Causal learning domain generalization model detection method for multi-center medical images

Abstract

The application discloses a multi-center medical image-oriented causal learning domain generalization model detection method, comprising the following steps: acquiring medical images and corresponding labeled data of a source domain; constructing a causal model, which decomposes the observation results of the medical images into stable anatomical content factors directly related to task labels and variable imaging acquisition factors related to medical centers; performing multi-layer causal intervention enhancement on the medical images, which at least includes appearance layer causal intervention, spatial layer causal intervention and frequency domain layer causal intervention for intervention on the variable imaging acquisition factors to generate intervention images simulating domain differences of different medical centers, wherein the intervention images share the same anatomical structure with the medical images; training an initial model using the medical images and the intervention images to learn anatomical structure representations robust to the variable imaging acquisition factors and sensitive to the stable anatomical content factors, and obtaining a generalization detection model for medical images of unknown medical centers.

Description

Technical Field

[0001] This application belongs to the field of image processing technology, specifically relating to a method for detecting causal learning domain generalization models for multicenter medical images. Background Technology

[0002] In recent years, with the development of deep learning technology, medical image segmentation and detection methods based on convolutional neural networks, Transformers, and generative models have made significant progress in tasks such as lesion identification, organ segmentation, anatomical structure localization, and assisted diagnosis. These methods typically rely on large amounts of labeled data for training and can achieve high recognition accuracy under the same data distribution. However, in real clinical applications, medical image data often comes from different medical centers, and there are significant differences between these centers in terms of imaging equipment models, scanning protocols, exposure parameters, image post-processing procedures, patient positioning, and operating habits. These differences can lead to variations in image grayscale distribution, contrast, noise levels, texture patterns, and structural clarity, resulting in a significant data domain shift between training and testing data.

[0003] In multi-center deployment scenarios, deep learning models typically face a challenge: while they perform well when trained on source center data, their performance can significantly degrade when directly applied to other unknown centers. This is because existing models readily learn appearance features related to the imaging conditions of the source center, such as texture patterns, brightness distributions, noise characteristics, or edge enhancement effects generated by specific equipment. However, these features do not necessarily have a stable relationship with the actual anatomical structures or lesion regions. When the imaging style changes at the test center, the appearance features the model relies on become ineffective, leading to false positives, false negatives, or inaccurate segmentation boundaries. Therefore, improving the model's generalization ability across different medical centers is crucial for the evolution of intelligent medical image analysis from single-center to multi-center clinical deployment.

[0004] To address the aforementioned issues, existing methods primarily include transfer learning, unsupervised domain adaptation, and domain generalization. Transfer learning typically trains the model using source center data and then fine-tunes it with partially labeled or unlabeled data from the target center to improve the model's adaptability to the target center. Unsupervised domain adaptation methods further attempt to utilize unlabeled data from the target center, reducing the distributional differences between the source and target domains through feature alignment, pseudo-label learning, or adversarial training. While these methods are effective when the target center is known, in real-world clinical scenarios, target center data is often difficult to obtain promptly, and data privacy, ethical approvals, and data sharing restrictions exist between different hospitals, making it difficult for the model to be refitted for each new center.

[0005] In contrast, domain generalization methods aim to train robust models capable of adapting to unknown centers using only source center data, without relying on target center data. This aligns better with the practical needs of multi-center deployments in medical imaging. Existing domain generalization methods typically expand the source domain data distribution through data augmentation, style perturbation, frequency domain transformation, or invariant feature learning, exposing the model to more diverse image appearances during training. However, most of these methods are statistical perturbation augmentations, primarily simulating domain differences by altering image brightness, contrast, texture, or frequency distribution, lacking explicit modeling of the medical image generation mechanism. Especially in medical imaging tasks, the model should truly rely on stable anatomical structures or lesion morphologies, rather than appearance differences introduced by equipment, protocols, or post-processing methods. If the augmentation strategy lacks constraints, it may disrupt the original anatomical structures, leading to inconsistencies between labels and images, and even reducing the model's ability to recognize key structural boundaries.

[0006] Taking spinal X-ray images as an example, images acquired by different medical centers may vary in terms of exposure intensity, image clarity, bone tissue contrast, noise distribution, and the patient's standing posture. Figure 1 For tasks involving the detection or segmentation of anatomical structures such as the vertebrae, sacrum, and femoral head, the model needs not only to identify local boundaries but also to maintain an understanding of the overall anatomical arrangement and spatial continuity. If the model relies excessively on the imaging style of a particular center, it is prone to problems such as structural localization shifts, blurred boundaries, or unstable sequence judgments when deployed to centers with different equipment, image quality, or significant differences in patient posture. Therefore, multi-center medical imaging generalization methods not only need to enhance the model's robustness to appearance changes but also need to ensure the consistency of anatomical structures and the stability of prediction results during enhancement. Summary of the Invention

[0007] The purpose of this application is to provide a multicenter medical image cross-center generalization method based on causal intervention enhancement, which reduces the model's dependence on center-related appearance cues, guides the model to learn more stable anatomical structural representations, and thus improves the generalization ability and deployment stability of medical image segmentation or detection models on unknown medical centers.

[0008] According to a first aspect of the embodiments of this application, a method for detecting causal learning domain generalization models for multi-center medical images is provided, including: Acquire medical images and corresponding annotation data from the source domain; A causal model is constructed, which decomposes the observation results of medical images into stable anatomical content factors directly related to the task label, and variable imaging acquisition factors related to the medical center. Multi-layer causal intervention enhancement is performed on medical images, which includes at least appearance-layer causal intervention, spatial-layer causal intervention, and frequency-domain-layer causal intervention that intervene in variable imaging acquisition factors to generate intervention images that simulate differences in different medical central domains, wherein the intervention images share the same anatomical structures as the medical images. The initial model is trained using medical images and intervention images to learn anatomical structural representations that are robust to variable imaging acquisition factors and sensitive to stable anatomical content factors, thereby obtaining a generalized detection model for medical images from unknown medical centers. When training the initial model, a consistency constraint loss function is used to constrain the model to output consistent prediction results for medical images and their corresponding intervention images.

[0009] In some optional embodiments of this application, appearance layer causal intervention includes: perturbing the overall brightness, contrast, texture distribution, or noise performance of medical images through a random style transformation module to simulate the overall imaging style differences between different medical centers.

[0010] In some optional embodiments of this application, spatial layer causal intervention includes: Based on the prior anatomical structures in medical images, a spatial weight map is generated to control the intervention intensity in different regions. Based on the spatial weight map, the results of causal intervention in the appearance layer are spatially non-uniformly fused to simulate local domain shifts in medical images caused by uneven exposure, local occlusion, or artifacts.

[0011] In some optional embodiments of this application, the step of generating a spatial weight map includes: The key point sequence of the target anatomical structure is obtained based on the labeled data, and the principal axis of the structure is obtained by fitting the key point sequence. One or more control points are randomly sampled along the main axis of the structure, and a random weight is assigned to each control point; Interpolation smoothing is performed on one or more control points and their random weights to generate a continuous spatial weight map covering the entire image.

[0012] In some optional embodiments of this application, frequency domain-level causal intervention includes: Perform a Fourier transform on the medical image to obtain its frequency domain representation; Frequency-dependent adaptive perturbations are applied to the amplitude spectrum in the frequency domain representation while keeping the phase spectrum unchanged to simulate differences in image edge sharpness, noise levels, and detail texture between different medical centers. The frequency domain representation after perturbation is subjected to inverse Fourier transform to obtain the frequency domain interference image.

[0013] In some optional embodiments of this application, frequency-dependent adaptive perturbation is applied to the amplitude spectrum, including setting different perturbation amplitudes according to the low-frequency, mid-frequency or high-frequency regions, so as to differentiate the overall brightness contrast change, texture change or edge detail change.

[0014] In some optional embodiments of this application, the method for detecting causal learning domain generalization models for multicenter medical images further includes a partitioning modeling step: Based on the characteristics of the anatomical structure, the target area to be detected or segmented is divided into multiple sub-regions; A model is constructed that includes a shared encoder and multiple region-specific decoding branches. The shared encoder is used to extract global context information, and each region-specific decoding branch is used to learn the local fine-grained structural features of the corresponding sub-region.

[0015] According to a second aspect of the embodiments of this application, a causal learning domain generalization model detection device for multi-center medical images is provided, comprising: The acquisition module is used to acquire medical images and corresponding annotation data from the source domain; The building module is used to construct causal models, which decompose the observation results of medical images into stable anatomical content factors directly related to the task labels and variable imaging acquisition factors related to the medical center. An enhancement module is used to perform multi-layer causal intervention enhancement on medical images. The multi-layer causal intervention enhancement includes at least appearance-layer causal intervention, spatial-layer causal intervention, and frequency-domain-layer causal intervention that intervene in variable imaging acquisition factors to generate intervention images that simulate differences in different medical central domains, wherein the intervention images share the same anatomical structures as the medical images. The training module is used to train the initial model using medical images and intervention images to learn anatomical structural representations that are robust to variable imaging acquisition factors and sensitive to stable anatomical content factors, thereby obtaining a generalized detection model for medical images from unknown medical centers. When training the initial model, a consistency constraint loss function is used to constrain the model to output consistent prediction results for medical images and their corresponding intervention images.

[0016] According to a third aspect of the embodiments of this application, an electronic device is provided, which may include: processor; Memory used to store processor-executable instructions; The processor is configured to execute instructions to implement the causal learning domain generalization model detection method for multicenter medical images as shown in any embodiment of the first aspect.

[0017] According to a fourth aspect of the embodiments of this application, a storage medium is provided, wherein when instructions in the storage medium are executed by a processor of an information processing device or a server, the information processing device or server implements the causal learning domain generalization model detection method for multicenter medical images as shown in any embodiment of the first aspect.

[0018] The above-mentioned technical solution of this application has the following beneficial technical effects: The method in this application addresses the problems of existing methods, such as limited enhancement methods, inability to fully understand the original anatomical structure, and strong dependence on central appearance features, by using a causal learning domain generalization model training method. This method can improve the cross-center generalization ability of spinal X-ray image segmentation or detection models. Attached Figure Description

[0019] Figure 1 This is a multicenter spinal X-ray image in an exemplary embodiment of this application; Figure 2 This is a schematic diagram of the causal learning domain generalization model detection method for multicenter medical images in an exemplary embodiment of this application; Figure 3 These are comparison diagrams of segmentation results at different centers in (a) and (b) of an exemplary embodiment of this application; Figure 4 This is a schematic diagram of the structure of a causal learning domain generalization model detection device for multicenter medical images in an exemplary embodiment of this application; Figure 5 This is a schematic diagram of an electronic device in an exemplary embodiment of this application; Figure 6 This is a schematic diagram of the hardware structure of an electronic device in an exemplary embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this application. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0021] The accompanying drawings illustrate layer structure diagrams according to embodiments of this application. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0022] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0024] Furthermore, the technical features involved in the different embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0025] The following description, in conjunction with the accompanying drawings, details the causal learning domain generalization model detection method for multi-center medical images provided in this application, through specific embodiments and application scenarios.

[0026] like Figure 2 As shown, in a first aspect of this application, a method for detecting causal learning domain generalization models for multi-center medical images is provided, comprising: S110: Obtain medical images and corresponding annotation data from the source domain; S120: Construct a causal model that decomposes the observation results of medical images into stable anatomical content factors directly related to the task label and variable imaging acquisition factors related to the medical center. S130: Perform multi-layer causal intervention enhancement on medical images, which includes at least appearance-layer causal intervention, spatial-layer causal intervention, and frequency-domain-layer causal intervention on variable imaging acquisition factors to generate intervention images that simulate differences in different medical central domains, wherein the intervention images share the same anatomical structures as the medical images. S140: The initial model is trained using medical images and intervention images to learn anatomical structural representations that are robust to variable imaging acquisition factors and sensitive to stable anatomical content factors, thereby obtaining a generalized detection model for medical images from unknown medical centers; when training the initial model, a consistency constraint loss function is used to constrain the model to output consistent prediction results for medical images and their corresponding intervention images.

[0027] This embodiment explores a generalized, cross-center whole-spine vertebral body detection model. Existing methods often expand training samples through brightness, contrast, noise, style transfer, or frequency domain perturbation, which are essentially still image statistical enhancements and struggle to distinguish between "task-related anatomical structural information" and "center-related imaging appearance information." This application, from a causal modeling perspective, attributes multi-center differences to variations in imaging acquisition factors and models vertebral body morphology, boundaries, and alignment relationships as stable anatomical content, thereby guiding the model to reduce its dependence on center-related appearance features.

[0028] Existing enhancement methods typically only perturb a single layer, such as changing only the overall style or performing only frequency domain transformation, which is insufficient to cover the complex variations in real multi-center images. This application designs three types of causal interventions simultaneously: appearance layer, spatial layer, and frequency domain layer, to simulate differences in overall imaging style, differences in local non-uniform imaging, and differences in frequency distribution, respectively, making the training samples closer to real multi-center scenes.

[0029] This application introduces a priori information about spinal structure into spatial layer intervention, generating a continuous spatial weight map using vertebral body center point sequences and the spinal principal axis to smoothly change local perturbations along the anatomical direction. Compared to random local enhancement, this approach can better simulate local exposure differences, occlusion differences, and artifact differences in spinal X-ray images, while reducing disruption to the continuity of spinal structure.

[0030] In addition, this application combines regional modeling to incorporate regional differences such as cervical, thoracic and lumbar vertebrae, sacrum, and femoral head into the training process. This design can enhance domain diversity while maintaining the rationality of anatomical structure and improve the predictive stability of the model on unknown centers.

[0031] In summary, this application addresses the problems of existing methods, such as their singular enhancement methods, inability to fully understand the original anatomical structure, and strong dependence on central appearance features, by using a causal learning domain generalization model training method. This method can improve the cross-center generalization ability of spinal X-ray image segmentation or detection models.

[0032] In some embodiments, appearance-layer causal intervention includes perturbing the overall brightness, contrast, texture distribution, or noise performance of medical images through a random style transformation module to simulate overall imaging style differences between different medical centers.

[0033] In some embodiments, spatial layer causal intervention includes: Based on the prior anatomical structures in medical images, a spatial weight map is generated to control the intervention intensity in different regions. Based on the spatial weight map, the results of causal intervention in the appearance layer are spatially non-uniformly fused to simulate local domain shifts in medical images caused by uneven exposure, local occlusion, or artifacts.

[0034] In some embodiments, the step of generating a spatial weight map includes: The key point sequence of the target anatomical structure is obtained based on the labeled data, and the principal axis of the structure is obtained by fitting the key point sequence. One or more control points are randomly sampled along the main axis of the structure, and a random weight is assigned to each control point; Interpolation smoothing is performed on one or more control points and their random weights to generate a continuous spatial weight map covering the entire image.

[0035] In some embodiments, frequency domain causal intervention includes: Perform a Fourier transform on the medical image to obtain its frequency domain representation; Frequency-dependent adaptive perturbations are applied to the amplitude spectrum in the frequency domain representation while keeping the phase spectrum unchanged to simulate differences in image edge sharpness, noise levels, and detail texture between different medical centers. The frequency domain representation after perturbation is subjected to inverse Fourier transform to obtain the frequency domain interference image.

[0036] In some embodiments, applying frequency-dependent adaptive perturbation to the amplitude spectrum includes setting different perturbation amplitudes for different low-frequency, mid-frequency, or high-frequency regions to differentiate and simulate overall brightness contrast changes, texture changes, or edge detail changes.

[0037] In some embodiments, the causal learning domain generalization model detection method for multicenter medical images further includes a partitioning modeling step: Based on the characteristics of the anatomical structure, the target area to be detected or segmented is divided into multiple sub-regions; A model is constructed that includes a shared encoder and multiple region-specific decoding branches. The shared encoder is used to extract global context information, and each region-specific decoding branch is used to learn the local fine-grained structural features of the corresponding sub-region.

[0038] For multi-center medical image segmentation or detection tasks, models typically need to be trained on source center data and be directly applicable to unknown medical centers. This application proposes a multi-center medical image cross-center generalization method based on causal intervention enhancement. This method starts from the medical image generation mechanism, distinguishing between stable anatomical structures and variable imaging conditions in the images, and enhances the simulation of domain differences that may occur between different medical centers through multi-layer causal intervention. During the training phase, the model learns from both the original image and the intervention image simultaneously, reducing its dependence on center-related appearance features, thereby improving the model's segmentation or detection generalization ability on unknown medical centers. Specifically, the method may include: (a) Causal Modeling The observation results of multicenter medical imaging can be understood as being formed by the combined effects of two types of factors: anatomical content factors and imaging acquisition factors. Anatomical content factors represent stable structural information directly related to the task label, such as target contours or anatomical sequence relationships; imaging acquisition factors represent appearance variations related to the medical center, such as equipment model, exposure parameters, noise level, contrast, texture mode, image sharpening, post-processing workflow, and local artifacts.

[0039] In medical image segmentation or detection tasks, the prediction results should be primarily determined by anatomical content factors, rather than being influenced by the imaging style specific to a particular medical center. In other words, even if the same anatomical structure presents different appearances under different equipment, exposure conditions, or post-processing procedures, the model should still output consistent segmentation or detection results. Based on this idea, this application transforms the cross-center offset problem into an intervention problem involving imaging acquisition factors. By actively changing center-related appearance factors during the training phase, the model learns feature representations that are insensitive to imaging conditions but provide stable responses to anatomical structures.

[0040] (ii) Multi-level causal intervention To simulate the complex domain differences in real multicenter medical imaging, this application intervenes in imaging acquisition factors from three aspects: appearance, spatial, and frequency domain. All three interventions are based on maintaining the correspondence between anatomical structures and labels, allowing the same anatomical content to appear in multiple forms under different simulated imaging conditions. In this way, the model can be exposed to richer intercenter differences during the training phase, thereby reducing its dependence on single-center imaging styles.

[0041] 1. Causal intervention at the appearance level The appearance layer causal intervention is used to simulate the overall intensity and texture changes caused by differences in imaging equipment, exposure strategies, image post-processing, and noise levels between different medical centers. For the source center image, this application constructs a random style transformation module to perturb the image brightness, contrast, texture distribution, noise performance, or overall grayscale style.

[0042] This random style transformation module can be implemented using lightweight convolutional networks, U-Net structures, random convolutional kernels, learnable filters, or other image style perturbation units. To prevent the generated images from deviating from the reasonable range of medical images, methods such as intensity normalization, linear fusion, or grayscale range cropping can be further employed to ensure that the enhanced images present an overall imaging style at different centers while maintaining the original anatomical structure.

[0043] By intervening at the appearance layer, the model can observe multiple global appearance forms corresponding to the same structure during the training phase, thereby reducing its dependence on specific center brightness, contrast, texture statistics, and noise patterns.

[0044] 2. Spatial-level causal intervention Causal interventions at the appearance layer can simulate variations in overall imaging style across different medical centers, such as differences in brightness, contrast, texture, and noise levels. However, in real clinical images, domain shift does not always manifest as a uniform change across the entire image. Different regions may be affected by uneven exposure, soft tissue occlusion, scattering noise, local artifacts, or regional post-processing differences, resulting in varying degrees of local appearance changes.

[0045] The core of spatial causal intervention is the construction of a spatial weight map, which controls the intensity of intervention received in different regions of the image. This spatial weight map can be generated based on prior target masks. In the model, when the input image is a spinal X-ray, the center points of each vertebra can be extracted using vertebral annotation information, and the spinal principal axis can be obtained by fitting the sequence of center points. Compared to randomly placing points in the entire image, using the spinal principal axis as a reference for spatial intervention is more consistent with the local offset characteristics in clinical images, because phenomena such as exposure gradation, scattering interference, and soft tissue occlusion often exhibit certain patterns along the body's longitudinal axis.

[0046] After obtaining the structural principal axis, this application can randomly sample several control points along the principal axis and assign random weights to each control point. The position, number, and weight of the control points can change randomly during training iterations, preventing the model from remembering fixed intervention areas or intervention patterns. Subsequently, through spline interpolation smoothing, the discrete control points are expanded into a continuous spatial weight map covering the entire image. Due to the spatial continuity of the weight map, the intervention intensity between adjacent regions can transition smoothly, avoiding unnatural abrupt boundaries.

[0047] Based on the aforementioned spatial weight map, this application can fuse different appearance intervention results, enabling different spatial regions within the same image to exhibit style variations of varying degrees and directions. This operation is equivalent to constructing non-uniform acquisition conditions within the image, allowing the model to face local domain shifts more closely resembling real-world multi-center scenes during the training phase. For spinal images, this spatial layer intervention can simulate exposure differences, occlusion differences, and image sharpness differences between different vertebral segments, while maintaining the continuity and anatomical rationality of vertebral alignment as much as possible under the constraints of the structural principal axis.

[0048] 3. Frequency Domain Causal Intervention Causal interventions at the appearance and spatial layers are primarily performed in the pixel domain, simulating overall intensity changes and local appearance differences. However, domain offsets between different medical centers may also manifest in the frequency space. For example, different imaging devices, exposure conditions, image sharpening, denoising algorithms, compression methods, and post-processing workflows can affect image edge sharpness, texture intensity, noise distribution, and detail representation. These differences often correspond to variations in different frequency components, which cannot be adequately simulated by pixel-domain enhancement alone.

[0049] The input image is subjected to a Fourier transform to obtain its frequency domain representation. Subsequently, the amplitude spectrum is perturbed by frequency correlation while the phase spectrum remains unchanged. Finally, an inverse transform is used to restore the image space, yielding a frequency-domain interfered image. The intensity of the amplitude perturbation can be adaptively adjusted according to the frequency location; for example, different perturbation amplitudes can be set for low-frequency, mid-frequency, or high-frequency regions to simulate differences in overall brightness, contrast, edge sharpness, and noise detail at different centers. This perturbation method can be implemented using polynomial weights, radial weights, piecewise weights, or other frequency-related functions.

[0050] Compared to pixel-level random perturbations, frequency-domain causal intervention can more effectively alter the frequency statistical characteristics of an image. Low-frequency components primarily affect overall brightness and contrast, while high-frequency components have a greater impact on edge sharpness, noise, and detail texture. By applying controlled perturbations to different frequency bands, this application can more fully simulate the domain shifts generated by different devices and post-processing procedures, while minimizing disruption to anatomical spatial structures.

[0051] (III) Partition Modeling Under multi-layered causal intervention, the model can reduce its dependence on centrally related appearance factors. However, for medical imaging tasks with strong structural dependence, appearance and frequency perturbations alone are insufficient. Therefore, this application introduces a partitioned modeling strategy. Based on the anatomical characteristics of the task, the target region is divided into several sub-regions with relatively stable morphological features, and corresponding decoding branches or local modeling units are set for different regions. Each region can share the same encoder to extract global contextual information, while simultaneously learning local fine-grained structural features through region-specific branches. This approach can improve the model's ability to distinguish between different anatomical regions while maintaining overall structural consistency.

[0052] The experimental results of the methods in the embodiments of this application are as follows: Figure 3 As shown, the original image, the ground truth labels, and the segmentation results under different methods are presented.

[0053] As can be seen from the image comparison, in different target center datasets, the causal-inspired multicenter medical image cross-center generalization method of this application effectively improves the model's generalization performance on unseen datasets.

[0054] The method described above divides influencing factors in medical imaging into stable anatomical content factors and variable imaging acquisition factors. Anatomical content factors are directly related to segmentation or detection labels, while imaging acquisition factors mainly correspond to differences in equipment, acquisition parameters, exposure strategies, noise levels, image post-processing procedures, and local artifacts between different medical centers. This modeling approach transforms the cross-center migration problem into a problem of controllable intervention in imaging acquisition factors, thereby guiding the model to learn anatomical structural representations that are stable and related to the task labels.

[0055] In some embodiments, the observations of medical images can be decomposed into a combination of stable content and variable factors: I = F ( C , S ) in, I For observed medical images (such as spinal X-ray images); C To stabilize anatomical content factors, representing intrinsic structural information directly related to task labels (such as vertebral morphology, boundaries, and arrangement sequence) remains unchanged in cross-center scenarios; S Variable imaging acquisition factors represent appearance attributes that vary from medical center to medical center (such as equipment model, exposure parameters, noise level, contrast, texture mode, post-processing procedures, etc.). F (⋅) is the image generation function, which combines the above two types of factors into an observed image. Medical image domain shift is attributed to imaging acquisition factors. S The changes clearly distinguish between "content" and "style," laying the foundation for subsequent... S This provides a theoretical basis for targeted interventions.

[0056] The appearance layer intervention simulates the differences in overall brightness, contrast, texture, and noise levels between different medical centers; the spatial layer intervention simulates spatial non-uniform domain shifts such as uneven local exposure, local occlusion, scattering artifacts, and regional post-processing differences; and the frequency domain intervention simulates frequency distribution changes caused by different devices, image sharpening, denoising, compression, and post-processing procedures. These three types of interventions work together at the overall appearance, local spatial, and frequency statistics levels, allowing the model to be exposed to imaging changes in real-world multi-center scenes more closely during the training phase.

[0057] Prior knowledge of anatomical structures can be used to generate spatially continuous weighted maps to simulate shifts in locally non-uniform domains.

[0058] in, For pixel coordinates ( x , y The spatial intervention weight at point ) ranges from 0 to 1; For the first i The location of each control point is obtained by random sampling based on the anatomical structural axis (such as the spinal axis obtained by fitting the center point of the vertebral body in a spinal X-ray image); To be assigned to the i The random weights of each control point change randomly in each training iteration; N is the number of control points. For interpolation smoothing functions (such as spline interpolation), the weights of discrete control points are expanded into a continuous weight map covering the entire image, ensuring a smooth transition of weights between adjacent regions.

[0059] By introducing structural priors to generate continuous spatial weight maps, the intervention intensity changes naturally along the anatomical direction. Compared with random local enhancement, it can more realistically simulate exposure gradation, scattering interference and soft tissue occlusion along the longitudinal axis in spinal images, while avoiding disruption of anatomical continuity.

[0060] Based on the aforementioned spatial weight map, non-uniform fusion of different appearance intervention results is performed to construct a local domain offset:

[0061] in, The image is generated after spatial layer intervention; , Intervention results for appearance layers with two different parameters (such as different brightness, contrast, and texture); This is a spatial weighting graph that controls the weights for fusion at each location.

[0062] By utilizing the spatial weight map within the same image, different regions exhibit style changes of varying degrees and directions, simulating non-uniform domain shifts such as uneven local exposure and regional artifacts in real clinical settings, thus overcoming the limitations of traditional global enhancement.

[0063] Methods for applying differentiated perturbations to different frequency bands in the frequency space:

[0064]

[0065] in, These are frequency domain coordinates; The frequency radius represents the distance from the center of the frequency spectrum. The amplitude spectrum of the original image (after Fourier transform); The amplitude spectrum after intervention (the phase spectrum remains unchanged); This is a frequency-dependent perturbation function, whose value varies with the frequency radius ρ; This refers to the overall disturbance intensity hyperparameter; For piecewise or continuous functions, used to define different disturbance amplitudes in the low-frequency, mid-frequency, and high-frequency regions (e.g., low-frequency band). Smaller to protect overall brightness and contrast, high frequency band (This is to simulate variations in edge sharpness and noise detail).

[0066] By applying frequency-adaptive perturbations to the amplitude spectrum, it is possible to specifically simulate the frequency statistical differences caused by image sharpening, denoising algorithms, compression methods, etc. in different medical centers, while keeping the phase spectrum unchanged to protect the spatial structure and making up for the shortcomings of pixel domain enhancement.

[0067] By integrating the model training objectives under multi-layered causal interventions, and guiding the model to learn stable representations through consistency constraints:

[0068] in, For the task loss function (such as Dice loss for segmentation or cross-entropy loss for detection), in the original image and real labels I Above calculation; The image obtained after multi-layered causal intervention (appearance layer, spatial layer, frequency domain layer); The consistency constraint loss is used to constrain the model to the original image. I and intervention images Output the same prediction results (e.g., using mean squared error or KL divergence); It serves as a balance coefficient, controlling the strength of consistency constraints.

[0069] By forcing the model to output consistent predictions under different intervention perspectives (simulating imaging conditions in different medical centers), the model is prompted to actively ignore factors related to imaging acquisition. S Related feature changes, thereby learning to rely solely on stable anatomical content factors. C The robust representation significantly improves the generalization ability to unknown centers.

[0070] To address the differences in morphological scale, imaging clarity, and occlusion among the cervical, thoracic, lumbar vertebrae, sacrum, and femoral head in spinal imaging, this application adopts a partitioned modeling strategy. It extracts overall contextual information through a shared encoder and sets regional branches to learn the local structural features of different vertebral segments.

[0071] The multicenter whole-spine X-ray vertebral body detection model proposed in this application helps improve cross-center generalization and facilitates the cross-center application of the vertebral body detection model in orthopedic clinical practice.

[0072] It should be noted that the method for detecting causal learning domain generalization models for multi-center medical images provided in this application can be executed by a device for detecting causal learning domain generalization models for multi-center medical images, or by a control module within that device for executing the method. This application uses the example of a device for detecting causal learning domain generalization models for multi-center medical images executing the method to illustrate the device for detecting causal learning domain generalization models for multi-center medical images provided in this application.

[0073] like Figure 4 As shown, in a second aspect of this application, a causal learning domain generalization model detection device for multi-center medical images is provided, comprising: The acquisition module 410 is used to acquire medical images and corresponding annotation data from the source domain; Module 420 is used to build a causal model that decomposes the observation results of medical images into stable anatomical content factors directly related to the task label and variable imaging acquisition factors related to the medical center. Enhancement module 430 is used to perform multi-layer causal intervention enhancement on medical images. The multi-layer causal intervention enhancement includes at least appearance-layer causal intervention, spatial-layer causal intervention, and frequency-domain-layer causal intervention that intervene in variable imaging acquisition factors to generate intervention images that simulate differences in different medical central domains, wherein the intervention images share the same anatomical structures as the medical images. Training module 440 is used to train an initial model using medical images and intervention images to learn anatomical structure representations that are robust to variable imaging acquisition factors and sensitive to stable anatomical content factors, thereby obtaining a generalized detection model for medical images from unknown medical centers. When training the initial model, a consistency constraint loss function is used to constrain the model to output consistent prediction results for medical images and their corresponding intervention images.

[0074] The causal learning domain generalization model detection for multicenter medical imaging in this application embodiment DeviceIt can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network attached storage (NAS), personal computers (PCs), televisions (TVs), ATMs, or self-service machines, etc. The embodiments in this application do not impose specific limitations.

[0075] The causal learning domain generalization model detection device for multi-center medical images in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0076] The causal learning domain generalization model detection device for multi-center medical imaging provided in this application embodiment can achieve... Figure 2 The various processes implemented in the method implementation examples will not be described again here to avoid repetition.

[0077] Optionally, such as Figure 5 As shown, this application embodiment also provides an electronic device 500, including a first processor 501, a first memory 502, and a program or instructions stored in the first memory 502 and executable on the first processor 501. When the program or instructions are executed by the first processor 501, they implement the various processes of the above-described embodiment of the causal learning domain generalization model detection method for multi-center medical images and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0078] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0079] Figure 6 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of this application.

[0080] The hardware structure 600 of the electronic device includes, but is not limited to, the following components: radio frequency unit 601, network module 602, audio output unit 603, input unit 604, sensor 605, display unit 606, user input unit 607, interface unit 608, second memory 609, and second processor 610.

[0081] Those skilled in the art will understand that the hardware structure 600 of the electronic device may also include a power supply (such as a battery) for supplying power to various components. The power supply can be logically connected to the second processor 610 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 6 The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.

[0082] It should be understood that, in this embodiment, the input unit 604 may include a graphics processing unit (GPU) 6041 and a microphone 6042. The GPU 6041 processes image data of still images or videos obtained by an image capture device (such as a camera) in video capture mode or image capture mode. The display unit 606 may include a display panel 6061, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, etc. The user input unit 607 includes a touch panel 6071 and other input devices 6072. The touch panel 6071 is also called a touch screen. The touch panel 6071 may include a touch detection device and a touch controller. Other input devices 6072 may include, but are not limited to, physical keyboards, function keys (such as volume control buttons, power buttons, etc.), trackballs, mice, joysticks, etc., which will not be described in detail here. The second memory 609 can be used to store software programs and various data, including but not limited to applications and operating systems. The second processor 610 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understandable that the aforementioned modem processor may not be integrated into the second processor 610.

[0083] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of the causal learning domain generalization model detection method for multi-center medical images and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0084] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0085] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described embodiment of the causal learning domain generalization model detection method for multi-center medical images, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0086] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0087] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0089] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for detecting causal learning domain generalization models for multi-center medical imaging, characterized in that, include: Acquire medical images and corresponding annotation data from the source domain; A causal model is constructed, which decomposes the observation results of medical images into stable anatomical content factors directly related to the task label and variable imaging acquisition factors related to the medical center. The medical images are subjected to multi-layer causal intervention enhancement, which includes at least appearance-layer causal intervention, spatial-layer causal intervention, and frequency-domain-layer causal intervention to intervene in the variable imaging acquisition factors, in order to generate intervention images that simulate differences in different medical center domains, wherein the intervention images share the same anatomical structures as the medical images; The initial model is trained using the medical images and the intervention images to learn anatomical structural representations that are robust to the variable imaging acquisition factors and sensitive to the stable anatomical content factors, thereby obtaining a generalized detection model for medical images from unknown medical centers. When training the initial model, a consistency constraint loss function is used to constrain the model to output consistent prediction results for the medical images and their corresponding intervention images.

2. The method for detecting causal learning domain generalization models for multi-center medical imaging according to claim 1, characterized in that, The appearance layer causal intervention includes: perturbing the overall brightness, contrast, texture distribution, or noise performance of the medical images through a random style transformation module to simulate the overall imaging style differences between different medical centers.

3. The method for detecting causal learning domain generalization models for multi-center medical imaging according to claim 1, characterized in that, The spatial layer causal intervention includes: Based on the prior anatomical structures in the medical images, a spatial weight map is generated to control the intervention intensity in different regions. Based on the spatial weight map, the results of the causal intervention of the appearance layer are spatially non-uniformly fused to simulate the local domain shift caused by uneven exposure, local occlusion or artifacts in medical images.

4. The method for detecting causal learning domain generalization models for multi-center medical imaging according to claim 1, characterized in that, The steps for generating the spatial weight map include: Based on the labeled data, the key point sequence of the target anatomical structure is obtained, and the principal axis of the structure is obtained by fitting the key point sequence. One or more control points are randomly sampled along the main axis of the structure, and a random weight is assigned to each control point; Interpolation smoothing is performed on the one or more control points and their random weights to generate a continuous spatial weight map covering the entire image.

5. The method for detecting causal learning domain generalization models for multi-center medical imaging according to claim 1, characterized in that, The frequency domain layer causal intervention includes: Perform a Fourier transform on the medical image to obtain its frequency domain representation; A frequency-dependent adaptive perturbation is applied to the amplitude spectrum of the frequency domain representation while keeping the phase spectrum unchanged to simulate the differences in image edge sharpness, noise level and detail texture between different medical centers; The frequency domain representation after perturbation is subjected to inverse Fourier transform to obtain the frequency domain interference image.

6. The method for detecting causal learning domain generalization models for multi-center medical imaging according to claim 5, characterized in that, Applying frequency-dependent adaptive perturbation to the amplitude spectrum includes setting different perturbation amplitudes according to the low-frequency, mid-frequency, or high-frequency regions to differentiate and simulate overall brightness contrast changes, texture changes, or edge detail changes.

7. The method for detecting causal learning domain generalization models for multi-center medical imaging according to claim 1, characterized in that, It also includes the partition modeling step: Based on the characteristics of the anatomical structure, the target area to be detected or segmented is divided into multiple sub-regions; A model is constructed that includes a shared encoder and multiple region-specific decoding branches. The shared encoder is used to extract global context information, and each region-specific decoding branch is used to learn the local fine-grained structural features of the corresponding sub-region.

8. A device for detecting causal learning domain generalization models for multi-center medical imaging, characterized in that, include: The acquisition module is used to acquire medical images and corresponding annotation data from the source domain; A construction module is used to construct a causal model that decomposes the observation results of medical images into stable anatomical content factors directly related to the task label and variable imaging acquisition factors related to the medical center. An enhancement module is used to perform multi-layer causal intervention enhancement on the medical image. The multi-layer causal intervention enhancement includes at least appearance-layer causal intervention, spatial-layer causal intervention, and frequency-domain-layer causal intervention that intervene in the variable imaging acquisition factors to generate intervention images that simulate differences in different medical center domains. The intervention images share the same anatomical structures as the medical image. The training module is used to train an initial model using the medical images and the intervention images to learn anatomical structural representations that are robust to the variable imaging acquisition factors and sensitive to the stable anatomical content factors, thereby obtaining a generalized detection model for medical images from unknown medical centers. When training the initial model, a consistency constraint loss function is used to constrain the model to output consistent prediction results for the medical images and their corresponding intervention images.

9. An electronic device, characterized in that, include: A processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the causal learning domain generalization model detection method for multicenter medical images as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the causal learning domain generalization model detection method for multicenter medical images as described in any one of claims 1-7.

Images