through five-layer parallel refinement + global joint downsampling, both small-scale ground-glass nodules and large-scale lung lobe structures are taken into account; the decoding layer suppresses the learned area, focuses on blood vessels, and lung parenchyma interfaces through interface focusing; the loss introduces boundary smoothing to optimize the overall segmentation contour and reduce false positive burr noise, thereby improving the accuracy of chest CT image processing.
[0053] Example 8, see Figure 1 This embodiment is based on the above embodiment. The chest CT image processing module collects chest CT images in real time, enhances them based on the multi-scale filtering processing module, the frequency domain edge enhancement module and the edge fusion correction module, and then inputs them into the chest CT image segmentation model. The segmentation result output by the model is used as the chest CT image processing result.
[0054] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that various changes, modifications, substitutions, and alterations can be made to the embodiments without departing from the principles and spirit of the invention.
[0055] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.
Claims
1. An intelligent chest CT image processing system, characterized by: The system includes an image acquisition module, a multi-scale filtering processing module, a frequency domain edge enhancement module, an edge fusion correction module, an enhancement loss design module, a chest CT image segmentation model establishment module, and a chest CT image processing module; The image acquisition module acquires historical chest CT images and constructs a chest CT sample set; The multi-scale filtering processing module calculates the change intensity map of the chest CT image at 1mm and 2mm voxel intervals respectively, and generates the multi-scale change intensity map by upsampling alignment and scale-weighted fusion; The frequency domain edge enhancement module performs 2DFFT separation on each layer of the multi-scale variation intensity map into three spectrum segments: low, medium, and high frequency. Multi-scale residual features are extracted using 5×5 and 1×1 convolutions. The vascular response is integrated into the medium frequency residual to reconstruct the restored chest CT gradient map. The edge fusion correction module combines the original input features of the chest CT image with the restored chest CT gradient map, and then performs weighted fusion of features through channel attention and attention map, and then outputs the chest CT correction feature map through three layers of 3×3 convolution and residual connection; The enhancement loss design module constructs a total loss function of the enhancement network based on high-quality chest CT reference images; The chest CT image segmentation model building module constructs an encoding-bottleneck-decoding network and trains the chest CT image segmentation model using normalized active contour and boundary length smoothing loss; The chest CT image processing module processes real-time chest CT images based on a multi-scale filtering processing module, a frequency domain edge enhancement module, an edge fusion correction module and a chest CT image segmentation model.
2. The intelligent chest CT image processing system according to claim 1, characterized in that: The multi-scale filtering processing module uses multi-scale Sobel filtering to estimate the gradient map for the chest CT image. ; The Sobel outputs with voxel spacing of 1mm and 2mm are upsampled, aligned and then weighted fused.
3. The intelligent chest CT image processing system according to claim 2, characterized in that: The frequency domain edge enhancement module adds multi-spectral segment analysis and vascular enhancement guidance based on the frequency domain amplitude spectrum correction; performs 2DFFT on each layer of CT gradient map and separates it into three spectrum segments: low frequency, medium frequency, and high frequency; and extracts multi-scale features in the spatial domain through a convolutional network; For the vascular structure, Frangi filter is used to generate the vascular response map V and compared with Fusion; reconstruct the fusion to obtain the final enhanced and repaired gradient map.
4. The intelligent chest CT image processing system according to claim 3, characterized in that: The edge fusion correction module adds space-slice attention in addition to channel attention to dynamically adjust the gradient and chest CT image features; U is obtained by combining the features and dual attention; U is the combined features; 3D convolution + Sigmoid processing is performed on U to generate a spatial attention map ; Finally, the final enhanced output feature map is obtained through three layers of 3×3 convolution + residual output.
5. The intelligent chest CT image processing system according to claim 4, characterized in that: The enhancement loss design module enhances the chest CT sample set based on a multi-scale filtering processing module, a frequency domain edge enhancement module and an edge fusion correction module, compares it with a high-quality chest CT reference image, and obtains a chest CT segmentation sample set based on loss training.
6. The intelligent chest CT image processing system according to claim 5, characterized in that: The chest CT image segmentation model establishment module specifically includes the following contents: Architecture Overview: After inputting an enhanced chest CT image, it undergoes an initial convolution and enters a five-layer encoder, followed by a bottleneck layer and a five-layer decoder. Five lossless transmission channels are set up between the corresponding encoding and decoding layers, with the blood vessel-lung parenchyma interface being enhanced at the deepest layer. Encoder design: Introducing dual-branch parallel feature extraction and downsampling after doubling the number of channels; Core compression layer design: Chest CT images are processed using a hybrid convolutional layer combined with an attention mechanism to obtain segmentation output. Boundary Enhancement: The decoding stage introduces a dual mechanism of skip connection feature fusion and reverse attention boundary. The first three levels of decoding are used to refine lung tissue microstructure; for the deepest two layers of decoding, a priori prediction and reverse attention mechanisms are introduced. Loss function design: Add a boundary length smoothing term to the boundary accuracy loss.
7. The intelligent chest CT image processing system according to claim 6, characterized in that: The image acquisition module first acquires historical chest CT images, then standardizes the intensity values; at the same time, it labels high-quality chest CT reference images, and uses them as a benchmark to label all historical chest CT images, ultimately obtaining a chest CT sample set.
8. The intelligent chest CT image processing system according to claim 7, characterized in that: The chest CT image processing module collects chest CT images in real time, enhances them based on the multi-scale filtering processing module, the frequency domain edge enhancement module and the edge fusion correction module, and inputs them into the chest CT image segmentation model. The segmentation results output by the model are used as the chest CT image processing results.