CN-122023175-A - Low-dose CT image denoising method integrating gradient extraction and mixing loss

CN122023175ACN 122023175 ACN122023175 ACN 122023175ACN-122023175-A

Abstract

The invention provides a low-dose CT image denoising method integrating gradient extraction and mixing loss. The method comprises the steps of S1, obtaining a low-dose CT image as input, carrying out numerical truncation and normalization pretreatment on input data to construct a denoising model based on a U-Net backbone network, S2, parallelly constructing a gradient extraction module at the output end of the U-Net backbone network, extracting high-frequency edge characteristics of the image in the horizontal and vertical directions by using a Sobel operator, mapping the image from an intensity domain to a gradient amplitude domain, S3, constructing a mixed loss function comprising pixel consistency loss, structural similarity loss and gradient perception loss, S4, calculating the difference between a predicted image and a gold standard image by using the mixed loss function, and optimizing network parameters through counter propagation to obtain the denoised CT image. The invention can effectively relieve the competition conflict between the mean square error and the structural index, and eliminate the edge artifact of the high-density bone region while suppressing noise.

Inventors

WU TINGTING
SONG PEIXUAN
ZHAO WUFAN
ZENG TIEYONG

Assignees

南京邮电大学

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. The low-dose CT image denoising method integrating gradient extraction and mixing loss is characterized by comprising the following steps of: s1, acquiring a low-dose CT image as input, carrying out numerical truncation and normalization pretreatment on input data, and constructing a denoising model based on a U-Net backbone network; S2, constructing a gradient extraction module at the output end of the U-Net backbone network in parallel, extracting high-frequency edge characteristics of the image in the horizontal and vertical directions by utilizing a Sobel operator, and mapping the image from an intensity domain to a gradient amplitude domain; S3, constructing a mixed loss function comprising pixel consistency loss, structural similarity loss and gradient sensing loss; and S4, calculating the difference between the predicted image and the gold standard image by using the mixed loss function, and obtaining the denoised CT image by optimizing network parameters through back propagation.
2. The method for denoising a low dose CT image with fusion gradient extraction and mixing loss according to claim 1, wherein the data preprocessing in step S1 comprises the following specific steps: Step1, setting the range of the clinical abdomen window to be [ -1000,1000] HU, and carrying out numerical truncation on an original CT image; step2, linearly mapping the cut data to the [0,1] interval to obtain floating point number tensor input of the network.
3. The method for denoising the low-dose CT image with fusion gradient extraction and mixing loss according to claim 1, wherein the U-Net backbone network in the step S1 is composed of a left multi-scale feature encoding module, a right image reconstruction and recovery module and a terminal mixing constraint module, and is integrally in a U-shaped symmetrical structure.
4. The method for denoising a low-dose CT image with fusion gradient extraction and mixing loss according to claim 3, wherein the multi-scale feature encoding module comprises stacked double convolution units and a maximum pooling layer, and the implementation process comprises the following steps: Step1, after an input image enters a coding block, carrying out two continuous convolution operations, and sequentially connecting a batch normalization layer and a linear rectification unit activation function after each convolution layer; Step2, after each coding block, executing a maximum pooling operation with the kernel size of 2 multiplied by 2 and the Step length of 2, and halving the spatial resolution of the feature map and doubling the channel number.
5. A fusion gradient extraction and mixing loss low-dose CT image denoising method according to claim 3, wherein the encoding module comprises four downsampling processes, the spatial dimension of the feature map is gradually compressed from the initial input dimension to 32 x 32 of the bottleneck layer, and the number of feature channels is gradually increased from 64 to 1024.
6. The low-dose CT image denoising method with fusion gradient extraction and mixing loss according to claim 3, wherein the image reconstruction and recovery module and the multi-scale feature coding module are symmetrically distributed, and the method comprises the following specific implementation processes: Step1, each decoding stage executes up-sampling operation by means of transposed convolution or bilinear interpolation and convolution, and the spatial resolution of the feature map is increased by two times and the number of channels is halved; step2, after the up-sampling is completed, splicing the same-resolution shallow feature map output by the corresponding coding module with the current up-sampling feature map in the channel dimension through a jump connection mechanism; Step3, the spliced feature images enter a double convolution unit, and the up-sampling-splicing-convolution process is repeated four times through two continuous convolutions, batch normalization and ReLU activation treatment until the feature images are restored to the original resolution; step4, mapping the multi-channel characteristics back to a single channel through a convolution layer with the core size of 1×1, and outputting a predicted image.
7. The method for denoising a low-dose CT image with fusion gradient extraction and mixing loss according to claim 1, wherein the gradient extraction module in step S2 performs spatial filtering by adopting a Sobel convolution check feature map with fixed weight, and the Sobel convolution kernel comprises a horizontal differential operator And vertical direction differential operator The kernel function is defined as follows: 。
8. The low-dose CT image denoising method with fusion gradient extraction and mixing loss according to claim 7, wherein the calculation formula of the output gradient edge map M of the gradient extraction module is: ; Wherein, the The input image is represented by a representation of the input image, Representing a convolution operation.
9. The method according to claim 1, wherein the step S3 is characterized in that the mixing loss function is a low-dose CT image denoising method The definition is as follows: ; Wherein, the In order to be a mean square error loss, In order to achieve a loss of structural similarity, Is a gradient perceived loss; And Is a balance weight coefficient.
10. The method for denoising a low dose CT image with fusion gradient extraction and mixing loss according to claim 9, wherein the weight coefficients are set as follows , 。

Description

Low-dose CT image denoising method integrating gradient extraction and mixing loss Technical Field The invention relates to the technical field of medical image processing and computer-aided diagnosis, in particular to a low-dose CT image denoising method integrating gradient extraction and mixing loss. Background Along with the rapid development of medical image technology and the continuous improvement of clinical diagnosis precision requirements, the large-scale landing of key diagnosis and treatment links such as disease screening, focus positioning, curative effect evaluation, operation planning and the like, strict requirements are put on the imaging quality, radiation safety and clinical suitability of CT images. As a core support technology for clinical disease diagnosis, computer Tomography (CT) is not only a key tool for visualizing an anatomical structure, but also a basic premise for realizing early lesion detection and accurate diagnosis by virtue of high spatial resolution and non-overlapping tomography advantages. The principle of 'reasonable minimum dose (ALARA)' is followed, the radiation dose is reduced as much as possible on the premise of meeting the diagnosis requirement, the low-dose CT (LDCT) technology has become the common imaging scheme in clinic because of the radiation safety and the diagnosis requirement, and the performance of the low-dose CT (LDCT) technology directly determines the accuracy of clinical diagnosis and the radiation safety of patients. However, current low-dose CT image denoising techniques face significant technical challenges in clinical applications. On one hand, the direct reduction of the current of an X-ray tube can lead to the sharp reduction of photon quantity in projection data, and according to the Poisson statistical distribution rule, the reconstructed LDCT image can generate serious quantum speckle noise and streak artifacts, the noise not only reduces the signal-to-noise ratio of the image, but also covers the anatomical details of low-contrast soft tissues such as micro tumor, blood vessel and the like, clinical missed diagnosis or misdiagnosis is extremely easy to be caused, which is the core contradiction of low-dose imaging, and meanwhile, the clinical diagnosis has extremely high requirements on high-frequency anatomical details (such as micro blood vessel textures and bone edges) of the image, and a denoising algorithm needs to realize accurate balance between noise suppression and reserved details, so that strict requirements are put forward on the fineness of a technical scheme. On the other hand, the existing mainstream denoising technology has obvious defects that although the traditional iterative reconstruction algorithm (such as a regularization method based on Total Variation (TV) and a BM3D algorithm based on non-local mean (NLM) and provided by Sidky and Dabov) utilize image self-similarity prior to achieve certain effect in denoising, the problems of high computational complexity, difficult parameter adjustment, easiness in generating a step effect and the like generally exist, image edge blurring is caused, reconstruction time is long, clinical real-time requirements are difficult to meet, and the image post-processing algorithm (such as a U-Net structure and a residual encoder-decoder structure (RED-CNN) provided by Ronneberger and Chen and the like) based on deep learning is superior to the traditional algorithm in performance, but has key limitations, and is difficult to consider the dual targets of noise suppression and detail preservation. In recent years, a data driving method represented by a Convolutional Neural Network (CNN) has been significantly advanced in the field of medical image processing, and has become a mainstream technology for LDCT image denoising. However, in practical application, most of the existing network models tend to use Mean Square Error (MSE) or L2 norm as a unique loss measure when training, and from the statistical perspective, minimizing MSE is equivalent to calculating the posterior mean value of all possible solutions, which forces the network to output a "statistical average" result, and such optimization strategies with pixel-level errors as a guide can obtain a higher peak signal-to-noise ratio (PSNR) index, but cause serious "smoothing effect" and "spectrum deviation" phenomena. As indicated by Blau et al in the theory of "perception-distortion trade-off", if the image restoration algorithm excessively pursues low mean square error, the perceived reality of the image is damaged, and the algorithm is particularly characterized in that when the algorithm removes noise, tiny focus points, blood vessel ends and high-frequency edges of bones are often erased together as noise, so that the denoised image presents a 'oil painting feel' with missing details, and the diagnosis confidence of doctors is directly influenced. In addition, the prior art has the common problems of noise suppression and d