A cancer image denoising method and system based on multi-scale neural network

Abstract

The present invention discloses a cancer image denoising method and system based on a multiscale neural network, including: adaptively segmenting cancer images, performing noise estimation and statistical analysis on each segmented region, and automatically adjusting denoising parameters; using wavelet transform and pyramid decomposition to extract multiscale features, analyzing and processing the image after adjusting parameters; performing quality assessment and distortion detection on the multiscale processed image, automatically identifying and locating artifacts and distortion, using point spread function inversion and Wiener filtering methods to correct and compensate interference to obtain a corrected image; applying histogram equalization and contrast-limited adaptive histogram equalization to the corrected image, using a convolutional neural network combined with a sparse representation and a dictionary learning algorithm to perform denoising processing to obtain a denoising result. The present invention can effectively remove noise from different types of cancer images while retaining key diagnostic information and improving image quality and clarity.

Description

Technical Field

[0001] The present invention belongs to the field of image denoising, and in particular relates to a cancer image denoising method and system based on a multi-scale neural network. Background Art

[0002] Cancer image denoising is a challenging technical problem. First, cancer images usually contain complex tissue structures and texture features. Noise can interfere with these key image information, leading to deviations in subsequent analysis and diagnosis. Second, different types of cancer images differ in the type, intensity, and distribution of noise. Denoising algorithms need to be able to adaptively handle these changes. Third, the size and resolution of cancer images vary. Denoising methods need to extract and fuse features at different scales to make full use of image information. At the same time, various artifacts and distortions may be introduced during the acquisition of cancer images. The denoising process needs to be able to effectively distinguish and eliminate these interferences. Finally, the quality of cancer images in practical applications varies, with overexposure, underexposure, motion blur, etc. Denoising systems need to have strong robustness and generalization capabilities. In summary, cancer image denoising needs to consider factors such as the diversity of noise, the complexity of the image, the variability of scale, the multi-source of interference, and the robustness of practical applications. This places high demands on the design of denoising methods and systems. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a cancer image denoising method and system based on a multi-scale neural network. Among them, a cancer image denoising method based on a multi-scale neural network includes:

[0004] Adaptively segment cancer images, divide the images into different regions, perform noise estimation and statistical analysis on each segmented region, and automatically adjust the denoising algorithm parameters;

[0005] Wavelet transform and pyramid decomposition are used to extract multi-scale features, and the images after parameter adjustment are analyzed and processed in different scale spaces;

[0006] By making full use of local and global information of the image through feature fusion, the quality assessment and distortion detection of the multi-scale processed image are carried out, the artifacts and distortion are automatically identified and located, and the point spread function inversion and Wiener filtering methods are used to correct and compensate for interference to obtain the corrected image;

[0007] After applying histogram equalization and contrast-limited adaptive histogram equalization to the corrected image, the image is used as a model training input to train a convolutional neural network model;

[0008] Use convolutional neural networks combined with sparse representation and dictionary learning algorithms to learn noise and interference feature patterns through training and optimization with cancer image samples;

[0009] The trained target convolutional neural network model is cross-validated and performance evaluated, and the optimal model is selected and applied to the enhanced image for denoising to obtain the denoising result.

[0010] Preferably, the process of adaptively segmenting the cancer image and dividing the image into different regions includes:

[0011] Obtain a cancer image, and use an adaptive segmentation algorithm to divide the image into different regions according to local features and texture information of the image to obtain segmented sub-regions;

[0012] For each sub-region, extract the texture feature, grayscale feature and contrast feature of the sub-region to obtain a feature vector of the sub-region;

[0013] According to the pre-established denoising method library, the feature vector of each sub-region is matched to obtain the denoising method of the corresponding sub-region.

[0014] Preferably, the process of performing noise estimation and statistical analysis on each segmented region and automatically adjusting the parameters of the denoising algorithm includes:

[0015] According to the image segmentation result, image data of each segmented area is obtained, and noise estimation is performed on each area to obtain the noise type, intensity and distribution characteristics of the area;

[0016] Conduct statistical analysis on the noise parameters of each area, calculate the differences in noise parameters between different areas, and determine whether there are obvious differences in noise distribution;

[0017] If there are obvious differences in noise distribution between regions, the optimal parameter combination is selected from the preset denoising algorithm parameter range according to the noise characteristics of different regions to obtain denoising algorithm parameters that are suitable for the noise characteristics of each region;

[0018] Adopting adaptive optimization strategy, the denoising algorithm parameters of each region are continuously adjusted through iterative optimization method to minimize the noise residue and detail loss of the denoised image and obtain the optimal denoising algorithm parameter combination of each region.

[0019] Preferably, wavelet transform and pyramid decomposition are used to extract multi-scale features, and the process of analyzing and processing the image after parameter adjustment in different scale spaces includes:

[0020] The wavelet transform algorithm is used to decompose the image into multiple scales to obtain wavelet coefficients at different scales.

[0021] At each scale, the wavelet coefficients are threshold processed according to the preset threshold to remove noise and redundant information and obtain the processed wavelet coefficients;

[0022] For the processed wavelet coefficients, a pyramid decomposition algorithm is used to extract local and global features of the image in different scale spaces to obtain a multi-scale feature set;

[0023] According to the preset feature selection rules, the most effective feature subset for image analysis and processing is screened out from the multi-scale feature set to obtain the optimized multi-scale features.

[0024] Preferably, by making full use of local and global information of the image through feature fusion, quality assessment and distortion detection are performed on the multi-scale processed image, and the process of automatically identifying and locating artifacts and distortion includes:

[0025] The optimized multi-scale features are input into the feature fusion model, and the local information and the global information are fully utilized and fused through feature weighting and feature splicing to obtain the fused image feature representation;

[0026] Based on the fused image feature representation, according to the preset image quality assessment index, a support vector machine algorithm is used to perform quality assessment on the fused image to obtain an image quality score;

[0027] According to the image quality score, determine whether the image is distorted. If distortion is present, further extract multi-scale features of the image, analyze the multi-scale features using a convolutional neural network algorithm, and automatically identify the type of distortion;

[0028] According to the identified distortion type, a multi-scale decomposition method based on wavelet transform is used to decompose the image into multiple scale sub-images, and the distorted area is detected on each scale sub-image to obtain the position coordinates of the distorted area.

[0029] Preferably, the process of correcting and compensating interference by using point spread function inversion and Wiener filtering methods to obtain a corrected image includes:

[0030] According to the position coordinates of the distorted area, the image block of the corresponding area is extracted, and the point spread function parameters of the area are estimated by solving the inverse problem of the point spread function to obtain the distorted point spread function model;

[0031] Performing deconvolution processing on the image block in the distorted area through the point spread function model to eliminate the distortion caused by the point spread effect and obtain a corrected image block;

[0032] Perform Wiener filtering on the corrected image block in the frequency domain to suppress noise interference and obtain a filtered image block;

[0033] The filtered image blocks are replaced back to the corresponding positions of the original images to complete the correction and compensation of the distorted areas, and a high-quality image after correction and compensation is output.

[0034] Preferably, the process of applying histogram equalization and contrast limited adaptive histogram equalization to the corrected image comprises:

[0035] Performing histogram equalization processing on the corrected image to obtain a first equalized image;

[0036] Using a contrast-limited adaptive histogram equalization algorithm to process the first equalized image to obtain a second equalized image;

[0037] By calculating the contrast and clarity indexes of the second equalized image, it is determined whether a preset contrast threshold and clarity threshold are met, and if so, the second equalized image is determined to be an optimized image;

[0038] If the contrast and clarity indicators of the second equalized image do not meet the preset thresholds, adaptively adjusting the parameters of the contrast-limited adaptive histogram equalization algorithm according to the histogram distribution characteristics of the second equalized image, and reprocessing the first equalized image until a third equalized image that meets the preset contrast and clarity thresholds is obtained;

[0039] The equalized image that meets the preset contrast and clarity thresholds is used as the optimized image, and the image quality evaluation index of the optimized image is obtained; the image quality evaluation index of the optimized image includes contrast, clarity, and color saturation.

[0040] Preferably, the process of learning noise and interference feature patterns using a convolutional neural network combined with a sparse representation and dictionary learning algorithm through cancer image sample training and optimization includes:

[0041] Design the model network structure and model parameters according to the characteristics of cancer images, and build a convolutional neural network model;

[0042] Based on the convolutional neural network model, sparse representation and dictionary learning algorithms are introduced to extract robust feature representation through sparse decomposition of image data and dictionary construction;

[0043] Apply the trained convolutional neural network model and sparse representation dictionary to new cancer image data to extract and represent image features;

[0044] Based on the extracted features, a classifier or regression model is used to identify and analyze cancer images, and the identification and analysis results are compared with the results of expert annotation to calculate the accuracy, sensitivity and specificity indicators of the model;

[0045] Based on the evaluation results, the model is iteratively optimized and improved, the network structure is adjusted, training data is increased, and new algorithms are introduced.

[0046] Preferably, the target convolutional neural network model obtained through training is cross-validated and performance evaluated, and the optimal model is selected to be applied to the enhanced image for denoising. The process of obtaining the denoising result includes:

[0047] Based on the trained target convolutional neural network model, multi-scale features of the image are extracted through convolution and pooling operations, and a denoising model is trained using the multi-scale features;

[0048] Apply the trained denoising model to each pixel of the image data, calculate the denoising result of each pixel through forward propagation, and obtain the denoised image;

[0049] The quality of the denoised image is evaluated, and the peak signal-to-noise ratio and structural similarity indicators are used to quantify the quality improvement of the denoised image;

[0050] Based on the quality assessment results, determine whether the denoising effect has reached the expected goal. If not, return the parameters of the denoising model that need to be adjusted urgently, retrain the model and process it until the quality requirements are met;

[0051] The denoised image that meets the quality requirements is output as the final high-quality denoising result to obtain a denoised image.

[0052] The present invention also provides a cancer image denoising system based on a multi-scale neural network, comprising:

[0053] An adaptive image segmentation module, used to adaptively segment cancer images and divide the images into different regions;

[0054] The noise estimation and parameter adjustment module is used to perform noise estimation and statistical analysis on each segmented area and automatically adjust the denoising algorithm parameters;

[0055] Multi-scale feature extraction and fusion module, used to extract multi-scale features using wavelet transform and pyramid decomposition, and analyze and process the image after parameter adjustment in different scale spaces;

[0056] Image quality assessment and correction module, which is used to make full use of local and global information of the image through feature fusion, perform quality assessment and distortion detection on the multi-scale processed image, automatically identify and locate artifacts and distortion, and use point spread function inversion and Wiener filtering methods to correct and compensate for interference to obtain a corrected image;

[0057] An image enhancement module, used for applying histogram equalization and contrast-limited adaptive histogram equalization to the corrected image and then using the image as a model training input to train a convolutional neural network model;

[0058] A machine learning model training module for learning noise and interference feature patterns using convolutional neural networks combined with sparse representation and dictionary learning algorithms, trained and optimized through cancer image samples;

[0059] The model evaluation and application module is used to cross-validate and evaluate the performance of the trained target convolutional neural network model, select the optimal model and apply it to the enhanced image to perform denoising and obtain the denoising result;

[0060] The adaptive closed-loop control module is used to evaluate the quality of the denoised image and dynamically adjust the model parameters based on the evaluation results.

[0061] Compared with the prior art, the present invention has the following advantages and technical effects:

[0062] The adaptive denoising method for cancer images disclosed in the present invention first adaptively segments the cancer images, and then selects a suitable denoising algorithm according to the characteristics of different regions. Through multi-scale feature extraction and fusion, the local and global information of the image is fully utilized. The noise model is trained by machine learning methods and applied to the enhanced image for denoising. The whole process forms a closed loop, and the parameters of each step can be dynamically adjusted according to the denoising effect. The present invention can effectively remove noise in different types of cancer images while retaining key diagnostic information and improving image quality and clarity. This adaptive and robust denoising method can provide more reliable image support for accurate diagnosis and treatment of cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0064] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention;

[0065] Figure 2 Schematic diagram of the system structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0066] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0067] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0068] Embodiment 1

[0069] like Figure 1 As shown, this embodiment provides a cancer image denoising method and system based on a multi-scale neural network, including:

[0070] Adaptively segment cancer images, divide the images into different regions, perform noise estimation and statistical analysis on each segmented region, and automatically adjust the denoising algorithm parameters;

[0071] Wavelet transform and pyramid decomposition are used to extract multi-scale features, and the images after parameter adjustment are analyzed and processed in different scale spaces;

[0072] By making full use of local and global information of the image through feature fusion, the quality assessment and distortion detection of the multi-scale processed image are carried out, the artifacts and distortion are automatically identified and located, and the point spread function inversion and Wiener filtering methods are used to correct and compensate for interference to obtain the corrected image;

[0073] After applying histogram equalization and contrast-limited adaptive histogram equalization to the rectified image, the convolutional neural network model is trained as the model training input;

[0074] Use convolutional neural networks combined with sparse representation and dictionary learning algorithms to learn noise and interference feature patterns through training and optimization with cancer image samples;

[0075] The trained target convolutional neural network model is cross-validated and performance evaluated, and the optimal model is selected and applied to the enhanced image for denoising to obtain the denoising result.

[0076] Furthermore, the process of adaptively segmenting the cancer image and dividing the image into different regions includes:

[0077] Obtain a cancer image, and use an adaptive segmentation algorithm to divide the image into different regions according to local features and texture information of the image to obtain segmented sub-regions;

[0078] For each sub-region, extract the texture feature, grayscale feature and contrast feature of the sub-region to obtain a feature vector of the sub-region;

[0079] According to the pre-established denoising method library, the feature vector of each sub-region is matched to obtain the denoising method of the corresponding sub-region.

[0080] Specifically, acquiring cancer images is the starting point of the entire processing flow. Taking breast cancer X-ray images as an example, the original images usually need to be preprocessed, such as contrast enhancement and noise suppression. The histogram equalization method can be used to increase the image contrast and make the lesion area more prominent. At the same time, methods such as median filtering can be used to remove salt and pepper noise in the image to prepare for subsequent segmentation.

[0081] Adaptive segmentation algorithm is an effective method for segmentation based on local features of the image. Taking the region growing method as an example, the breast tissue region can be selected as the seed point, and the region can be gradually expanded according to the similarity of grayscale values, and finally the breast tissue can be separated from the background. This method can adapt to the differences in breast tissue morphology of different patients and improve the accuracy of segmentation.

[0082] For the segmented sub-regions, feature vectors need to be extracted for subsequent processing. Taking texture features as an example, the gray-level co-occurrence matrix can be calculated to extract statistics such as energy, contrast, and correlation as texture features. For grayscale features, statistics such as the average grayscale value and variance of the sub-region can be calculated. These features together constitute a feature vector that describes the characteristics of the sub-region.

[0083] The choice of denoising method is crucial to preserving image details. Taking wavelet transform denoising as an example, the appropriate wavelet basis and decomposition level can be selected according to the texture complexity of the sub-region. For areas with more complex textures, a wavelet basis with good directional selectivity, such as curvelet transform, can be selected to better preserve edge information. For smoother areas, a simple wavelet basis such as Haar wavelet can be selected, and a larger threshold can be used for coefficient shrinkage to achieve better denoising effect.

[0084] Through this processing step, the quality of cancer images can be effectively improved, providing a more reliable image basis for subsequent image denoising. For example, in early screening for breast cancer, preprocessed images can more clearly display tiny calcifications, helping doctors to detect potential malignant lesions early. It is worth noting that the entire processing flow needs to be appropriately adjusted according to different types of cancer images. For example, for CT images, more attention may need to be paid to removing artifacts and interference caused by metal implants; while for PET images, more attention may need to be paid to noise suppression and contrast enhancement. Therefore, in practical applications, it is necessary to flexibly adjust the preprocessing strategy according to the specific imaging modality and clinical needs to obtain the best image processing effect.

[0085] Furthermore, the process of performing noise estimation and statistical analysis on each segmented region and automatically adjusting the parameters of the denoising algorithm includes:

[0086] According to the image segmentation results, the image data of each segmented area is obtained, and the noise type, intensity and distribution characteristics of the area are obtained for each area.

[0087] Conduct statistical analysis on the noise parameters of each area, calculate the differences in noise parameters between different areas, and determine whether there are obvious differences in noise distribution;

[0088] If there are obvious differences in noise distribution between regions, the optimal parameter combination is selected from the preset denoising algorithm parameter range according to the noise characteristics of different regions to obtain denoising algorithm parameters that are suitable for the noise characteristics of each region;

[0089] Adopting adaptive optimization strategy, the denoising algorithm parameters of each region are continuously adjusted through iterative optimization method to minimize the noise residue and detail loss of the denoised image and obtain the optimal denoising algorithm parameter combination of each region.

[0090] Specifically, in medical image processing, regional adaptive denoising is an effective method. First, the image data of each region is obtained according to the image segmentation results to estimate the noise. For example, for a lung CT image, it may be divided into lung parenchyma, trachea, blood vessels and other regions. By analyzing the grayscale distribution, texture characteristics, etc. of each region, different types of noise parameters can be estimated. For example, there may be Gaussian noise in the lung parenchyma region, while there may be salt and pepper noise in the trachea region. Next, the noise parameters of each region are statistically analyzed to calculate the difference. This step can help determine whether different denoising strategies need to be adopted for different regions. For example, if it is found that the noise intensity in the lung parenchyma region is significantly higher than that in other regions, a more powerful denoising algorithm can be selected for this region. If there is a significant difference in noise distribution between regions, it is necessary to select the optimal denoising algorithm parameters for different regions. For example, for the lung parenchyma region with strong Gaussian noise, a Gaussian filter can be selected, and the standard deviation of the filter can be adjusted according to the noise intensity; while for the vascular region with more edges, an edge-preserving filter, such as a bilateral filter, can be selected. Adaptive optimization strategy is adopted to continuously adjust the denoising algorithm parameters of each region through iterative optimization method. This process can use an objective function, such as minimizing the weighted sum of noise residue and detail loss. For example, the initial parameters can be set, and then the filter window size, weight and other parameters can be gradually adjusted until the best effect is achieved.

[0091] Furthermore, wavelet transform and pyramid decomposition are used to extract multi-scale features, and the process of analyzing and processing the image after adjusting parameters in different scale spaces includes:

[0092] The wavelet transform algorithm is used to decompose the image into multiple scales to obtain wavelet coefficients at different scales.

[0093] At each scale, the wavelet coefficients are threshold processed according to the preset threshold to remove noise and redundant information and obtain the processed wavelet coefficients;

[0094] For the processed wavelet coefficients, a pyramid decomposition algorithm is used to extract local and global features of the image in different scale spaces to obtain a multi-scale feature set;

[0095] According to the preset feature selection rules, the most effective feature subset for image analysis and processing is screened out from the multi-scale feature set to obtain the optimized multi-scale features.

[0096] Specifically, wavelet transform is a powerful multi-scale analysis tool. For chest X-ray images, Haar wavelet can be used for three-level decomposition to obtain approximate coefficients and detail coefficients at different scales. This can separate different frequency components in the image, which is beneficial for subsequent feature extraction.

[0097] Threshold processing is an effective method to remove noise. Taking brain MRI images as an example, the wavelet coefficients can be processed using the soft threshold method. By setting a suitable threshold (such as using the VisuShrink method to calculate), noise can be effectively suppressed while retaining important structural information of the image.

[0098] Pyramid decomposition algorithms can extract multi-scale features. For example, applying Gaussian-Laplacian pyramid decomposition to skin lesion images can obtain image representations at different scales. This multi-scale representation can capture the local texture features and global shape features of the lesion, which is helpful for subsequent lesion classification.

[0099] Feature selection is crucial to improving analysis efficiency. Taking breast cancer pathology images as an example, the most discriminative feature subset can be selected from the multi-scale feature set. The feature selection method based on mutual information can be used to select the features most relevant to cancer classification, such as morphological features of cell nuclei, tissue structure features, etc.

[0100] Furthermore, feature fusion is used to make full use of local and global information of the image, and quality assessment and distortion detection are performed on the multi-scale processed image. The process of automatically identifying and locating artifacts and distortion includes:

[0101] The optimized multi-scale features are input into the feature fusion model, and the local information and the global information are fully utilized and fused through feature weighting and feature splicing to obtain the fused image feature representation;

[0102] Based on the feature representation of the fused image and the preset image quality evaluation index, the support vector machine algorithm is used to evaluate the quality of the fused image to obtain the image quality score;

[0103] According to the image quality score, determine whether the image is distorted. If distortion exists, further extract the multi-scale features of the image, use the convolutional neural network algorithm to analyze the multi-scale features, and automatically identify the distortion type;

[0104] According to the identified distortion type, a multi-scale decomposition method based on wavelet transform is used to decompose the image into multiple scale sub-images, and the distorted area is detected on each scale sub-image to obtain the position coordinates of the distorted area.

[0105] Specifically, feature fusion can make full use of information at different scales. In the task of lung nodule detection, the shape features, texture features, and density features extracted at different scales can be weighted and fused. By setting reasonable weights, local details and global structural information can be balanced to improve the accuracy of nodule detection. The feature fusion of this embodiment can adapt to different types of medical images and effectively suppress noise and redundant information while retaining important diagnostic information.

[0106] Image quality assessment is also an important part of image processing. The support vector machine algorithm can be used to build an image quality scoring model. By extracting features such as image contrast, clarity, and color saturation as input, the model is trained to output a quality score. For example, for a high-definition image with a resolution of 1920x1080, its local contrast, edge clarity, color richness and other features can be extracted and input into the trained support vector machine model to obtain a quality score of 0-100. If distortion is detected in the image, the type of distortion needs to be further analyzed.

[0107] Convolutional neural networks perform well in image recognition and can be used to automatically identify various types of distortion, such as blur, noise, and compression artifacts. For example, this embodiment constructs a network structure including a convolutional layer, a pooling layer, and a fully connected layer, inputs multi-scale image features, and outputs a probability distribution of distortion types. For an image with Gaussian blur, the network may output a probability of 0.95 of Gaussian blur type, which is much higher than other distortion types.

[0108] Locating the distorted area is the basis of correction. Wavelet transform can decompose the image into sub-bands of different frequencies and directions. Detecting abnormal coefficients in each sub-band can locate the distorted area. For example, for an image with local motion blur, detecting the area with abnormal coefficients in the high-frequency sub-band can determine the position coordinates of the blurred area.

[0109] Furthermore, the point spread function inversion and Wiener filtering method are used to correct and compensate for interference, and the process of obtaining the corrected image includes:

[0110] According to the position coordinates of the distorted area, the image block of the corresponding area is extracted, and the point spread function parameters of the area are estimated by solving the inverse problem of the point spread function to obtain the distorted point spread function model;

[0111] Deconvolution is performed on the image block in the distorted area through the point spread function model to eliminate the distortion caused by the point spread effect and obtain the corrected image block;

[0112] Perform Wiener filtering on the corrected image block in the frequency domain to suppress noise interference and obtain a filtered image block;

[0113] The filtered image blocks are replaced back to the corresponding positions of the original images to complete the correction and compensation of the distorted areas, and a high-quality image after correction and compensation is output.

[0114] Specifically, the point spread function describes the response of the imaging system to a point light source. By solving the inverse problem of the point spread function, the point spread function model of the distorted area can be estimated. For example, for Gaussian blur, the variance parameter of the Gaussian kernel can be estimated by the least squares method. After obtaining the point spread function, deconvolution using methods such as Wiener filtering can effectively eliminate the blur effect. Frequency domain filtering is an effective means of suppressing noise. Wiener filtering takes into account the power spectrum of the signal and noise, and can retain image details as much as possible while removing noise. For example, for an image block contaminated by Gaussian white noise, applying Wiener filtering in the frequency domain can significantly improve the signal-to-noise ratio while avoiding the loss of details caused by over-smoothing. Finally, the processed image block is replaced back to the original position to complete the correction of the entire image. This local processing method can adopt different correction strategies for different regions, which is more flexible and effective than global processing. Through the above steps, a low-quality image with local blur, noise and other distortions can be converted into a high-quality image with higher clarity and richer details, laying the foundation for subsequent image analysis and application.

[0115] Further, the process of applying histogram equalization and contrast limited adaptive histogram equalization to the corrected image includes:

[0116] Performing histogram equalization processing on the corrected image to obtain a first equalized image;

[0117] Using a contrast-limited adaptive histogram equalization algorithm to process the first equalized image to obtain a second equalized image;

[0118] By calculating the contrast and clarity indexes of the second equalized image, it is determined whether a preset contrast threshold and clarity threshold are met, and if so, the second equalized image is determined to be an optimized image;

[0119] If the contrast and clarity indicators of the second equalized image do not meet the preset thresholds, adaptively adjust the parameters of the contrast-limited adaptive histogram equalization algorithm according to the histogram distribution characteristics of the second equalized image, and reprocess the first equalized image until a third equalized image that meets the preset contrast and clarity thresholds is obtained;

[0120] The equalized image that meets the preset contrast and clarity thresholds is used as the optimized image, and the image quality evaluation index of the optimized image is obtained; the image quality evaluation index of the optimized image includes contrast, clarity, and color saturation.

[0121] Specifically, this embodiment uses the image enhancement technology of histogram equalization to improve the contrast by redistributing the gray value distribution of the image. Taking a low-contrast cancer photo as an example, after the histogram equalization processing, the originally blurred image contour and image details become clearer and more visible.

[0122] As an additional embodiment, this embodiment can also use the contrast limited adaptive histogram equalization (CLAHE) algorithm for image enhancement to avoid the problem of noise amplification caused by over-enhancement. Suppose there is a medical X-ray image. Ordinary equalization may make the bone structure too prominent and cover up the soft tissue details, while CLAHE can retain more diagnostically valuable information while enhancing the contrast. In practical applications, the contrast threshold can be set to 1.5 and the clarity threshold to 0.8 (assuming that some standardized measurement is used). If the processed image does not reach these thresholds, the system automatically adjusts the parameters of CLAHE, such as the contrast limit value or the block size, to obtain better results. This adaptive adjustment ensures that different types of images can achieve the best enhancement effect.

[0123] The calculation of image quality evaluation index is crucial to judging the processing effect. This embodiment uses the structural similarity index (SSIM) to evaluate the clarity, and the color saturation can be measured by calculating the average value of the S channel in the HSV color space. These indicators provide an objective basis for subsequent processing.

[0124] Furthermore, the process of learning noise and interference feature patterns by using convolutional neural networks combined with sparse representation and dictionary learning algorithms through cancer image sample training and optimization includes:

[0125] Design the model network structure and model parameters according to the characteristics of cancer images, and build a convolutional neural network model;

[0126] Based on the convolutional neural network model, sparse representation and dictionary learning algorithms are introduced to extract robust feature representation through sparse decomposition of image data and dictionary construction;

[0127] Apply the trained convolutional neural network model and sparse representation dictionary to new cancer image data to extract and represent image features;

[0128] Based on the extracted features, a classifier or regression model is used to identify and analyze cancer images, and the identification and analysis results are compared with the results of expert annotation to calculate the accuracy, sensitivity and specificity indicators of the model;

[0129] Based on the evaluation results, the model is iteratively optimized and improved, the network structure is adjusted, training data is increased, and new algorithms are introduced.

[0130] Specifically, for cancer images, this embodiment designs a network structure consisting of multiple layers of convolution, pooling and fully connected layers. For example, a small 3×3 convolution kernel is used to extract local features, the maximum pooling layer reduces the number of parameters, and the dropout layer prevents overfitting. The depth and width of the network need to be adjusted according to the specific task. For example, identifying lung nodules may require a deeper network to capture complex features. Introducing sparse representation and dictionary learning on the basis of convolutional networks can extract more refined features. For example, for breast cancer mammography images, an overcomplete dictionary containing typical structures such as normal tissue, benign and malignant masses can be constructed. The image is sparsely decomposed using algorithms such as orthogonal matching pursuit to obtain a more discriminative feature representation. This method is particularly suitable for situations where the sample size is insufficient. When the trained model is applied to a new cancer image, a multi-scale, multi-level feature map is first extracted through a convolutional network, and then the learned dictionary is used for sparse coding to obtain a compact representation of the image. This representation method retains the powerful feature extraction capability of the convolutional network and utilizes the excellent properties of sparse representation.

[0131] The model evaluation of this embodiment adopts the cross-validation method, and the data set is divided into a training set, a validation set, and a test set. Indicators such as accuracy, sensitivity, and specificity are calculated, or the area under the ROC curve is used as a comprehensive evaluation indicator. For example, a good lung nodule detection system should have a sensitivity of more than 90% and a specificity of not less than 85%. According to the evaluation results, a variety of methods can be used to optimize the model. If overfitting occurs, the regularization strength can be increased or more data enhancement techniques can be introduced. If the model performs poorly, a deeper or wider network structure can be tried, or new algorithms such as the attention mechanism can be introduced.

[0132] Furthermore, the trained target convolutional neural network model is cross-validated and performance evaluated, and the optimal model is selected and applied to the enhanced image for denoising. The process of obtaining the denoising result includes:

[0133] Based on the trained target convolutional neural network model, multi-scale features of the image are extracted through convolution and pooling operations, and the multi-scale features are used to train the denoising model;

[0134] Apply the trained denoising model to each pixel of the image data, calculate the denoising result of each pixel through forward propagation, and obtain the denoised image;

[0135] The quality of the denoised image is evaluated, and the peak signal-to-noise ratio and structural similarity indicators are used to quantify the quality improvement of the denoised image;

[0136] Based on the quality assessment results, determine whether the denoising effect has reached the expected goal. If not, return the parameters of the denoising model that need to be adjusted urgently, retrain the model and process it until the quality requirements are met;

[0137] The denoised image that meets the quality requirements is output as the final high-quality denoising result to obtain a denoised image.

[0138] Specifically, the convolutional neural denoising network structure of this embodiment includes multiple convolutional layers, pooling layers and upsampling layers. The convolutional layer is responsible for extracting features, the pooling layer reduces the size of the feature map, and the upsampling layer restores the image resolution. For example, a 3x3 convolution kernel, 64 filters per layer, a ReLU activation function, maximum pooling and transposed convolution can be used for upsampling. This structure can effectively learn noise patterns and generate clear images. After the denoising model is trained, it is applied to each pixel. Assuming there is a 512x512 cancer image, the model will process all 262144 pixels. Each pixel and its surrounding area are used as input, and the denoised pixel value is obtained through the network forward propagation. This process is performed pixel by pixel, and a complete denoised image is finally generated. Image quality assessment is a key step in measuring the denoising effect.

[0139] The quality improvement of the denoised image is quantified by two indicators: peak signal-to-noise ratio (PSNR) and structural similarity (SSIM); PSNR measures the overall quality of the image, and the higher the value, the better. Generally, 30 dB or more indicates good quality. SSIM focuses on the degree of retention of structural information, and the value range is 0 to 1, and the closer to 1, the better. For example, if the PSNR of the denoised image reaches 35 dB and the SSIM is 0.95, it can be considered that the denoising effect is good. If the denoising effect is not ideal, you need to return to adjust the model parameters. You can try to increase the network depth, change the convolution kernel size, adjust the learning rate, etc. For example, increase the original 8-layer network to 12 layers, change the convolution kernel from 3x3 to 5x5, and reduce the learning rate from 0.001 to 0.0001. These adjustments may bring better denoising effects, but they also need to weigh the computational complexity and the risk of overfitting. Finally, high-quality denoised images will be output. These images not only have effective noise suppression, but also retain the key features of cancer tissue. Doctors can make more accurate diagnoses based on these clear images, improving the accuracy and timeliness of cancer detection. At the same time, these denoised images can also be used for subsequent image analysis tasks, such as tumor segmentation and lesion classification, providing strong support for cancer research and treatment.

[0140] Embodiment 2

[0141] like Figure 2 As shown, based on the same inventive concept, this embodiment also provides a cancer image denoising system based on a multi-scale neural network, including:

[0142] An adaptive image segmentation module, used to adaptively segment cancer images and divide the images into different regions;

[0143] The noise estimation and parameter adjustment module is used to perform noise estimation and statistical analysis on each segmented area and automatically adjust the denoising algorithm parameters;

[0144] Multi-scale feature extraction and fusion module, used to extract multi-scale features using wavelet transform and pyramid decomposition, and analyze and process the image after parameter adjustment in different scale spaces;

[0145] Image quality assessment and correction module, which is used to make full use of local and global information of the image through feature fusion, perform quality assessment and distortion detection on the multi-scale processed image, automatically identify and locate artifacts and distortion, and use point spread function inversion and Wiener filtering methods to correct and compensate for interference to obtain a corrected image;

[0146] An image enhancement module, used for applying histogram equalization and contrast-limited adaptive histogram equalization to the corrected image and then using the image as a model training input to train a convolutional neural network model;

[0147] A machine learning model training module for learning noise and interference feature patterns using convolutional neural networks combined with sparse representation and dictionary learning algorithms, trained and optimized through cancer image samples;

[0148] The model evaluation and application module is used to cross-validate and evaluate the performance of the trained target convolutional neural network model, select the optimal model and apply it to the enhanced image to perform denoising and obtain the denoising result;

[0149] The adaptive closed-loop control module is used to evaluate the quality of the denoised image and dynamically adjust the model parameters based on the evaluation results.

[0150] The cancer image denoising system based on a multi-scale neural network provided in this embodiment has all the advantages of the cancer image denoising method based on a multi-scale neural network provided in the first embodiment.

[0151] Embodiment 3

[0152] This embodiment further discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in the first embodiment.

[0153] Embodiment 4

[0154] This embodiment further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in the first embodiment are implemented.

[0155] Embodiment 5

[0156] This embodiment also discloses a computer program product, including a computer program, which implements the steps of the method described in the first embodiment when executed by a processor.

[0157] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A cancer image denoising method based on a multi-scale neural network, characterized in that: include: Adaptively segment cancer images, divide the images into different regions, perform noise estimation and statistical analysis on each segmented region, and automatically adjust the denoising algorithm parameters; The wavelet transform algorithm is used to decompose the image into multiple scales to obtain wavelet coefficients at different scales. At each scale, the wavelet coefficients are threshold processed according to the preset threshold to remove noise and redundant information and obtain the processed wavelet coefficients; For the processed wavelet coefficients, a pyramid decomposition algorithm is used to extract local and global features of the image in different scale spaces to obtain a multi-scale feature set; According to the preset feature selection rules, the most effective feature subset for image analysis and processing is selected from the multi-scale feature set to obtain the optimized multi-scale features; By making full use of local and global information of the image through feature fusion, the quality assessment and distortion detection of the multi-scale processed image are carried out, the artifacts and distortion are automatically identified and located, and the point spread function inversion and Wiener filtering methods are used to correct and compensate for interference to obtain the corrected image; After applying histogram equalization and contrast-limited adaptive histogram equalization to the corrected image, the image is used as a model training input to train a convolutional neural network model; Use convolutional neural networks combined with sparse representation and dictionary learning algorithms to learn noise and interference feature patterns through training and optimization with cancer image samples; The trained target convolutional neural network model is cross-validated and performance evaluated, and the optimal model is selected and applied to the enhanced image for denoising to obtain the denoising result.

2. The method according to claim 1, characterized in that Adaptive segmentation of cancer images to divide the image into different regions involves: Obtain a cancer image, and use an adaptive segmentation algorithm to divide the image into different regions according to local features and texture information of the image to obtain segmented sub-regions; For each sub-region, extract the texture feature, grayscale feature and contrast feature of the sub-region to obtain a feature vector of the sub-region; According to the pre-established denoising method library, the feature vector of each sub-region is matched to obtain the denoising method of the corresponding sub-region.

3. The method according to claim 1, characterized in that The process of performing noise estimation and statistical analysis on each segmented area and automatically adjusting the parameters of the denoising algorithm includes: According to the image segmentation result, image data of each segmented area is obtained, and noise estimation is performed on each area to obtain the noise type, intensity and distribution characteristics of the area; Conduct statistical analysis on the noise parameters of each area, calculate the differences in noise parameters between different areas, and determine whether there are obvious differences in noise distribution; If there are obvious differences in noise distribution between regions, the optimal parameter combination is selected from the preset denoising algorithm parameter range according to the noise characteristics of different regions to obtain denoising algorithm parameters that are suitable for the noise characteristics of each region; Adaptive optimization strategy is adopted to continuously adjust the denoising algorithm parameters of each region through iterative optimization method to minimize the noise residue and detail loss of the denoised image and obtain the optimal denoising algorithm parameter combination for each region.

4. The method according to claim 1, characterized in that By making full use of local and global information of the image through feature fusion, quality assessment and distortion detection are performed on the multi-scale processed image. The process of automatically identifying and locating artifacts and distortion includes: The optimized multi-scale features are input into the feature fusion model, and the local information and the global information are fully utilized and fused through feature weighting and feature splicing to obtain the fused image feature representation; Based on the fused image feature representation, according to the preset image quality assessment index, a support vector machine algorithm is used to perform quality assessment on the fused image to obtain an image quality score; According to the image quality score, determine whether the image is distorted. If distortion is present, further extract multi-scale features of the image, analyze the multi-scale features using a convolutional neural network algorithm, and automatically identify the type of distortion; According to the identified distortion type, a multi-scale decomposition method based on wavelet transform is used to decompose the image into multiple scale sub-images, and the distorted area is detected on each scale sub-image to obtain the position coordinates of the distorted area.

5. The method according to claim 1, characterized in that: The process of correcting and compensating interference by using point spread function inversion and Wiener filtering to obtain a corrected image includes: According to the position coordinates of the distorted area, the image block of the corresponding area is extracted, and the point spread function parameters of the area are estimated by solving the inverse problem of the point spread function to obtain the distorted point spread function model; Performing deconvolution processing on the image block in the distorted area through the point spread function model to eliminate the distortion caused by the point spread effect and obtain a corrected image block; Perform Wiener filtering on the corrected image block in the frequency domain to suppress noise interference and obtain a filtered image block; The filtered image blocks are replaced back to the corresponding positions of the original images to complete the correction and compensation of the distorted areas, and a high-quality image after correction and compensation is output.

6. The method according to claim 1, characterized in that The process of applying histogram equalization and contrast limited adaptive histogram equalization to the corrected image includes: Performing histogram equalization processing on the corrected image to obtain a first equalized image; Using a contrast-limited adaptive histogram equalization algorithm to process the first equalized image to obtain a second equalized image; By calculating the contrast and clarity indexes of the second equalized image, it is determined whether a preset contrast threshold and clarity threshold are met, and if so, the second equalized image is determined to be an optimized image; If the contrast and clarity indicators of the second equalized image do not meet the preset thresholds, adaptively adjusting the parameters of the contrast-limited adaptive histogram equalization algorithm according to the histogram distribution characteristics of the second equalized image, and reprocessing the first equalized image until a third equalized image that meets the preset contrast and clarity thresholds is obtained; The equalized image that meets the preset contrast and clarity thresholds is used as the optimized image, and the image quality evaluation index of the optimized image is obtained; the image quality evaluation index of the optimized image includes contrast, clarity, and color saturation.

7. The method according to claim 1, characterized in that Using convolutional neural networks combined with sparse representation and dictionary learning algorithms, the process of learning noise and interference feature patterns through cancer image sample training and optimization includes: Design the model network structure and model parameters according to the characteristics of cancer images, and build a convolutional neural network model; Based on the convolutional neural network model, sparse representation and dictionary learning algorithms are introduced to extract robust feature representation through sparse decomposition of image data and dictionary construction; Apply the trained convolutional neural network model and sparse representation dictionary to new cancer image data to extract and represent features of the images; Based on the extracted features, a classifier or regression model is used to identify and analyze cancer images, and the identification and analysis results are compared with the results of expert annotation to calculate the accuracy, sensitivity and specificity indicators of the model; Based on the evaluation results, the model is iteratively optimized and improved, the network structure is adjusted, training data is increased, and new algorithms are introduced.

8. The method according to claim 1, characterized in that The trained target convolutional neural network model is cross-validated and performance evaluated, and the optimal model is selected and applied to the enhanced image for denoising. The process of obtaining the denoising result includes: Based on the trained target convolutional neural network model, multi-scale features of the image are extracted through convolution and pooling operations, and a denoising model is trained using the multi-scale features; Apply the trained denoising model to each pixel of the image data, calculate the denoising result of each pixel through forward propagation, and obtain the denoised image; The quality of the denoised image is evaluated, and the peak signal-to-noise ratio and structural similarity indicators are used to quantify the quality improvement of the denoised image; Based on the quality assessment results, determine whether the denoising effect has reached the expected goal. If not, return the parameters of the denoising model that need to be adjusted urgently, retrain the model and process it until the quality requirements are met; The denoised image that meets the quality requirements is output as the final high-quality denoising result to obtain a denoised image.

9. A cancer image denoising system based on a multi-scale neural network, characterized in that: include: An adaptive image segmentation module, used to adaptively segment cancer images and divide the images into different regions; The noise estimation and parameter adjustment module is used to perform noise estimation and statistical analysis on each segmented area and automatically adjust the denoising algorithm parameters; The multi-scale feature extraction and fusion module is used to use the wavelet transform algorithm to perform multi-scale decomposition on the image to obtain wavelet coefficients at different scales; at each scale, the wavelet coefficients are threshold processed according to the preset threshold value to remove noise and redundant information to obtain the processed wavelet coefficients; for the processed wavelet coefficients, the pyramid decomposition algorithm is used to extract the local and global features of the image at different scales to obtain a multi-scale feature set; according to the preset feature selection rules, the most effective feature subset for image analysis and processing is selected from the multi-scale feature set to obtain the optimized multi-scale features; Image quality assessment and correction module, which is used to make full use of local and global information of the image through feature fusion, perform quality assessment and distortion detection on the multi-scale processed image, automatically identify and locate artifacts and distortion, and use point spread function inversion and Wiener filtering methods to correct and compensate for interference to obtain a corrected image; An image enhancement module, used for applying histogram equalization and contrast-limited adaptive histogram equalization to the corrected image and then using the image as a model training input to train a convolutional neural network model; A machine learning model training module for learning noise and interference feature patterns using convolutional neural networks combined with sparse representation and dictionary learning algorithms, trained and optimized through cancer image samples; The model evaluation and application module is used to cross-validate and evaluate the performance of the trained target convolutional neural network model, select the optimal model and apply it to the enhanced image to perform denoising and obtain the denoising result; The adaptive closed-loop control module is used to evaluate the quality of the denoised image and dynamically adjust the model parameters based on the evaluation results.

Images