Medical image processing method and device, equipment and medium

Abstract

The embodiment of the invention provides a medical image processing method and device, equipment and a medium, and the method comprises the steps: obtaining a medical image, determining a sensitive region mask in the medical image, and enabling the sensitive region mask to be used for marking a to-be-desensitized pixel region in the medical image; determining a desensitized medical image according to the sensitive area mask and the medical image; judging whether the desensitized medical image meets a preset evaluation condition or not, wherein the preset evaluation condition comprises at least one of an image quality condition and a privacy protection condition; if the preset evaluation condition is met, outputting a desensitized medical image, clearly distinguishing sensitive and non-sensitive areas, and preventing blind processing of diagnosis features of the non-sensitive area from being damaged; according to the sensitive area mask and the medical image, privacy information is removed as much as possible, and the image quality and the diagnostic value of a non-sensitive area are reserved to the maximum extent; it is ensured that the output image has no privacy leakage risk, the scientific research and clinical requirements for image quality are met, and the balance between privacy protection and data availability is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, and in particular to a medical image processing method, device, equipment and medium. BACKGROUND

[0002] With the rapid development of medical imaging technology, medical images have become the core data support for clinical diagnosis, scientific research innovation and multi-center collaboration. However, medical images contain not only key anatomical structures for disease diagnosis (such as tumor lesions, organ textures), but also a large amount of sensitive information that needs to be strictly protected (such as patient facial features, identity text, private organ areas, etc.). In the medical data sharing scenario, if the original medical image is directly used, it may lead to patient privacy leakage, cause legal risks and ethical disputes, and damage patient trust and medical institution reputation; if it is over-processed, it will destroy the diagnostic features in the image, resulting in loss of clinical and scientific research value of the data. Therefore, how to completely remove sensitive information while maximizing the preservation of the diagnostic value of medical images, and achieve the balance between privacy protection and data usability, is the core challenge faced by existing medical image desensitization technology. SUMMARY

[0003] In view of the above problems, the embodiments of the present application are proposed to provide a medical image processing method, device, equipment and medium which overcome the above problems or at least partially solve the above problems.

[0004] In order to solve the above problems, the embodiments of the present application disclose a medical image processing method, which comprises:

[0005] obtaining a medical image and determining a sensitive region mask in the medical image, the sensitive region mask being used to mark a pixel region in the medical image that needs to be desensitized;

[0006] determining a desensitized medical image according to the sensitive region mask and the medical image;

[0007] judging whether the desensitized medical image meets a preset evaluation condition, the preset evaluation condition comprising at least one of an image quality condition and a privacy protection condition;

[0008] if the preset evaluation condition is met, outputting the desensitized medical image.

[0009] Optionally, the determination of the sensitive region mask in the medical image comprises:

[0010] sensitive region detection is performed on the medical image to determine sensitive boundary information of at least one sensitive region;

[0011] determining the sensitive region mask according to the sensitive boundary information.

[0012] Optionally, the step of performing sensitive region detection on the medical image to determine the sensitive boundary information of at least one sensitive region includes:

[0013] A sensitive boundary information detection model is obtained, which includes a multi-scale feature extraction module, a multi-scale position encoding module, a semantic vector extraction module, a dynamic fusion module, and a detection module;

[0014] The medical image is input into the sensitive boundary information detection model, and the multi-scale feature extraction module determines the multi-layer feature map of the medical image. The multi-layer feature map is then input into the multi-scale position encoding module.

[0015] The multi-scale position encoding module embeds position information into each layer of the multi-layer feature map to determine the multi-layer enhanced feature map; the position information includes the horizontal and vertical coordinate information of the medical image.

[0016] The semantic vector extraction module determines a global semantic vector from the deepest enhanced feature map of the multi-layer enhanced feature map; the global semantic vector is used to characterize the overall anatomical semantics of the medical image.

[0017] The dynamic fusion module applies the global semantic vector to each layer of the enhanced feature map in the multi-layer enhanced feature map to perform feature fusion and determine the multi-layer fused feature map.

[0018] The detection module determines sensitive boundary information for at least one sensitive region for each layer of the multi-layer fusion feature map.

[0019] Optionally, the dynamic fusion module includes a perceptron submodule. The step of using the dynamic fusion module to perform feature fusion by applying the global semantic vector to each layer of the multi-layer enhanced feature map to determine the multi-layer fused feature map includes:

[0020] For each layer of the multi-layer enhanced feature map, local statistical features are extracted. These local statistical features are used to characterize the local texture and structural information of the current layer of enhanced feature map.

[0021] The local statistical features are concatenated with the global semantic vector to determine the gated input vector;

[0022] The gated input vector is input into the perceptron submodule, and the perceptron submodule determines the dynamic fusion weight. The dynamic fusion weight is a scalar in the interval [0,1], which is used to adaptively adjust the contribution ratio of the current layer enhanced feature map and the high layer enhanced feature map.

[0023] The multi-layer fusion feature map is determined based on the multi-layer enhanced feature map and the dynamic fusion weights corresponding to each layer enhanced feature map.

[0024] Optionally, the sensitive boundary information includes the coordinate range of the region enclosed by the sensitive boundary, and determining the sensitive region mask based on the sensitive boundary information includes:

[0025] Obtain a blank mask; the blank mask is a blank image of a size corresponding to the medical image;

[0026] For each pixel in the blank mask, if the pixel coordinates belong to the coordinate range within the area enclosed by the sensitive boundary, then the pixel is marked as 1; otherwise, it is marked as 0.

[0027] Optionally, the method further includes:

[0028] Obtain auxiliary information corresponding to the medical image;

[0029] The step of determining the desensitized medical image based on the sensitive region mask and the medical image includes:

[0030] Based on the sensitive area mask, the auxiliary information, and the medical image, a desensitized medical image is determined.

[0031] Optionally, the auxiliary information includes image label information and structural prior map information, and the step of determining the desensitized medical image based on the sensitive region mask, the auxiliary information, and the medical image includes:

[0032] A medical image reconstruction model is obtained, which includes a multimodal fusion module, a forward diffusion perturbation module, an inverse diffusion reconstruction module, and a control fusion module;

[0033] The sensitive region mask, the image label information, the structural prior map information, and the medical image are input into the multimodal fusion module. The multimodal fusion module then splices the sensitive region mask, the image label information, the structural prior map information, and the medical image to determine the input data. Finally, the input data is input into the forward diffusion perturbation module.

[0034] The forward diffusion perturbation module applies Gaussian noise perturbation to the pixel regions in the medical image that are marked by the sensitive region mask, thereby determining the noisy image.

[0035] The inverse diffusion reconstruction module performs stepwise inverse diffusion restoration on the noisy image to determine the reconstructed pixels of the pixel region that needs to be desensitized.

[0036] The control fusion module replaces the pixels in the desensitized pixel region of the medical image with the reconstructed pixels to determine the desensitized medical image.

[0037] Optionally, the image quality conditions include a peak signal-to-noise ratio threshold condition and a structural similarity condition, and the step of determining whether the desensitized medical image meets the preset evaluation conditions includes:

[0038] Determine the peak signal-to-noise ratio (PSNR) of the desensitized medical image; the PSNR is used to characterize the similarity between the pixel values ​​of the non-sensitive regions of the desensitized medical image and the original medical image;

[0039] Determine the structural similarity index between the desensitized medical image and the original medical image; the similarity index is used to characterize the consistency between the non-sensitive region and the original medical image in terms of brightness, contrast, and anatomical structure.

[0040] If the peak signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold and the structural similarity index is greater than a preset similarity threshold, the desensitized medical image is determined to meet the image quality conditions.

[0041] Optionally, determining whether the desensitized medical image meets preset evaluation conditions includes:

[0042] The confidence level of the sensitive region in the desensitized medical image is determined using the sensitive boundary information detection model.

[0043] If the confidence level of the sensitive area is less than a preset confidence threshold, the desensitized medical image is output to the user for manual verification.

[0044] If the manual verification passes, the desensitized medical image is deemed to meet the privacy protection conditions.

[0045] Optionally, it also includes:

[0046] Before determining the sensitive region mask in the medical image, the medical image is preprocessed; the preprocessing includes converting the raw medical image data into a standardized two-dimensional image.

[0047] On the other hand, embodiments of the present invention also disclose a medical image processing apparatus, the apparatus comprising:

[0048] An image acquisition module is used to acquire medical images and determine a sensitive region mask in the medical images, wherein the sensitive region mask is used to mark the pixel regions in the medical images that need to be desensitized;

[0049] The desensitized image acquisition module is used to determine the desensitized medical image based on the sensitive region mask and the medical image;

[0050] An image evaluation module is used to determine whether the desensitized medical image meets preset evaluation conditions, which include at least one of image quality conditions and privacy protection conditions.

[0051] An image output module is used to output the desensitized medical image if the preset evaluation conditions are met.

[0052] Optionally, the image acquisition module includes:

[0053] The region detection submodule is used to perform sensitive region detection on the medical image and determine the sensitive boundary information of at least one sensitive region.

[0054] The mask determination submodule is used to determine the mask of the sensitive region based on the sensitive boundary information.

[0055] Optionally, the region detection submodule includes:

[0056] The model acquisition unit is used to acquire a sensitive boundary information detection model, which includes a multi-scale feature extraction module, a multi-scale position encoding module, a semantic vector extraction module, a dynamic fusion module, and a detection module.

[0057] The feature extraction unit is used to input the medical image into the sensitive boundary information detection model, determine the multi-layer feature map of the medical image through the multi-scale feature extraction module, and input the multi-layer feature map into the multi-scale position encoding module;

[0058] The feature enhancement unit is used to embed position information into each layer of the multi-layer feature map through the multi-scale position encoding module, thereby determining the multi-layer enhanced feature map; the position information includes the horizontal and vertical coordinate information of the medical image;

[0059] A semantic extraction unit is used to determine a global semantic vector from the deepest enhanced feature map of the multi-layer enhanced feature map through the semantic vector extraction module; the global semantic vector is used to characterize the overall anatomical semantics of the medical image.

[0060] The fusion unit is used to perform feature fusion by applying the global semantic vector to each layer of the enhanced feature map in the multi-layer enhanced feature map through the dynamic fusion module, and to determine the multi-layer fused feature map.

[0061] The sensitive boundary determination unit is used to determine the sensitive boundary information of at least one sensitive region for each layer of the multi-layer fusion feature map through the detection module.

[0062] Optionally, the dynamic fusion module includes a sensing machine submodule, and the fusion unit includes:

[0063] The local feature acquisition unit is used to extract local statistical features for each layer of the multi-layer enhanced feature map, and the local statistical features are used to characterize the local texture and structural information of the current layer of enhanced feature map.

[0064] A gated vector determination unit is used to concatenate the local statistical features with the global semantic vector to determine the gated input vector;

[0065] The weight determination unit is used to input the gated input vector into the perceptron submodule and determine the dynamic fusion weight through the perceptron submodule. The dynamic fusion weight is a scalar in the interval [0,1] and is used to adaptively adjust the contribution ratio of the current layer enhanced feature map and the high layer enhanced feature map.

[0066] The fusion feature determination unit is used to determine the multi-layer fusion feature map based on the multi-layer enhanced feature map and the dynamic fusion weight corresponding to each layer enhanced feature map.

[0067] Optionally, the sensitive boundary information includes the coordinate range of the region enclosed by the sensitive boundary, and the mask determination submodule includes:

[0068] The blank mask acquisition unit is used to acquire a blank mask; the blank mask is a blank image of a size corresponding to the medical image.

[0069] The mask marking unit is used to mark each pixel in the blank mask as 1 if the pixel coordinates belong to the coordinate range of the area surrounded by the sensitive boundary, and otherwise mark it as 0.

[0070] Optionally, the device further includes:

[0071] An auxiliary information acquisition submodule is used to acquire auxiliary information corresponding to the medical image. The desensitized image acquisition module includes:

[0072] The integrated desensitized image acquisition submodule is used to determine the desensitized medical image based on the sensitive area mask, the auxiliary information, and the medical image.

[0073] Optionally, the auxiliary information includes image label information and structural prior map information, and the comprehensive desensitized image acquisition submodule includes:

[0074] The reconstruction model acquisition unit is used to acquire a medical image reconstruction model, which includes a multimodal fusion module, a forward diffusion perturbation module, an inverse diffusion reconstruction module, and a control fusion module.

[0075] The data fusion unit is used to input the sensitive region mask, the image label information, the structural prior map information, and the medical image into the multimodal fusion module, and then use the multimodal fusion module to stitch together the sensitive region mask, the image label information, the structural prior map information, and the medical image to determine the input data, and then input the input data into the forward diffusion perturbation module.

[0076] The perturbation unit is used to apply Gaussian noise perturbation to the pixel region that needs to be desensitized and is marked by the sensitive region mask in the medical image through the forward diffusion perturbation module, thereby determining the noisy image;

[0077] The reconstruction unit is used to perform stepwise reverse diffusion restoration on the noisy image through the reverse diffusion reconstruction module to determine the reconstructed pixels of the pixel region that needs to be desensitized;

[0078] The final image determination unit is used to replace the pixels of the pixel region to be desensitized in the medical image with the reconstructed pixels through the control fusion module, thereby determining the desensitized medical image.

[0079] Optionally, the image quality conditions include a peak signal-to-noise ratio threshold condition and a structural similarity condition, and the image evaluation module includes:

[0080] The signal-to-noise ratio (SNR) determination submodule is used to determine the peak SNR of the desensitized medical image; the peak SNR is used to characterize the similarity between the pixel values ​​of the non-sensitive areas of the desensitized medical image and the original medical image;

[0081] The similarity determination submodule is used to determine the structural similarity index between the desensitized medical image and the medical image; the similarity index is used to characterize the consistency between the non-sensitive region and the original medical image in terms of brightness, contrast and anatomical structure.

[0082] The quality determination submodule is used to determine that the desensitized medical image meets the image quality conditions when the peak signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold and the structural similarity index is greater than a preset similarity threshold.

[0083] Optionally, the image evaluation module includes:

[0084] The model reuse submodule is used to determine the confidence level of the sensitive region of the desensitized medical image through the sensitive boundary information detection model;

[0085] The manual verification submodule is used to output the desensitized medical image to the user when the confidence level of the sensitive area is less than a preset confidence threshold, so that the user can perform manual verification.

[0086] The privacy determination submodule is used to determine that the desensitized medical image meets the privacy protection conditions if the manual verification passes.

[0087] Optionally, the device further includes:

[0088] An image preprocessing submodule is used to preprocess the medical image before determining the sensitive region mask in the medical image; the preprocessing includes converting the raw medical image data into a standardized two-dimensional image.

[0089] Accordingly, this invention discloses an electronic device, including: a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the various steps of the above-described medical image processing method embodiment.

[0090] Accordingly, this invention discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the various steps of the above-described medical image processing method embodiment.

[0091] The embodiments of this invention include the following advantages: By acquiring medical images and determining sensitive region masks in the medical images, the sensitive region masks are used to mark the pixel areas in the medical images that need to be desensitized, clearly distinguishing between sensitive and non-sensitive areas, and preventing blind processing that would damage the diagnostic features of non-sensitive areas; by determining the desensitized medical image based on the sensitive region masks and the medical image, privacy information can be removed as much as possible while preserving the image quality and diagnostic value of non-sensitive areas to the greatest extent; by determining whether the desensitized medical image meets preset evaluation conditions, the preset evaluation conditions include at least one of image quality conditions and privacy protection conditions; if the preset evaluation conditions are met, the desensitized medical image is output, ensuring that the output image has no risk of privacy leakage and meets the image quality requirements of scientific research and clinical practice, achieving a balance between privacy protection and data usability. Attached Figure Description

[0092] Figure 1 This is a flowchart illustrating the steps of an embodiment of the medical image processing method of the present invention;

[0093] Figure 2 This is a flowchart illustrating the steps of another embodiment of the medical image processing method of the present invention;

[0094] Figure 3 This is a schematic diagram of a sensitive boundary information detection model according to an embodiment of a medical image processing method of the present invention;

[0095] Figure 4 This is a schematic diagram of a medical image reconstruction model according to an embodiment of a medical image processing method of the present invention;

[0096] Figure 5 This is a structural block diagram of an embodiment of a medical image processing device according to the present invention. Detailed Implementation

[0097] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0098] One of the core concepts of this invention is to accurately locate and process the areas to be desensitized by using a sensitive area mask, and then output the processed images after evaluation and verification, thereby achieving a balance between medical image privacy protection and diagnostic value.

[0099] Reference Figure 1 The diagram illustrates a flowchart of an embodiment of a medical image processing method according to the present invention, which may specifically include the following steps:

[0100] Step 101: Acquire a medical image and determine a sensitive region mask in the medical image. The sensitive region mask is used to mark the pixel regions in the medical image that need to be desensitized.

[0101] The main application scenarios of this invention include: scientific research and academic cooperation scenarios, such as when hospitals, universities, or enterprises share medical image data, it is necessary to remove patient privacy information to ensure that the data is used compliantly for AI model training, disease statistical analysis, etc.; clinical teaching and training scenarios, when displaying typical case images in medical teaching, it is necessary to hide patient privacy information to avoid leakage of sensitive personal data during the teaching process; or cross-institutional data transmission scenarios, when medical images are transferred across institutions such as hospitals, physical examination centers, and medical insurance institutions, it is necessary to reduce the risk of privacy leakage through de-identification processing while retaining diagnostic reference value. Specifically, it can run on medical image processing equipment or systems, specifically referring to hardware devices or integrated hardware and software systems with medical image acquisition, processing, and storage functions, such as medical imaging workstations, medical cloud servers, or embedded medical devices.

[0102] Medical images contain both sensitive information, such as the patient's face and medical record number, and diagnostic information, such as lung texture or tumor lesions. Sharing these images directly without processing would compromise patient privacy; however, blurring the entire image would destroy the diagnostic information. Therefore, it is crucial to clearly define the areas that need to be hidden (privacy) and the areas that need to be preserved (diagnostic information) to provide precise guidance for subsequent processing and avoid blind operations that could lead to a loss of diagnostic value.

[0103] Medical images are images generated using medical imaging techniques such as CT, MRI, and ultrasound. They serve as fundamental data for diagnosis and research. Examples include two-dimensional slices of chest CT scans and three-dimensional volumetric data from brain MRI stored in hospital radiology departments.

[0104] Sensitive region masks are structured marking tools used to label privacy areas that need to be hidden. Their core function is to define the processing scope; only the marked areas are modified in subsequent processing, while unmarked areas remain unchanged. For example, if a patient's face in a medical image needs to be hidden, a sensitive region mask will mark the specific location range of the face in the image using contours or bounding boxes, and subsequent processing will only be performed on this range.

[0105] Step 102: Determine the desensitized medical image based on the sensitive area mask and the medical image;

[0106] By modifying only the range of the mask markings in sensitive areas, operations such as blurring and replacement can be hidden, while the original pixels of unmarked areas are completely preserved. This ensures that the processed image has no risk of privacy leakage and can still be used in diagnostic, scientific research and other scenarios.

[0107] Desensitized medical images refer to medical images that have been processed in which sensitive information in private areas has been removed, while key features in diagnostic areas have been fully preserved. For example, in an original chest CT image, the patient's face is masked with sensitive areas. After processing, these facial areas are replaced with non-identifying feature pixels that match the texture of the surrounding skin, while the texture and contours of diagnostic areas such as the lungs and heart remain completely consistent with the original image. Such an image is a desensitized medical image.

[0108] Step 103: Determine whether the desensitized medical image meets the preset evaluation conditions, which include at least one of image quality conditions and privacy protection conditions;

[0109] Generally, processed images may have two types of problems: 1. Incomplete processing of privacy areas, such as a face that is still recognizable after being blurred; 2. Accidental damage to diagnostic areas, such as blurred lung texture in a lung CT scan. Therefore, it is necessary to verify the processing effect through preset standards to ensure that the image simultaneously meets the requirements of privacy and security, that is, no sensitive information is left behind, and data is usable, that is, to ensure that diagnostic features are not damaged, and to prevent unqualified images from flowing into downstream scenarios.

[0110] Preset evaluation conditions refer to pre-defined verification standards used to determine whether desensitized images are qualified. These standards include at least two aspects: image quality conditions, ensuring that key features of the diagnostic area are not damaged. For example, by evaluating the clarity of the diagnostic area (e.g., the structural similarity between non-sensitive areas and the original image reaches a preset level) and the integrity of texture details (e.g., no obvious blurring at lesion edges), the image can still be used for diagnostic reference or scientific research analysis. Privacy protection conditions, ensuring that sensitive information in privacy areas has been completely removed. For example, by checking whether identifiable features such as facial contours are unrecognizable or medical record number text is unreadable in privacy areas, or by verifying with privacy information detection tools, the amount of residual privacy content is confirmed to be below a preset threshold.

[0111] Step 104: If the preset evaluation conditions are met, output the desensitized medical image.

[0112] Qualified desensitized images that have passed evaluation have both privacy compliance (i.e., no sensitive information remains) and data usability (i.e., complete diagnostic features). They can be safely output to the target scenario, enabling the effective application of medical data under the premise of protecting privacy, and supporting the needs of scientific research, teaching, and cross-institutional collaboration.

[0113] For example, the desensitized CT images that pass the evaluation can be exported in a common medical image format and sent to the research team to train the AI ​​model for detecting lung nodules, or uploaded to a clinical teaching platform for medical students to learn the anatomical features of typical cases. Throughout the process, patient privacy is not compromised, and the diagnostic and research value of the images is not affected.

[0114] By acquiring medical images and identifying sensitive region masks within them—which are used to mark pixel areas requiring desensitization—the system clearly distinguishes between sensitive and non-sensitive areas, preventing indiscriminate processing that could damage the diagnostic features of non-sensitive regions. Based on the sensitive region masks and the medical images, a desensitized medical image is determined, minimizing privacy information while preserving the image quality and diagnostic value of non-sensitive areas to the greatest extent possible. The system then assesses whether the desensitized medical image meets pre-defined evaluation conditions, including at least one of image quality and privacy protection criteria. If the pre-defined evaluation conditions are met, the desensitized medical image is output, ensuring that the output image poses no risk of privacy leakage while meeting the image quality requirements for research and clinical use, thus achieving a balance between privacy protection and data usability.

[0115] Reference Figure 2 The diagram illustrates a flowchart of another embodiment of the medical image processing method of the present invention, which may specifically include the following steps:

[0116] Step 201: Obtain the medical image and corresponding auxiliary information, and determine the sensitive region mask in the medical image. The sensitive region mask is used to mark the pixel regions in the medical image that need to be desensitized.

[0117] Ancillary information associated with medical images refers to supplementary information related to the medical image, used to assist in identifying sensitive areas or optimizing subsequent desensitization processing. It is typically determined based on the content characteristics of the medical image or the needs of the application scenario. By providing additional context or reference, it helps to more accurately locate sensitive areas or provides structural constraints for subsequent desensitization processing, such as ensuring that the desensitized area is consistent with the surrounding anatomical structures. For example, ancillary information may include: image labeling information, used to label the categories of sensitive content in the medical image, such as text regions, face regions, and private organ regions, to help clarify the specific type of sensitive area; and structural reference map information, such as organ outline maps and tissue segmentation maps corresponding to the medical image, used to help define the boundaries between sensitive and non-sensitive areas.

[0118] One exemplary implementation is to acquire CT image sequence data within a cardiac cycle from a CT scanner or PACS system. After acquisition, the system generates a set of DICOM format data files with the .IMA extension. Taking cardiac CT images as an example, the acquired dataset contains all DICOM sequence image data from 1% to 100% of a complete cardiac cycle, fully recording the dynamic changes of the heart throughout the entire beat cycle.

[0119] In one embodiment, prior to step 201, the method further includes the following:

[0120] Before determining the sensitive region mask in the medical image, the medical image is preprocessed; the preprocessing includes converting the raw medical image data into a standardized two-dimensional image.

[0121] Raw medical image data may have issues such as inconsistent formats, mismatched dimensions, or quality differences. Directly using it for sensitive region detection may lead to low algorithm recognition accuracy or poor processing efficiency. The purpose of preprocessing is to convert the raw data into a uniform, standardized two-dimensional image, providing a stable and consistent input basis for the accurate generation of subsequent sensitive region masks.

[0122] Raw medical imaging data refers to the unprocessed, raw data directly generated by medical imaging equipment, which may be three-dimensional volumetric data or non-standardized format. For example, CT equipment typically generates three-dimensional volumetric data containing multiple slices, such as 128 consecutive slices of a 128-slice CT scan, rather than a single, directly usable two-dimensional image.

[0123] Standardized 2D images refer to 2D images that conform to a unified standard after being converted from raw medical image data through preprocessing. Standardization typically includes: dimensional transformation, cutting the 3D volume data into single 2D slices along anatomical planes such as transverse or sagittal planes; size standardization, adjusting the image to a preset size, such as 512×512 pixels, to ensure consistent input size for subsequent algorithms; and pixel value normalization, adjusting pixel values ​​to a fixed range, such as the [0,1] interval, to reduce the impact of brightness and contrast differences between different devices on detection accuracy. Through the above transformations, the raw medical image data is processed into 2D images with a unified format, consistent size, and stable quality, which can be directly input into sensitive area detection processes, improving the accuracy and efficiency of mask generation.

[0124] For example, one possible implementation process is as follows: After reading the aforementioned .IMA format image data, each frame of 3D data is cut into two-dimensional images along different anatomical planes such as transverse, sagittal, and coronal planes. The output size of one anatomical plane, such as the transverse plane, is specified as (H, W, C), where H represents the image height, W represents the width, and C represents the number of channels. Then, a series of preprocessing operations are performed on the generated two-dimensional slices to ensure data consistency and quality: First, normalization is performed, adjusting pixel values ​​to the range of [0, 1] to reduce the impact of data range differences on subsequent processing; finally, the image size is uniformly adjusted, adjusting all slices to the same (H, W, C) size to meet the input requirements of subsequent models.

[0125] In one embodiment, step 201 may include the following sub-steps:

[0126] Sub-step S11: Perform sensitive region detection on the medical image to determine sensitive boundary information of at least one sensitive region;

[0127] From preprocessed medical images, all sensitive regions that need to be hidden are automatically identified and located. The specific location and extent of these regions, i.e., sensitive boundary information, are output, providing a precise basis for generating sensitive region masks. Only by clearly defining the boundaries of sensitive regions can the pixel range requiring desensitization be marked using a mask. In medical images, the location and shape of sensitive regions are diverse; for example, text labels may be small areas, faces may be medium-sized areas, and private organs may be large structural areas, often adjacent to non-sensitive diagnostic areas such as lesions or normal organs. Manually marking sensitive regions is inefficient and prone to omissions; relying on simple rules such as fixed-position cropping may result in mis-cropping of diagnostic areas or omission of sensitive regions. Therefore, automated and precise location detection of sensitive regions is necessary to ensure that subsequent desensitization processing neither leaks privacy nor damages diagnostic information.

[0128] The core logic of this step is to extract features from the image, then distinguish between sensitive and non-sensitive regions, and finally locate the boundaries of the sensitive regions. Specific technical methods can be adjusted according to the application scenario. Feature extraction refers to extracting features that characterize sensitive regions from medical images, such as texture features, structural features, and semantic features; region classification refers to distinguishing between sensitive and non-sensitive regions in the image based on the extracted features; boundary localization refers to accurately calculating the position range of the classified sensitive regions in the image, such as bounding boxes and contours.

[0129] Sensitive boundary information refers to information used to characterize the specific location and extent of sensitive regions in medical images, and it serves as the direct basis for subsequently generating sensitive region masks. Its core function is to define the desensitization scope. For example, the types of sensitive boundary information can include: bounding box coordinates, such as the pixel coordinates of the upper left and lower right corners of the sensitive region, used to mark sensitive regions within a rectangular area; region contours, such as the vertex coordinates of polygons within the sensitive region, used to mark irregularly shaped sensitive regions; and region categories, such as labels for text regions, face regions, and private organ regions, used to distinguish different types of sensitive content.

[0130] In one embodiment, sub-step S11 may include the following sub-steps:

[0131] Sub-step S111: Obtain the sensitive boundary information detection model, which includes a multi-scale feature extraction module, a multi-scale position encoding module, a semantic vector extraction module, a dynamic fusion module, and a detection module;

[0132] A sensitive boundary information detection model was developed for performing sensitive region detection. This model is a deep learning network integrating multiple functional modules, specifically designed for the precise localization of sensitive regions in medical images. The model comprises five core modules, each with a clearly defined function, working together to complete the detection task: a multi-scale feature extraction module for extracting features at different scales from the image; a multi-scale location encoding module for enhancing the spatial location awareness of features; a semantic vector extraction module for extracting the overall anatomical semantics of the image; a dynamic fusion module for fusing multi-layer features and aligning local details with global semantics; and a detection module for outputting the specific boundary information of the sensitive region.

[0133] Sub-step S112: Input the medical image into the sensitive boundary information detection model, determine the multi-layer feature map of the medical image through the multi-scale feature extraction module, and input the multi-layer feature map into the multi-scale position encoding module;

[0134] The preprocessed standardized two-dimensional medical image is then subjected to layer-by-layer feature extraction using a multi-scale feature extraction module to generate multi-layer feature maps. Specifically, these can include: shallow feature maps, which are generally high-resolution and contain detailed image information such as the edges of text labels, small region textures, and organ surface contours; and deep feature maps, which are generally low-resolution and contain high-level semantic information such as the overall structure of organs, anatomical region classification, and image modality judgment.

[0135] Sub-step S113 involves embedding position information into each layer of the multi-layer feature map using the multi-scale position encoding module, thereby determining the multi-layer enhanced feature map; the position information includes the horizontal and vertical coordinate information of the medical image.

[0136] In medical images, the same texture features can correspond to completely different semantics at different locations. For example, a white edge texture in the upper left corner might be a text label, while in the center it might be the edge of a heart. Traditional feature extraction only focuses on pixel values ​​and lacks spatial location information, causing models to be unable to distinguish positional differences, leading to positioning errors, such as misidentifying the edge of a lung as the boundary of a heart. In this embodiment of the invention, the multi-scale positional encoding module embeds positional information into each layer of the feature map to generate an enhanced feature map to alleviate this problem. The positional information is the pixel coordinates of the medical image; for example, when the image resolution is 512×512, the x-range is 0–511, and the y-range is 0–511. The embedding method involves using the x and y coordinates as two additional positional channels, concatenating them into each layer of the multi-layer feature map. For example, if the original feature map has C channels, the enhanced feature map has C+2 channels, adding x and y channels. Each feature map in the multi-layer enhanced feature map contains both the original texture features and the spatial location information of that feature in the image, thus solving the problem of missing positional awareness.

[0137] For example, the implementation could be: in each layer of feature F i The model incorporates a position encoding module to enable spatial location awareness. It employs a coordinate channel embedding method to encode the position channel X. i ,Y i By splicing the images together with the original feature map, the following is formed:

[0138] F i =Concat(F) i ,X i ,Y i )

[0139] in This is the enhanced feature map after embedding location information.

[0140] Sub-step S114: The semantic vector extraction module determines a global semantic vector from the deepest enhanced feature map of the multi-layer enhanced feature map; the global semantic vector is used to characterize the overall anatomical semantics of the medical image.

[0141] Sensitive areas in medical images, especially large structural regions such as genital organs, rely on overall anatomical relationships. For example, the location of the uterus needs to be considered in relation to the relative positions of the bladder and rectum, while the location of the heart needs to be considered in relation to the distribution of the lungs and aorta. Relying solely on local features such as texture can easily lead to misinterpretation; therefore, global semantic vectors are needed to provide overall anatomical semantic guidance.

[0142] The global semantic vector (hereinafter referred to as G) is extracted from the deepest enhanced feature map in the multi-layer enhanced feature map, which is the feature map with the richest semantic information and contains the overall structural features of the image:

[0143] The enhanced feature map is compressed into a vector of a specified dimension through global pooling, and then mapped to a vector of fixed dimension, namely the global semantic vector G, through a fully connected layer. G encodes the overall anatomical semantics of the entire medical image, such as high-level information like the image modality being the heart as the main organ in MRI, the anatomical region being the thoracic cavity, and sensitive areas possibly being distributed in the upper half of the image. This information is used to guide the feature fusion process of the subsequent dynamic fusion module.

[0144] Sub-step S115: Through the dynamic fusion module, feature fusion is performed on each layer of the enhanced feature map in the multi-layer enhanced feature map using the global semantic vector to determine the multi-layer fused feature map;

[0145] Multi-layer augmented feature maps exhibit a semantic gap: shallow augmented feature maps are rich in detail but weak in semantics, while deep augmented feature maps are strong in semantics but lack detail. Direct concatenation and fusion can lead to feature conflicts. Dynamic fusion, guided by global semantics, adaptively adjusts the contribution ratio of features from each layer to resolve this contradiction.

[0146] For example, the enhanced feature map F' of the i-th layer i For example:

[0147] For F' i Perform global pooling to generate a 1×1×C local statistical feature S. i Characterize the texture and structural statistics of the current layer, such as S1 of F'1: many high-frequency edges, possibly containing text; S4 of F'4: many low-frequency structures, possibly containing organs; and combine local statistical features S i The global semantic vector G is concatenated with the global semantic vector G along its channels to generate the gated input vector V. i (V i =[S i [,G], integrating local detail statistics with global anatomical semantics; gating input V iInput the perceptron module and output the dynamic fusion weight W in the interval [0,1]. i W i A value close to 1 indicates that the current layer's details are important and more should be retained; a value close to 0 indicates that the current layer's semantics are weak and needs to rely on higher-level features; then, the features are weighted according to W. i Merge current layer F' i By fusing the feature map with the high-level feature map, a fused feature map F is generated. i :

[0148] Sub-step S116: The detection module determines sensitive boundary information of at least one sensitive region for each layer of the multi-layer fusion feature map.

[0149] Based on the fused feature map, the specific boundaries of sensitive regions are located and output, determining information such as location, range, and category, providing a direct basis for generating sensitive region masks. Specifically, multi-layer fused feature maps can be input into a detection module such as an improved FPN detection head or a YOLO detection head. Each layer of the feature map is responsible for detecting sensitive regions at the corresponding scale: predicting the bounding box coordinates of the sensitive region, predicting the category label of the sensitive region, and predicting the confidence score; at the same time, overlapping and redundant bounding boxes are removed through techniques such as non-maximum suppression, retaining the optimal detection results. Thus, sensitive boundary information of at least one sensitive region is output, such as: (bounding box (100,200)-(300,400), face region, confidence 0.95).

[0150] Reference Figure 3 The diagram illustrates a sensitive boundary information detection model according to an embodiment of a medical image processing method of the present invention, including:

[0151] The backbone network is the foundational part of the entire network, responsible for extracting image features. It includes multiple convolutional layers (Conv), pooling layers (Pool), normalization layers (BN), and activation function layers (ReLU, etc.). These layers progressively process the input image data, extracting features at different levels. Specifically, the input layer of the backbone network receives medical image data, such as preprocessed CT scan images; the convolutional layers extract local features of the image through a series of convolutional operations; the pooling layers downsample the feature maps to reduce the amount of data while preserving important features; and the normalization and activation function layers normalize and nonlinearly transform the features, enhancing the model's expressive power.

[0152] The Neck network, located between the backbone and head networks, further processes and integrates features extracted from the backbone network. It includes multiple multi-scale feature fusion modules to fuse features from different levels, improving feature richness and robustness.

[0153] The Head (head network) is the final part of the model, used to make the final prediction or classification based on the features processed earlier. It includes multiple multi-scale feature fusion modules and the final prediction output (Detection).

[0154] In addition, the model includes a gated weight generation module and a multi-scale guided fusion module. The gated weight generation module receives a global semantic vector as input to provide global semantic information, and local statistical features as input to provide local statistical information. AvgPool (average pooling) is used to perform average pooling on the local statistical features to reduce their dimensionality. Concat (concatenation) is used to concatenate the global semantic vector and the average-pooled local statistical features. MLP (Multilayer Perceptron) is used to process the concatenated features to extract more complex features. Sigmoid is used to generate the gated weight α, with a value between 0 and 1. The multi-scale guided fusion module receives local statistical features from the previous layer as input to provide local statistical information, and local statistical features from the current layer as input to provide local statistical information. Concat (concatenation) is used to concatenate the local statistical features from the previous layer and the local statistical features from the current layer. F_fuse (fusion function) is used to perform a fusion operation on the concatenated features, typically a convolution or fully connected operation. Upsampling is used to upsample the fused features to match the dimensions of other layers. Finally, the upsampled features are multiplied by α generated by the gate weight generation module to achieve weighted feature fusion. The gate weight generation module generates a gate weight by combining global semantic information and local statistical information to control the degree of fusion of multi-scale features. The multi-scale guided fusion module utilizes the local statistical features of the previous and current layers, combined with the gate weight, to achieve multi-scale feature fusion, thereby integrating information at different levels and improving model performance.

[0155] In one embodiment, the dynamic fusion module includes a sensing machine submodule, and sub-step S115 may include the following sub-steps:

[0156] Sub-step S1151: For each layer of the enhanced feature map in the multi-layer enhanced feature map, extract local statistical features. The local statistical features are used to characterize the local texture and structural information of the current layer enhanced feature map.

[0157] Information representing the local details and structural characteristics of each layer's enhanced feature map is extracted to determine the importance of the current layer's features. For each layer of the multi-layer enhanced feature map, local statistical features are extracted. By performing global statistics on the feature map of that layer, such as calculating the mean, variance, maximum, and minimum values ​​of pixel values, the high-dimensional feature map is compressed into a low-dimensional vector. These local statistical features characterize the local texture and structural information of the current layer. For example, the local statistical features of a shallow enhanced feature map may show regions with high-frequency components and dense edge textures; the local statistical features of a deep enhanced feature map may show regions with high-frequency components and clear structural outlines.

[0158] Sub-step S1152: Concatenate the local statistical features with the global semantic vector to determine the gated input vector;

[0159] By combining local detail information from the current layer with global anatomical and semantic information of the entire image, a comprehensive input is formed that reflects both local characteristics and overall structure, providing a complete basis for subsequent dynamic adjustment of fusion weights. Local statistical features and the global semantic vector are concatenated along the channel direction to obtain the gated input vector. Through concatenation, the gated input vector simultaneously contains two aspects of information: local perspective information, such as the texture and structural statistics of the current layer; and global perspective information, such as the anatomical structure and modality of the entire image. This combination avoids misjudgments caused by relying solely on local details.

[0160] Sub-step S1153: Input the gated input vector into the perceptron submodule, and determine the dynamic fusion weight through the perceptron submodule. The dynamic fusion weight is a scalar in the interval [0,1], which is used to adaptively adjust the contribution ratio of the current layer enhanced feature map and the high layer enhanced feature map.

[0161] Based on the gated input vector, dynamic fusion weights are automatically calculated to control the contribution of the current layer's enhanced feature map and the higher-level enhanced feature maps during fusion. The perceptron submodule can be implemented using a multilayer perceptron, but since multilayer perceptrons are existing technology, they will not be elaborated upon here.

[0162] The gated input vector is input into the perceptron submodule, which learns the feature patterns of medical images and outputs a dynamic fusion weight in the range [0,1]. If the weight is close to 1, such as 0.8, it means that the local details of the current layer's enhanced feature map are more important, such as shallow layers containing text labels, and more features of the current layer are retained during fusion.

[0163] If the weight is close to 0, such as 0.3, it means that the semantics of the current layer's enhanced feature map are relatively weak, and the fusion relies more on the semantic guidance of higher-level features. Through this adaptive adjustment, it is ensured that the fused feature map can clearly retain the details of small sensitive areas and accurately align with the global structure of large sensitive areas.

[0164] Sub-step S1154: Determine the multi-layer fusion feature map based on the multi-layer enhanced feature map and the dynamic fusion weight corresponding to each layer enhanced feature map.

[0165] By utilizing the generated dynamic fusion weights, multi-layer enhanced feature maps are fused in an adaptive ratio between current layer details and higher-level semantics, generating multi-layer fused feature maps that balance local details and global semantics. These fused feature maps preserve the details of sensitive regions in the shallow layers while aligning with the semantics of the anatomical structures in the deep layers, providing high-quality feature input for subsequent detection modules to accurately locate sensitive regions. For example, the fusion process can be performed layer by layer from deep to shallow. The generation of each fused feature map depends on the current layer's enhanced feature map, the corresponding dynamic fusion weights, and the feature map already fused in the previous layer. Deep fused feature maps have lower resolution and need to be upsampled, such as through bilinear interpolation or transposed convolution, to adjust their resolution to match that of the current layer's enhanced feature maps. Figure 1 For the current layer's enhanced feature map, it is weighted and summed with the upsampled higher-level fused feature map according to its corresponding dynamic fusion weights, thus iteratively generating multi-layer fused feature maps layer by layer:

[0166] Sub-step S12: Determine the sensitive region mask based on the sensitive boundary information.

[0167] The output sensitive boundary information is converted into a sensitive region mask. By marking whether each pixel in the blank mask belongs to the sensitive region, a binary marker map with the same size as the medical image is generated, thus achieving accurate positioning of the desensitized region.

[0168] In one embodiment, the sensitive boundary information includes the coordinate range of the region enclosed by the sensitive boundary, and sub-step S12 may further include the following sub-steps:

[0169] Sub-step 121: Obtain a blank mask; the blank mask is a blank image of a size corresponding to the medical image;

[0170] Obtain a blank image with the exact same size as the medical image to serve as a canvas for marking sensitive areas. Only with the same size can we ensure that each pixel in the mask corresponds one-to-one with the pixel in the medical image, avoiding misalignment during subsequent fusion. Blank image: refers to an image in which all pixels have a value of 0 or other uniform background value in the initial state, containing no valid content and serving only as a marker carrier. Corresponding size means that the medical image is 512×512 pixels, and the blank mask must also be 512×512 pixels to ensure that the number of pixels in the horizontal and vertical directions is exactly the same.

[0171] Sub-step 122: For each pixel in the blank mask, if the pixel coordinates belong to the coordinate range within the area surrounded by the sensitive boundary, then the pixel is marked as 1; otherwise, it is marked as 0.

[0172] Iterate through each pixel of the blank mask. Based on the coordinate range of the sensitive boundary region (e.g., from the top left corner (x1, y1) to the bottom right corner (x2, y2) of the bounding box), determine whether the pixel belongs to the sensitive region. Use 1 and 0 markers to distinguish between the region to be desensitized and the region not to be desensitized. Specifically, check each pixel in the blank mask row by row or column by column. For pixel (i, j), determine whether its coordinates are within the coordinate range of the sensitive boundary region. For example, if the sensitive boundary is a rectangular region with the top left corner (100, 200) and the bottom right corner (300, 400), then the coordinate range is i ∈ [100, 300] and j ∈ [200, 400].

[0173] If the pixel coordinates fall within the above range, meaning the pixel is part of a sensitive area, then the pixel value of that pixel in the blank mask is marked as 1, indicating that this is an area to be desensitized and needs to be replaced later.

[0174] If the pixel coordinates do not fall within the above range, meaning the pixel is part of a non-sensitive area, then the pixel value is kept at 0 to represent a non-desensitized area, and the original pixel must be retained subsequently. After pixel-by-pixel marking, the blank mask is converted into a sensitive area mask, where areas with a value of 1 correspond to the pixel range in the medical image that needs to be desensitized, and areas with a value of 0 correspond to the diagnostic areas that need to be retained.

[0175] For example, one possible form of representation is:

[0176]

[0177] Where B represents the set of points within the area enclosed by the sensitive boundary, and M... i，j This represents the pixel value of the blank mask.

[0178] Step 202: Determine the desensitized medical image based on the sensitive area mask, the auxiliary information, and the medical image;

[0179] In one embodiment, the auxiliary information includes image label information and structural prior map information, and step 202 may include the following sub-steps:

[0180] Sub-step S21: Obtain a medical image reconstruction model, which includes a multimodal fusion module, a forward diffusion perturbation module, a reverse diffusion reconstruction module, and a control fusion module.

[0181] A dedicated model for performing desensitization and reconstruction is obtained: a medical image reconstruction model. This model comprises four functional modules that work collaboratively to complete the entire desensitization and reconstruction process, ensuring that the desensitized areas retain no privacy remnants while conforming to the anatomical structure of the medical image. Specifically, the multimodal fusion module stitches together sensitive area masks, image labels, structural prior maps, and medical images, integrating multi-source constraint information; the forward diffusion perturbation module applies noise to sensitive areas, providing a starting point for inverse diffusion reconstruction; the inverse diffusion reconstruction module gradually removes noise from sensitive areas under multimodal constraints, generating reconstructed pixels that conform to the anatomical structure; and the control fusion module replaces sensitive areas with reconstructed pixels while retaining the original pixels in non-sensitive areas.

[0182] Sub-step S22: Input the sensitive region mask, the image label information, the structural prior map information, and the medical image into the multimodal fusion module. The multimodal fusion module then stitches together the sensitive region mask, the image label information, the structural prior map information, and the medical image to determine the input data. Finally, input the input data into the forward diffusion perturbation module.

[0183] This system integrates four key information types—sensitive region mask, image labels, structural prior maps, and medical images—into unified input data. This provides multi-dimensional constraints for subsequent directional perturbation and reconstruction, preventing semantic errors in reconstructed sensitive regions. For example, desensitizing a face might result in textures unrelated to a chest CT scan, or structural conflicts might occur, such as desensitized areas of private organs not matching the surrounding tissue morphology. Image labels represent the category labels of sensitive regions, such as faces, text, and private organs. These are transformed into feature maps through an embedding layer, providing semantic category constraints. For instance, face regions should generate skin textures, not bone textures. Structural prior maps represent organ segmentation maps or anatomical reference maps corresponding to the medical images, such as organ contour maps or skeletal structure maps. These provide anatomical constraints, such as ensuring desensitized areas conform to the morphology of surrounding organs and do not exceed anatomical boundaries. Medical images refer to the original medical image pixels, used to provide the original pixels for non-sensitive areas and the initial pixels for sensitive areas.

[0184] Sub-step S23: Through the forward diffusion perturbation module, Gaussian noise perturbation is applied to the pixel regions in the medical image that are marked by the sensitive region mask to be desensitized, and the noisy image is determined.

[0185] Gaussian noise is applied to the masked pixel regions that require desensitization, generating a locally noisy image. This provides a noisy starting point for subsequent inverse diffusion reconstruction, ensuring that the reconstruction process targets only the sensitive regions and that the generated reconstructed pixels transition naturally with the surrounding non-sensitive pixels. During this process, the pixel values ​​in the sensitive regions are contaminated by noise, making original sensitive information such as facial textures and text content completely invisible; conversely, the pixel values ​​in the non-sensitive regions are completely consistent with the medical image, and diagnostic features such as lesions and organ textures remain undamaged.

[0186] For example, the classic formula for noise addition is:

[0187]

[0188] Where t∈[1,T] is a randomly selected diffusion time step, and x0 represents the original image data. It is the starting point of the entire perturbation process, i.e., the original image vector without added noise. t The image data α obtained after a forward diffusion perturbation process at time step t represents the image data obtained at time step t. t代表 A parameter related to time step t, used to control the intensity or proportion of noise added at each step. Its value affects the amount of noise added, α. t The larger the value of , the less noise is added. The specific value can be set according to the business scenario. ∈ represents random noise sampled from the standard normal distribution N(0,1). This noise is added to the original image x0 to construct perturbation samples xt at different time steps.

[0189] Sub-step S24 involves using the reverse diffusion reconstruction module to perform stepwise reverse diffusion restoration on the noisy image to determine the reconstructed pixels of the pixel region that needs to be desensitized.

[0190] The inverse diffusion reconstruction module takes the noisy image and multimodal constraint information as input and gradually removes noise from the sensitive area through a preset number of inverse diffusion steps. After the preset number of steps, the noise in the sensitive area is completely removed, generating reconstructed pixels with the same texture, brightness, and contrast as the surrounding non-sensitive areas.

[0191] For example, the implementation can be shown in the following equation:

[0192]

[0193] Where ∈ θ (xt, t, M, ...) represents the noise predicted based on the neural network θ. The input includes the current noisy image xt, time step t, sensitive region mask M, etc., αˉ t σ represents the cumulative diffusion parameter. t Represents the parameters in the reverse diffusion process, z ~ N(0,1)

[0194] In sub-step S25, the pixels of the pixel region to be desensitized in the medical image are replaced with the reconstructed pixels through the control fusion module, thereby determining the desensitized medical image.

[0195] For pixels in medical images whose sensitive regions are masked as 1, replace them with the generated reconstructed pixels;

[0196] For pixels with a mask mark of 0 in sensitive areas, the original pixels of the medical image are directly preserved;

[0197] The replaced image is a desensitized medical image. In this image, the privacy information of sensitive areas has been completely removed, while the anatomical structure, lesion characteristics and other diagnostic information of non-sensitive areas have been fully preserved. It can be safely used for sharing or circulation.

[0198] Reference Figure 4 The diagram illustrates a medical image reconstruction model according to an embodiment of a medical image processing method of the present invention, including:

[0199] The encoder block includes: a convolutional layer (Conv) for performing convolution operations on the input image to extract features; timestep modulation for adjusting the time step or processing the temporal dimension of features; a ReLU activation function; and downsampling for downsampling the feature map to reduce resolution and increase the receptive field.

[0200] The MembaBlock includes: SSIKKernel, which indicates processing using a specific kernel function (SSIK); Conv1D+Sigmoid, which indicates a one-dimensional convolution followed by a Sigmoid activation function for gating; Linear+GeLU, which indicates a linear transformation followed by a GeLU activation function; and Output, which indicates the generation of the final output.

[0201] The forward diffusion part involves adding noise multiple times to the output feature map X0 of the encoder block and then performing forward diffusion (Skipconnection) to obtain feature maps X1, X2, ..., XT at different time steps. After the feature maps have been diffused, they are reverse reconstructed and then stitched together to generate the final medical image.

[0202] The network training objective of this model is to minimize the difference between the predicted noise and the actual noise.

[0203]

[0204] Meanwhile, to improve the medical usability of desensitized images, structural consistency loss, such as anatomical mask DiceLoss and SSIM structural similarity loss, is also introduced:

[0205]

[0206] The final total loss is:

[0207]

[0208] Where A represents the structural prior image and A′ is the image after model processing, and L is the image label.

[0209] Step 203: Determine whether the desensitized medical image meets the preset evaluation conditions, which include at least one of image quality conditions and privacy protection conditions;

[0210] By ensuring at least one of the image quality and privacy protection conditions, desensitized medical images can retain the diagnostic value of non-sensitive areas while completely removing privacy information from sensitive areas, thus avoiding data invalidation or privacy leaks caused by substandard images.

[0211] In a first exemplary embodiment, the image quality conditions include a peak signal-to-noise ratio threshold condition and a structural similarity condition, and step 203 may include the following:

[0212] Determine the peak signal-to-noise ratio (PSNR) of the desensitized medical image; the PSNR is used to characterize the similarity between the pixel values ​​of the non-sensitive regions of the desensitized medical image and the original medical image;

[0213] Determine the structural similarity index between the desensitized medical image and the original medical image; the similarity index is used to characterize the consistency between the non-sensitive region and the original medical image in terms of brightness, contrast, and anatomical structure.

[0214] If the peak signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold and the structural similarity index is greater than a preset similarity threshold, the desensitized medical image is determined to meet the image quality conditions.

[0215] Peak Signal-to-Noise Ratio (PSNR) is an objective metric for measuring the degree of image distortion. It is based on the mean square error (MSE) of pixel values ​​and is measured in decibels (dB). It quantifies the similarity between the pixel values ​​of non-sensitive areas in a desensitized medical image and the corresponding pixel values ​​in the original medical image. A higher PSNR value indicates a smaller difference in pixel values, meaning the image is closer to its original quality and has not been distorted by the desensitization process. This applies only to non-sensitive areas in desensitized medical images where the sensitive areas are masked as 0.

[0216] A PSNR acceptable standard, such as 30dB, can be preset according to business needs. When PSNR≥30dB, the human eye and subsequent algorithms have difficulty perceiving the difference between non-sensitive areas and the original image, and the diagnostic value can be considered unaffected.

[0217] The Structural Similarity Index (SSIM) is a metric that measures the similarity between two images across three dimensions: brightness, contrast, and structure. A value ranging from 0 to 1 indicates better structural similarity; values ​​closer to 1 indicate better structural consistency. Compared to PSNR, which focuses solely on pixel differences, SSIM emphasizes the consistency of anatomical structure, texture details, and brightness / contrast in non-sensitive areas. A pre-defined SSIM threshold, such as 0.85, can be set to ensure structural consistency based on business needs. When SSIM ≥ 0.85, the anatomical structure and texture details in non-sensitive areas are highly consistent with the original image, supporting subsequent diagnosis or research.

[0218] When the PSNR of the non-sensitive area of ​​the desensitized medical image is greater than the preset signal-to-noise ratio threshold and the SSIM is greater than the preset similarity threshold, the image quality condition is deemed to be met and the diagnostic value of the non-sensitive area is not compromised.

[0219] In another exemplary embodiment, step 204 may include the following:

[0220] The confidence level of the sensitive region in the desensitized medical image is determined using the sensitive boundary information detection model.

[0221] If the confidence level of the sensitive area is less than a preset confidence threshold, the desensitized medical image is output to the user for manual verification.

[0222] If the manual verification passes, the desensitized medical image is deemed to meet the privacy protection conditions.

[0223] After generating the de-identified image, the confidence level of sensitive areas can be determined by reusing the sensitive boundary information detection model to determine whether the privacy protection of the de-identified image is up to standard. The confidence threshold can be dynamically set according to business needs. Model verification may leave some privacy residues, and manual verification involves professionals subjectively judging whether there is sensitive information that can be recognized by the human eye in the reconstructed area.

[0224] When the confidence level of a sensitive area is less than the threshold and manual verification passes, it is determined that the privacy information in the sensitive area has been completely removed.

[0225] Step 204: If the preset evaluation conditions are met, output the desensitized medical image.

[0226] Non-sensitive areas in the desensitized region are ensured to remain intact through PSNR / SSIM; sensitive areas are ensured to have privacy information removed through model reuse and manual review; at this point, the desensitized medical image can be safely output for use in scientific research, teaching, and other scenarios, achieving a balance between privacy protection and data usability.

[0227] By acquiring medical images and identifying sensitive region masks within them—which are used to mark pixel areas requiring desensitization—the system clearly distinguishes between sensitive and non-sensitive areas, preventing indiscriminate processing that could damage the diagnostic features of non-sensitive regions. Based on the sensitive region masks and the medical images, a desensitized medical image is determined, minimizing privacy information while preserving the image quality and diagnostic value of non-sensitive areas to the greatest extent possible. The system then assesses whether the desensitized medical image meets pre-defined evaluation conditions, including at least one of image quality and privacy protection criteria. If the pre-defined evaluation conditions are met, the desensitized medical image is output, ensuring that the output image poses no risk of privacy leakage while meeting the image quality requirements for research and clinical use, thus achieving a balance between privacy protection and data usability.

[0228] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0229] Reference Figure 5 The diagram shows a structural block diagram of an embodiment of a medical image processing device according to the present invention, which may specifically include the following modules:

[0230] Image acquisition module 301 is used to acquire medical images and determine sensitive region masks in the medical images, wherein the sensitive region masks are used to mark pixel regions in the medical images that need to be desensitized;

[0231] The desensitized image acquisition module 302 is used to determine the desensitized medical image based on the sensitive area mask and the medical image;

[0232] Image evaluation module 303 is used to determine whether the desensitized medical image meets preset evaluation conditions, the preset evaluation conditions including at least one of image quality conditions and privacy protection conditions;

[0233] The image output module 304 is used to output the desensitized medical image if the preset evaluation conditions are met.

[0234] Optionally, the image acquisition module includes:

[0235] The region detection submodule is used to perform sensitive region detection on the medical image and determine the sensitive boundary information of at least one sensitive region.

[0236] The mask determination submodule is used to determine the mask of the sensitive region based on the sensitive boundary information.

[0237] Optionally, the region detection submodule includes:

[0238] The model acquisition unit is used to acquire a sensitive boundary information detection model, which includes a multi-scale feature extraction module, a multi-scale position encoding module, a semantic vector extraction module, a dynamic fusion module, and a detection module.

[0239] The feature extraction unit is used to input the medical image into the sensitive boundary information detection model, determine the multi-layer feature map of the medical image through the multi-scale feature extraction module, and input the multi-layer feature map into the multi-scale position encoding module;

[0240] The feature enhancement unit is used to embed position information into each layer of the multi-layer feature map through the multi-scale position encoding module, thereby determining the multi-layer enhanced feature map; the position information includes the horizontal and vertical coordinate information of the medical image;

[0241] A semantic extraction unit is used to determine a global semantic vector from the deepest enhanced feature map of the multi-layer enhanced feature map through the semantic vector extraction module; the global semantic vector is used to characterize the overall anatomical semantics of the medical image.

[0242] The fusion unit is used to perform feature fusion by applying the global semantic vector to each layer of the enhanced feature map in the multi-layer enhanced feature map through the dynamic fusion module, and to determine the multi-layer fused feature map.

[0243] The sensitive boundary determination unit is used to determine the sensitive boundary information of at least one sensitive region for each layer of the multi-layer fusion feature map through the detection module.

[0244] Optionally, the dynamic fusion module includes a sensing machine submodule, and the fusion unit includes:

[0245] The local feature acquisition unit is used to extract local statistical features for each layer of the multi-layer enhanced feature map, and the local statistical features are used to characterize the local texture and structural information of the current layer of enhanced feature map.

[0246] A gated vector determination unit is used to concatenate the local statistical features with the global semantic vector to determine the gated input vector;

[0247] The weight determination unit is used to input the gated input vector into the perceptron submodule and determine the dynamic fusion weight through the perceptron submodule. The dynamic fusion weight is a scalar in the interval [0,1] and is used to adaptively adjust the contribution ratio of the current layer enhanced feature map and the high layer enhanced feature map.

[0248] The fusion feature determination unit is used to determine the multi-layer fusion feature map based on the multi-layer enhanced feature map and the dynamic fusion weight corresponding to each layer enhanced feature map.

[0249] Optionally, the sensitive boundary information includes the coordinate range of the region enclosed by the sensitive boundary, and the mask determination submodule includes:

[0250] The blank mask acquisition unit is used to acquire a blank mask; the blank mask is a blank image of a size corresponding to the medical image.

[0251] The mask marking unit is used to mark each pixel in the blank mask as 1 if the pixel coordinates belong to the coordinate range of the area surrounded by the sensitive boundary, and otherwise mark it as 0.

[0252] Optionally, the device further includes:

[0253] An auxiliary information acquisition submodule is used to acquire auxiliary information corresponding to the medical image. The desensitized image acquisition module includes:

[0254] The integrated desensitized image acquisition submodule is used to determine the desensitized medical image based on the sensitive area mask, the auxiliary information, and the medical image.

[0255] Optionally, the auxiliary information includes image label information and structural prior map information, and the comprehensive desensitized image acquisition submodule includes:

[0256] The reconstruction model acquisition unit is used to acquire a medical image reconstruction model, which includes a multimodal fusion module, a forward diffusion perturbation module, an inverse diffusion reconstruction module, and a control fusion module.

[0257] The data fusion unit is used to input the sensitive region mask, the image label information, the structural prior map information, and the medical image into the multimodal fusion module, and then use the multimodal fusion module to stitch together the sensitive region mask, the image label information, the structural prior map information, and the medical image to determine the input data, and then input the input data into the forward diffusion perturbation module.

[0258] The perturbation unit is used to apply Gaussian noise perturbation to the pixel region that needs to be desensitized and is marked by the sensitive region mask in the medical image through the forward diffusion perturbation module, thereby determining the noisy image;

[0259] The reconstruction unit is used to perform stepwise reverse diffusion restoration on the noisy image through the reverse diffusion reconstruction module to determine the reconstructed pixels of the pixel region that needs to be desensitized;

[0260] The final image determination unit is used to replace the pixels of the pixel region to be desensitized in the medical image with the reconstructed pixels through the control fusion module, thereby determining the desensitized medical image.

[0261] Optionally, the image quality conditions include a peak signal-to-noise ratio threshold condition and a structural similarity condition, and the image evaluation module includes:

[0262] The signal-to-noise ratio (SNR) determination submodule is used to determine the peak SNR of the desensitized medical image; the peak SNR is used to characterize the similarity between the pixel values ​​of the non-sensitive areas of the desensitized medical image and the original medical image;

[0263] The similarity determination submodule is used to determine the structural similarity index between the desensitized medical image and the medical image; the similarity index is used to characterize the consistency between the non-sensitive region and the original medical image in terms of brightness, contrast and anatomical structure.

[0264] The quality determination submodule is used to determine that the desensitized medical image meets the image quality conditions when the peak signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold and the structural similarity index is greater than a preset similarity threshold.

[0265] Optionally, the image evaluation module includes:

[0266] The model reuse submodule is used to determine the confidence level of the sensitive region of the desensitized medical image through the sensitive boundary information detection model;

[0267] The manual verification submodule is used to output the desensitized medical image to the user when the confidence level of the sensitive area is less than a preset confidence threshold, so that the user can perform manual verification.

[0268] The privacy determination submodule is used to determine that the desensitized medical image meets the privacy protection conditions if the manual verification passes.

[0269] Optionally, the device further includes:

[0270] An image preprocessing submodule is used to preprocess the medical image before determining the sensitive region mask in the medical image; the preprocessing includes converting the raw medical image data into a standardized two-dimensional image.

[0271] By acquiring medical images and identifying sensitive region masks within them—which are used to mark pixel areas requiring desensitization—the system clearly distinguishes between sensitive and non-sensitive areas, preventing indiscriminate processing that could damage the diagnostic features of non-sensitive regions. Based on the sensitive region masks and the medical images, a desensitized medical image is determined, minimizing privacy information while preserving the image quality and diagnostic value of non-sensitive areas to the greatest extent possible. The system then assesses whether the desensitized medical image meets pre-defined evaluation conditions, including at least one of image quality and privacy protection criteria. If the pre-defined evaluation conditions are met, the desensitized medical image is output, ensuring that the output image poses no risk of privacy leakage while meeting the image quality requirements for research and clinical use, thus achieving a balance between privacy protection and data usability.

[0272] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0273] This invention also provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the various processes of the above-described medical image processing method embodiment and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0274] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described medical image processing method embodiment and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0275] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0276] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0277] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0278] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0279] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0280] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0281] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device 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 terminal device. Without further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0282] The above provides a detailed description of a medical image processing method, apparatus, device, and medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A medical image processing method characterized by, The method comprises: obtaining a medical image and determining a sensitive region mask in the medical image, the sensitive region mask being used to mark a pixel region in the medical image that needs to be desensitized; determining a desensitized medical image according to the sensitive region mask and the medical image; judging whether the desensitized medical image meets preset evaluation conditions, the preset evaluation conditions including at least one of an image quality condition and a privacy protection condition; if the preset evaluation conditions are met, outputting the desensitized medical image.

2. The medical image processing method according to claim 1, characterized in that, The determination of the sensitive region mask in the medical image comprises: sensitive region detection is performed on the medical image to determine sensitive boundary information of at least one sensitive region; the sensitive region mask is determined according to the sensitive boundary information.

3. The medical image processing method according to claim 2, characterized in that, The sensitive region detection performed on the medical image to determine the sensitive boundary information of at least one sensitive region comprises: obtaining a sensitive boundary information detection model, the sensitive boundary information detection model comprising a multi-scale feature extraction module, a multi-scale position encoding module, a semantic vector extraction module, a dynamic fusion module and a detection module; inputting the medical image into the sensitive boundary information detection model, determining multi-layer feature maps of the medical image through the multi-scale feature extraction module, and inputting the multi-layer feature maps into the multi-scale position encoding module; embedding position information for each layer of the multi-layer feature maps through the multi-scale position encoding module, thereby determining multi-layer enhanced feature maps; the position information comprises horizontal and vertical coordinate information of the medical image; determining a global semantic vector from the deepest layer of the multi-layer enhanced feature maps through the semantic vector extraction module; the global semantic vector is used to represent the overall anatomical semantics of the medical image; applying the global semantic vector to perform feature fusion for each layer of the multi-layer enhanced feature maps through the dynamic fusion module, thereby determining multi-layer fusion feature maps; determining the sensitive boundary information of at least one sensitive region for each layer of the multi-layer fusion feature maps through the detection module.

4. The medical image processing method according to claim 3, characterized by, The dynamic fusion module comprises a perception machine sub-module, and the application of the global semantic vector to perform feature fusion for each layer of the multi-layer enhanced feature maps through the dynamic fusion module to determine the multi-layer fusion feature maps comprises: extracting local statistical features for each layer of the multi-layer enhanced feature maps, respectively; the local statistical features are used to represent local texture and structure information of the current layer of the enhanced feature maps; concatenating the local statistical features and the global semantic vector to determine a gating input vector; inputting the gating input vector into the perception machine sub-module to determine a dynamic fusion weight through the perception machine sub-module; the dynamic fusion weight is a scalar in the interval [0, 1] and is used to adaptively adjust the contribution proportion of the current layer of the enhanced feature maps and the higher layer of the enhanced feature maps; determining the multi-layer fusion feature maps according to the multi-layer enhanced feature maps and the dynamic fusion weight corresponding to each layer of the enhanced feature maps.

5. The medical image processing method according to claim 2, characterized by, The sensitive boundary information includes a coordinate range of a sensitive boundary surrounding area, and the determining the sensitive area mask according to the sensitive boundary information includes: obtaining a blank mask; the blank mask is a blank image corresponding to the size of the medical image; for each pixel point in the blank mask, if the pixel point coordinate belongs to the coordinate range within the sensitive boundary surrounding area, the pixel point is marked as 1, otherwise it is marked as 0.

6. The medical image processing method according to claim 1, characterized by, The method further comprises: obtaining auxiliary information corresponding to the medical image; the determining the desensitized medical image according to the sensitive area mask and the medical image comprises: determining the desensitized medical image according to the sensitive area mask, the auxiliary information and the medical image.

7. The medical image processing method according to claim 6, characterized by, The auxiliary information includes image label information and structure prior graph information, and the determining the desensitized medical image according to the sensitive area mask, the auxiliary information and the medical image comprises: obtaining a medical image reconstruction model, the medical image reconstruction model including a multi-modal fusion module, a forward diffusion disturbance module, an inverse diffusion reconstruction module and a control fusion module; inputting the sensitive area mask, the image label information, the structure prior graph information and the medical image into the multi-modal fusion module, splicing the sensitive area mask, the image label information, the structure prior graph information and the medical image through the multi-modal fusion module to determine input data, and inputting the input data into the forward diffusion disturbance module; applying Gaussian noise disturbance to the pixel region needing desensitization marked by the sensitive area mask in the medical image through the forward diffusion disturbance module to determine a noised image; gradually inversely diffusing and restoring the noised image through the inverse diffusion reconstruction module to determine the reconstructed pixels of the pixel region needing desensitization; replacing the pixels of the pixel region needing desensitization in the medical image with the reconstructed pixels through the control fusion module to determine the desensitized medical image.

8. The medical image processing method according to claim 1, characterized by, The image quality condition includes a peak signal-to-noise ratio threshold condition and a structure similarity condition, and the determining whether the desensitized medical image meets the preset evaluation condition comprises: determining the peak signal-to-noise ratio of the desensitized medical image; the peak signal-to-noise ratio is used to represent the similarity between the pixel value of the non-sensitive area of the desensitized medical image and the original medical image; determining the structural similarity index of the desensitized medical image and the medical image; the similarity index is used to represent the consistency of the non-sensitive area in brightness, contrast and anatomical structure with the original medical image; in the case that the peak signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, and the structural similarity index is greater than the preset similarity threshold, it is determined that the desensitized medical image meets the image quality condition.

9. The medical image processing method according to claim 3, characterized by, The determining whether the desensitized medical image meets the preset evaluation condition comprises: determining the sensitive area confidence of the desensitized medical image through the sensitive boundary information detection model; in the case that the sensitive area confidence is less than the preset confidence threshold, outputting the desensitized medical image to the user for manual verification; If the artificial verification passes, it is determined that the desensitized medical image satisfies the privacy protection condition.

10. The medical image processing method according to claim 2, characterized by, Further comprising: Before determining the sensitive region mask in the medical image, pre-processing the medical image; The pre-processing includes converting original medical image data into standardized two-dimensional images.

11. A medical image processing apparatus characterized by comprising: The device comprises: An image acquisition module configured to acquire a medical image and determine a sensitive region mask in the medical image, the sensitive region mask being used to mark a pixel region in the medical image that needs to be desensitized; A desensitized image acquisition module configured to determine a desensitized medical image according to the sensitive region mask and the medical image; An image evaluation module configured to determine whether the desensitized medical image satisfies a preset evaluation condition, the preset evaluation condition including at least one of an image quality condition and a privacy protection condition; An image output module configured to output the desensitized medical image if the preset evaluation condition is satisfied.

12. An electronic device, comprising: Further comprising: A processor, a memory, and a computer program stored on the memory and capable of running on the processor, the computer program being executed by the processor to implement the steps of the medical image processing method according to any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, A computer program is stored on the computer readable storage medium, and the computer program is executed by the processor to implement the steps of the medical image processing method according to any one of claims 1-10.

Images