A low-dose CT image deep learning adaptive denoising reconstruction system

Abstract

The application relates to the technical field of medical image processing, and discloses a low-dose CT image deep learning self-adaptive denoising reconstruction system which comprises a low-dose CT image acquisition module, an adjacent layer structure analysis module, a scanning stability index generation module, a noise level estimation module, a scanning stability constraint generation module, a partition denoising reconstruction module and a boundary consistency correction module. By generating a gray continuity index, an edge continuity index and a structure offset stability index, a scanning stability graph is established, and a scanning stability constraint relationship is established in combination with a local noise estimation value, so that the basic denoising result is controlled to participate in the fusion reconstruction of a normalized image sequence, boundary consistency correction is finally completed, and a reconstruction result is output. According to the application, the gray continuity index, the edge continuity index and the structure offset stability index are generated, and the scanning stability graph is generated, so that the scanning stable state participates in the denoising reconstruction, and the condition that a structure change is mistaken for noise is reduced.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a deep learning adaptive noise reduction and reconstruction system for low-dose CT images. Background Technology

[0002] With the development of medical imaging technology, computed tomography (CT) has been widely used in clinical disease screening, lesion localization, and treatment evaluation due to its advantages such as fast imaging speed, high spatial resolution, and clear display of tissue structures. To reduce the radiation dose to patients undergoing CT examinations, low-dose CT is gradually becoming an important development direction in medical imaging.

[0003] Existing low-dose CT image denoising and reconstruction methods typically employ deep learning models to denoise low-dose CT images. By inputting low-dose CT images, corresponding denoising and reconstruction results are obtained. Some schemes combine local noise estimation, texture features, or edge information to adjust the denoising intensity. However, in cases where patients exhibit slight voluntary movement, respiratory fluctuations, or cardiac pulsation, subtle positional changes in local tissue structures can occur between adjacent scan layers. Current methods primarily rely on single-layer image information or local noise distribution for denoising and reconstruction, lacking an analysis of the structural continuity between adjacent scan layers. They often incorporate structural changes between adjacent layers along with random noise into the denoising process.

[0004] However, in the process of realizing the technical solution of this application, the inventors of this application discovered that in scanning scenarios where slight voluntary movements, breathing fluctuations, or heartbeats of the patient cause subtle changes in the local tissue structure of adjacent scanning layers, the existing low-dose CT image deep learning noise reduction and reconstruction methods lack an analysis mechanism for the steady state of the scan. They cannot distinguish between the differences between local structural changes generated during the scanning process and random noise, and are prone to misjudging the structural changes caused by the patient's slight movements as noise and uniformly processing them for noise reduction, thereby affecting the preservation of tissue structure information during the noise reduction and reconstruction process. Summary of the Invention

[0005] To overcome the above shortcomings, this invention provides a low-dose CT image deep learning adaptive noise reduction and reconstruction system, which aims to improve the problem that existing technologies easily misjudge structural changes caused by slight patient movements as noise and perform uniform noise reduction processing, affecting the preservation of tissue structure information during the noise reduction and reconstruction process.

[0006] This invention provides the following technical solution: a low-dose CT image deep learning adaptive noise reduction and reconstruction system, comprising: The low-dose CT image acquisition module acquires low-dose CT image sequences and generates normalized image sequences. The adjacent layer structure analysis module is connected to the low-dose CT image acquisition module and generates grayscale continuity index, edge continuity index and structural displacement stability index based on the normalized image sequence. The scanning stability index generation module is connected to the adjacent layer structure analysis module and generates a scanning stability map based on the grayscale continuity index, the edge continuity index, and the structure offset stability index. The noise level estimation module is connected to the low-dose CT image acquisition module and generates local noise estimates based on the normalized image sequence. The scan stability constraint generation module is connected to the scan stability index generation module and the noise level estimation module. It establishes scan stability constraint relationships based on the scan stability map and the local noise estimation value, and generates a noise reduction intensity control map based on the scan stability constraint relationships. The partitioned noise reduction and reconstruction module is data-connected to the scan stability constraint generation module. It controls the basic noise reduction results to participate in the fusion reconstruction of the normalized image sequence according to the noise reduction intensity control map, and generates a noise-reconstructed image. The boundary consistency correction module is connected to the partitioned noise reduction and reconstruction module. It performs boundary consistency correction based on the noise reduction and reconstruction image to generate low-dose CT image deep learning adaptive noise reduction and reconstruction results.

[0007] Preferably, the adjacent layer structure analysis module generates grayscale continuity index, edge continuity index, and structural offset stability index based on the normalized image sequence, specifically including: A grayscale continuity index is generated based on the grayscale differences between normalized images of adjacent layers. An edge continuity index is generated based on the edge gradient changes between normalized images of adjacent layers; A structural offset stability index is generated based on the local structural matching relationship between normalized images of adjacent layers. The adjacent layer structure analysis results are generated based on the grayscale continuity index, the edge continuity index, and the structural offset stability index.

[0008] Preferably, the scan stability index generation module generates a scan stability map based on the adjacent layer structure analysis results, specifically including: The scanning stability index is calculated based on the grayscale continuity index, the edge continuity index, and the structural offset stability index. Establish interlayer continuity relationships based on the current layer scan stability index and the scan stability indices of adjacent layers; A scan stability map is generated based on the scan stability index and the interlayer continuity relationship.

[0009] Preferably, the noise level estimation module generates local noise estimates based on the normalized image sequence, specifically including: A base reference image is generated based on the normalized image sequence; Local noise residuals are generated based on the grayscale residuals between the normalized image sequence and the base reference image; A local noise estimate is generated based on the local noise residual.

[0010] Preferably, the scan stability constraint generation module establishes scan stability constraint relationships based on the scan stability map and the local noise estimate, specifically including: Stability constraint weights are generated based on the scan stability map; Noise response weights are generated based on the local noise estimates; Establish a scan stability constraint relationship based on the stability constraint weights and the noise response weights; A noise reduction intensity control map is generated based on the scan stability constraint relationship.

[0011] Preferably, the partitioned denoising and reconstruction module controls the basic denoising results to participate in the fusion and reconstruction of the normalized image sequence according to the denoising intensity control map, specifically including: Establish the noise reduction intensity distribution based on the noise reduction intensity control chart; A corresponding noise reduction control region is generated based on the noise reduction intensity distribution; The participation ratio of the basic noise reduction results is controlled according to the noise reduction intensity corresponding to each noise reduction control area; Based on the participation ratio, the basic denoising results are fused with the normalized image sequence to generate a denoised and reconstructed image.

[0012] Preferably, the partitioned noise reduction and reconstruction module generates a noise reduction control region based on the noise reduction intensity distribution, specifically including: Establish the region boundary based on the continuous variation of noise reduction intensity; Multiple noise reduction control zones are divided according to the region boundaries; Establish regional transition relationships based on the noise reduction intensity variation between adjacent noise reduction control areas.

[0013] Preferably, the partitioned noise reduction and reconstruction module controls the participation ratio of the basic noise reduction results according to each noise reduction control region, specifically including: Generate corresponding model output participation coefficients based on each noise reduction control region; The fusion relationship between the basic denoising results and the normalized image sequence is established based on the participation coefficients output by the model. Based on the fusion relationship, the fusion results corresponding to each noise reduction control region are generated.

[0014] Preferably, the boundary consistency correction module performs boundary consistency correction based on the denoised reconstructed image, specifically including: Boundary transition weights are generated based on the regional transition relationship between adjacent noise reduction control regions. The fusion ratio at the boundary of the region is adjusted according to the boundary transition weight; Generate boundary consistency correction results based on the corrected fusion ratio.

[0015] Preferably, the scan stability constraint generation module generates a noise reduction intensity control map based on the scan stability constraint relationship, specifically including: Establish a scan stability constraint mapping relationship based on the aforementioned scan stability constraint relationship; The noise reduction intensity corresponding to different positions is determined based on the scan stability constraint mapping relationship; A noise reduction intensity control chart is generated based on the noise reduction intensity corresponding to different locations.

[0016] The present invention has the following beneficial effects: 1. This invention generates grayscale continuity index, edge continuity index, and structural offset stability index, and generates a scan stability map, enabling the scan stability state to participate in the noise reduction and reconstruction process. This solves the technical problem in the prior art that it is easy to misjudge the structural changes caused by slight patient movements as noise and perform uniform noise reduction processing.

[0017] 2. This invention establishes a scanning stability constraint relationship based on the scanning stability map and the estimated local noise value, and generates a noise reduction intensity control map, so that the noise reduction process can simultaneously consider the scanning stability state and the local noise distribution, thereby achieving adaptive control of the noise reduction intensity.

[0018] 3. This invention controls the basic noise reduction results to participate in the fusion and reconstruction of normalized image sequences according to the noise reduction intensity control map, and generates reconstruction results by combining boundary consistency correction, so that different noise reduction control areas can maintain a continuous transition. Attached Figure Description

[0019] Figure 1 This is a block diagram of a low-dose CT image deep learning adaptive noise reduction and reconstruction system proposed in this invention; Figure 2 This is a flowchart of the adjacent layer structure analysis proposed in this invention; Figure 3 The flowchart for generating the scanning stability map proposed in this invention is as follows; Figure 4 The flowchart for generating the scanning stability constraint relationship proposed in this invention is shown below. Figure 5 This is a flowchart of the partitioned noise reduction and reconstruction process proposed in this invention; Figure 6 This is a flowchart of the boundary consistency correction process proposed in this invention; Figure 7 This is a flowchart of the low-dose CT image deep learning adaptive noise reduction and reconstruction process proposed in this invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 and Figure 7 This invention provides a low-dose CT image deep learning adaptive noise reduction and reconstruction system, comprising: The low-dose CT image acquisition module acquires low-dose CT image sequences and generates normalized image sequences. The adjacent layer structure analysis module is connected to the low-dose CT image acquisition module to generate grayscale continuity index, edge continuity index and structural displacement stability index based on the normalized image sequence. The scanning stability index generation module is connected to the adjacent layer structure analysis module to generate a scanning stability map based on grayscale continuity index, edge continuity index, and structural offset stability index. The noise level estimation module is connected to the low-dose CT image acquisition module and generates local noise estimates based on the normalized image sequence. The scan stability constraint generation module is connected to the scan stability index generation module and the noise level estimation module. It establishes scan stability constraint relationships based on the scan stability map and local noise estimates, and generates a noise reduction intensity control map based on the scan stability constraint relationships. The partitioned noise reduction and reconstruction module is connected to the scan stability constraint generation module. Based on the noise reduction intensity control map, it controls the basic noise reduction results to participate in the fusion and reconstruction of the normalized image sequence to generate a noise-reconstructed image. The boundary consistency correction module is connected to the partitioned noise reduction and reconstruction module. It performs boundary consistency correction based on the noise reduction and reconstruction image to generate deep learning adaptive noise reduction and reconstruction results for low-dose CT images.

[0022] Specifically, the low-dose CT image acquisition module acquires continuous slice images formed by low-dose CT scans and performs grayscale normalization on the low-dose CT image sequence to generate a normalized image sequence; the adjacent layer structure analysis module calculates the grayscale changes, edge changes, and local structural shifts between adjacent layers based on the normalized image sequence, generating grayscale continuity index, edge continuity index, and structural shift stability index; the scan stability index generation module fuses the above indices to generate a scan stability map; the noise level estimation module calculates local noise estimates based on the normalized image sequence; the scan stability constraint generation module establishes scan stability constraint relationships by combining the scan stability map and local noise estimates, and generates a noise reduction intensity control map based on the scan stability constraint relationships; the zonal noise reduction and reconstruction module controls the basic noise reduction results to participate in the fusion and reconstruction of the normalized image sequence based on the noise reduction intensity control map, generating a noise-reconstructed image; and the boundary consistency correction module performs boundary consistency correction based on the fusion relationship between adjacent regions in the noise-reconstructed image, generating a low-dose CT image deep learning adaptive noise reduction and reconstruction result.

[0023] This invention completes the deep learning adaptive noise reduction and reconstruction of low-dose CT images by following the processing sequence of normalized image sequence generation, adjacent layer structure analysis, scan stability map generation, local noise estimation, scan stability constraint relationship establishment, noise reduction intensity control map generation, zonal noise reduction and reconstruction, and boundary consistency correction. The modules are connected sequentially through data streams, and the output of the previous module serves as the input data for the next module. All modules work together to complete the deep learning adaptive noise reduction and reconstruction process of low-dose CT images.

[0024] Reference Figure 2 Furthermore, the adjacent layer structure analysis module generates grayscale continuity indicators, edge continuity indicators, and structural offset stability indicators based on the normalized image sequence, specifically including: A grayscale continuity index is generated based on the grayscale differences between normalized images of adjacent layers. An edge continuity index is generated based on the edge gradient changes between normalized images of adjacent layers; A structural offset stability index is generated based on the local structural matching relationship between normalized images of adjacent layers. The analysis results of adjacent layer structures are generated based on the grayscale continuity index, edge continuity index, and structural offset stability index.

[0025] Specifically, the adjacent layer structure analysis module first performs corresponding analysis on two adjacent normalized layers in the normalized image sequence according to the CT image acquisition order. Let the first layer be... Layer normalized image , No. Layer normalized image Calculate the grayscale difference at the corresponding pixel location, and generate a grayscale continuity index based on the grayscale difference. The grayscale continuity index can be expressed as: ; in, Indicates the continuity of grayscale; Indicates the first In a layer-normalized image, coordinates represent the grayscale value corresponding to the location. Indicates the first The gray value at the corresponding location in the layer-normalized image; This represents the grayscale normalization adjustment coefficient. The smaller the grayscale difference, the greater the grayscale continuity index at the corresponding position.

[0026] After obtaining the grayscale continuity index, edge gradient information is calculated for the normalized images of adjacent layers, and an edge continuity index is generated based on the changes in edge gradients at corresponding locations. The edge gradient is calculated using the local grayscale change rate of the image. When the edge gradient change in the corresponding region of the adjacent layer is small, the location is determined to have high edge continuity; when the edge gradient change is large, the edge continuity at the corresponding location decreases, thus forming the edge continuity index.

[0027] Subsequently, using local image patches in the current layer's normalized image as reference regions, corresponding local structural regions are searched in adjacent layer normalized images. The correspondence between local structures is determined based on the matching error between the reference region and the candidate regions, and a structural migration stability index is generated based on the offset distance between the matching positions. The smaller the local structural migration distance, the larger the structural migration stability index at the corresponding position; conversely, the larger the local structural migration distance, the smaller the structural migration stability index at the corresponding position.

[0028] Finally, the grayscale continuity index, edge continuity index, and structural offset stability index are uniformly organized to generate the adjacent layer structure analysis results corresponding to the current normalized image layer. These results are then used as data input for generating the scan stability map, providing structural continuity information for subsequent scan stability analysis. This ensures that grayscale changes, edge changes, and local structural changes between adjacent layers can uniformly participate in the subsequent scan stability analysis process.

[0029] Reference Figure 3 Furthermore, the scan stability index generation module generates a scan stability map based on the structural analysis results of adjacent layers, specifically including: The scanning stability index is calculated based on the grayscale continuity index, edge continuity index, and structural offset stability index. Establish interlayer continuity relationships based on the current layer scan stability index and the scan stability indices of adjacent layers; A scan stability map is generated based on the scan stability index and the interlayer continuity relationship.

[0030] Specifically, the scan stability index generation module receives the grayscale continuity index, edge continuity index, and structural offset stability index output by the adjacent layer structure analysis module, and performs a unified fusion calculation on the three indices to obtain the scan stability index corresponding to the current normalized image layer. Let the grayscale continuity index at the corresponding position of the current normalized image layer be... The edge continuity index is The structural offset stability index is Then scan stability index It can be represented as: ; in, Indicates the first The scanning stability index of the corresponding position in the slice-normalized image; Indicates the continuity of grayscale; Indicates the continuity index of the edge; Indicates the structural offset stability index; , and These represent the fusion weights of the corresponding indicators, and satisfy the following conditions: .

[0031] After obtaining the current layer's scan stability index, a correlation analysis is performed between the current layer's scan stability index and the scan stability indices of adjacent layers. An inter-layer continuity relationship is established based on the continuous changes in the scan stability indices of adjacent layers. When the scan stability indices of corresponding positions across multiple consecutive image layers maintain continuous change, the inter-layer continuity at those positions is maintained. When there is a sudden change in the scan stability index between adjacent image layers, the inter-layer continuity relationship at the corresponding position of the current layer is corrected based on the changing trends of the scan stability indices of adjacent layers, ensuring that scan stability maintains continuous change along the CT scan layer direction.

[0032] Subsequently, the scan stability index and inter-layer continuity relationship are fused to form a unified scan stability map for the corresponding positions in the normalized images of each layer. The scan stability map is used to characterize the scan stability state corresponding to each position in the normalized image sequence and maintain the continuous correspondence between the scan stability states of adjacent image layers, so as to form the basic data required for establishing subsequent scan stability constraints.

[0033] In this way, grayscale continuity, edge continuity, and structural offset stability are collectively converted into a unified scan stability map, so that the scan stability state between adjacent layers can participate in the establishment of subsequent scan stability constraint relationships, providing scan stability information for the generation of noise reduction intensity control map.

[0034] Furthermore, the noise level estimation module generates local noise estimates based on the normalized image sequence, specifically including: Generate a base reference image from the normalized image sequence; Local noise residuals are generated based on the grayscale residuals between the normalized image sequence and the base reference image. Local noise estimates are generated based on local noise residuals.

[0035] Specifically, the noise level estimation module receives the normalized image sequence output by the low-dose CT image acquisition module. Taking the current normalized image layer as the processing object, it performs local smoothing on the normalized image while maintaining the continuity of local tissue structure, generating a corresponding baseline reference image. The baseline reference image is used to characterize the basic gray-level distribution in the current normalized image layer and maintains the main tissue structure information to reduce the influence of tissue edges on the noise estimation process. Subsequently, a local noise residual is generated based on the gray-level difference at corresponding positions between the normalized image sequence and the baseline reference image. It can be represented as: ; in, Indicates coordinates as Local noise residuals corresponding to the location; This represents the gray value at the corresponding position in the normalized image sequence; This represents the grayscale value at the corresponding location in the base reference image.

[0036] After obtaining the local noise residuals, a local statistical region is established centered on the current location. Statistical analysis is performed on the local noise residuals corresponding to each pixel within the local statistical region to generate a local noise estimate for the current location. The local noise estimate reflects the degree of noise distribution within the current region. When the local noise residuals corresponding to each location within the local statistical region maintain a high degree of consistency, it is determined that the region has a corresponding local noise estimate. When the local noise residuals within the local statistical region change significantly, the local noise estimate for the corresponding location is recalculated based on the statistical results to ensure that the local noise estimate corresponds to the actual noise distribution.

[0037] Using the above method, a base reference image is generated based on the normalized image sequence, and a local noise estimate is generated by combining the local noise residual. This allows the local noise information to participate in the establishment of the scan stability constraint relationship together with the subsequently generated scan stability map, providing noise distribution information for the generation of the subsequent noise reduction intensity control map.

[0038] Reference Figure 4 Furthermore, the scan stability constraint generation module establishes scan stability constraint relationships based on the scan stability map and local noise estimates, specifically including: Generate stability constraint weights based on the scan stability map; Noise response weights are generated based on local noise estimates; Establish scan stability constraint relationships based on stability constraint weights and noise response weights; A noise reduction intensity control map is generated based on the scanning stability constraint relationship.

[0039] The scan stability constraint generation module generates a noise reduction intensity control map based on the scan stability constraint relationship, specifically including: Establish a scan stability constraint mapping relationship based on the scan stability constraint relationship; The noise reduction intensity corresponding to different positions is determined based on the scanning stability constraint mapping relationship; A noise reduction intensity control chart is generated based on the noise reduction intensity corresponding to different locations.

[0040] Specifically, the scan stability constraint generation module receives the scan stability map and local noise estimates, and generates stability constraint weights and noise response weights, respectively. The stability constraint weights characterize the degree to which the scan stability state at the corresponding location constrains the denoising process, while the noise response weights characterize the degree to which the noise distribution at the corresponding location responds to the denoising process. Let the scan stability index at the corresponding location on the scan stability map be... The local noise estimate is Then scan the stable constraint relationship. It can be represented as: ; in, Indicates coordinates as The scanning stability constraints corresponding to the position; This indicates the scan stability index at the corresponding location; This represents the estimated local noise value at the corresponding location; This represents the stability coefficient to prevent the denominator from being zero. The scan stability index and the local noise estimate together determine the scan stability constraint relationship at the corresponding location, establishing a correspondence between the scan stability state and the local noise distribution.

[0041] After establishing scan stabilization constraints, a scan stabilization constraint mapping relationship is generated based on these constraints. This mapping relationship describes the correspondence between scan stabilization constraints and noise reduction intensity, and determines the noise reduction intensity at each location based on the continuous distribution of the scan stabilization constraints in the image space. When the scan stabilization constraints at adjacent locations change continuously, the noise reduction intensity at those locations also changes continuously. However, when the scan stabilization constraints change significantly, the noise reduction intensity at the corresponding location is redefined based on the scanning stabilization constraint mapping relationship, ensuring that the noise reduction intensity changes synchronously with the scan stabilization constraints.

[0042] After obtaining the denoising intensity corresponding to each location based on the scanning stability constraint mapping relationship, a denoising intensity control map is established for the entire normalized image. The denoising intensity control map records the denoising intensity corresponding to each location in the normalized image and maintains a continuous correspondence between the denoising intensity of adjacent locations, serving as the basis for controlling the participation degree of the basic denoising results in the subsequent partitioned denoising and reconstruction module.

[0043] In the above manner, the scan stability map and the local noise estimate are jointly converted into a scan stability constraint relationship, and a noise reduction intensity control map is further generated. This allows the scan stability state and the local noise distribution to jointly participate in the determination of the noise reduction intensity, providing a unified noise reduction control basis for the subsequent fusion and reconstruction of the basic noise reduction results and the normalized image sequence.

[0044] Reference Figure 5 Furthermore, the partitioned noise reduction and reconstruction module controls the basic noise reduction results to participate in the fusion and reconstruction of the normalized image sequence based on the noise reduction intensity control map, specifically including: Establish the noise reduction intensity distribution based on the noise reduction intensity control chart; Generate corresponding noise reduction control areas based on the noise reduction intensity distribution; The participation ratio of the basic noise reduction results is controlled according to the noise reduction intensity corresponding to each noise reduction control area; Based on the participation ratio, the basic denoising results are fused with the normalized image sequence to generate a denoised and reconstructed image.

[0045] The zone-based noise reduction and reconstruction module generates noise reduction control regions based on the noise reduction intensity distribution, specifically including: Establish the region boundary based on the continuous variation of noise reduction intensity; Multiple noise reduction control zones are divided according to the area boundaries; Establish regional transition relationships based on the noise reduction intensity variation between adjacent noise reduction control areas.

[0046] The zone-based noise reduction and reconstruction module controls the participation ratio of the basic noise reduction results according to each noise reduction control area, specifically including: Generate corresponding model output participation coefficients based on each noise reduction control region; The fusion relationship between the basic denoising results and the normalized image sequence is established based on the participation coefficients output by the model. The fusion results for each noise reduction control region are generated based on the fusion relationship.

[0047] Specifically, the partitioned noise reduction and reconstruction module receives the noise reduction intensity control map output by the scan stability constraint generation module, and establishes the noise reduction intensity distribution of the entire normalized image based on the noise reduction intensity corresponding to each position in the noise reduction intensity control map. Subsequently, it establishes region boundaries based on the continuous spatial variation of noise reduction intensity. When the noise reduction intensity variation of adjacent positions remains continuous, they are assigned to the same noise reduction control region; when the noise reduction intensity variation exceeds a preset range, a new region boundary is established, and multiple noise reduction control regions are divided based on the region boundaries.

[0048] Meanwhile, a correlation analysis is conducted on the changes in noise reduction intensity between adjacent noise reduction control areas to establish regional transition relationships, so that the noise reduction intensity changes between adjacent noise reduction control areas remain continuous.

[0049] After completing the noise reduction control region division, the model output participation coefficients are generated based on the noise reduction intensity corresponding to each noise reduction control region. Let the th... The noise reduction intensity corresponding to each noise reduction control region is: The model output participation coefficient is The model output participation coefficient can then be expressed as: ; in, Indicates the first The model output participation coefficients corresponding to each noise reduction control region; Indicates the first The noise reduction intensity corresponding to each noise reduction control area; This represents the maximum noise reduction intensity in the noise reduction intensity control chart. The model output participation coefficient is used to determine the proportion of the base noise reduction result that participates in the fusion reconstruction.

[0050] After obtaining the model output participation coefficients, a fusion relationship is established between the basic denoising results and the normalized image sequence based on the model output participation coefficients. The fusion calculation is then performed based on the model output participation coefficients at corresponding locations, resulting in the fused reconstruction. It can be represented as: ; in, Indicates coordinates as The grayscale value of the denoised and reconstructed image corresponding to the location; This represents the model output participation coefficient at the corresponding position; This represents the grayscale value at the corresponding position in the basic noise reduction result; This represents the grayscale value at the corresponding position in the normalized image sequence. Based on the above fusion relationship, corresponding fusion results are generated for each noise reduction control region, and adjacent regions are continuously stitched together according to the region transition relationship to form a complete noise-reconstructed image.

[0051] Using the above method, a noise reduction control region is established based on the noise reduction intensity control map. The participation coefficient of the model output corresponding to each noise reduction control region is used to control the participation of the basic noise reduction result in the fusion and reconstruction of the normalized image sequence. This allows the participation degree of the basic noise reduction result to correspond to the noise reduction intensity of different regions, providing continuous noise reduction and reconstruction images for subsequent boundary consistency correction.

[0052] Reference Figure 6 Furthermore, the boundary consistency correction module performs boundary consistency correction based on the denoised reconstructed image, specifically including: Boundary transition weights are generated based on the regional transition relationship between adjacent noise reduction control regions. The fusion ratio at the boundary of the region is adjusted according to the boundary transition weight; Generate boundary consistency correction results based on the corrected fusion ratio.

[0053] Specifically, the boundary consistency correction module receives the denoised and reconstructed image output by the partitioned denoising and reconstruction module, and performs boundary consistency correction on the boundary locations of adjacent denoising control regions based on the established regional transition relationships. First, it extracts the regional transition relationships on both sides of the boundary and generates boundary transition weights based on these relationships. Let the regional transition weights corresponding to the boundary locations be... The model output participation coefficients on both sides of the boundary are respectively and Then the boundary transition weight can be expressed as: ; in, Indicates coordinates as Boundary transition weights corresponding to the location; This represents the model output participation coefficient at the corresponding location on one side of the boundary. This represents the model output participation coefficient at the corresponding position on the other side of the boundary.

[0054] After obtaining the boundary transition weights, these weights are used as the basis for correcting the fusion ratio at the region boundaries. The fusion ratio at the region boundaries is redefined based on these weights, ensuring that the fusion ratio at the boundaries maintains a continuous relationship with the adjacent noise reduction control regions. Subsequently, the reconstructed grayscale values ​​at the region boundaries are recalculated based on the corrected fusion ratios, and the reconstruction results at the region boundaries are updated point by point, ensuring that the grayscale changes between adjacent noise reduction control regions remain continuous.

[0055] After correcting the fusion ratio at the region boundaries, the corrected reconstruction results of each region are recombined to generate a boundary consistency correction result. The boundary consistency correction result keeps the reconstruction results within each noise reduction control region unchanged, and only corrects the fusion ratio at the region boundaries according to the boundary transition weight to ensure that the entire noise reduction and reconstruction image maintains continuous change at the region boundaries.

[0056] Using the above method, boundary transition weights are generated based on the regional transition relationship, and the fusion ratio at the regional boundary is corrected according to the boundary transition weights, so that the reconstruction results between adjacent noise reduction control regions can maintain a continuous correspondence, and generate low-dose CT image deep learning adaptive noise reduction reconstruction results after boundary consistency correction.

[0057] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning adaptive noise reduction and reconstruction system for low-dose CT images, characterized in that, include: The low-dose CT image acquisition module acquires low-dose CT image sequences and generates normalized image sequences. The adjacent layer structure analysis module is connected to the low-dose CT image acquisition module and generates grayscale continuity index, edge continuity index and structural displacement stability index based on the normalized image sequence. The scanning stability index generation module is connected to the adjacent layer structure analysis module and generates a scanning stability map based on the grayscale continuity index, the edge continuity index, and the structure offset stability index. The noise level estimation module is connected to the low-dose CT image acquisition module and generates local noise estimates based on the normalized image sequence. The scan stability constraint generation module is connected to the scan stability index generation module and the noise level estimation module. It establishes scan stability constraint relationships based on the scan stability map and the local noise estimation value, and generates a noise reduction intensity control map based on the scan stability constraint relationships. The partitioned noise reduction and reconstruction module is data-connected to the scan stability constraint generation module. It controls the basic noise reduction results to participate in the fusion reconstruction of the normalized image sequence according to the noise reduction intensity control map, and generates a noise-reconstructed image. The boundary consistency correction module is connected to the partitioned noise reduction and reconstruction module. It performs boundary consistency correction based on the noise reduction and reconstruction image to generate low-dose CT image deep learning adaptive noise reduction and reconstruction results.

2. The low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 1, characterized in that, The adjacent layer structure analysis module generates grayscale continuity index, edge continuity index, and structural offset stability index based on the normalized image sequence, specifically including: A grayscale continuity index is generated based on the grayscale differences between normalized images of adjacent layers. An edge continuity index is generated based on the edge gradient changes between normalized images of adjacent layers; A structural offset stability index is generated based on the local structural matching relationship between normalized images of adjacent layers. The adjacent layer structure analysis results are generated based on the grayscale continuity index, the edge continuity index, and the structural offset stability index.

3. The low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 1, characterized in that, The scan stability index generation module generates a scan stability map based on the adjacent layer structure analysis results, specifically including: The scanning stability index is calculated based on the grayscale continuity index, the edge continuity index, and the structural offset stability index. Establish interlayer continuity relationships based on the current layer scan stability index and the scan stability indices of adjacent layers; A scan stability map is generated based on the scan stability index and the interlayer continuity relationship.

4. The low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 1, characterized in that, The noise level estimation module generates local noise estimates based on the normalized image sequence, specifically including: A base reference image is generated based on the normalized image sequence; Local noise residuals are generated based on the grayscale residuals between the normalized image sequence and the base reference image; A local noise estimate is generated based on the local noise residual.

5. The low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 1, characterized in that, The scan stability constraint generation module establishes scan stability constraint relationships based on the scan stability map and the local noise estimate, specifically including: Stability constraint weights are generated based on the scan stability map; Noise response weights are generated based on the local noise estimates; Establish a scan stability constraint relationship based on the stability constraint weights and the noise response weights; A noise reduction intensity control map is generated based on the scan stability constraint relationship.

6. The low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 1, characterized in that, The partitioned noise reduction and reconstruction module controls the basic noise reduction results to participate in the fusion and reconstruction of the normalized image sequence according to the noise reduction intensity control map, specifically including: Establish the noise reduction intensity distribution based on the noise reduction intensity control chart; A corresponding noise reduction control region is generated based on the noise reduction intensity distribution; The participation ratio of the basic noise reduction results is controlled according to the noise reduction intensity corresponding to each noise reduction control area; Based on the participation ratio, the basic denoising results are fused with the normalized image sequence to generate a denoised and reconstructed image.

7. A low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 6, characterized in that, The partitioned noise reduction and reconstruction module generates a noise reduction control region based on the noise reduction intensity distribution, specifically including: Establish the region boundary based on the continuous variation of noise reduction intensity; Multiple noise reduction control zones are divided according to the region boundaries; Establish regional transition relationships based on the noise reduction intensity variation between adjacent noise reduction control areas.

8. A low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 6, characterized in that, The partitioned noise reduction and reconstruction module controls the participation ratio of the basic noise reduction results according to each noise reduction control region, specifically including: Generate corresponding model output participation coefficients based on each noise reduction control region; The fusion relationship between the basic denoising results and the normalized image sequence is established based on the participation coefficients output by the model. Based on the fusion relationship, the fusion results corresponding to each noise reduction control region are generated.

9. A low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 1, characterized in that, The boundary consistency correction module performs boundary consistency correction based on the denoised and reconstructed image, specifically including: Boundary transition weights are generated based on the regional transition relationship between adjacent noise reduction control regions. The fusion ratio at the boundary of the region is adjusted according to the boundary transition weight; Generate boundary consistency correction results based on the corrected fusion ratio.

10. A low-dose CT image deep learning adaptive noise reduction and reconstruction system according to claim 5, characterized in that, The scan stability constraint generation module generates a noise reduction intensity control map based on the scan stability constraint relationship, specifically including: Establish a scan stability constraint mapping relationship based on the aforementioned scan stability constraint relationship; The noise reduction intensity corresponding to different positions is determined based on the scan stability constraint mapping relationship; A noise reduction intensity control chart is generated based on the noise reduction intensity corresponding to different locations.

Images