CN-121981914-A - CT image automatic noise reduction processing method based on multi-scale feature fusion

CN121981914ACN 121981914 ACN121981914 ACN 121981914ACN-121981914-A

Abstract

The invention relates to the technical field of medical image processing, in particular to an automatic CT image noise reduction processing method based on multi-scale feature fusion. The method comprises the steps of carrying out frequency domain analysis on a CT image, dividing the CT image into a plurality of scale spaces, carrying out edge detection on each scale space, determining the coincident edges of adjacent scale spaces, determining the real tissue edges by calculating the gradient direction consistency and the cross-scale gradient stability of pixel points at two sides of the coincident edges, marking the other edges as noise, determining the optimal scale combination according to the noise coverage rate and the noise priority, carrying out noise reduction processing by adopting different noise reduction algorithms, and finally merging the noise-reduced scale space with the non-noise-reduced scale space to obtain the noise-reduced CT image. The invention can effectively distinguish the real tissue structure from noise, can inhibit noise and simultaneously can keep the tissue structure details to the maximum extent, thereby improving the reliability of CT image diagnosis.

Inventors

CUI XIAOJUN
WU JUNCHAO
YAN YANZHANG
HAN XUEHUA
MENG YALI
YAN HONG
DU LIMIN
WANG LONG
ZHANG GUANGFEI
Zhang Shuchi

Assignees

深圳市亿康医疗技术有限公司

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. An automatic CT image noise reduction processing method based on multi-scale feature fusion is characterized by comprising the following steps: performing frequency domain analysis on the CT image, and dividing the CT image into a plurality of scale spaces based on a frequency range; Analyzing gradient direction distribution characteristics of the overlapped edges in different scale spaces, determining real tissue edges in the scale spaces, and taking edges except the real tissue edges in each scale space as noise; and determining an optimal scale value according to the noise distribution condition of each scale space, and carrying out noise reduction treatment on the CT image based on the optimal scale value.
2. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 1, wherein the method for dividing the scale space comprises: acquiring a body part to which the CT image belongs, and determining the highest frequency component and the lowest frequency component of frequency domain analysis according to the frequency domain characteristics corresponding to the body part; Calculating a frequency step based on the highest frequency component, the lowest frequency component and the preset scale number, generating a plurality of center frequencies corresponding to the scale number based on the frequency step, and constructing a plurality of scale spaces corresponding to the center frequencies by utilizing a band-pass filter.
3. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 2, wherein the method for determining the coincident edges comprises: Sequencing all scale spaces according to the magnitude of the center frequency, selecting a target scale space except for a head scale space and a tail scale space, and determining a previous scale space and a next scale space of the target scale space as adjacent scale spaces; and identifying the edge of the target scale space and the space in the adjacent scale space, and taking the edge as the coincident edge of the target scale space.
4. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 1, wherein the method for calculating the gradient direction distribution feature comprises: Calculating the gradient direction consistency of pixel points at two sides of the coincident edge in each scale space, calculating the gradient direction stability of the coincident edge between adjacent scale spaces, and normalizing the product of the gradient direction consistency and the gradient direction stability to obtain gradient direction distribution characteristics.
5. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 4, wherein the method for calculating the gradient direction consistency comprises: And taking any one pixel point on the coincident edge as a central pixel point, respectively selecting a preset number of pixel points at two sides of the normal direction of the central pixel point, respectively calculating the gradient direction mean value of the pixel points at two sides, and determining gradient direction consistency according to the discrete degree of the gradient direction mean value.
6. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 5, wherein the method for calculating the gradient direction stability comprises: Calculating the difference degree of the gradient direction mean value of the central pixel point in one scale space and the gradient direction mean value of the central pixel point in an adjacent scale space, and determining gradient direction stability based on the difference degree.
7. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 4, wherein the method for determining the edge of the real tissue comprises: and if the gradient direction distribution characteristic is larger than a preset threshold value, the coincident edge is a real tissue edge.
8. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 1, wherein the method for determining the optimal scale value comprises: Taking a communication area formed by the edges of the real tissues as a real tissue area, and counting the quantity of noise falling into the real tissue area in each scale space to obtain noise coverage rate; Calculating the shortest distance between noise and the edge of the real tissue in each scale space, and calculating the noise priority based on the product of the inverse of the distance and the noise quantity; and scoring the plurality of scale spaces based on the noise coverage rate and the noise priority, and taking the frequency range corresponding to the scale space with the highest score as the optimal scale value.
9. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 8, wherein the scoring method comprises: And selecting three scale spaces in different frequency ranges, and calculating a weighted result of the sum of the noise coverage rates and the sum of the noise priorities of all the three scale spaces to obtain a scoring result.
10. The method for automatically denoising a CT image based on multi-scale feature fusion according to claim 9, wherein the denoising method comprises: Marking three frequency ranges corresponding to the optimal scale values as a fine scale, a middle scale and a coarse scale respectively; Performing noise reduction processing by adopting an edge-preserving filtering algorithm for the scale space corresponding to the fine scale, performing noise reduction processing by adopting a multi-scale transformation denoising algorithm for the scale space corresponding to the middle scale, and performing noise reduction processing by adopting a scale space smoothing algorithm for the scale space corresponding to the coarse scale; And fusing the scale space after the noise reduction treatment with the scale space without the noise reduction treatment to obtain a CT image after the noise reduction.

Description

CT image automatic noise reduction processing method based on multi-scale feature fusion Technical Field The invention relates to the technical field of medical image processing, in particular to an automatic CT image noise reduction processing method based on multi-scale feature fusion. Background CT image (Computed Tomography ) is a medical imaging technique in which a portion of the human body is scanned in cross-section by an X-ray beam, the X-rays transmitted through the slice are received by a detector, and the cross-sectional anatomy of the human body is reconstructed after computer processing. The CT image inevitably contains various types of noise components, which can reduce the image quality, mask fine lesions, blur tissue boundaries, affect density resolution, and particularly in diagnostic scenes with high requirements on image details, such as lung nodule detection, vascular imaging, early tumor screening, etc., the presence of noise can significantly reduce the diagnostic accuracy and increase the risks of missed diagnosis and misdiagnosis, so that noise reduction treatment is required on the CT image to improve the image quality and the diagnostic reliability. In the prior art, when the noise reduction processing is performed on a CT image, a multi-scale noise reduction method based on fixed-scale decomposition or frequency characteristics is generally adopted, the image noise is divided into fine-scale noise, middle-scale noise and coarse-scale noise, and a fixed corresponding relation exists between different scale components and noise types by default, and then the corresponding noise reduction strategies are adopted for processing respectively. For example, a CT image is decomposed into different frequency bands using a fixed decomposition method such as wavelet transform or laplacian pyramid, threshold denoising is used for high-frequency subbands, and smoothing filtering is used for low-frequency subbands. However, the method fails to fully consider the dynamic performance characteristics of the real tissue texture and the anatomical structure in the CT image under different scales, so that random noise and the real tissue structure are difficult to effectively distinguish, and particularly when the noise and the real tissue structure are overlapped on the spatial position and the frequency distribution, the fine structures such as the edge of a real blood vessel, the boundary of a soft tissue and the like are easily misjudged as noise to be removed, or the structured noise is misjudged as the real structure to be reserved, so that the problems of detail loss, structural blurring or artifact residues of the CT image after noise reduction appear, and the diagnosis reliability and the clinical application value of the CT image are seriously affected. Disclosure of Invention In order to solve the technical problems that the prior art is difficult to effectively distinguish random noise from a real tissue structure and the detail is lost or the structure is fuzzy due to insufficient consideration of dynamic performance characteristics of the real tissue structure under different scales, the invention aims to provide an automatic CT image noise reduction processing method based on multi-scale feature fusion, and the adopted technical scheme is as follows: The invention provides a CT image automatic noise reduction processing method based on multi-scale feature fusion, which comprises the following steps: performing frequency domain analysis on the CT image, and dividing the CT image into a plurality of scale spaces based on a frequency range; Analyzing gradient direction distribution characteristics of the overlapped edges in different scale spaces, determining real tissue edges in the scale spaces, and taking edges except the real tissue edges in each scale space as noise; and determining an optimal scale value according to the noise distribution condition of each scale space, and carrying out noise reduction treatment on the CT image based on the optimal scale value. Further, the method for dividing the scale space comprises the following steps: acquiring a body part to which the CT image belongs, and determining the highest frequency component and the lowest frequency component of frequency domain analysis according to the frequency domain characteristics corresponding to the body part; Calculating a frequency step based on the highest frequency component, the lowest frequency component and the preset scale number, generating a plurality of center frequencies corresponding to the scale number based on the frequency step, and constructing a plurality of scale spaces corresponding to the center frequencies by utilizing a band-pass filter. Further, the method for determining the coincident edges comprises the following steps: Sequencing all scale spaces according to the magnitude of the center frequency, selecting a target scale space except for a head scale space and a tail scale space, an