A method, apparatus, and storage medium for denoising and enhancing low-dose CT images
By employing a deep learning method that adaptively adjusts denoising intensity and robust noise estimation, the problem of insufficient fixed denoising intensity or excessive smoothing in low-dose CT images is solved, achieving efficient denoising enhancement in the image domain, suitable for clinical applications with different equipment and noise levels.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE SECOND XIANGYA HOSPITAL OF CENT SOUTH UNIV
- Filing Date
- 2026-06-01
- Publication Date
- 2026-06-26
AI Technical Summary
Existing deep learning-based low-dose CT image denoising methods often suffer from insufficient denoising intensity when noise levels are high, and excessive smoothing when noise levels are low. Furthermore, many methods require the original projection domain data, making them difficult to apply in clinical workflows.
By calculating the noise level parameters of CT images, the denoising intensity is adaptively adjusted. Robust noise estimation is performed in the image domain using high-pass residual and median absolute deviation methods. A deep learning denoising network is used to generate denoised images, which are then weighted and fused to output an enhanced image.
It achieves adaptive denoising and enhancement, automatically adjusts noise levels, reduces manual operation, improves image quality, adapts to different devices and noise levels, and is easy to deploy in existing CT post-processing pipelines.
Smart Images
Figure CN122289095A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing and deep learning inference technology, specifically relating to a method, device and storage medium for denoising and enhancing low-dose CT images. Background Technology
[0002] Low-dose CT is an imaging technique that reduces radiation dose; typically, the radiation dose per examination is 1 / 5 to 1 / 10 of that of conventional-dose CT. Ultra-low-dose CT is an imaging technique that is further optimized from low-dose CT, with a radiation dose per examination that is less than 1 / 10 of that of conventional-dose CT.
[0003] Under low-dose or ultra-low-dose scanning conditions, quantum noise and reconstruction noise in CT images typically increase significantly. Although algorithms such as iterative reconstruction can reduce some noise, problems such as blurred details and insufficient edge sharpness may still occur, affecting the visibility of small target structures and the stability of subsequent measurements.
[0004] Existing deep learning-based denoising post-processing methods often employ a fixed denoising intensity in engineering practice. When image noise is high, fixed-intensity denoising may be insufficient; when image noise is low, fixed-intensity denoising may cause over-smoothing, resulting in loss of edge details or texture distortion. Meanwhile, many deep learning reconstruction schemes require raw projection domain data, but in common clinical workflows, only image domain DICOM images are often available, making it difficult to widely implement these schemes.
[0005] Therefore, it is necessary to design a method, device, and storage medium for denoising and enhancing low-dose CT images. Summary of the Invention
[0006] The purpose of this invention is to provide a method, apparatus, and storage medium for denoising and enhancing low-dose CT images, in order to solve the problems mentioned in the background art. The fixed denoising intensity used in the existing deep learning-based denoising post-processing methods may be insufficient when the image noise is large; when the image noise is small, fixed intensity denoising may cause over-smoothing, resulting in loss of edge details or texture distortion. Furthermore, many deep learning reconstruction schemes require the original projection domain data, but in common clinical workflows, only image domain DICOM images are often available, making it difficult for related schemes to be widely implemented.
[0007] To achieve the above objectives, the present invention provides a method for denoising and enhancing low-dose CT images, comprising the following steps:
[0008] S1. Obtain the CT image I to be processed. The CT image I to be processed is a DICOM image reconstructed from low-dose or ultra-low-dose CT.
[0009] S2. Calculate the noise level parameter N of the CT image I to be processed; specifically as follows:
[0010] S2.1 Perform threshold segmentation, connected component analysis and morphological processing on the CT image I to be processed to obtain a CT image with at least two types of uniform reference region masks for noise estimation.
[0011] S2.2. Robust noise estimation within the mask is achieved using high-pass residuals and median absolute deviation.
[0012] S2.3, Perform robust fusion of noise level parameter N;
[0013] S3. A convolutional neural network is trained based on paired data of conventional dose CT images and low dose CT images to obtain a deep learning denoising network; a denoised image I_dn is generated based on the CT image to be processed I using the deep learning denoising network.
[0014] S4. Calculate the denoising intensity coefficient s based on the noise level parameter N, and perform weighted fusion of the CT image I to be processed and the denoised image I_dn to obtain the output enhanced image.
[0015] In one specific implementation, in step S2.1, the obtained uniform reference region mask includes at least an external air mask M_air_out.
[0016] In one specific implementation, the process of obtaining the external air mask M_air_out is as follows:
[0017] First, select regions whose pixel count is less than -950HU as air candidates;
[0018] Next, connected regions connected to the image boundary are retained in the air candidates, and the air cavity region inside the patient is excluded to obtain the external air region;
[0019] Next, open / close operations are performed on the external air region to remove small connected components; erosion operations are performed to remove boundary artifacts, resulting in the external air mask M_air_out.
[0020] In one specific implementation, the uniform reference region mask obtained in step S2.1 further includes an in vivo air mask M_air_in and a subcutaneous fat mask M_fat.
[0021] In one specific implementation, the process of obtaining the in-body air mask M_air_in is as follows:
[0022] First, the human body contour region is obtained by selecting using thresholds and connected components;
[0023] Then, within the human body contour area, select regions that satisfy the condition of having a pixel value less than -900HU as candidates for internal airways.
[0024] Candidate airways in the body are selected based on the connected region area threshold and position constraints to obtain the in vivo air mask M_air_in.
[0025] The process of obtaining the subcutaneous fat mask M_fat is as follows:
[0026] First, select regions that satisfy the condition of having a pixel value greater than or equal to -150HU and less than or equal to -50HU as fat candidates;
[0027] Next, only retain the area close to the skin contour to avoid including blood vessels;
[0028] After morphological cleaning, the subcutaneous fat mask M_fat was obtained.
[0029] In one specific implementation, step S2.2, the specific steps for robust noise estimation within the mask are as follows:
[0030] First, perform in-mask Gaussian smoothing on each mask of the CT image with at least two types of uniform reference region masks to obtain a smoothed base map I_lp;
[0031] Next, calculate the high-pass residual;
[0032] Then, the standard deviation of each mask class is estimated using the high-pass residual robustness estimation.
[0033] For each type of mask, the validity is determined. If the number of valid pixels of a mask is less than a preset threshold, the mask is determined not to participate in the subsequent robust fusion of the noise level parameter N.
[0034] In step S2.3, when performing robust fusion of the noise level parameter N, if at least two types of masks are effective, the median of the standard deviation estimates of each effective mask is taken as the noise level parameter N; if only one type of mask is effective, the standard deviation estimate of that mask is taken as the noise level parameter N.
[0035] In one specific implementation, in step S4, the noise reduction intensity coefficient s is calculated based on the noise level parameter N: ;
[0036] In the formula, N0 is the low noise threshold, N1 is the high noise threshold, and clip indicates that the calculation result is restricted to [0,1].
[0037] The CT image I to be processed is weighted and fused with the denoised image I_dn to obtain the output enhanced image I_out: .
[0038] In one specific implementation, the CT image denoising and enhancement method further includes:
[0039] S5. Convert the output enhanced image to DICOM format, and record the noise level parameter N and the denoising intensity coefficient s.
[0040] The present invention also provides a device for denoising and enhancing low-dose CT images, the device comprising an image acquisition module, a mask generation module, a noise estimation module, a denoising network inference module, a fusion module and an output module;
[0041] The image acquisition module is used to acquire the CT image to be processed, which is a DICOM image reconstructed from low-dose or ultra-low-dose CT.
[0042] The mask generation module is used to perform threshold segmentation, connected component analysis and morphological processing on the CT image I to be processed, so as to obtain a CT image with at least two types of uniform reference region masks.
[0043] The noise estimation module is used to perform robust noise estimation within the mask using high-pass residuals and median absolute deviation methods, and to perform robust fusion of the noise level parameter N.
[0044] The denoising network inference module is used to generate a denoised image I_dn based on the CT image to be processed I using a deep learning denoising network.
[0045] The fusion module is used to calculate the denoising intensity coefficient s based on the noise level parameter N, and to perform weighted fusion of the CT image to be processed I and the denoised image I_dn to obtain the output enhanced image;
[0046] The output module is used to convert the output enhanced image into DICOM format, while recording the noise level parameter N and the denoising intensity coefficient s.
[0047] The present invention also provides a readable storage medium for denoising and enhancing low-dose CT images, wherein the readable storage medium contains a computer program for implementing the steps in the denoising and enhancing method for low-dose CT images as described above.
[0048] Compared with the prior art, the present invention has the following beneficial effects:
[0049] (1) Adaptability: By automatically calculating the noise level parameter N and mapping N to the noise reduction strength coefficient s, the adaptive output is achieved so that the greater the noise, the stronger the noise reduction and the less noise, the more details are preserved.
[0050] (2) No need for manual selection of ROI (Region of Interest): Uniform reference regions are automatically obtained through threshold segmentation, connected component constraints and morphological cleaning, reducing the differences and instability caused by manual operation.
[0051] (3) Robustness: High-pass residuals and median absolute deviation (MAD) are used to estimate noise, and the impact of edge structure and single region failure on estimation is reduced by fusing multiple reference regions through median.
[0052] (4) Easy to deploy: It can be implemented by relying only on DICOM images in the image domain, which is easy to embed into existing CT post-processing links and maintains a consistent output trend under different devices and different noise levels.
[0053] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The present invention will now be described in further detail. Attached Figure Description
[0054] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:
[0055] Figure 1 This is a flowchart of one embodiment of the present invention. Detailed Implementation
[0056] The embodiments of the present invention will be described in detail below. The specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.
[0057] See Figure 1 The present invention provides a method for denoising and enhancing low-dose CT images, comprising the following steps:
[0058] S1. Obtain the CT image I to be processed. The CT image I to be processed is a DICOM image reconstructed from low-dose or ultra-low-dose CT. This invention operates in the image domain and does not require the original data in the projection domain.
[0059] S2. Calculate the noise level parameter N of the CT image I to be processed; the computer finds a reference area and calculates the noise without the doctor selecting a region of interest (ROI).
[0060] Specifically as follows:
[0061] S2.1 Perform threshold segmentation, connected component analysis and morphological processing on the CT image I to be processed to obtain a CT image with at least two types of uniform reference region masks for noise estimation.
[0062] S2.2. Robust noise estimation within the mask is achieved using high-pass residuals and median absolute deviation.
[0063] S2.3, Perform robust fusion of noise level parameter N;
[0064] S3. A convolutional neural network is trained based on paired data of conventional dose CT images and low dose CT images to obtain a deep learning denoising network; a denoised image I_dn is generated based on the CT image to be processed I using the deep learning denoising network.
[0065] S4. Calculate the denoising intensity coefficient s based on the noise level parameter N, and perform weighted fusion of the CT image I to be processed and the denoised image I_dn to obtain the output enhanced image.
[0066] In step S2.1, the obtained uniform reference region mask includes at least an external air mask M_air_out; preferably, it also includes an internal air mask M_air_in and a subcutaneous fat mask M_fat.
[0067] The process of obtaining the external air mask M_air_out is as follows:
[0068] First, select regions whose pixel count is less than -950HU as air candidates;
[0069] Next, connected regions connected to the image boundary are retained in the air candidates, and the air cavity region inside the patient is excluded to obtain the external air region;
[0070] Next, open / close operations are performed on the external air region to remove small connected components; erosion operations are performed to remove boundary artifacts, resulting in the external air mask M_air_out.
[0071] The process of obtaining the in-body air mask M_air_in is as follows:
[0072] First, the human body contour region is obtained by selecting using thresholds and connected components;
[0073] Then, within the human body contour area, select regions that satisfy the condition of having a pixel value less than -900HU as candidates for internal airways.
[0074] Candidate airways in the body are selected based on the connected region area threshold and position constraints to obtain the in vivo air mask M_air_in.
[0075] The process of obtaining the subcutaneous fat mask M_fat is as follows:
[0076] First, select regions that satisfy the condition of having a pixel value greater than or equal to -150HU and less than or equal to -50HU as fat candidates;
[0077] Next, only retain the area close to the skin contour to avoid including blood vessels;
[0078] After morphological cleaning, the subcutaneous fat mask M_fat was obtained.
[0079] In step S2.2, the specific steps for robust noise estimation within the mask are as follows;
[0080] First, perform in-mask Gaussian smoothing on each mask of the CT image with at least two types of uniform reference region masks to obtain a smoothed base map I_lp;
[0081] Next, calculate the high-pass residual;
[0082] Then, the standard deviation of each mask class is estimated using the high-pass residual robustness estimation.
[0083] For each type of mask, the validity is determined. If the number of valid pixels of a mask is less than a preset threshold, the mask is determined not to participate in the subsequent robust fusion of the noise level parameter N.
[0084] If you directly calculate pixel value fluctuations, you will be affected by edge structures such as blood vessels and lesions, resulting in inaccurate noise calculations.
[0085] In step S2.3, when performing robust fusion of the noise level parameter N, if at least two types of masks are effective, the median of the standard deviation estimates of each effective mask is taken as the noise level parameter N; if only one type of mask is effective, the standard deviation estimate of that mask is taken as the noise level parameter N.
[0086] In step S4, the noise reduction intensity coefficient s is calculated based on the noise level parameter N: ;
[0087] In the formula, N0 is the low noise threshold, N1 is the high noise threshold, and clip indicates that the calculation result is restricted to [0,1].
[0088] The CT image I to be processed is weighted and fused with the denoised image I_dn to obtain the output enhanced image I_out: .
[0089] When N is large, s increases, and the output is closer to I_dn, thus automatically enhancing the denoising effect; when N is small, s decreases, and the output retains more details of I, reducing the risk of over-smoothing. High noise (s≈1): The final image is almost the denoised image, with maximum denoising effect; Low noise (s≈0): The final image is almost the original image, with only slight denoising, perfectly preserving details; Medium noise: The original image and the denoised image are mixed proportionally, balancing denoising and detail.
[0090] The CT image denoising and enhancement method further includes:
[0091] S5. Convert the output enhanced image to DICOM format, and record the noise level parameter N and the denoising intensity coefficient s, so as to facilitate doctors' verification and quality control.
[0092] The present invention also provides a device for denoising and enhancing low-dose CT images, the device comprising an image acquisition module, a mask generation module, a noise estimation module, a denoising network inference module, a fusion module and an output module;
[0093] The image acquisition module is used to acquire the CT image to be processed, which is a DICOM image reconstructed from low-dose or ultra-low-dose CT.
[0094] The mask generation module is used to perform threshold segmentation, connected component analysis and morphological processing on the CT image I to be processed, so as to obtain a CT image with at least two types of uniform reference region masks.
[0095] The noise estimation module is used to perform robust noise estimation within the mask using high-pass residuals and median absolute deviation methods, and to perform robust fusion of the noise level parameter N.
[0096] The denoising network inference module is used to generate a denoised image I_dn based on the CT image to be processed I using a deep learning denoising network.
[0097] The fusion module is used to calculate the denoising intensity coefficient s based on the noise level parameter N, and to perform weighted fusion of the CT image to be processed I and the denoised image I_dn to obtain the output enhanced image;
[0098] The output module is used to convert the output enhanced image into DICOM format, while recording the noise level parameter N and the denoising intensity coefficient s.
[0099] The present invention also provides a readable storage medium for denoising and enhancing low-dose CT images, wherein the readable storage medium contains a computer program for implementing the steps in the denoising and enhancing method for low-dose CT images as described above.
[0100] Example 1
[0101] An implementation of single-mask noise estimation and adaptive fusion based on external air mask.
[0102] In this embodiment, axial DICOM image sequences from low-dose chest CT scans are used as input images. The input images are HU value images after conventional reconstruction, with a slice thickness of 1.25 mm and a matrix size of 512×512. The computing device is an image processing server deployed on the CT workstation, with a multi-core CPU processor, and a GPU to perform deep learning network inference.
[0103] In this embodiment, the following processing is performed on each input slice.
[0104] First, the image to be processed, I, is acquired and preprocessed. Preferably, the pixel values are limited to a preset HU window width, which in this embodiment is limited to [-1024HU, 1024HU], to reduce the impact of outliers on subsequent noise estimation. Then, the original HU values are retained for subsequent fusion output.
[0105] Subsequently, an external air mask M_air_out is automatically generated. Specifically, the image is first binarized using a threshold T_out = -950HU, and pixels satisfying I < -950HU are selected as candidate air regions to obtain a binary image. Then, connected components connected to the image boundary are retained in this binary image to exclude air cavity regions within the patient's body. Next, an opening operation with a radius of 2 pixels and a closing operation with a radius of 3 pixels are performed to remove scattered artifacts and small holes. Finally, an erosion operation of 1 pixel size is performed to reduce the interference of edge artifacts, scattering artifacts, or reconstruction ring artifacts near the patient's surface on noise estimation, thus obtaining the external air mask M_air_out. Preferably, the mask is considered valid when the number of effective pixels in M_air_out is not less than 2000 pixels.
[0106] After obtaining M_air_out, the input image I is low-pass smoothed to obtain the smoothed baseline image I_lp. A Gaussian filter with a standard deviation of 1.0 pixel and a kernel size of 5×5 is preferably used for smoothing. Then, the high-frequency residual image R = I - I_lp is calculated. Further, residual pixel values are extracted only within the M_air_out coverage area, and the noise estimate is calculated using the following formula: .
[0107] In this embodiment, since only a single reference region is used, σ_out is directly used as the global noise level parameter N.
[0108] Further, the denoising intensity coefficient s is adaptively determined based on N. Preferably, the low noise threshold N0 is set to 18 HU, the high noise threshold N1 is set to 55 HU, and calculated according to the following formula: .
[0109] When an input slice is calculated to have N=42HU, we get s=(42-18) / (55-18)=24 / 37≈0.65. This indicates that the noise level of the current slice is at a medium-to-high level, and the output image should be more inclined towards denoising while retaining some of the original texture.
[0110] In this embodiment, the deep learning denoising network f(·) is preferably a two-dimensional convolutional neural network. The network input is a single axial CT image, and the output is the corresponding denoised image I_dn=f(I). Preferably, the network includes a first-layer convolutional feature extraction module, multiple residual dense blocks, and an end-level reconstruction convolutional layer. Network training is completed based on paired data of conventional dose CT images and low-dose CT images, but only image domain DICOM images need to be input during the inference stage.
[0111] Finally, the original image I and the denoised image I_dn are weighted and fused to obtain the output image I_out: .
[0112] When s≈0.65, the proportion of denoised image in the output image is about 65%, and the proportion of original image is about 35%. The resulting output image is re-encoded into DICOM format, and N and s are written together as private labels or quality control parameters.
[0113] By relying solely on the external air region to obtain stable noise estimation, the ROI selection can be avoided, making it suitable for rapid deployment in existing chest CT post-processing workflows. At the same time, due to the adoption of an N-based adaptive fusion strategy, compared with the fixed-intensity denoising method, it can achieve more significant noise reduction on noisy images and reduce over-smoothing on noisy images, thereby better preserving lung texture, vascular boundaries and small nodule edge details.
[0114] Example 2
[0115] A robust noise fusion implementation based on three types of masks: external air, internal air, and subcutaneous fat.
[0116] In this embodiment, a sequence of ultra-low-dose chest CT images is used as input, preferably for lung nodule screening. Related studies have shown that image-domain deep learning post-processing can significantly reduce noise in ultra-low-dose CT images and improve small nodule detection performance; however, noise distribution may differ across body types and anatomical levels. This embodiment improves the stability of noise estimation by constructing multiple uniform reference regions and performing robust fusion.
[0117] In this embodiment, an external air mask M_air_out, an internal air mask M_air_in, and a subcutaneous fat mask M_fat are constructed for each input image I.
[0118] The external air mask M_air_out is generated using the method described in Example 1.
[0119] For the body air mask M_air_in, the image is first coarsely segmented using a threshold T_body = -300 HU, and the largest connected component is selected as the human body contour region M_body. Then, pixels within M_body that satisfy I < -900 HU are further selected as candidate body air regions. To maximize the selection of trachea or main bronchus regions, connected component filtering is performed on the candidate regions, preferentially retaining targets with an area between 100 and 3000 pixels and whose connected component centroids are located within the image center region. The image center region is defined as a circular region with the image center as the center and a radius 0.25 times the image width. After filtering, the body air mask M_air_in is obtained.
[0120] For the subcutaneous fat mask M_fat, pixels satisfying -150HU≤I≤-50HU are first selected within the human body contour region M_body as fat candidate regions. Then, the boundary of M_body is eroded inward by 5 pixels and inward by 25 pixels, forming a ring region M_ring with a width of approximately 20 pixels. Only pixels located simultaneously within both the fat candidate region and the ring region are retained as subcutaneous fat candidates. A final opening operation is performed to remove small abnormal areas caused by blood vessels, skin folds, or local artifacts, ultimately yielding the subcutaneous fat mask M_fat.
[0121] In the noise estimation stage, this embodiment still employs the robust estimation method of high-pass residual plus MAD. Preferably, image I is first processed with a Gaussian filter with a standard deviation of 1.2 pixels and a kernel size of 7×7 to obtain I_lp, and then the residual image R=I-I_lp is calculated. Next, calculations are performed in the corresponding regions of M_air_out, M_air_in, and M_fat respectively: ; ; .
[0122] To ensure statistical stability, the preferred rule is: if the number of effective pixels in a mask is less than 500, the mask is considered invalid; if the number of effective pixels is not less than 500, the mask is considered valid. If at least two types of masks are valid, the median of all valid noise estimates is used as the global noise level parameter N, i.e.: .
[0123] In a given input slice, if σ_out = 46.2HU, σ_in = 43.5HU, and σ_fat = 40.8HU are calculated, then N = 43.5HU. Since median fusion is insensitive to outliers, even if one type of mask is affected by local structures or artifacts, its perturbation to the final N is relatively small.
[0124] In this embodiment, to further enhance the adjustment flexibility in different noise ranges, the denoising intensity coefficient s is determined using a piecewise linear function. Preferably, N0 = 15 HU, N1 = 30 HU, and N2 = 55 HU are set. When N ≤ N0, s = 0.15; when N0 < N < N1, s linearly increases from 0.15 to 0.45; when N1 ≤ N < N2, s linearly increases from 0.45 to 0.90; when N ≥ N2, s = 0.90. This avoids the situation where s is completely 0 under low-noise conditions and the loss of the minor texture optimization effect, and also avoids the risk of structural distortion caused by over-relying on the deep learning output under extremely high-noise conditions.
[0125] In this embodiment, the deep learning denoising network is a residual dense network. The network input is a single CT image, and the network output is the denoised image I_dn. Preferably, the network includes 3 convolutional feature extraction layers, 8 residual dense blocks, and 2 reconstruction convolutional layers, and long skip connections are set to retain the original structural features. Subsequently, through: ;
[0126] The enhanced output image is obtained.
[0127] Furthermore, this embodiment also performs smoothing constraints at the image sequence level. Specifically, for the N value calculated from three adjacent slices, one-dimensional median filtering is used for in-sequence smoothing to obtain the smoothed noise level parameter , and the corresponding s is calculated based on . This can reduce the sudden change in the fusion intensity caused by local anatomical differences between slices and improve the visual continuity during sequence browsing.
[0128] Compared with only using a single external air region for noise estimation, multi-mask robust fusion can simultaneously utilize the information of relatively uniform regions such as external air, airway air, and subcutaneous fat. When there are patient position changes, scan field changes, boundary artifacts, or local missing parts, it can still relatively stably characterize the image noise level; at the same time, the piecewise linear fusion function can achieve a smoother and more clinically visually habitual denoising intensity adjustment between low-noise, medium-noise, and high-noise images, thereby improving the image quality consistency and being beneficial to the retention of small nodule edges, subsolid nodule contours, and fine blood vessel textures.
[0129] Embodiment 3
[0130] Embodiment of sequence-level adaptive denoising enhancement with confidence evaluation and fallback strategy.
[0131] To address issues such as small external air areas, unclear internal airways, and insufficient influence of anatomical location on fat areas in some slices, a noise-based confidence level C is introduced. Combined with a sequence-level fallback strategy, a more stable enhancement output is generated. This approach is particularly suitable for top and bottom slices of chest CT sequences, as well as complex scenarios such as patients with thin body types or scan truncation.
[0132] In this embodiment, a complete CT image sequence is used as the processing object. Preferably, for each slice, M_air_out, M_air_in, and M_fat are generated according to the method in Example 2, and the corresponding noise estimates σ_out, σ_in, and σ_fat are calculated respectively. Then, the confidence parameter C is further calculated for each slice. The confidence parameter can comprehensively reflect the following factors: the number of effective masks, the total number of pixels in each effective mask, and the consistency between different noise estimates.
[0133] Preferably, the confidence level C is calculated as follows:
[0134] First, let the number of effective masks be n_valid, where n_valid can take the value of 0, 1, 2 or 3;
[0135] Second, let the total number of effective pixels be P_total;
[0136] Third, if n_valid≥2, then calculate the degree of variation D of each valid noise estimate: ;
[0137] Where σ_i represents the noise estimate corresponding to each effective mask.
[0138] Based on this, we define: ;
[0139] Wherein, C_num represents the score based on the number of effective masks, and is preferably set as follows: C_num=1.0 when n_valid=3, C_num=0.7 when n_valid=2, C_num=0.4 when n_valid=1, and C_num=0 when n_valid=0;
[0140] C_pix represents the score based on the total number of effective pixels. It is preferably 1.0 when the total number of effective pixels P_total≥4000 and 0 when P_total≤500, with linear interpolation in between.
[0141] C_cons represents the score based on multi-mask consistency, preferably 1.0 when D≤0.10, 0 when D≥0.50, and linear interpolation in between;
[0142] w1, w2, and w3 are weighting coefficients, preferably 0.4, 0.3, and 0.3.
[0143] If a slice has three effective masks with a total of 5200 effective pixels, and the noise estimates for the three masks are 39.8 HU, 41.2 HU, and 43.0 HU respectively, then D = (43.0 - 39.8) / 41.2 ≈ 0.078, which yields a high confidence level, C ≈ 0.95. Conversely, if only one mask is effective and the number of pixels is small, then C may be less than 0.5.
[0144] In this embodiment, when C ≥ 0.75, the current slice noise estimation result N_cur is directly used to calculate the current slice fusion intensity; when 0.4 ≤ C < 0.75, the weighted average of the current slice noise estimation result N_cur and the noise estimation results of adjacent slices is used as the final noise level parameter N_final, preferably: ;
[0145] When C < 0.4, the noise estimation reliability of the current slice is considered insufficient, and a sequence-level fallback strategy is adopted. Preferably, the median of the noise level parameters of the five consecutive high-confidence slices within the same sequence is taken as the N_final of that slice; if the number of high-confidence slices in the neighborhood is still less than three, then the median of the noise level parameters of all high-confidence slices in the entire sequence is used as N_final.
[0146] Furthermore, in this embodiment, the denoising intensity coefficient s can also be adjusted by both N_final and C. Preferably, the basic fusion coefficient s0 is first calculated based on N_final as follows: .
[0147] Then, the confidence level is used to suppress and correct s0, in order to avoid outputting overly aggressive denoising results under low confidence conditions. Preferably, the correction formula is: .
[0148] Therefore, when C is low, the actual fusion coefficient s will be slightly lower than s0, thus more conservatively preserving the original image information; when C is close to 1, s is close to s0, which does not affect the normal adaptive denoising effect.
[0149] For example, when a slice N_final=48HU, then s0=(48-20) / (60-20)=0.70. If the slice C=0.90, then s≈0.7×0.70+0.3×0.70×0.90=0.679; if the same N_final is used but C=0.30, then s≈0.553. It can be seen that when the confidence level is low, the system automatically reduces its reliance on the denoised image, lowering the risk of oversmoothing due to local misjudgments.
[0150] In this embodiment, the deep learning network adopts a two-dimensional or 2.5-dimensional structure. A 2.5-dimensional input method is preferred, i.e., a three-channel input consisting of the current slice and one slice before and after it, to enhance the representation ability of the continuous structure between layers. The network outputs the denoised image I_dn corresponding to the current slice. Finally, the images are fused according to the following formula: .
[0151] Simultaneously, N_final, C, and s corresponding to each slice are recorded for quality control display. Preferably, when C < 0.4, "low confidence estimate" is marked in the system log for subsequent algorithm tuning or manual review.
[0152] By introducing noise estimation confidence and sequence-level fallback mechanisms, stable enhancement results can still be output even when the reference area is incomplete, local anatomical structures are missing, or image boundaries are abnormal, reducing the fluctuation of fusion intensity between slices and local over-denoising. Especially in slices of the top of the thoracic cavity, the top of the diaphragm, or irregular positions, it can better ensure image continuity and structural authenticity, improving the stability of engineering implementation and clinical usability.
[0153] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions and substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A method for denoising and enhancing low-dose CT images, characterized in that, Includes the following steps: S1. Obtain the CT image I to be processed. The CT image I to be processed is a DICOM image reconstructed from low-dose or ultra-low-dose CT. S2. Calculate the noise level parameter N of the CT image I to be processed; specifically as follows: S2.1 Perform threshold segmentation, connected component analysis and morphological processing on the CT image I to be processed to obtain a CT image with at least two types of uniform reference region masks for noise estimation. S2.
2. Robust noise estimation within the mask is achieved using high-pass residuals and median absolute deviation. S2.3, Perform robust fusion of noise level parameter N; S3. A deep learning denoising network is obtained by training a convolutional neural network based on paired data of conventional dose CT images and low dose CT images. Based on the CT image to be processed, a deep learning denoising network is used to generate a denoised image I_dn; S4. Calculate the denoising intensity coefficient s based on the noise level parameter N, and perform weighted fusion of the CT image I to be processed and the denoised image I_dn to obtain the output enhanced image.
2. The method for denoising and enhancing low-dose CT images according to claim 1, characterized in that, In step S2.1, the obtained uniform reference region mask includes at least the external air mask M_air_out.
3. The method for denoising and enhancing low-dose CT images according to claim 2, characterized in that, The process of obtaining the external air mask M_air_out is as follows: First, select regions whose pixel count is less than -950HU as air candidates; Next, connected regions connected to the image boundary are retained in the air candidates, and the air cavity region inside the patient is excluded to obtain the external air region; Next, open / close operations are performed on the external air region to remove small connected components; erosion operations are performed to remove boundary artifacts, resulting in the external air mask M_air_out.
4. The method for denoising and enhancing low-dose CT images according to claim 2, characterized in that, In step S2.1, the obtained uniform reference region mask also includes an in vivo air mask M_air_in and a subcutaneous fat mask M_fat.
5. The method for denoising and enhancing low-dose CT images according to claim 4, characterized in that, The process of obtaining the in-body air mask M_air_in is as follows: First, the human body contour region is obtained by selecting using thresholds and connected components; Then, within the human body contour area, select regions that satisfy the condition of having a pixel value less than -900HU as candidates for internal airways. Candidate airways in the body are screened based on the area threshold of connected components and position constraints to obtain the in vivo air mask M_air_in. The process of obtaining the subcutaneous fat mask M_fat is as follows: First, select regions that satisfy the condition of having a pixel value greater than or equal to -150HU and less than or equal to -50HU as fat candidates; Next, only retain the area close to the skin contour to avoid including blood vessels; After morphological cleaning, the subcutaneous fat mask M_fat was obtained.
6. The method for denoising and enhancing low-dose CT images according to claim 1, characterized in that, In step S2.2, the specific steps for robust noise estimation within the mask are as follows; First, perform in-mask Gaussian smoothing on each mask of the CT image with at least two types of uniform reference region masks to obtain a smoothed base map I_lp; Next, calculate the high-pass residual; Then, the standard deviation of each mask class is estimated using the high-pass residual robustness estimation. For each type of mask, the validity is determined. If the number of valid pixels of a mask is less than a preset threshold, the mask is determined not to participate in the subsequent robust fusion of the noise level parameter N. In step S2.3, when performing robust fusion of the noise level parameter N, if at least two types of masks are effective, the median of the standard deviation estimates of each effective mask is taken as the noise level parameter N; if only one type of mask is effective, the standard deviation estimate of that mask is taken as the noise level parameter N.
7. The method for denoising and enhancing low-dose CT images according to claim 1, characterized in that, In step S4, the noise reduction intensity coefficient s is calculated based on the noise level parameter N: ； In the formula, N0 is the low noise threshold, N1 is the high noise threshold, and clip indicates that the calculation result is restricted to [0,1]. The CT image I to be processed is weighted and fused with the denoised image I_dn to obtain the output enhanced image I_out: 。 8. The method for denoising and enhancing low-dose CT images according to claim 1, characterized in that, The CT image denoising and enhancement method further includes: S5. Convert the output enhanced image to DICOM format, and record the noise level parameter N and the denoising intensity coefficient s.
9. A device for denoising and enhancing low-dose CT images, characterized in that, The device for denoising and enhancing low-dose CT images includes an image acquisition module, a mask generation module, a noise estimation module, a denoising network inference module, a fusion module, and an output module. The image acquisition module is used to acquire the CT image to be processed, which is a DICOM image reconstructed from low-dose or ultra-low-dose CT. The mask generation module is used to perform threshold segmentation, connected component analysis and morphological processing on the CT image I to be processed, so as to obtain a CT image with at least two types of uniform reference region masks. The noise estimation module is used to perform robust noise estimation within the mask using high-pass residuals and median absolute deviation methods, and to perform robust fusion of the noise level parameter N. The denoising network inference module is used to generate a denoised image I_dn based on the CT image to be processed I using a deep learning denoising network. The fusion module is used to calculate the denoising intensity coefficient s based on the noise level parameter N, and to perform weighted fusion of the CT image to be processed I and the denoised image I_dn to obtain the output enhanced image; The output module is used to convert the output enhanced image into DICOM format, while recording the noise level parameter N and the denoising intensity coefficient s.
10. A readable storage medium for denoising and enhancing low-dose CT images, characterized in that, The readable storage medium contains a computer program for implementing the steps of the denoising and enhancement method for low-dose CT images as described in any one of claims 1 to 8.