Focus evolution simulation method and device, computer equipment and storage medium

Abstract

The invention provides a focus evolution simulation method and device, computer equipment and a storage medium, and belongs to the field of image processing, and the method comprises the steps: obtaining an SPECT tumor imaging image; constructing a conditional generative adversarial network cGAN; inputting the SPECT tumor imaging image into cGAN, and obtaining fusion features through a multi-scale feature fusion network; in the first stage, a preliminary lesion image is generated through feature fusion, in the second stage, lesion features and health features are weighted based on a semantic regulatory factor alpha based on a gradient control mechanism of the semantic regulatory factor alpha, and the morphological change gradient of the lesion image from minimal lesion to severe lesion is controlled by adjusting the numerical range of alpha. Thus, through multi-scale feature fusion, progressive generation and introduction of semantic regulatory factors, interaction of tissues around a focus area is highlighted, and fine evolution and smooth transition of lesion features are realized.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and in particular relates to a method, device, computer equipment and storage medium for simulating lesion evolution. Background Art

[0002] As a core pillar of modern clinical diagnosis and treatment, medical imaging technology uses multimodal, high-precision imaging equipment to obtain information on the internal anatomical structure and functional metabolism of the human body, providing an objective basis for disease diagnosis and treatment. Computer-aided diagnosis (CAD) systems use advanced algorithms to analyze complex medical imaging data, assisting radiologists in achieving efficient and accurate lesion detection and quantitative analysis, significantly improving diagnostic efficiency and reducing misdiagnosis rates. Nuclear medicine molecular imaging technology, particularly single-photon emission computed tomography (SPECT) and positron emission tomography (PET), with its unique functional metabolic imaging advantages, can capture molecular abnormalities in tissues and organs before morphological changes occur, providing key information for early detection and precise intervention of diseases.

[0003] However, existing medical imaging technologies have significant limitations, primarily in the detection and characterization of static lesions, and the lack of the ability to dynamically predict the progression of lesions from early features to severe features. Current technologies primarily rely on qualitative diagnosis of the lesion area, locating the lesion through image segmentation algorithms, and performing feature extraction or image enhancement on the lesion area. While these methods can improve the accuracy of single-phase diagnosis, they cannot effectively capture the progression of lesions from early features to severe features, and cannot simulate the natural progression trajectory of the disease. This technical shortcoming makes it difficult for clinicians to accurately assess the development trend of lesions and to formulate individualized treatment plans for different stages of the disease, ultimately affecting treatment effectiveness and long-term prognosis for patients.

[0004] In the field of nuclear medicine imaging, such as SPECT bone imaging, existing technical solutions mostly focus on the identification and analysis of static lesions, and rarely involve the dynamic simulation of lesions progressing from early features to severe features. In clinical practice, reference area comparison methods or historical image comparison methods are mainly used to assess lesion changes, that is, the patient's current image is compared with a standard template or previous examination results to infer the extent of the lesion. This type of static comparison method cannot predict the dynamic evolution characteristics of the lesion and cannot provide a complete evolutionary view of the lesion from early features to severe features. Most current technologies rely on extracting features from healthy or standard images and identifying abnormal areas through pattern recognition algorithms, while ignoring the gradual change of the lesion's transformation from initial features to severe features and its related physiological characteristics.

[0005] Conditional generative adversarial networks (cGANs), as an important technology in the field of deep learning, can generate highly realistic medical imaging data through an adversarial training mechanism between generators and discriminators. However, existing cGAN applications mainly focus on the repair and generation of static lesions, and lack the ability to dynamically simulate the progression of lesions from early features to severe features. Existing methods usually generate repaired images of lesions without or with mild lesions by training on healthy images or images of mild lesions, but they cannot accurately reproduce the dynamic changes of lesions over time. This results in the generated images being less accurate in terms of detail, and unable to truly show the subtle changes in the various intermediate stages of the lesion's progression from early features to severe features, which in turn affects doctors' understanding of the progression of the disease and their formulation of treatment decisions.

[0006] Existing cGAN-based lesion simulation methods have three key technical bottlenecks: first, the feature extraction architecture is single and cannot fully capture the complex interaction patterns between the lesion area and the surrounding tissues; second, the generative model lacks a fine-grained control mechanism and cannot achieve continuous and gradual control of lesion features; third, the model training and evaluation indicators are insufficient to accurately assess the authenticity of lesion feature changes. Specifically, existing methods usually rely on the standard cGAN framework to randomly remove healthy image areas and train the network to generate repaired images. Such methods focus more on restoring healthy images rather than the dynamic changes of lesions. Therefore, they cannot fully simulate the process of lesions developing from early features to severe features. Existing technical solutions lack a comprehensive description of the dynamic changes of lesions, ignore the interaction of tissues around the lesion area and the gradual evolution of lesion characteristics, and cannot accurately reflect the deterioration rate and scope of the lesion and its impact on surrounding tissues. As a result, the generated images cannot truly show the entire process of lesion evolution from early to late stages, which in turn affects doctors' judgment of the disease and the formulation of treatment decisions.

[0007] Therefore, there is an urgent need to develop a new image generation technology that can accurately simulate the progression of lesions from early features to severe features to meet the urgent clinical needs for disease progression prediction and individualized treatment plans. Summary of the Invention

[0008] In order to solve the problem that existing methods ignore the interaction of tissues around the lesion area and the gradual evolution of lesion characteristics, resulting in the inability to simulate the formation, development and evolution process of the lesion, the present invention provides a simulation method, device, computer equipment and storage medium for lesion evolution.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A method for simulating lesion evolution, comprising:

[0011] Obtain SPECT tumor imaging images;

[0012] Constructing a conditional generative adversarial network (cGAN), wherein the cGAN includes a progressive generator and a multi-scale feature fusion network; the progressive generator includes a two-stage generation module;

[0013] The SPECT tumor imaging image is input into the cGAN, and the multi-scale features of the lesion area and surrounding tissue are extracted and fused through the multi-scale feature fusion network to obtain fused features. The fused features are used as input to the progressive generator. In the first stage, a preliminary lesion image is generated by the fused features, and the preliminary lesion image contains morphological features of the minimal lesion area. In the second stage, the morphological features are weighted according to the semantic adjustment factor α, and by adjusting the numerical range of α, a morphological evolution gradient image from minimal lesions to severe lesions is generated.

[0014] Optionally, the multi-scale features of the lesion area and surrounding tissues are extracted and fused by the multi-scale feature fusion network to obtain fused features including:

[0015] The multi-scale feature fusion network extracts multi-scale features of the lesion area and surrounding tissues through a parallel branch network, and adopts a cross-scale feature map fusion strategy to weightedly superimpose feature maps of different scales to generate fused features that contain local details and global structural information.

[0016] Optionally, the parallel branch network includes multiple convolution branches, each using a convolution kernel of a different scale to extract lesion features of the corresponding scale;

[0017] The weighted superposition of feature maps of different resolutions includes:

[0018] The multi-scale feature maps are normalized and feature fusion is performed through weighted fusion, where the weight values ​​are dynamically adjusted according to the contrast between the lesion area and the surrounding tissue.

[0019] Optionally, before inputting the SPECT tumor imaging image into the cGAN, the SPECT tumor imaging image is further preprocessed, including noise suppression, region positioning, standardized filling, and image cropping;

[0020] The noise suppression is achieved through filtering technology, the region positioning determines the region of interest through extreme coordinates, the standardized filling aligns the center of the cropped image to a preset canvas size, and the image cropping crops the standardized image to a fixed size.

[0021] Optionally, before inputting the SPECT tumor imaging image into the cGAN, the progressive generator is further trained, including:

[0022] The first stage uses a joint loss function, the specific formula is:

[0023] L stagel =λ1·||G(lesion)-healthy||1+λ2·SSIM(G(lesion),healthy);

[0024] The second stage adopts the balanced loss function, the specific formula is:

[0025] L balance =α·L lesion +(1-α)·L healthy ;

[0026] Among them, λ1 and λ2 are weight coefficients, G(lesion) is the generator, healthy represents healthy images, SSIM represents the structural similarity index, L lesion represents the loss of lesion image, L healthy represents the loss of de-sickness and reconstructed image;

[0027] The goal of the progressive generator is to minimize the loss function, and the training is completed when the loss functions of the two stages are both less than the preset loss threshold.

[0028] Optionally, weighting the morphological features according to the semantic adjustment factor α includes:

[0029] The α value is embedded in the feature space, and the lesion features and normal tissue features are weighted by α and (1-α) respectively. The weight distribution of lesion features and normal tissue features is adjusted based on the proportional relationship between α and (1-α) to form a continuous and smooth evolution control function.

[0030] A device for simulating lesion evolution, comprising:

[0031] An acquisition module, used for acquiring SPECT tumor imaging images;

[0032] A building block for constructing a conditional generative adversarial network (cGAN), which includes a progressive generator and a multi-scale feature fusion network.

[0033] The simulation module is used to input the SPECT tumor imaging image into the cGAN, extract the multi-scale features of the lesion area and surrounding tissues through the multi-scale feature fusion network, and fuse them to obtain fusion features; the progressive generator includes a two-stage generation module, in the first stage, a preliminary lesion image is generated by fusing features, and the preliminary lesion image contains morphological features of the minimal lesion area; in the second stage, the preliminary lesion image is used as input, the morphological features are weighted according to the semantic adjustment factor α, and a morphological evolution gradient image from minimal lesions to severe lesions is generated by adjusting the numerical range of α.

[0034] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for simulating lesion evolution.

[0035] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the above-mentioned method for simulating lesion evolution is implemented.

[0036] The present invention provides a method for simulating lesion evolution, which has the following beneficial effects:

[0037] First, a multi-scale feature extraction network captures and adaptively fuses multi-level features of the lesion region and surrounding tissue, accurately characterizing the interaction between the lesion and its surrounding tissue. This multi-scale feature integration not only enhances lesion recognition but also simulates the dynamic interaction between the lesion and surrounding tissue, thereby more realistically reflecting the lesion's evolution. Second, a two-stage generation module is employed to progressively simulate the lesion's evolution from minimal to severe lesions, avoiding the limitations of lesion feature mutations in existing methods. The introduction of a semantic adjustment factor α provides fine-grained control over lesion morphological changes. Morphological features are weighted according to the semantic adjustment factor α. This feature modulation mechanism enables a smooth transition of lesion morphological changes. Thus, through multi-scale feature fusion, progressive generation, and conditional embedding of semantic adjustment factors, not only the interaction between the lesion region and surrounding tissue is highlighted, but also the fine-grained evolution and smooth transition of lesion features are achieved, providing a more realistic, continuous, and comprehensive simulation of the formation, development, and evolution of tumor lesions. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0039] Figure 1 The figure is a flow chart of a method for simulating lesion evolution provided by the present invention according to an exemplary embodiment.

[0040] Figure 2 The present invention provides a flowchart of a model implementation according to an exemplary embodiment.

[0041] Figure 3 A diagram of a multi-scale network architecture provided according to an exemplary embodiment of the present invention.

[0042] Figure 4A flowchart of generating a dynamic lesion evolution image based on cGAN according to an exemplary embodiment of the present invention is provided.

[0043] Figure 5 This is a block diagram of a device for simulating lesion evolution according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the technical solution of the present invention and to be able to implement it, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0045] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0046] First, the present invention provides a method for simulating the evolution of lesions, specifically Figure 1 As shown, the following steps are included:

[0047] S101. Acquire a SPECT tumor imaging image.

[0048] In this step, a SPECT image is first acquired, which serves as the fundamental data source for the entire process. The SPECT tumor imaging image is then preprocessed, including noise suppression, region localization, normalized padding, and image cropping. Noise suppression is achieved through filtering techniques to reduce noise interference in the image. Region localization determines the region of interest (ROI) using extreme coordinates. Normalized padding aligns the center of the cropped image to a preset canvas size, such as 256×256. Image cropping reduces the normalized image to a fixed size for subsequent processing.

[0049] For example, the steps of implementing the data preprocessing process can be defined according to the following mathematical expressions.

[0050] Noise suppression is expressed as follows:

[0051]

[0052] Where I(x,y) is the pixel value of the image at the (x,y) position, and θ is the preset pixel threshold, which is set to 5 here.

[0053] If I(x, y) is less than the preset pixel threshold θ, the pixel value is set to 0, that is, black.

[0054] If I(x,y) is greater than or equal to the preset pixel threshold θ, the original pixel value remains unchanged.

[0055] This method can be used to remove noise from images, especially single-luminance noise. By setting a threshold, only pixels exceeding this threshold are retained, while pixels below the threshold are suppressed or removed, thus achieving the effect of noise reduction.

[0056] The chest area is located and the ROI area is determined by the extreme coordinates as shown in the following formula:

[0057]

[0058] By using the above formula, determine the minimum x value where the sum of the pixel values ​​of all pixels in a column is greater than 0 and the pixel values ​​of all pixels in a row are greater than the threshold N. thresh The maximum y value of the image can be used to determine the left and lower boundaries of the chest area in the image, thereby helping to locate the ROI area of ​​interest.

[0059] Normalize the padding to align the center of the cropped image to the 256×256 canvas as shown in the following formula:

[0060]

[0061] Among them, the variable P is defined ij , i is used as the pixel index in the image to locate a specific pixel position in the image; in the formula, when the value of i is When the pixel P is within the interval ij =I ij (ie original pixel value), when index i is not within this range, P ij =0.

[0062] M represents the side length of the cropped image. In image processing, cropping usually means extracting a smaller region from the original image. Here, M refers to the size of the cropped image. For example, if M = 256, the cropped image is a 256*256 square.

[0063] S102. Construct a conditional generative adversarial network (cGAN), which includes a progressive generator and a multi-scale feature fusion network.

[0064] Specifically, the conditional generative adversarial network (cGAN) includes a progressive generator and a multi-scale feature fusion network. The progressive generator generates a sequence of lesion evolution images through a two-stage generation module. The first stage generates the initial lesion image, and the second stage dynamically enhances the lesion morphology based on the semantic adjustment factor α.

[0065] In the present invention, the simulation of lesion evolution is mainly achieved through the following four parts:

[0066] Part I: Modeling the Dynamic Evolution of Early Lesions into Advanced Lesions: The core of this invention is to use deep learning methods to predict the progression of lesions from minimal to severe. By introducing a conditional generative adversarial network (cGAN), the network is trained to generate image sequences of different stages, from the early stages of lesions to the later stages. This process accurately simulates the evolution of lesions, providing doctors with a complete sequence of lesion images to aid diagnosis.

[0067] Part II, Multi-scale Feature Fusion Network: This invention adopts a multi-scale feature fusion strategy, using multiple branches to extract features of different scales. The feature maps of each branch path have different focuses on scales, covering information extraction from dense to sparse, so that the model can understand image content at multiple scales. Through the feature map fusion strategy, it not only promotes the fusion of information of different scales, but also ensures that each pixel contains rich contextual information. By adding the feature maps of the four paths, the information of each pixel is integrated with the features of the neighboring areas, forming a comprehensive feature representation.

[0068] Part III, a two-stage progressive generator: This invention is based on the cGAN framework. In the first stage, the generator generates images of lesions that initialize minimal lesions. In the second stage, the semantic adjustment factor (α) is controlled to gradually enhance the lesion features, generating an image sequence that depicts the lesion's evolution, simulating the progressive deterioration of the disease from early to late stages.

[0069] In the fourth part, the present invention introduces a semantic regulation factor (α) as a precise control parameter for the dynamic evolution process of lesions, with a value range of 0 to 1, corresponding to the states of minimal lesions and severe lesions, respectively. In the conditional generative adversarial network architecture, the semantic regulation factor is used to weight the lesion characteristics and normal morphological characteristics, and a nonlinear mapping mechanism is used to control the gradual changes in the lesion morphological characteristics (such as size and density). This design can accurately simulate the evolutionary characteristics of lesions observed clinically: a complete evolutionary cycle of slow development in the early stage, rapid deterioration in the middle stage, and slowing changes in the late stage. At the same time, a multi-level regulatory mechanism is used to ensure that the generated lesion image sequence conforms to medical laws, providing strong support for clinical diagnosis.

[0070] This study pioneered the construction of a medical imaging lesion progression simulation system, the core architecture of which integrates multi-path feature capture and staged generation technology. Figure 2As shown, the system first extracts image information at different scales through a four-channel parallel architecture, comprehensively covering everything from microscopic details to macroscopic structures. Each path is equipped with a combination of convolutional layers with different receptive fields, capturing comprehensive lesion characterization, from local texture to overall morphology. The feature integration stage utilizes an adaptive weighting mechanism, dynamically adjusting the fusion ratio based on the relevance of information across each dimension to ensure that key pathological features are prominently represented. During data preprocessing, the system performs noise filtering, region of interest localization, and normalization on the acquired SPECT tumor images, laying the foundation for subsequent feature extraction. Notably, the first feature extraction path focuses on subtle pixel-level changes, continuously extracting them using a small filter; the second path focuses on tissue boundaries and localized features; the third path emphasizes capturing medium-range contextual associations; and the fourth path utilizes dilated convolution techniques to achieve a wide receptive field, ensuring the system understands the lesion's appearance within a broader tissue context. The generation module employs an innovative two-stage design: the initial stage focuses on basic lesion construction, while the secondary stage continuously adjusts the lesion's severity through parameter manipulation, accurately reproducing the complete progression of the disease, from mild early symptoms to severe late-stage changes.

[0071] The entire system utilizes specialized morphology-preserving constraints and temporal consistency mechanisms to ensure that the generated lesion evolution sequences conform to medical principles, avoiding the problem of sudden changes in lesion features common in traditional methods. The discriminative component employs a regional evaluation strategy, focusing on verifying the quality of local image patches and incorporating stabilization techniques to prevent fluctuations during training. In terms of implementation details, control parameters are embedded at multiple levels of the generative network, influencing the statistical distribution of feature maps through conditional normalization operations, enabling fine-grained control of lesion morphology. The training strategy employs a progressive learning approach, initially focusing on developing basic lesion generation capabilities before transitioning to comprehensive optimization of the complete evolution sequence. The loss function design incorporates three complementary components: an adversarial loss ensures the authenticity of generated images; a feature matching loss maintains semantic consistency; and a morphology-preserving loss specifically constrains the medical plausibility of lesion evolution. This approach addresses the limitations of existing technologies in capturing the interactions between lesions and surrounding tissues and the detailed evolutionary process, providing clinicians with a visualization tool to help understand disease progression and aid in early diagnosis and prognosis assessment. The system is evaluated using multi-dimensional metrics, in addition to traditional image quality assessments, including medical expert ratings and clinical relevance verification. Experimental results demonstrate that the lesion sequences generated by this system achieve high levels of morphological continuity, detail fidelity, and medical plausibility. The system particularly excels in the transition from minimal to severe lesions, a weakness of traditional methods. This innovative system enables medical professionals to gain a more comprehensive understanding of the underlying disease trajectory, providing a scientific basis for precision medicine and personalized treatment plans, and opening up new technological paths for intelligent diagnosis in related fields.

[0072] In addition, the present invention also provides a progressive generation training process, which is as follows:

[0073] In the first stage (training to obtain minimal lesions), a joint loss function is used, and the training objective function is as shown in the following formula:

[0074] L stagel =λ1·||G(lesion)-healthy||1+λ2·SSIM(G(lesion),healthy);

[0075] Among them, L stage1 is the symbolic representation of the first-stage tameness objective function, used to train the model for generating minimal lesion lesions. G(lesion) is the generator, part of the Generative Adversarial Network (GAN), used to generate images containing lesions. Lesion represents the lesion region; healthy represents a healthy image (i.e., an image without lesions) used for comparison with the generated lesion image; λ1 and λ2 are weight coefficients, ||·|| represents the L1 norm, and SSIM represents the structural similarity index.

[0076] In the second stage (gradual process generation), the microlesion-exacerbation balance loss function is introduced as shown in the following formula:

[0077] L balance =·L lesion +(1-α)·L healthy ;

[0078] Among them, L balance is the symbolic representation of the lesion-exacerbation balance loss function introduced in the second stage, which is used to balance the severity of lesions during the generation process. lesion represents the loss of lesion image, L healthy represents the loss of de-lesioned and reconstructed images, and α∈[0,1] is an adjustable semantic factor used to control the severity of the lesion.

[0079] In summary, the training process is as follows Figure 4As shown in the figure, a multi-stage strategy is adopted to achieve accurate simulation of the dynamic evolution of lesions. In the initial stage, a lesion region localization algorithm is first implemented to accurately identify the spatial distribution characteristics of the target lesion in the SPECT image. Subsequently, the system uses this localization information to train the basic generative model and construct a characteristic expression of early lesions. Next, the initial generation results are introduced into the conditional adversarial generative network architecture. Through the dynamic game mechanism between the generator and the discriminator, a continuous change sequence of gradually worsening lesions is generated. During the generation process, the control parameters are gradually adjusted according to medical laws to ensure that the progressive changes in lesion morphology are consistent with clinical observation experience. The final output result is a set of complete evolutionary images from early minor symptoms to late severe lesions. This visual presentation method provides medical professionals with a full dynamic view of disease progression, significantly enhances the ability to predict the development trajectory of lesions, and provides richer information support for clinical decision-making.

[0080] S103 , inputting the SPECT tumor imaging image into cGAN to obtain an output gradient image of morphological evolution from minimal lesions to severe lesions.

[0081] In this step, multi-scale features of the lesion area and surrounding tissues are extracted and fused through the multi-scale feature fusion network to obtain fusion features; the progressive generator includes a two-stage generation module, in the first stage, a preliminary lesion image is generated by fusion features, and the preliminary lesion image contains the morphological features of the small lesion area; in the second stage, the preliminary lesion image is used as input, and the morphological features are weighted according to the semantic adjustment factor α. In this way, combined with the gradient control mechanism of the semantic adjustment factor α, the α value is used to control the key feature parameters such as the volume, density and metabolic activity of the lesion through a nonlinear mapping function, accurately simulating the progressive changes of the lesion morphological features, and generating a morphological evolution gradient image from small lesions to severe lesions.

[0082] Specifically, the preprocessed image is input into a multi-scale feature fusion network, which contains multiple branch paths, each of which focuses on different scales, covering information extraction from dense to sparse. Figure 2As shown in the figure, in the feature extraction stage, the network extracts image features from different scales. In the feature fusion stage, feature maps at different scales are fused together through feature map fusion, ensuring that each pixel contains rich contextual information, ultimately forming a comprehensive feature representation. Next, the feature map is input into a two-stage generator. The first-stage generator generates preliminary lesion images through adversarial training between the generator and the discriminator. The second-stage generator gradually increases the severity of the lesion by controlling the semantic adjustment factor (α), generating images of more severe lesions. This process simulates the dynamic evolution of lesions from minor to severe. Finally, the generated lesion image sequence is output, providing doctors with a gradual progression of lesions from minor to severe, assisting in diagnosis and treatment planning. Through multi-scale feature fusion and the two-stage generator, the entire process accurately predicts the dynamic evolution of lesions, improving the accuracy and detail of the lesions and surrounding tissues in the generated images, providing strong support for clinical diagnosis and treatment.

[0083] For example, after the SPECT tumor imaging image is input into cGAN, features of different scales are extracted through multiple feature extraction branches. After the output of each branch is unified to 256×256 resolution through bilinear interpolation, feature fusion is achieved through pixel-level addition as shown in the following formula:

[0084]

[0085] Among them, F fused is the fusion feature, F k These are the different scale features output by each branch.

[0086] In another embodiment, the parallel branch network includes multiple convolution branches, each using a convolution kernel of different scales to extract lesion features of corresponding scales; the multi-scale feature map is normalized, and feature fusion is performed by weighted fusion, wherein the weight value is dynamically adjusted according to the contrast between the lesion area and the surrounding tissue. Figure 3 As shown, it can include 4 convolution branches, which are the first convolution branch: 3*3 convolution, padding 1, step size 1; the second convolution branch: 3*3 convolution, 3*3 convolution, 3*3 convolution, padding 1, step size 1; the third convolution branch: 3*3 convolution, 3*1 convolution, 1*3 convolution, padding 1, step size 1; the fourth convolution branch: 1*1 convolution, padding 0, step size 1.

[0087] The purpose of normalization is to unify feature maps of different scales into the same numerical range to avoid imbalance in feature weights caused by scale differences. The specific steps are as follows:

[0088] In one embodiment, the network structure of the multi-scale network architecture is as follows: Figure 3As shown, it contains four feature extraction branches. The multi-scale information integration component of the present invention is cleverly arranged at the front end of the model, aiming to significantly improve the model's recognition and analysis efficiency of image content by aggregating image information at different resolutions. The component innovatively designs four parallel processing paths, each of which is ingeniously optimized for detail features at different levels, ensuring that everything from the subtlest local texture to the most macroscopic overall structure can be accurately captured and processed. Each path has been carefully adjusted to focus on a specific level of detail, so that when the model faces an image, it can start from multiple scales and conduct a full-scale, multi-level interpretation. This design not only makes the model more handy when processing images, but also effectively avoids information loss or misjudgment caused by single-scale processing. By cleverly merging feature maps of different scales at the pixel level, the component achieves seamless fusion of multi-scale information, ensuring that each pixel contains rich contextual information, providing a solid foundation for subsequent image analysis and processing. Finally, by superimposing the feature maps of the four paths, a feature representation with a huge amount of information is formed. Each pixel integrates feature information from the surrounding area, making the feature representation of the entire image more comprehensive and detailed. This multi-scale feature fusion method not only improves the model's ability to capture image details, but also enhances its understanding of the overall image structure, providing strong support for more accurate image recognition and analysis.

[0089] Extracting multi-scale feature maps: Feature maps of different scales are extracted from SPECT tumor imaging images through a multi-scale feature fusion network (such as a convolutional neural network). These feature maps capture the local details and global structural information of the lesion area and its surrounding tissues. Then, normalization is performed to normalize the feature maps at each scale. This normalization can use common methods, such as Min-Max normalization: linearly mapping the pixel values ​​of the feature map to the interval [0,1]; Z-Score normalization: converting the pixel values ​​of the feature map to a distribution with a mean of 0 and a standard deviation of 1; L2 normalization: performing L2 norm normalization on each channel of the feature map to ensure that the feature vectors of each channel have the same scale.

[0090] The normalized feature map has a consistent numerical range, which facilitates subsequent weighted fusion operations.

[0091] The core of weighted fusion is to dynamically adjust the weight value based on the contrast between the lesion area and the surrounding tissue to highlight the characteristics of the lesion area while retaining the contextual information of the surrounding tissue. The specific steps are as follows:

[0092] Calculate contrast: For each scale feature map, calculate the contrast between the lesion area and the surrounding tissue. The contrast can be defined as follows:

[0093] Grayscale contrast: Calculate the difference in average grayscale values ​​between the lesion area and the surrounding tissue.

[0094] Texture contrast: The texture difference between the lesion area and the surrounding tissue is calculated by local binary pattern (LBP) or gray level co-occurrence matrix (GLCM).

[0095] Structural contrast: Evaluate the structural difference between the lesion area and the surrounding tissue through edge detection or gradient calculation.

[0096] Dynamic weight assignment: Weights are assigned dynamically based on contrast values. Areas with higher contrast (i.e., lesion areas) are assigned larger weights, while areas with lower contrast (i.e., surrounding tissues) are assigned smaller weights.

[0097] The normalized multi-scale feature map is weightedly summed with the corresponding weight value to obtain the fused feature map.

[0098] In addition, in order to further improve the effect of weighted fusion, the following optimization strategies can be adopted:

[0099] Adaptive contrast calculation: An adaptive contrast calculation method is introduced to dynamically adjust the range and scale of contrast calculation according to the morphology and size of the lesion area.

[0100] Multimodal information fusion: Combines information from other modalities of medical images (such as CT or MRI) to further optimize contrast calculation and weight distribution.

[0101] Deep learning optimization: By training a lightweight neural network, the weight distribution strategy of multi-scale feature maps is automatically learned to more accurately highlight the characteristics of the lesion area.

[0102] Through the above steps, the lesion area can be highlighted, and dynamic weight adjustment can significantly enhance the characteristics of the lesion area, making it more prominent in the fused feature map; retaining contextual information, the characteristics of the surrounding tissue are properly retained during the fusion process, providing important contextual information for the positioning and evolution simulation of the lesion; multi-scale information integration, feature maps of different scales are effectively integrated, which not only retains local details but also captures the global structure, providing comprehensive feature support for subsequent lesion generation and evolution simulation.

[0103] In another embodiment, the progressive generator adopts a two-stage architecture design to achieve accurate simulation of lesion evolution. The first stage constructs an initial lesion image based on multi-scale fusion features. This image accurately represents the morphological characteristics and boundary information of the minimal lesion. The second stage uses the initial lesion image as a conditional input and introduces a semantic adjustment factor α (0≤α≤1) as a control parameter for the degree of lesion evolution. The α value is mapped to a feature channel weighting coefficient through a feature space modulation mechanism to achieve dynamic regulation of the lesion morphological characteristics. In specific implementation, the system converts the α value into a modulation parameter acting on each feature dimension through a nonlinear transformation function, thereby accurately controlling key evolutionary characteristics such as the lesion volume expansion rate, boundary clarity changes, and metabolic activity enhancement. Ultimately, a complete evolutionary image sequence is generated that shows the continuous change of the lesion from mild to severe state.

[0104] The semantic adjustment factor α serves as a precise control parameter for the lesion evolution process, with a value range of [0, 1], where 0 corresponds to a minimal early lesion and 1 indicates progression to a severe late lesion. This parameter achieves fine-tuning of lesion morphological characteristics through a feature modulation mechanism: the system embeds the α value into the feature space, weights lesion and normal tissue features using α and (1-α), respectively, and dynamically adjusts the weight distribution of lesion and normal tissue features based on the proportional relationship between α and (1-α), forming a continuous and smooth evolution control function. Specifically, when the α value is small, the system retains more normal tissue features, presenting only mild lesion signs. As the α value increases, the influence of lesion features in the generation process gradually increases, and characteristics such as lesion volume expansion, blurred boundaries, and increased metabolic activity gradually emerge. When α approaches 1, lesion features completely dominate, indicating a significant deterioration. Through this parametric control mechanism, the system accurately simulates the entire disease progression cycle and generates medically consistent and continuous lesion evolution image sequences.

[0105] Alternatively, a dynamic weight matrix is ​​constructed based on the semantic modulation factor α. This matrix is ​​mapped to the α value through a carefully designed function and then applied to the feature representation of the lesion region through a feature space multiplication operation. This matrix multiplication mechanism enables precise interaction between the weight value and each spatial location and channel dimension of the feature map, achieving dynamic and fine-tuned control of lesion characteristics. The construction of the weight matrix ensures a nonlinear mapping relationship between the α value and key lesion parameters (such as volume, boundary definition, and metabolic activity): as the α value gradually increases, the intensity distribution of the corresponding weight matrix increases accordingly, especially in the lesion core and extended edge areas, resulting in a significant amplification of the lesion signal in the feature map, thus simulating the gradual process of lesion volume expansion, boundary blurring, and increased metabolic activity. Conversely, as the α value decreases, the overall intensity of the weight matrix decreases, and the lesion signal in the feature map is correspondingly suppressed, presenting an early or mild lesion state. This parameterized matrix-based feature modulation mechanism overcomes the limitations of simple linear interpolation in traditional methods and can more accurately simulate the nonlinear changes in lesion evolution, making the generated image sequence more medically realistic and continuous.

[0106] In this way, by embedding the semantic adjustment factor α into the feature representation space, we achieve fine control over the morphological evolution of the lesion. This technology not only enhances the realism and continuity of the generated images, but also provides important support for medical research and clinical diagnosis.

[0107] Using this method, a multi-scale feature extraction network captures and adaptively fuses multi-level features of the lesion region and surrounding tissue, accurately characterizing the interactions between the lesion and its surrounding tissue. This multi-scale feature integration not only enhances lesion recognition but also simulates the dynamic interactions between the lesion and surrounding tissue, thereby more realistically reflecting the lesion's evolution. Secondly, a two-stage generation module is employed to progressively simulate the lesion's evolution from minimal to severe lesions, avoiding the limitations of existing methods in the abrupt changes in lesion features. The introduction of a semantic adjustment factor α provides fine-grained control over lesion morphological changes. Morphological features are weighted according to the semantic adjustment factor α. This feature modulation mechanism enables a smooth transition of lesion morphological changes. Thus, through multi-scale feature fusion, progressive generation, and conditional embedding of the semantic adjustment factor, not only the interactions between the lesion region and surrounding tissue are highlighted, but also the fine-grained evolution and smooth transition of lesion features are achieved, providing a more realistic, continuous, and comprehensive simulation of the formation, development, and evolution of tumor lesions.

[0108] Secondly, the present invention also provides a device for simulating the evolution of lesions, such as Figure 5 As shown, including:

[0109] The acquisition module 501 is used to acquire SPECT tumor imaging images.

[0110] Construction module 502 is used to construct a conditional generative adversarial network (cGAN), which includes a progressive generator and a multi-scale feature fusion network.

[0111] The simulation module 503 is used to input the SPECT tumor imaging image into the cGAN, extract the multi-scale features of the lesion area and surrounding tissues through the multi-scale feature fusion network, and fuse them to obtain fusion features. The progressive generator includes a two-stage generation module. In the first stage, a preliminary lesion image is generated by fusing features. The preliminary lesion image contains morphological features of the minimal lesion area. In the second stage, the preliminary lesion image is used as input, the morphological features are weighted according to the semantic adjustment factor α, and a morphological evolution gradient image from minimal lesions to severe lesions is generated by adjusting the numerical range of α.

[0112] Using the above-mentioned device, a multi-scale feature extraction network captures and adaptively fuses multi-level features of the lesion region and surrounding tissue, accurately characterizing the interaction between the lesion and its surrounding tissue. This multi-scale feature integration not only enhances the ability to identify the lesion region but also simulates the dynamic interaction between the lesion and surrounding tissue, thereby more realistically reflecting the lesion's evolution. Secondly, a two-stage generation module is used to progressively simulate the lesion's evolution from minimal to severe lesions, avoiding the limitations of lesion feature mutations in existing methods. The introduction of a semantic adjustment factor α provides fine-grained control over lesion morphological changes. Morphological features are weighted according to the semantic adjustment factor α. This feature modulation mechanism enables a smooth transition of lesion morphological changes. Thus, through multi-scale feature fusion, progressive generation, and conditional embedding of semantic adjustment factors, not only the interaction between the lesion region and surrounding tissue is highlighted, but also the fine-grained evolution and smooth transition of lesion features are achieved, providing a more realistic, continuous, and comprehensive simulation of the formation, development, and evolution of tumor lesions.

[0113] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 The steps of a method for simulation of lesion evolution are provided.

[0114] The present invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 The steps of a method for simulation of lesion evolution are provided.

[0115] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0117] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0119] It should be noted that the specific embodiments described above can enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although this specification has described the present invention in detail, those skilled in the art should understand that the present invention can still be modified or replaced with equivalents; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are included in the scope of protection of the patent for the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. A method for simulating lesion evolution, characterized in that: The method comprises: Obtain SPECT tumor imaging images; Constructing a conditional generative adversarial network (cGAN), wherein the cGAN includes a progressive generator and a multi-scale feature fusion network; the progressive generator includes a two-stage generation module; The SPECT tumor imaging image is input into the cGAN, and the multi-scale features of the lesion area and surrounding tissue are extracted and fused through the multi-scale feature fusion network to obtain fused features. The fused features are used as input to the progressive generator. In the first stage, a preliminary lesion image is generated by the fused features, and the preliminary lesion image contains morphological features of the minimal lesion area. In the second stage, the morphological features are weighted according to the semantic adjustment factor α, and by adjusting the numerical range of α, a morphological evolution gradient image from minimal lesions to severe lesions is generated.

2. The method for simulating lesion evolution according to claim 1, wherein: The multi-scale feature fusion network is used to extract and fuse the multi-scale features of the lesion area and surrounding tissues, and the fusion features obtained include: The multi-scale feature fusion network extracts multi-scale features of the lesion area and surrounding tissues through a parallel branch network, and adopts a cross-scale feature map fusion strategy to weightedly superimpose feature maps of different scales to generate fused features that contain local details and global structural information.

3. The method for simulating lesion evolution according to claim 2, wherein: The parallel branch network includes multiple convolution branches, each of which uses convolution kernels of different scales to extract lesion features of corresponding scales; . The weighted superposition of feature maps of different resolutions includes: The multi-scale feature maps are normalized and feature fusion is performed through weighted fusion, where the weight values ​​are dynamically adjusted according to the contrast between the lesion area and the surrounding tissue.

4. The method for simulating lesion evolution according to claim 1, wherein: Before inputting the SPECT tumor imaging image into the cGAN, the SPECT tumor imaging image is also preprocessed, including noise suppression, region positioning, standardized filling, and image cropping; The noise suppression is achieved through filtering technology, the region positioning determines the region of interest through extreme coordinates, the standardized filling aligns the center of the cropped image to a preset canvas size, and the image cropping crops the standardized image to a fixed size.

5. The method for simulating lesion evolution according to claim 1, characterized in that: Before inputting the SPECT tumor imaging image into the cGAN, the progressive generator is trained, including: The first stage uses a joint loss function, the specific formula is: L stagel =λ1·||G(lesion)-healthy||1+λ2·SSIM(G(lesion),healthy); The second stage adopts the balanced loss function, the specific formula is: L balance =α·K lcsion +(1-a)·L healthy ; Among them, λ1 and λ2 are weight coefficients, G(lesion) is the generator, healthy represents healthy images, SSIM represents the structural similarity index, L lesion represents the loss of lesion image, L healthy represents the loss of de-sickness and reconstructed image; The goal of the progressive generator is to minimize the loss function, and the training is completed when the loss functions of the two stages are both less than the preset loss threshold.

6. The method for simulating lesion evolution according to claim 1, characterized in that: The weighting of the morphological features according to the semantic adjustment factor α includes: The α value is embedded in the feature space, and the lesion features and normal tissue features are weighted by α and (1-α) respectively. The weight distribution of lesion features and normal tissue features is adjusted based on the proportional relationship between α and (1-α) to form a continuous and smooth evolution control function.

7. A device for simulating lesion evolution, characterized in that: The device comprises: An acquisition module, used for acquiring SPECT tumor imaging images; A building block for constructing a conditional generative adversarial network (cGAN), which includes a progressive generator and a multi-scale feature fusion network. The simulation module is used to input the SPECT tumor imaging image into the cGAN, extract the multi-scale features of the lesion area and surrounding tissues through the multi-scale feature fusion network, and fuse them to obtain fusion features; the progressive generator includes a two-stage generation module, in the first stage, a preliminary lesion image is generated by fusing features, and the preliminary lesion image contains morphological features of the minimal lesion area; in the second stage, the preliminary lesion image is used as input, the morphological features are weighted according to the semantic adjustment factor α, and a morphological evolution gradient image from minimal lesions to severe lesions is generated by adjusting the numerical range of α.

8. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 6 when executing the program.

Images