CN-122023173-A - Low-dose scanning image noise reduction detection method for medical image

CN122023173ACN 122023173 ACN122023173 ACN 122023173ACN-122023173-A

Abstract

The invention discloses a low-dose scanning image noise reduction detection method for medical images, which relates to the technical field of medical diagnosis, monitoring and treatment equipment manufacture, and comprises the specific steps of acquiring low-dose images through preprocessing, and parallelly extracting structure and noise characteristics by a dual-path attention collaborative fusion network; the method comprises the steps of combining gradient and texture characteristics to dynamically determine fusion weights, weighting and fusing the dual characteristics, recovering image size through deconvolution to output a noise reduction image, finally scanning and positioning a pathological structure through a pathological detection model and generating a report, realizing full-flow processing from noise reduction to diagnosis, and improving image quality and diagnosis accuracy.

Inventors

Pan Anguang
HE YONGFANG
MENG LEI
MA BAOXIANG
JIA ZHIJUN
HU SHIJUN

Assignees

温州市第七人民医院

Dates

Publication Date: 20260512
Application Date: 20260126

Claims (10)

1. A low-dose scanning image noise reduction detection method for medical images is characterized by comprising the following specific steps: S1, acquiring and preprocessing a low-dose medical image to be processed, preprocessing the low-dose medical image, and inputting the preprocessed low-dose medical image into a pre-constructed dual-path attention collaborative fusion network; S2, extracting features of the preprocessed low-dose medical image by using a structural path and a noise estimation path of the dual-path attention collaborative fusion network, wherein the structural path is extracted to obtain structural features, and the noise estimation path is extracted to obtain noise features; S3, dynamic attention fusion, namely receiving structural features and noise features, extracting gradient features and texture complexity features of a local area of the low-dose medical image, determining fusion weights through a dual-feature collaborative dynamic weight algorithm, and fusing the structural features and the noise features through a weighted enhanced dual-path fusion algorithm to obtain fusion features; S4, outputting the noise-reduced image, namely inputting the fusion characteristics into an output layer by a dual-path attention cooperative fusion network, recovering to the original low-dose medical image size through deconvolution operation, and outputting the noise-reduced medical image; S5, detecting the pathological structure, namely carrying out region-by-region scanning on the noise reduction medical image by using a pathological structure detection model, identifying and positioning the pathological structure, outputting a detection report and finishing the noise reduction detection.
2. The method for detecting the noise reduction of the low-dose scanning image for the medical image according to claim 1 is characterized in that in S1, in image acquisition and preprocessing, a dual-path attention cooperative fusion network is a deep learning network for low-dose medical image feature extraction, dynamic weight distribution and feature fusion processing and comprises a structural path, a noise estimation path and dynamic attention fusion, the dual-path attention cooperative fusion network construction process comprises the steps of constructing a network main structure according to a PyTorch or TensorFlow deep learning frame, selecting the low-dose medical image and a normal-dose medical image as a training data set, dividing the training data set into a training set, a verification set and a test set, performing iterative training by adopting weighted sum of mean square error and structural similarity loss as a training loss function, determining training termination conditions according to the performance of the verification set, and storing network parameters.
3. The method for detecting noise reduction of the low-dose scanning image for the medical image according to claim 1 is characterized in that in S2, the dual-path parallel feature extraction, the structural path extraction obtains structural features, wherein the structural features comprise anatomic prior information contained in the low-dose medical image and edge texture information in different directions, the structural path carries out sliding window convolution filtering processing on the preprocessed low-dose medical image by adopting Gaussian direction filters in different directions, the Gaussian direction filters comprise 8 directions, the size of a filtering window is 3 x 3 pixels, the standard deviation of the filters is set to be 0.8-1.2, the edge texture information in different directions of the image is captured through the filtering processing, the edge texture features obtained through the filtering are sequentially convolved, normalized in batches and activated, the features after the filtering processing are further extracted, and finally the edge texture related features and the anatomic prior related features are integrated to obtain the structural features.
4. The method for detecting noise reduction of the low-dose scanning image for the medical image according to claim 1 is characterized in that in S2, the dual-path parallel feature extraction, the specific content of noise characteristics obtained by noise estimation path extraction is that the noise characteristics are integral distribution rule characteristics of noise in a frequency domain and non-uniform distribution characteristics of the noise in different scales of a wavelet domain in the low-dose medical image, the noise estimation path is used for carrying out channel splicing on the characteristics extracted in the frequency domain and the characteristics extracted in the wavelet domain, preserving amplitude spectrum and phase spectrum of the frequency domain, carrying out characteristic extraction on the amplitude spectrum through convolution operation to capture high-frequency distribution rules of the noise, carrying out 3-layer wavelet decomposition on the preprocessed low-dose medical image, extracting high-frequency detail coefficients after each layer decomposition to capture the non-uniform distribution characteristics of the noise, and carrying out channel splicing on the characteristics extracted in the frequency domain and the characteristics extracted in the wavelet domain, carrying out feature dimension reduction through convolution operation, and integrating to obtain the noise characteristics.
5. The method for detecting the noise reduction of the low-dose scanning image facing the medical image according to claim 4 is characterized in that the convolution operation is a sliding window operation according to a preset convolution kernel, and specifically comprises a first convolution operation for extracting frequency domain amplitude spectrum characteristics and a second convolution operation for characteristic dimension reduction, wherein the first convolution operation adopts a convolution kernel with the size of 1 multiplied by 1, the step length is set to be 1, filling is not carried out, a single pixel is used as the step length to traverse the frequency domain amplitude spectrum, characteristic response values are obtained through multiplication and summation of elements in a region corresponding to the amplitude spectrum through the convolution kernel, 64-dimensional frequency domain characteristics are output, the second convolution operation adopts a convolution kernel with the size of 3 multiplied by 3, the step length is set to be 1, the filling mode is to fill 1 pixel at edges, the 128-dimensional characteristic diagram after channel splicing is traversed, characteristic integration and dimension reduction are completed through multiplication and summation of the convolution kernel and elements in the region corresponding to the characteristic diagram, and finally 64-dimensional integrated characteristics are output.
6. The method for detecting noise reduction of a low-dose scanned image for medical imaging according to claim 1, wherein in the step S3, dynamic attention fusion, the gradient features are intensity and direction related features of pixel gray value change in a local area of the low-dose medical image, including amplitude features and change direction features of pixel gray value change, and specifically cover horizontal gradient features, vertical gradient features and diagonal gradient features, the texture complexity features are complexity related features of spatial distribution rules of pixel gray values in the local area of the low-dose medical image, including contrast, entropy, energy and correlation, wherein the contrast reflects the difference degree of different pixel gray values in the local area, the entropy reflects the disorder degree of gray value distribution in the local area, the energy reflects the uniformity of gray value distribution in the local area, and the correlation reflects the correlation degree of gray values of adjacent pixels in the local area.
7. The method for detecting noise reduction of a low-dose scanned image for medical imaging according to claim 1, wherein the mathematical expression of the dual-feature collaborative dynamic weighting algorithm is: Wherein, the Fusion weights for structural features; Fusion weights for noise features; Activating a function for Sigmoid; And Is a learnable adjustment coefficient; Gradient feature values of local areas of the low-dose medical image; texture complexity characteristic values of local areas of the low-dose medical image.
8. The method for detecting noise reduction of a low-dose scanned image for medical imaging according to claim 1, wherein in the step S3, a mathematical expression of a weighted enhancement dual-path fusion algorithm in dynamic attention fusion is as follows: Wherein, the Is the final fusion feature; Fusion weights for structural features; Extracting structural features for the structural path; is a structural feature enhancement factor; Fusion weights for noise features; Extracting noise characteristics for a noise estimation path; is a noise suppression factor.
9. The method for detecting the noise reduction of the low-dose scanning image facing the medical image according to claim 1 is characterized in that in S4, in the noise reduction image output, deconvolution operation is operation for realizing feature image size amplification and space information recovery and is used for restoring the size of a fusion feature to the original low-dose medical image size, and the specific operation process comprises the steps of firstly presetting deconvolution kernel size to be 3 multiplied by 3 and step size to be 2, adopting a zero filling mode to avoid the loss of feature image edge information, then inputting the fusion feature into an output layer, carrying out sliding coverage on the fusion feature image according to the preset step size by deconvolution kernel, integrating the feature information of a region corresponding to the feature image to generate a high-dimensional intermediate feature, carrying out dimension calibration and dimension adjustment on the intermediate feature, and finally outputting the feature image consistent with the original low-dose medical image size.
10. The low-dose scanned image noise reduction detection method for the medical image is characterized in that in S5, a pathological structure detection model is a special deep learning detection model designed for the medical image after noise reduction, the pathological structure detection model is used for accurately identifying and positioning pathological structures in the image, the pathological structure detection model adopts an end-to-end convolutional neural network architecture and mainly comprises an input layer, a feature extraction layer, an identification positioning layer and an output layer which are sequentially connected, the input layer is responsible for receiving the noise reduction medical image and finishing image preprocessing to improve the subsequent detection precision, the feature extraction layer performs multi-dimensional feature mining on the preprocessed image to extract key features of textures, gray scales and forms related to the pathological structures, the identification positioning layer performs region-by-region scanning analysis on the image according to the extracted features to judge whether the pathological structures exist in each region and determine the specific positions of the pathological structures, and the output layer is used for finishing the identification positioning result as a standardized detection report and outputting the standardized detection report.

Description

Low-dose scanning image noise reduction detection method for medical image Technical Field The invention relates to the technical field of medical diagnosis, monitoring and treatment equipment manufacturing, in particular to a low-dose scanning image noise reduction detection method for medical images. Background The medical image is used as a core basis for clinical diagnosis, disease monitoring and treatment evaluation, the quality of the medical image is directly related to the accuracy of a diagnosis result, along with the continuous improvement of the radiation protection importance degree in the medical health field, the low-dose scanning technology can effectively reduce the radiation exposure of patients and reduce the health risk caused by radiation accumulation, the medical image is widely popularized in various medical image inspections, and is particularly suitable for people sensitive to radiation, such as chronic patients, children, the elderly and the like, which need to review for many times, however, the low-dose scanning can cause noise increase in the image while reducing the radiation dose, the noise can interfere the structural detail presentation of the image, the boundary of pathological tissues and normal tissues is blurred, great challenges are brought to the identification and positioning of the pathological structures, in clinical diagnosis, the low-dose scanning is required to ensure the safety of the patients, and the image has enough definition and detail integrity to support the accurate diagnosis, so the research and development can balance the radiation protection and the image quality, and the technology for realizing the efficient noise reduction and the accurate pathological detection become the key requirements in the medical image processing field. In the feature extraction step, most techniques adopt a single path extraction mode, only feature information of a certain dimension of an image can be captured, structural details and noise distribution rules cannot be comprehensively obtained at the same time, so that the problems of losing the structural details or incomplete noise suppression easily occur in the noise reduction process, and in the aspect of feature fusion, the conventional method mostly adopts a fixed weight fusion strategy, lacks dynamic adaptation capability to the feature difference of the local area of the image, the fusion proportion is difficult to adjust according to the structural complexity and the noise intensity of different areas, suitability and flexibility are not enough, in addition, the traditional technology often splits noise reduction and pathology detection into independent processes, the image conversion of an intermediate link is easy to cause information loss, and the detection model is not specially optimized for the image characteristics after noise reduction, so that the two are poor in cooperativity, the processes are redundant, the processing efficiency is low, the accuracy of a detection result is influenced, and the double high requirements of clinical diagnosis on the image quality and the detection reliability are difficult to meet. Disclosure of Invention The invention aims to make up the defects of the prior art and provides a low-dose scanning image noise reduction detection method for medical images, which comprises the steps of preprocessing low-dose medical images, adopting a dual-path attention collaborative fusion network to extract structure and noise characteristics, combining a dynamic attention fusion mechanism, determining fusion weights according to gradient and texture complexity, the method has the advantages that the effective fusion of the dual-path characteristics is realized, the image size is restored through deconvolution operation, the noise-reduction medical image is output, the pathological structure is identified and positioned by using the pathological structure detection model, and finally the detection report is generated. The invention provides a low-dose scanning image noise reduction detection method for medical images, which aims to solve the technical problems and comprises the following specific steps: S1, acquiring and preprocessing a low-dose medical image to be processed, preprocessing the low-dose medical image, and inputting the preprocessed low-dose medical image into a pre-constructed dual-path attention collaborative fusion network; S2, extracting features of the preprocessed low-dose medical image by using a structural path and a noise estimation path of the dual-path attention collaborative fusion network, wherein the structural path is extracted to obtain structural features, and the noise estimation path is extracted to obtain noise features; S3, dynamic attention fusion, namely receiving structural features and noise features, extracting gradient features and texture complexity features of a local area of the low-dose medical image, determining fusion weights through a dual-fe