Adaptive Enhancement Method and System for Low-Contrast Regions of Medical Images

Abstract

The present invention discloses an adaptive enhancement method and system for low-contrast regions of medical images, which relates to the technical field of medical image processing. The method includes: receiving a set of target medical images; performing deformation measurement, and performing image registration according to the measurement result to generate a first registered image and a second registered image; performing multi-level fusion on the first registered image and the second registered image to generate a fused medical image; performing local contrast calculation under a sliding window to identify the fused low-contrast regions; performing multi-feature analysis on the fused low-contrast regions to generate a fused enhanced image. The present invention solves the technical problem in the prior art that it is difficult to effectively identify and enhance low-contrast regions when processing multi-modal medical images, resulting in low quality and contrast of medical images, and achieves the technical effect of realizing adaptive enhancement of low-contrast regions in multi-modal medical images and significantly improving the quality and contrast of medical images.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to an adaptive enhancement method and system for low-contrast regions of medical images. Background Art

[0002] In medical diagnosis, technologies such as CT, MRI, and ultrasonic imaging are important means for obtaining human body information. However, due to the influence of imaging principles, equipment performance, human tissue characteristics, and imaging environment, low-contrast regions often appear in medical images. For example, in CT, the difference in X-ray absorption of some tissues is small, resulting in similar gray values and making it difficult to distinguish structures and lesions; the signal intensities of soft tissues in MRI are similar, increasing the difficulty of identification; ultrasonic imaging also has poor contrast due to problems in sound wave propagation. Existing medical image processing technologies have deficiencies in dealing with low-contrast regions. Traditional enhancement methods such as histogram equalization are prone to amplifying noise and lack strong targeting for local regions. Local enhancement algorithms are difficult to adapt to complex image data of different modalities. When fusing multi-modal images, how to identify and enhance low-contrast regions remains a difficult problem.

[0003] There is a technical problem in the prior art that when processing multi-modal medical images, it is difficult to effectively identify and enhance low-contrast regions, resulting in low quality and contrast of medical images. Summary of the Invention

[0004] The present application provides an adaptive enhancement method and system for low-contrast regions of medical images, which are used to solve the technical problem in the prior art that when processing multi-modal medical images, it is difficult to effectively identify and enhance low-contrast regions, resulting in low quality and contrast of medical images.

[0005] In view of the above problems, the present application provides an adaptive enhancement method and system for low-contrast regions of medical images.

[0006] In the first aspect of the present application, there is provided an adaptive enhancement method for low-contrast regions of medical images, and the method includes:

[0007] Receive a set of target medical images, where the set of target medical images includes at least a first medical image corresponding to a first modality and a second medical image corresponding to a second modality; perform deformation measurement on the first medical image and the second medical image, and perform image registration according to the measurement result to generate a first registered image and a second registered image, where image registration includes rigid registration or elastic registration; perform multi-level fusion on the first registered image and the second registered image through multi-scale fusion to generate a fused medical image; perform local contrast calculation under a sliding window for the fused medical image to identify a fused low-contrast region; perform multi-feature analysis of the gray-scale distribution dynamics, gray-scale gradient, and frequency-domain features of the fused low-contrast region, and based on the multi-feature analysis result, call a pre-trained adaptive enhancement module to perform joint enhancement processing of multiple contrast enhancement models to generate a fused enhanced image.

[0008] In the second aspect of the present application, a system for adaptively enhancing low-contrast regions of medical images is provided. The system includes:

[0009] A medical image set acquisition module for receiving a set of target medical images, where the set of target medical images includes at least a first medical image corresponding to a first modality and a second medical image corresponding to a second modality; a registered image generation module for performing deformation measurement on the first medical image and the second medical image, and performing image registration according to the measurement result to generate a first registered image and a second registered image, where image registration includes rigid registration or elastic registration; a fused medical image generation module for performing multi-level fusion on the first registered image and the second registered image through multi-scale fusion to generate a fused medical image; a fused low-contrast region identification module for performing local contrast calculation under a sliding window for the fused medical image to identify a fused low-contrast region; a fused enhanced image generation module for performing multi-feature analysis of the gray-scale distribution dynamics, gray-scale gradient, and frequency-domain features of the fused low-contrast region, and based on the multi-feature analysis result, calling a pre-trained adaptive enhancement module to perform joint enhancement processing of multiple contrast enhancement models to generate a fused enhanced image.

[0010] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0011] Receive a set of target medical images; perform deformation measurement on the first medical image and the second medical image, perform image registration according to the measurement result to generate a first registered image and a second registered image; perform multi-level fusion on the first registered image and the second registered image through multi-scale fusion to generate a fused medical image; perform local contrast calculation under a sliding window on the fused medical image to identify a fused low-contrast region; perform multi-feature analysis on the fused low-contrast region, and based on the multi-feature analysis result, call a pre-trained adaptive enhancement module to perform joint enhancement processing of multiple contrast enhancement models to generate a fused enhanced image. It achieves the technical effect of realizing the adaptive enhancement of low-contrast regions in multi-modal medical images and significantly improving the quality and contrast of medical images. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0013] Figure 1 Schematic flowchart of the method for adaptively enhancing low-contrast regions of medical images provided by the embodiments of the present application;

[0014] Figure 2 Schematic structural diagram of the system for adaptively enhancing low-contrast regions of medical images provided by the embodiments of the present application.

[0015] Explanation of reference numerals: Medical image set acquisition module 10, registered image generation module 20, fused medical image generation module 30, fused low-contrast region identification module 40, fused enhanced image generation module 50. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The present application provides a method and system for adaptively enhancing low-contrast regions of medical images, which are used to solve the technical problem that in the prior art, when processing multi-modal medical images, it is difficult to effectively identify and enhance low-contrast regions, resulting in low quality and contrast of medical images.

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0018] Embodiment 1, as Figure 1As shown, the present application provides a method for adaptively enhancing low-contrast regions of medical images, and the method includes:

[0019] Step S100: Receive a set of target medical images, where the set of target medical images includes at least a first medical image corresponding to a first modality and a second medical image corresponding to a second modality.

[0020] Specifically, the received set of target medical images covers at least two different modalities of medical images, namely a first medical image corresponding to a first modality and a second medical image corresponding to a second modality, such as a first medical image in the CT modality and a second medical image in the PET modality. These images are all from the same user at the same location, and they carry different types of medical information. For example, CT images are good at presenting anatomical structures, while PET images highlight functional and metabolically active regions. The differences and complements between the two provide a rich data basis for subsequent processing. These images are the core data sources for subsequent operations such as image registration, fusion, and enhancement of low-contrast regions.

[0021] Step S200: Perform a deformation metric on the first medical image and the second medical image, and perform image registration according to the metric result to generate a first registered image and a second registered image, where the image registration includes rigid registration or elastic registration.

[0022] Specifically, for the first medical image and the second medical image, a deformation metric needs to be calculated first. The specific approach is to calculate the mutual information coefficient between the two, which can reflect the similarity degree and information correlation between the two images. When calculating the mutual information coefficient, first separately count the gray-level histograms of the first and second medical images to obtain the frequency of each gray value; then count the joint gray-level histogram of these two images to determine the frequency of the occurrence of specific gray value pairs; finally, according to the formula, divide the frequency of the joint gray-level histogram by the product of the gray-level histogram frequencies of the two images respectively, take the logarithm and then accumulate to calculate the mutual information coefficient. The larger this coefficient is, the higher the similarity between the two images. Based on this mutual information coefficient, the degree of deformation is analyzed. By comparing with a pre-set mutual information coefficient threshold obtained by training with historical rigid registration image samples, a suitable registration method is selected. If the mutual information coefficient meets the constraints of the rigid registration deformation condition, rigid registration is selected, that is, the translational alignment is initialized using the centroids of the first and second medical images, and then the translational and rotational parameters are continuously adjusted using the gradient descent method to maximize the mutual information between the two images; if not, elastic registration is adopted. First, the rigid registration is used to perform a preliminary alignment of the images, and then a deformation field model is constructed based on B-splines. Similarly, the gradient descent method is used to calculate the deformation field to minimize the difference between the first medical image and the second medical image. After the above rigid registration or elastic registration process, the first registered image and the second registered image are finally generated, laying a foundation for subsequent multi-scale fusion and enhancement processing of the low-contrast regions of medical images, ensuring that the information in different modality images can be accurately corresponding and integrated.

[0023] Step S300: Perform multi-level fusion on the first registered image and the second registered image through multi-scale fusion to generate a fused medical image.

[0024] Specifically, the two registered images are fused at multiple levels through multi-scale fusion technology to generate a fused medical image. First, the first registered image and the second registered image are respectively subjected to pyramid decomposition to construct Gaussian pyramids, and the images are decomposed into multiple levels of different resolutions. Each level contains different degrees of detailed information of the images. Then, at each resolution level of the Gaussian pyramid, pixel selection is performed by taking the maximum value. More representative pixel values are selected from the pixels of the corresponding levels of the two images to generate the fused image corresponding to each resolution level, and the features of the two images are initially integrated. Then, starting from the fused image of the bottom layer, the fusion result of each layer of the image is synthesized with the upper-level image layer by layer to construct a Laplacian pyramid image, further refining and optimizing the fusion effect. Finally, based on the Laplacian pyramid image, on the image of the highest resolution layer, the details of all low-resolution levels are gradually superimposed, and the information at different resolutions is completely combined to reconstruct the fused medical image, realizing the comprehensive integration of information of different modality medical images and providing richer and more accurate image data for subsequent analysis and processing.

[0025] Step S400: Calculate the local contrast under a sliding window for the fused medical image to identify the fused low-contrast regions.

[0026] Specifically, in order to identify the low-contrast regions in the fused medical image, local contrast calculation is performed by means of a sliding window. First, a sliding window of a preset size is constructed. The size of this window is preset according to a large amount of experimental data and the characteristics of medical images to ensure that low-contrast regions can be effectively detected. Then, let this sliding window slide pixel by pixel on the fused medical image. At each sliding position, calculate the contrast index of the pixels within the window, which is calculated based on factors such as the gray value difference of the pixels within the window and can quantify the contrast situation of the image region within the window. Finally, compare the calculated pixel contrast index with the preset contrast index, and extract the image blocks corresponding to the sliding windows where the pixel contrast index is less than or equal to the preset contrast index. These extracted image blocks constitute the fused low-contrast regions, laying a foundation for enhancing the contrast of these regions specifically in the future.

[0027] Step S500: Perform multi-feature analysis on the fused low-contrast regions in terms of dynamic gray distribution, gray gradient, and frequency domain features. Based on the results of the multi-feature analysis, call the pre-trained adaptive enhancement module to perform joint enhancement processing of multiple contrast enhancement models to generate a fused enhanced image.

[0028] Specifically, multi-feature analysis of dynamic gray-scale distribution, gray-scale gradient, and frequency-domain features is carried out on these regions. Dynamic gray-scale distribution analysis can understand the distribution change law of pixel gray-scale values. Gray-scale gradient analysis can determine the intensity and direction of gray-scale changes in the image. Frequency-domain feature analysis helps to grasp the features of the image under different frequency components. Based on historical data statistics, a preset low-contrast factor set is determined in advance. Multiple groups of contrast enhancement samples including original medical image samples, enhanced medical image samples, and corresponding multi-feature analysis samples are also collected. Multiple contrast enhancement models are constructed and combined into a model network. Through these samples, block fusion enhancement training is carried out on the model network to obtain a pre-trained adaptive enhancement module. According to the multi-feature analysis results, this module is called to enable multiple contrast enhancement models to work together and perform joint enhancement processing on the fused low-contrast regions. These models improve the display effect of low-contrast regions from different angles according to their respective characteristics, and finally generate a fused enhanced image to improve the overall visibility of the image.

[0029] In a possible implementation manner, step S100 further includes:

[0030] Step S110: The first modality and the second modality are different medical imaging modalities, and the first medical image and the second medical image are medical images of the same user in different medical imaging modalities at the same position.

[0031] Specifically, the first modality and the second modality must be different medical imaging modalities. Common combinations include CT (Computed Tomography) and PET (Positron Emission Tomography). CT can clearly present the anatomical structure details of the human body, while PET is good at highlighting functional and metabolically active regions. There is also CT and MRI (Magnetic Resonance Imaging). MRI has unique advantages in soft tissue imaging and can provide information complementary to CT. The first medical image and the second medical image both originate from the same user at the same position. This condition ensures that the images reflect the same body part or region of interest. Imaging the same part with different modalities can obtain medical information from multiple angles and provide rich and comprehensive data support for subsequent image fusion, low-contrast region enhancement, and accurate disease diagnosis.

[0032] In a possible implementation manner, step S200 further includes:

[0033] Step S210: Calculate the mutual information between the first medical image and the second medical image to generate a mutual information coefficient.

[0034] Step S220: Perform deformation degree analysis based on the mutual information coefficient, and screen for rigid registration and elastic registration according to the analysis results to determine the target registration scheme.

[0035] Step S230: Register the first medical image and the second medical image based on the target registration scheme to generate the first registered image and the second registered image.

[0036] Specifically, perform digital discrete processing on the first medical image and the second medical image to convert them into a digital image form based on a pixel matrix. For the first medical image, traverse its pixel matrix, count the frequency n of each gray value i i , and combine it with the total number of pixels N in the image to calculate the probability distribution of each gray value , thereby constructing the gray histogram of the first medical image; similarly, perform the same operation on the second medical image to obtain its gray value probability distribution q(j). Then, synchronously traverse the pixel matrices of the two images, count the frequency n of pixel pairs where the gray value of the first medical image is i and the gray value of the second medical image is j ij , and calculate its joint probability distribution , and then construct the joint gray histogram. Use the mutual information calculation formula (where represent the first and second medical images respectively, are the gray levels of the first and second medical images respectively), and calculate the mutual information coefficient.

[0037] Based on the generated mutual information coefficient, perform deformation degree analysis, and construct a rigid registration deformation condition constraint, which is a threshold of the mutual information coefficient obtained by training with historical rigid registration image samples. Compare the calculated mutual information coefficient with this threshold. If the mutual information coefficient satisfies the rigid registration deformation condition constraint, it means that the deformation degree of the two images is relatively small and rigid registration is suitable; if not, it indicates that the image deformation is more complex and elastic registration is required. Thus, determine the target registration scheme to ensure the accuracy and effectiveness of subsequent registration operations.

[0038] Register the first medical image and the second medical image according to the determined target registration scheme. If rigid registration is selected, the center of gravity of the image will be used for initial translational alignment, and then the translational and rotational parameters will be adjusted by the gradient descent method to maximize the mutual information between the two images; if elastic registration is adopted, first perform preliminary alignment using rigid registration, and then construct a deformation field model based on B-splines, and calculate the deformation field by the gradient descent method to minimize the difference between the images. After this registration process, the first registered image and the second registered image are successfully generated, laying a solid foundation for subsequent multi-scale fusion and enhancement processing of low-contrast regions in medical images.

[0039] In a possible implementation manner, step S220 further includes:

[0040] Step S221: Construct a rigid registration deformation condition constraint, where the rigid registration deformation condition constraint is a mutual information coefficient threshold trained by historical rigid registration image samples.

[0041] Step S222: Determine whether the mutual information coefficient satisfies the rigid registration deformation condition constraint.

[0042] Step S223: If so, generate the target registration scheme by rigid registration.

[0043] Step S224: If not, generate the target registration scheme by elastic registration.

[0044] Specifically, to construct a rigid registration deformation condition constraint, the support vector machine (SVM) algorithm is used to determine the mutual information coefficient threshold based on historical rigid registration image samples. First, a large number of historical rigid registration image samples in different scenarios are collected, and the mutual information coefficient is calculated for each sample. These samples are divided into two categories: suitable for rigid registration and not suitable for rigid registration, and labels are assigned respectively. Then, the mutual information coefficient is used as a feature, and the sample label is used as the target variable and input into the SVM model. The SVM will find an optimal hyperplane in the feature space to separate the two categories of samples as clearly as possible. During the training process, by adjusting the parameters of the model, such as the penalty factor and the kernel function parameters, methods such as cross-validation are used to evaluate the model performance, and the model is continuously optimized to achieve the best classification effect. After training, a specific mutual information coefficient value can be determined at the intersection or near the intersection of the hyperplane and the feature axis, and it is used as the threshold of the rigid registration deformation condition constraint. For subsequent new image registration tasks, it can be judged whether rigid registration is suitable based on this threshold.

[0045] Accurately calculate the mutual information coefficient of these two images, which reflects the correlation and overlap degree of information between the two images. Then, compare the calculated mutual information coefficient with the mutual information coefficient threshold in the previously constructed rigid registration deformation condition constraint. If the calculated mutual information coefficient is greater than or equal to the threshold, it indicates that these two images satisfy the rigid registration deformation condition constraint under the current conditions, meaning that the feature correspondence between them is relatively stable, and it is more likely to achieve accurate registration by rigid registration; conversely, if the calculated mutual information coefficient is less than the threshold, it means that the rigid registration deformation condition constraint is not satisfied. At this time, the feature correspondence between the images is relatively complex, and rigid registration may not achieve the ideal registration effect, and other more flexible registration methods such as elastic registration need to be considered.

[0046] If the mutual information coefficient satisfies the constraints of the rigid registration deformation condition, that is, the calculated mutual information coefficient is greater than or equal to the pre-constructed mutual information coefficient threshold, it indicates that the feature correspondence between the first medical image and the second medical image is relatively stable, and rigid registration is suitable. At this time, a target registration scheme is generated by rigid registration. The rigid registration process first uses the centroids of the first medical image and the second medical image for initial translational alignment to ensure a preliminary match in their approximate positions. Then, with the help of the gradient descent method, the translational and rotational parameters are continuously adjusted. During this process, the mutual information between the images is continuously calculated, and with the goal of maximizing the mutual information, the registration parameters are gradually optimized until the combination of translational and rotational parameters that maximizes the mutual information between the two images is found, thereby achieving precise spatial alignment of the images and completing the generation of the target registration scheme.

[0047] If the mutual information coefficient does not satisfy the constraints of the rigid registration deformation condition, that is, the calculated mutual information coefficient is less than the pre-constructed mutual information coefficient threshold, it indicates that the feature correspondence between the images is relatively complex, and rigid registration may not achieve the ideal registration effect. At this time, a target registration scheme needs to be generated by elastic registration. However, before performing elastic registration, rigid registration still needs to be carried out first. The first medical image and the second medical image are initially aligned by rigid registration, and the two images are made close in their approximate positions through translational and rotational operations. Then, a deformation field model is constructed based on B-splines, which can describe the local deformation of the image. Next, the deformation field is calculated by the gradient descent method, and the parameters of the deformation field are continuously adjusted during the calculation to minimize the difference between the two images, thereby achieving more precise image registration and finally generating the target registration scheme.

[0048] In a possible implementation manner, step S223 further includes:

[0049] Step S2231: The rigid registration uses the centroids of the first medical image and the second medical image for initial translational alignment, and uses the gradient descent method to maximize the mutual information between the first medical image and the second medical image by adjusting the translational and rotational parameters.

[0050] Step S2232: The elastic registration first uses rigid registration to preliminarily align the first medical image and the second medical image, constructs a deformation field model based on B-splines, and calculates the deformation field by the gradient descent method to minimize the difference between the first medical image and the second medical image.

[0051] Specifically, when performing rigid registration to generate a target registration scheme, the centroid calculations of the first medical image and the second medical image are first carried out. By analyzing information such as the pixel distribution of these two images respectively, the centroid positions of each of them are determined. Then, based on these two centroid positions, an initial translation operation is performed on the images to preliminarily align the first medical image and the second medical image in space, making their positional relationship closer. After that, the gradient descent method is used to further optimize the registration effect. The gradient descent method continuously adjusts the translation and rotation parameters of the images. After each adjustment of the parameters, the mutual information between the first medical image and the second medical image is calculated. The mutual information can reflect the correlation and information overlap degree between these two images. With the goal of maximizing the mutual information, the parameters are continuously iteratively adjusted until a set of translation and rotation parameters is found such that the mutual information between the two images reaches the maximum value. At this time, the rigid registration process is completed, and a relatively accurate registration result is obtained.

[0052] When performing elastic registration, first use rigid registration (including translation and rotation operations) to preliminarily align the first medical image and the second medical image. The purpose of this step is to reduce the computational amount of subsequent non-rigid registration, bring the two images closer in the overall position first, and lay a foundation for subsequent more refined registration. After completing the preliminary alignment, construct a deformation field model based on B-splines. B-splines (basis spline functions) are a commonly used mathematical model that describes the local deformation of an image through a set of control points. These control points can be flexibly distributed in the image as needed. Each control point has a corresponding weight, and they jointly determine the deformation mode of the local area of the image. B-splines have good smoothness and local controllability, and can well adapt to the complex shape changes in medical images. For example, in medical images, the shape of an organ may undergo local distortion and deformation due to physiological conditions, diseases, etc., and the B-splines deformation field model can accurately capture and describe these changes. After constructing the deformation field model, calculate the deformation field by the gradient descent method. The gradient descent method is an optimization algorithm that uses an energy function to measure the similarity between the first medical image and the second medical image, such as mean square error, mutual information, structural similarity index, etc. During the calculation process, continuously adjust the parameters of the deformation field, and update the parameters along the direction that makes the value of the energy function decrease fastest (i.e., the opposite direction of the gradient), so that the difference between the two images is gradually minimized. In each iteration, a new deformation field will be calculated according to the current deformation field parameters, and this deformation field will be applied to the target image to gradually align it with the reference image. In order to make the deformation field be smoothly applied to the image, interpolation methods such as cubic interpolation or B-spline interpolation are usually used. Through interpolation, continuous deformation can be generated between the control points to ensure the smoothness and naturalness of the image deformation. During the entire elastic registration process, multiple iterations and optimizations are also required to gradually adjust the deformation field until the alignment error between the first medical image and the second medical image reaches the minimum value, thereby achieving accurate elastic registration.

[0053] In a possible implementation manner, step S300 further includes:

[0054] Step S310: Perform pyramid decomposition on the first registered image and the second registered image respectively to construct a Gaussian pyramid.

[0055] Step S320: Select the maximum value for the pixel points at each resolution level in the Gaussian pyramid to generate a fused image corresponding to each resolution level.

[0056] Step S330: Based on the fused images corresponding to each resolution level, starting from the fused image of the bottommost layer, layer by layer, synthesize the fusion result of each layer of the image with the upper-level image to generate a high-resolution image until the highest resolution level is reached, and construct a Laplacian pyramid image.

[0057] Step S340: Based on the Laplacian pyramid image, on the image of the highest resolution level, gradually superimpose the details of all low-resolution levels to reconstruct the fused medical image.

[0058] Specifically, pyramid decomposition is performed on the first registered image and the second registered image respectively to construct a Gaussian pyramid. A Gaussian pyramid is an image pyramid obtained by downsampling through Gaussian filtering, which can present image information at different scales. During specific operations, for the first registered image, first apply a Gaussian filter to it. This filter will perform weighted averaging on each pixel in the image and its neighborhood, playing a role in smoothing the image and reducing noise interference in the image. Then, downsample the image after Gaussian filtering. Usually, the size of the image is reduced by half, for example, from the original width and height of 2n pixels to width and height of n pixels, so as to obtain the first-level lower-resolution image, which retains the global structure of the first registered image. Then, repeat the above Gaussian filtering and downsampling operations on this lower-resolution image, and a lower-resolution image at a lower level will be obtained. Iterate continuously in this way until the preset number of pyramid levels is reached. For the second registered image, it is also processed in the same way. Finally, a complete Gaussian pyramid is constructed for the first registered image and the second registered image respectively, laying a foundation for subsequent image fusion operations.

[0059] For each resolution level of the Gaussian pyramid, a pixel-based fusion operation will be performed. For the first registered image and the second registered image at the same resolution level, compare the pixel points at their corresponding positions. At each pixel position, take out the pixel value of the first registered image at this position, and at the same time take out the pixel value of the second registered image at the same position, and then compare the sizes of these two pixel values. Select the larger of the two pixel values as the pixel value at this position after fusion. In this way, perform the above maximum value selection operation on all pixel points at the same resolution level, and a new image can be obtained. This new image is the fused image corresponding to this resolution level. For each resolution level of the Gaussian pyramid, repeat the operation of selecting the maximum value for pixel points, and finally generate the fused image corresponding to each resolution level. These fused images retain as much as possible the more significant feature information in the first registered image and the second registered image at their respective resolution levels, preparing for the subsequent further synthesis of a high-resolution fused image.

[0060] After obtaining the fused images corresponding to each resolution level, the bottom layer refers to the layer with the highest resolution in the Gaussian pyramid, and the fused image of this layer is used as the starting point. The Laplacian pyramid is constructed to better retain the detailed information of the image, so as to finally reconstruct a high-quality fused medical image. Take the fused image of the bottom layer (the highest resolution) as the basis. Then, take the fused image of its upper level (one level lower in resolution), and perform an upsampling operation on the image at this level, that is, enlarge the size of the image to make its resolution close to that of the bottom layer image. Upsampling usually makes the image blurred, and at this time, it is necessary to combine the information of the bottom layer image to restore the details. By performing a difference operation between the upsampled image and the bottom layer image, the detailed information between this layer image and the bottom layer image is obtained, that is, one layer of the Laplacian image. Then, the obtained Laplacian image of this layer is synthesized with the bottom layer image to obtain a new image. After that, take the fused image of the next higher level (one level lower in resolution), and repeat the above operations of upsampling, calculating the difference, and synthesizing, continuously synthesize the fusion result of each layer image with the upper level image, and gradually restore the details of the image. As the processing progresses layer by layer like this, continuously fuse the image information of the low-resolution layer with the high-resolution layer until reaching the highest resolution layer (that is, the top layer of the Gaussian pyramid, the layer with the lowest resolution). Through such a layer-by-layer synthesis process, the Laplacian pyramid image is finally constructed. Each layer of the Laplacian pyramid contains the detailed information of the image at different resolutions, providing an important data basis for reconstructing the fused medical image based on these detailed information in the future.

[0061] The Laplacian pyramid image contains the detailed information of the image at different resolution levels, and these information are the key to reconstructing the fused medical image. At this time, based on the image of the highest resolution layer in the Laplacian pyramid, because the image of this layer retains the finest part of the fused image. Starting from the next lower resolution level adjacent to it, gradually superimpose the detailed information contained in the Laplacian pyramid image of this level on the image of the highest resolution layer. These detailed information represent the differences between the images at adjacent resolution levels. By superimposing, the information lost due to downsampling is gradually restored, making the image become more rich and complete. After superimposing the details of each lower resolution level, continue to perform the same operation on the next lower resolution level, and so on in turn until all the details of the lower resolution levels are superimposed on the image of the highest resolution layer. In this process, the details of the image are continuously enriched, and features such as texture and edges gradually become clear, and finally the fused medical image is reconstructed. The fused medical image integrates the features of the first registered image and the second registered image at different resolutions, providing a comprehensive and clear image basis for medical diagnosis and analysis.

[0062] In a possible implementation manner, step S500 further includes:

[0063] Step S510: Determine a preset low-contrast factor set, where the preset low-contrast factor set is based on historical data statistics.

[0064] Step S520: Collect multiple groups of contrast enhancement samples based on the preset low-contrast factor set. Any group of contrast enhancement samples includes an original medical image sample, an enhanced medical image sample, and corresponding multi-feature analysis samples.

[0065] Step S530: Construct multiple model networks for the multiple contrast enhancement models, and perform block fusion enhancement training on the multiple model networks with the multiple groups of contrast enhancement samples to generate the multiple contrast enhancement models, and construct the adaptive enhancement module.

[0066] Step S540: Invoke the adaptive enhancement module. The multiple contrast enhancement models perform enhancement processing on the fusion low-contrast regions in the fusion medical image based on the multi-feature analysis results to generate the fusion enhanced image.

[0067] Specifically, determine a preset low-contrast factor set. To accurately identify and process low-contrast regions in medical images, statistical analysis needs to be performed based on a large amount of historical medical image data. By sorting out and summarizing various relevant factors that appear in low-contrast regions in this historical data, such as parameter differences of different imaging devices, different physiological states of patients, and the impact of disease types on image contrast, etc., a factor set containing various possible factors leading to low-contrast situations is determined, and this set is the preset low-contrast factor set.

[0068] Collect multiple groups of contrast enhancement samples based on the preset low-contrast factor set. In actual operation, according to different factors in the set, eligible samples are specifically selected from a large number of medical images. Each group of contrast enhancement samples includes an original medical image sample, an enhanced medical image sample after professional enhancement processing, and corresponding multi-feature analysis samples obtained by performing multi-feature analysis such as dynamic gray-scale distribution, gray-scale gradient, and frequency domain features on the original medical image sample. These samples provide a rich data basis for subsequent training of the contrast enhancement model.

[0069] Multiple model networks for constructing multiple contrast enhancement models are based on four existing models, specifically covering CLAHE-based local enhancement, Retinex multi-scale enhancement (using the multi-scale Retinex (MSR) method to enhance local contrast), generative adversarial network (GAN) enhancement, and deep learning super-resolution reconstruction. First, according to the principles and structural characteristics of these models, the corresponding model networks are built respectively. For the CLAHE-based local enhancement model network, it can effectively improve the contrast of local regions of the image; the Retinex multi-scale enhancement model network can enhance local contrast at multiple scales using the MSR method; the generative adversarial network (GAN) enhancement model network optimizes the image contrast by means of the adversarial training mechanism of the generator and discriminator; the deep learning super-resolution reconstruction model network focuses on improving the resolution of the image to improve the contrast. Then, multiple groups of contrast enhancement samples collected are used to carry out block fusion enhancement training on these model networks. The sample blocks are input into each model network, and during the training process, the parameters of the models are continuously adjusted, and at the same time, the fusion strategy between the model networks is considered to make them cooperate with each other and complement each other's advantages. After multiple rounds of training and optimization, multiple contrast enhancement models with excellent performance are generated. Finally, these models are integrated together to construct an adaptive enhancement module for subsequent enhancement processing of low-contrast regions in the fused medical images.

[0070] The adaptive enhancement module is called to enhance the fused low-contrast regions in the fused medical images. After obtaining the fused medical images, multi-feature analysis is first performed on them to obtain information such as the dynamic gray-scale distribution, gray-scale gradient, and frequency-domain features of the low-contrast regions. These multi-feature analysis results are input into the adaptive enhancement module, and multiple contrast enhancement models in the module will cooperate to enhance the fused low-contrast regions according to the knowledge and methods they have learned respectively. Each model optimizes the low-contrast region from different angles. Some models may focus on enhancing the edge contrast, and some models may be better at adjusting the gray-scale distribution. Through their joint action, a fused enhanced image is finally generated, making the low-contrast regions in the fused medical images significantly improved.

[0071] In a possible implementation manner, step S530 further includes:

[0072] Step S531: Using the original medical image sample and the multi-feature analysis sample in any group of contrast enhancement samples as the training input, and using the enhanced medical image sample as the output supervision truth value, perform iterative adjustment of the training parameters of the multiple model networks, where the training parameters include the learning parameters corresponding to the multiple model networks respectively, and the enhancement fusion weights of the multiple model networks, and train the multiple model networks until convergence to generate the adaptive enhancement module.

[0073] Specifically, to construct the adaptive enhancement module, iterative adjustment of the training parameters of multiple model networks is required. Select any set of contrast enhancement samples, use the original medical image samples and multi-feature analysis samples therein as training inputs, and the enhanced medical image samples as the output supervision ground truth. During the training process, for multiple model networks, the training parameters to be adjusted include two parts: one is the learning parameters corresponding to each model network respectively, and these parameters control the ability of the model to capture and understand the features of the input data during the learning process; the other is the enhancement fusion weights of multiple model networks, and this weight determines the size of the role played by each model network in the final fusion result. After the training starts, multiple model networks process the input data according to the current training parameters and generate their respective output results. These output results are compared with the enhanced medical image samples used as the supervision ground truth, and the loss function value is calculated to measure the difference between the model output and the real result. Subsequently, based on the loss function value, the training parameters are adjusted using an optimization algorithm (stochastic gradient descent method). This process will be repeated continuously, and each iteration will make the model network closer to the optimal state. As the iteration continues, the loss function value will gradually decrease. When the loss function value decreases to a minimum value and does not change significantly in subsequent multiple iterations, it indicates that multiple model networks have been trained to convergence. At this time, the learning parameters and enhancement fusion weights of each model network have reached a relatively stable and optimal state. Finally, these trained model networks are integrated together to generate the adaptive enhancement module, which can effectively enhance the low-contrast regions in the image according to the input medical image and its multi-feature analysis results.

[0074] In a possible implementation manner, step S400 further includes:

[0075] Step S410: Construct a preset sliding window.

[0076] Step S420: Slide the preset sliding window on the fused medical image, and calculate the pixel contrast index under each sliding window.

[0077] Step S430: Extract the image block corresponding to the sliding window with the pixel contrast index equal to or equal to the preset contrast index to generate the fused low-contrast region.

[0078] Specifically, a preset sliding window is constructed. When constructing it, the characteristics of medical images and the requirements of subsequent calculations should be fully considered. First, determine the shape of the window. Due to its regularity and computational convenience, the rectangular window is a commonly used choice, which can facilitate the row-by-row and column-by-column scanning of the image. Then, set the window size, which requires comprehensive consideration of various factors. If the window size is too small, although the computational speed is fast, it may not cover enough pixel information to accurately reflect the contrast characteristics of the local area, and it is easy to miss low-contrast areas. If the window size is too large, although more information can be obtained, the computational amount will increase significantly, and it may also include the surrounding normal contrast areas in the judgment range of low-contrast areas, resulting in misjudgment. Generally, the size of common tissues and lesions in medical images is referred to, and the appropriate side length or radius is determined through multiple experiments and experience. For example, for lung CT fusion images, after a large amount of image analysis and testing, a rectangular window with a side length of 10-20 pixels may be selected. In addition, it can be appropriately adjusted according to the resolution of the image. Larger windows can be selected for high-resolution images, and smaller windows can be adapted for low-resolution images to ensure that the window can effectively cover the target area and does not lose key information, thereby constructing a preset sliding window that meets the requirements of subsequent low-contrast area recognition.

[0079] After the construction of the preset sliding window is completed, the preset sliding window is slid on the fused medical image. Here, the sliding window method is adopted, and the window moves pixel by pixel or at a certain step size on the image, covering different regions of the image. For each region covered by the sliding window, it is necessary to calculate its pixel contrast index. Here, the local contrast measurement (LCM) is used to calculate the pixel contrast index. According to the formula , where represents the pixel value within the local window, is a decimal number used to avoid the denominator being zero. Through this formula, the LCM value under each sliding window is calculated. This LCM value reflects the degree of difference in pixel values within the sliding window, that is, the contrast situation. A larger LCM value indicates a higher pixel contrast within the window, and the image details and features are relatively obvious; while a smaller LCM value indicates a lower pixel contrast within the window, and there may be low-contrast areas. Subsequently, this LCM value can be compared with the preset threshold (LCM < 0.1) to determine whether the region corresponding to the sliding window is a low-contrast area, providing data support for the subsequent accurate extraction of low-contrast areas.

[0080] Compare the calculated pixel contrast metric with a preset contrast metric (LCM < 0.1). For those sliding windows whose pixel contrast metrics are less than or equal to the preset contrast metric, extract the corresponding image blocks. These extracted image blocks exhibit relatively low contrast features in the fused medical image. Integrate all the image blocks that meet the conditions to generate a fused low-contrast region. This region contains the parts of the image with insufficient contrast and unclear details, providing a clear target and scope for subsequent targeted enhancement processing (such as enhancement through an adaptive enhancement module) of these regions, which helps improve the overall quality of the medical image.

[0081] Embodiment 2, based on the same inventive concept as the method for adaptively enhancing low-contrast regions in medical images in the foregoing embodiment, as Figure 2 shown, the present application provides a system for adaptively enhancing low-contrast regions in medical images. The system in the embodiments of the present application and the method embodiments are based on the same inventive concept. Among them, the system includes:

[0082] A medical image set acquisition module 10, configured to receive a target medical image set, where the target medical image set includes at least a first medical image corresponding to a first modality and a second medical image corresponding to a second modality.

[0083] A registered image generation module 20, configured to perform deformation measurement on the first medical image and the second medical image, perform image registration according to the measurement result, and generate a first registered image and a second registered image, where image registration includes rigid registration or elastic registration.

[0084] A fused medical image generation module 30, configured to perform multi-level fusion on the first registered image and the second registered image through multi-scale fusion to generate a fused medical image.

[0085] A fused low-contrast region identification module 40, configured to perform local contrast calculation under a sliding window for the fused medical image to identify a fused low-contrast region.

[0086] A fused enhanced image generation module 50, configured to perform multi-feature analysis of the gray distribution dynamics, gray gradient, and frequency domain features of the fused low-contrast region, and based on the multi-feature analysis result, call a pre-trained adaptive enhancement module to perform joint enhancement processing of multiple contrast enhancement models to generate a fused enhanced image.

[0087] Further, the system is also used to implement the following functions:

[0088] The first modality and the second modality are different medical imaging modalities, and the first medical image and the second medical image are medical images of the same user at the same location under different medical imaging modalities.

[0089] Furthermore, the system is also used to implement the following functions:

[0090] Calculate the mutual information between the first medical image and the second medical image to generate a mutual information coefficient; perform deformation degree analysis based on the mutual information coefficient, screen for rigid registration and elastic registration according to the analysis results, and determine the target registration scheme; register the first medical image and the second medical image based on the target registration scheme to generate the first registered image and the second registered image.

[0091] Furthermore, the system is also used to implement the following functions:

[0092] Construct a rigid registration deformation condition constraint, where the rigid registration deformation condition constraint is a mutual information coefficient threshold trained by historical rigid registration image samples; determine whether the mutual information coefficient satisfies the rigid registration deformation condition constraint; if so, generate the target registration scheme by rigid registration; if not, generate the target registration scheme by elastic registration.

[0093] Furthermore, the system is also used to implement the following functions:

[0094] The rigid registration uses the center of gravity of the first medical image and the second medical image for initial translational alignment, and uses the gradient descent method to maximize the mutual information between the first medical image and the second medical image by adjusting the translational and rotational parameters; the elastic registration first performs a preliminary alignment of the first medical image and the second medical image using rigid registration, constructs a deformation field model based on B-splines, and calculates the deformation field by the gradient descent method to minimize the difference between the first medical image and the second medical image.

[0095] Furthermore, the system is also used to implement the following functions:

[0096] Perform pyramid decomposition on the first registered image and the second registered image respectively to construct a Gaussian pyramid; perform maximum value selection on the pixel points at each resolution level in the Gaussian pyramid to generate a fused image corresponding to each resolution level; based on the fused images corresponding to each resolution level, starting from the fused image at the bottom layer, layer by layer synthesize the fusion result of each layer of the image with the previous-level image to generate a high-resolution image until the highest resolution layer is reached, and construct a Laplacian pyramid image; based on the Laplacian pyramid image, on the image at the highest resolution layer, gradually superimpose the details of all low-resolution levels to reconstruct the fused medical image.

[0097] Further, the system is also used to implement the following functions:

[0098] Determine a preset low-contrast factor set, where the preset low-contrast factor set is based on historical data statistics; collect multiple groups of contrast enhancement samples based on the preset low-contrast factor set, where any group of contrast enhancement samples includes an original medical image sample and an enhanced medical image sample, as well as corresponding multi-feature analysis samples; construct multiple model networks of the multiple contrast enhancement models, and perform block fusion enhancement training on the multiple model networks with the multiple groups of contrast enhancement samples to generate the multiple contrast enhancement models, and construct the adaptive enhancement module; call the adaptive enhancement module, and the multiple contrast enhancement models perform enhancement processing on the fused low-contrast regions in the fused medical image based on the multi-feature analysis results to generate the fused enhanced image.

[0099] Further, the system is also used to implement the following functions:

[0100] Use the original medical image sample and the multi-feature analysis sample in any group of contrast enhancement samples as training inputs, and use the enhanced medical image sample as the output supervision truth value to perform iterative adjustment of the training parameters of the multiple model networks, where the training parameters include learning parameters corresponding to the multiple model networks respectively, and the enhancement fusion weights of the multiple model networks, and train the multiple model networks until convergence to generate the adaptive enhancement module.

[0101] Further, the system is also used to implement the following functions:

[0102] Construct a preset sliding window; slide the preset sliding window on the fused medical image, calculate the pixel contrast index under each sliding window; extract the image block corresponding to the sliding window with the pixel contrast index equal to or greater than the preset contrast index to generate the fused low-contrast region.

[0103] It should be noted that the above order of the embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. In addition, the specific embodiments of this specification have been described. Moreover, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0104] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

[0105] This specification and the drawings are only exemplary descriptions of the present application and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application is intended to include these changes and modifications.

Claims

1. An adaptive enhancement method for low-contrast regions of medical images, characterized in that, Including: Receiving a set of target medical images, where the set of target medical images includes at least a first medical image corresponding to a first modality and a second medical image corresponding to a second modality; Performing deformation measurement on the first medical image and the second medical image, and performing image registration according to the measurement result to generate a first registered image and a second registered image, where image registration includes rigid registration or elastic registration; Performing multi-level fusion on the first registered image and the second registered image through multi-scale fusion to generate a fused medical image; Calculating local contrast under a sliding window for the fused medical image to identify fused low-contrast regions; Performing multi-feature analysis on the fused low-contrast regions for dynamic gray-scale distribution, gray-scale gradient, and frequency-domain features. Based on the multi-feature analysis result, calling a pre-trained adaptive enhancement module to perform joint enhancement processing on multiple contrast enhancement models to generate a fused enhanced image; The first modality and the second modality are different medical imaging modalities, and the first medical image and the second medical image are medical images of the same user in different medical imaging modalities at the same position; Performing deformation measurement on the first medical image and the second medical image, and performing image registration according to the measurement result to generate a first registered image and a second registered image, including: Calculating mutual information between the first medical image and the second medical image to generate a mutual information coefficient; Performing deformation degree analysis based on the mutual information coefficient, screening rigid registration and elastic registration according to the analysis result, and determining a target registration scheme; Performing registration on the first medical image and the second medical image based on the target registration scheme to generate the first registered image and the second registered image; Performing deformation degree analysis based on the mutual information coefficient, screening rigid registration and elastic registration according to the analysis result, and determining a target registration scheme, including: Constructing a rigid registration deformation condition constraint, where the rigid registration deformation condition constraint is a mutual information coefficient threshold trained by historical rigid registration image samples; Judging whether the mutual information coefficient satisfies the rigid registration deformation condition constraint; If so, generating the target registration scheme by rigid registration; If not, generating the target registration scheme by elastic registration; The rigid registration uses the centroids of the first medical image and the second medical image for initial translational alignment, and uses the gradient descent method to maximize the mutual information between the first medical image and the second medical image by adjusting translational and rotational parameters; The elastic registration first performs preliminary alignment on the first medical image and the second medical image by rigid registration, constructs a deformation field model based on B-splines, and calculates the deformation field by the gradient descent method to minimize the difference between the first medical image and the second medical image.

2. The medical image low-contrast region adaptive enhancement method according to claim 1, characterized in that Performing multi-level fusion on the first registered image and the second registered image through multi-scale fusion to generate a fused medical image, including: Performing pyramid decomposition on the first registered image and the second registered image respectively to construct a Gaussian pyramid; Perform maximum selection on the pixel points of each resolution level in the Gaussian pyramid to generate a fused image corresponding to each resolution level; Based on the fused images corresponding to each resolution level, starting from the fused image of the bottom layer, layer by layer, synthesize the fusion result of each layer of the image with the upper-level image to generate a high-resolution image until reaching the highest resolution layer, and construct a Laplacian pyramid image; Based on the Laplacian pyramid image, on the image of the highest resolution layer, gradually stack the details of all low-resolution levels to reconstruct the fused medical image.

3. The medical image low-contrast region adaptive enhancement method according to claim 1, wherein Perform multi-feature analysis on the gray-scale distribution dynamics, gray-scale gradient, and frequency-domain features of the fused low-contrast region. Based on the multi-feature analysis results, call a pre-trained adaptive enhancement module to perform joint enhancement processing of multiple contrast enhancement models to generate a fused enhanced image, including: Determine a preset set of low-contrast factors, where the preset set of low-contrast factors is based on historical data statistics; Collect multiple groups of contrast enhancement samples based on the preset set of low-contrast factors. Any group of contrast enhancement samples includes an original medical image sample, an enhanced medical image sample, and corresponding multi-feature analysis samples; Construct multiple model networks of the multiple contrast enhancement models, and perform block fusion enhancement training on the multiple model networks with the multiple groups of contrast enhancement samples to generate the multiple contrast enhancement models, and construct the adaptive enhancement module; Call the adaptive enhancement module, and the multiple contrast enhancement models perform enhancement processing on the fused low-contrast region in the fused medical image based on the multi-feature analysis results to generate the fused enhanced image.

4. The medical image low-contrast region adaptive enhancement method according to claim 3, wherein, Construct the adaptive enhancement module, including: Use the original medical image sample and multi-feature analysis sample in any group of contrast enhancement samples as training inputs, and use the enhanced medical image sample as the output supervision truth value to perform iterative adjustment of the training parameters of the multiple model networks. The training parameters include the learning parameters corresponding to the multiple model networks respectively, and the enhancement fusion weights of the multiple model networks. Train the multiple model networks until convergence to generate the adaptive enhancement module.

5. The method for adaptively enhancing low-contrast regions of medical images according to claim 1, characterized in that, Perform local contrast calculation under a sliding window for the fused medical image to identify the fused low-contrast region, including: Construct a preset sliding window; Slide the preset sliding window on the fused medical image to calculate the pixel contrast index under each sliding window; Extract the image block corresponding to the sliding window with the pixel contrast index equal to or equal to the preset contrast index to generate the fused low-contrast region.

6. An adaptive enhancement system for low-contrast regions of medical images, characterized in that, Include: A medical image set acquisition module, configured to receive a target medical image set, where the target medical image set includes at least a first medical image corresponding to a first modality and a second medical image corresponding to a second modality; A registered image generation module, configured to perform deformation measurement on the first medical image and the second medical image, and perform image registration according to the measurement result to generate a first registered image and a second registered image, where the image registration includes rigid registration or elastic registration; A fused medical image generation module, which is used to perform multi-level fusion on the first registered image and the second registered image through multi-scale fusion to generate a fused medical image; A fused low-contrast region recognition module, which is used to calculate the local contrast under a sliding window for the fused medical image to identify the fused low-contrast region; A fused enhanced image generation module, which is used to perform multi-feature analysis of the gray-scale distribution dynamics, gray-scale gradient, and frequency-domain features of the fused low-contrast region. Based on the multi-feature analysis results, a pre-trained adaptive enhancement module is called to perform joint enhancement processing of multiple contrast enhancement models to generate a fused enhanced image.

Images