Tampering detection method and system for CT medical images, tampering detection model training method and system

Abstract

The application provides a tampering detection method, a tampering detection model training method and a system for CT medical images, which are realized based on a pre-trained feature extraction model and include the following steps: inputting a to-be-detected CT medical image into a multi-scale edge feature extraction module to output edge feature maps of multiple scales, and inputting the to-be-detected CT medical image into a multi-scale noise feature extraction module to output noise feature maps of multiple scales; inputting the edge feature maps and the noise feature maps into multi-scale fusion modules respectively to obtain edge fusion feature maps and noise fusion feature maps; performing weighted fusion on the edge fusion feature maps and the noise fusion feature maps by using a double-feature fusion module to obtain a fusion feature map corresponding to the to-be-detected CT medical image; and inputting the fusion feature map into a pre-trained tampering detection classification model to output a tampering detection result corresponding to the to-be-detected CT medical image. The application can improve the accuracy of tampering detection of CT medical images.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, and in particular to a method for tamper detection of CT medical images, a method for training a tamper detection model, and a system. Background Technology

[0002] With the development of hospital image archiving and communication systems, remote image consultation systems, cloud-based image reading systems, and AI-assisted diagnostic platforms, computed tomography (CT) images in medicine are no longer just static images stored on a single machine. Instead, they serve as crucial medical data assets, frequently accessed in multiple stages including acquisition, transmission, storage, retrieval, and analysis. While this improves medical efficiency, it also introduces new security risks. Especially in the context of image digitization and networking, CT medical images can be maliciously tampered with. Examples of malicious tampering include inserting fake lesions into lung CT images, deleting or erasing existing lesion areas, copying a lesion to another location, splicing abnormal tissue from other images into the current CT image, and repairing or redrawing local areas to conceal real lesions. These acts of tampering can mislead doctors' judgments, potentially leading to missed diagnoses, misdiagnoses, or incorrect treatments. Therefore, researching CT medical image tampering detection technology has significant clinical safety implications and application value.

[0003] While there is some existing research on medical image tampering, most methods still suffer from limitations such as limited applicability, insensitivity to small lesions, high false positive rates, and limited generalization ability. This is particularly true for CT medical images, where subtle textures, smooth boundary transitions, and high local structural similarity make it difficult to accurately distinguish normal changes from tampering traces using only a single feature.

[0004] While existing CT medical image tampering detection technologies can identify forgery traces to a certain extent, they still have the following significant shortcomings: ① Lack of hierarchical modeling capability for different types of tampering traces: Existing methods tend to use a single network to uniformly extract features (such as automatically learning tampering features using convolutional neural networks or other deep models), resulting in insufficient adaptability to complex tampering scenarios; ② Insufficient sensitivity to small-scale, low-contrast tampering areas: Tampering in medical images often occurs in small lesion areas or tissue areas with weak edge transitions. These areas are small in size, have slight grayscale changes, and are easily confused with normal anatomical structures. Existing methods, if overly reliant on global semantic information, are prone to missed detections, especially in scenarios such as lesion insertion, lesion erasure, and local stitching.

[0005] Therefore, there is an urgent need for a method for detecting tampering in CT medical images that can overcome the shortcomings of existing technologies and improve detection accuracy. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a tampering detection method, a tampering detection model training method and system for CT medical images, which can improve the ability to identify tampered regions in CT medical images by combining edge enhancement, noise residual analysis or multi-scale feature fusion.

[0007] One aspect of the present invention provides a tampering detection method for CT medical images. This method is based on a pre-trained feature extraction model, which includes a multi-scale edge feature extraction module, a multi-scale noise feature extraction module, a multi-scale fusion module, and a dual-feature fusion module. The method includes the following steps: The CT medical image to be detected is input into the multi-scale edge feature extraction module, which outputs edge feature maps at multiple scales. The CT medical image to be detected is also input into the multi-scale noise feature extraction module, which outputs noise feature maps at multiple scales. The edge feature map and the noise feature map are respectively input into the multi-scale fusion module to obtain the edge fusion feature map and the noise fusion feature map; The edge fusion feature map and the noise fusion feature map are weighted and fused using a dual feature fusion module to obtain the fusion feature map corresponding to the CT medical image to be detected. The fused feature map is input into a pre-trained tamper detection classification model, and the tamper detection result corresponding to the CT medical image to be detected is output.

[0008] In some embodiments of the present invention, after feature fusion using a multi-scale fusion module, the method further includes: A prior knowledge graph is generated based on unaltered CT medical images; wherein the unaltered CT medical images and the CT medical images to be detected belong to the same anatomical region; the prior knowledge graph is used to represent the location of organ regions, tissue structure regions or lesion boundary regions in the unaltered CT medical images; The prior knowledge graph and specific feature maps are used as inputs to the attention mechanism, and the output is the attention weight. The attention weight is then used to adjust the specific feature map. The specific feature map is an edge fusion feature map, a noise fusion feature map, or a fusion feature map corresponding to the CT medical image to be detected.

[0009] In some embodiments of the present invention, the multi-scale fusion module is used to perform the following operations: The multi-scale feature map of the input multi-scale fusion module is downsampled by a bottom-up feature encoding path to obtain the first sampled feature map of the multi-scale; The first multi-scale sampled feature map is upsampled by a top-down feature encoding path to obtain the second multi-scale sampled feature map. Align the first and second sampled feature maps, and obtain the fused feature map corresponding to the multi-scale feature map of the input multi-scale fusion module through weighted fusion.

[0010] In some embodiments of the present invention, the fusion feature map corresponding to the CT medical image to be detected is obtained in the following manner: Align the edge fusion feature map and the noise fusion feature map, and use the attention mechanism to generate the corresponding first attention map and second attention map respectively; By combining the first and second attention maps through element-wise multiplication, and using the resulting joint attention map as the joint weight; By performing weighted fusion processing on the edge fusion feature map and the noise fusion feature map through joint weights, the fusion feature map corresponding to the CT medical image to be detected is obtained.

[0011] In some embodiments of the present invention, the CT medical image to be detected is obtained by performing grayscale normalization and Gaussian filtering on the acquired original CT medical image to be detected. The tampering detection results corresponding to the CT medical image to be detected include a judgment result indicating whether tampering has occurred and a tampering probability map; The multi-scale edge feature extraction module is built based on the U-Net model, while the multi-scale noise feature extraction module is used to extract noise features in the multi-scale transform domain.

[0012] In some embodiments of the present invention, before inputting the CT medical image to be detected into the multi-scale edge feature extraction module, the method further includes: performing edge enhancement processing on the CT medical image to be detected by gradient response.

[0013] In some embodiments of the present invention, before inputting the CT medical image to be detected into the multi-scale noise feature extraction module, the method further includes: performing noise residual enhancement processing on the CT medical image to be detected.

[0014] Another aspect of the present invention provides a training method for a tampering detection model for CT medical images. The tampering detection model includes a feature extraction model and a tampering detection classification model. The trained tampering detection model is used to execute the tampering detection method described in any of the above embodiments. The method includes the following steps: Multiple training samples and their corresponding labels are obtained; the training samples include unaltered samples and altered samples obtained by adding edge perturbation and noise perturbation to the unaltered samples. Both the unaltered and altered samples in the training samples are CT medical images. Input the unaltered sample and the tampered sample into the tamper detection model to be trained, and output the tamper detection results corresponding to the unaltered sample and the tamper detection results corresponding to the tampered sample. The training loss of the feature extraction model is formed by the feature maps output by the multi-scale edge feature extraction module, multi-scale noise feature extraction module, multi-scale fusion module, and dual feature fusion module for untampered samples and tampered samples, respectively. The training loss of the tamper detection classification model is formed by the tamper detection results and corresponding labels of the untampered samples and tampered samples. The tamper detection model is then iteratively trained using the training losses of the feature extraction model and the tamper detection classification model to obtain the trained tamper detection model.

[0015] In some embodiments of the present invention, the tampered sample and its corresponding label are obtained by inputting the untampered sample into a pre-trained generative model; The generative model and the tamper detection classification model are trained together.

[0016] Another aspect of the present invention provides a tamper detection system for CT medical images, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method described in any of the above embodiments.

[0017] The tamper detection method, tamper detection model training method, and system proposed in this invention for CT medical images can extract features from structural and statistical anomalies through different branches. Furthermore, it can simultaneously integrate shallow detail information and deep semantic information, taking into account both fine-grained tamper localization and the identification of larger-area anomalies, and then jointly modeling to identify tampered regions. The method proposed in this application performs layered modeling of tamper traces through dual-branch feature extraction and feature fusion. Because the feature extraction is complete, it can accurately identify small-scale, low-contrast tampered regions in the image, achieving high tamper detection accuracy.

[0018] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0019] Those skilled in the art will understand that the objectives and advantages achievable with this invention are not limited to those specifically described above, and that the above and other objectives achievable with this invention will become clearer from the following detailed description. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a schematic diagram of a tamper detection framework in one embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating a tamper detection method in one embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the feature fusion process performed by the multi-scale fusion module in one embodiment of the present invention.

[0023] Figure 4 This is a flowchart illustrating a tamper detection model training method in one embodiment of the present invention.

[0024] Figure 5 This is a flowchart illustrating the tamper detection model training method in another embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0026] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0027] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0028] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0029] Existing methods for tampering detection in CT medical images mostly employ a single-channel, uniform feature modeling approach, which lacks sufficient sensitivity to small-scale, low-contrast tampering areas. Considering that CT medical images are grayscale images with fine textures and smooth boundary transitions, the grayscale differences between many lesion areas and normal tissues are not obvious. For medical image scenarios, this application proposes improvements to existing technologies in the following directions: First, enhance image boundary, texture, or artifact information to improve sensitivity to local abnormal areas; second, utilize residual features, high-frequency features, or statistical features to enhance the ability to identify tampering traces; and third, improve the model's adaptability to tampering areas of different sizes through multi-scale feature integration.

[0030] Based on this, and addressing the issues of subtle tampering traces, concealed structural changes, and inconspicuous local statistical anomalies in CT medical image tampering detection, this application proposes a CT medical image tampering detection method based on edge-noise dual-branch collaborative modeling. The core of this method lies in classifying tampering traces in CT medical images into two categories: edge anomaly information and noise anomaly information. Feature extraction models incorporating edge and noise anomaly branches are used to specifically model the edge and noise feature maps respectively, thereby improving the ability to identify complex tampering traces. Furthermore, multi-scale feature fusion and dual-branch feature fusion enhance the detection capability for tampering regions of different sizes.

[0031] like Figure 1 As shown, in this application, the edge anomaly branch can use the multi-scale edge feature extraction module in the feature extraction model to extract multi-scale edge feature maps, and use the multi-scale fusion module to obtain an edge fusion feature map; the noise anomaly branch can use the multi-scale noise feature extraction module in the feature extraction model to extract multi-scale noise feature maps, and use the multi-scale fusion module to obtain an edge fusion feature map.

[0032] The basic idea behind this application for constructing a dual-branch feature extraction model is that after CT medical images are tampered with, they usually leave not only structural abnormalities in organ boundaries, lesion contours, and tissue transition areas, but also statistically abnormal traces in local noise distribution, texture residuals, and high-frequency responses. Therefore, it is necessary to analyze both edge information and noise information simultaneously to fully describe the tampering behavior.

[0033] Furthermore, the overall framework for tamper detection proposed in this application can be as follows: Figure 1As shown, this framework is based on a tampering detection model, which can include a feature extraction model and a tampering detection classification model. The feature extraction model can include a multi-scale edge feature extraction module, a multi-scale noise feature extraction module, a multi-scale fusion module, and a dual-feature fusion module. Specifically, the edge anomaly branch is used to extract boundary discontinuities and contour anomalies from CT medical images, the noise anomaly branch is used to extract statistical inconsistencies in noise from CT medical images, and the dual-feature fusion module (also called the dual-branch feature fusion module) is used to jointly model the features (containing anomaly information) extracted from the edge anomaly branch and the noise anomaly branch, enabling the overall tampering detection framework to output tampering detection results for CT medical images.

[0034] Figure 2 This is a flowchart illustrating a tamper detection method for CT medical images according to one embodiment of the present invention. Figure 2 As shown, this method is based on a pre-trained feature extraction model and includes steps S110 to S140, as detailed below: Before executing step S110, it is necessary to acquire the CT medical image to be detected. Whether the CT medical image to be detected is acquired during the inference phase or the unaltered and tampered samples used for model training during the training phase, this application can directly use public datasets or collect the corresponding CT images itself.

[0035] It should be noted that, in the research and development and implementation of the technical solutions involved in this application, all user personal information (such as CT medical images) was processed in strict accordance with the principles of legality, legitimacy, necessity, and good faith. Specifically, the relevant data was obtained through one or more of the following compliant methods: ① Before collecting the user's personal information, the user was clearly informed of the purpose, method, and scope of the collection, and the user's separate and explicit authorization and consent were obtained; ② The personal data used came from publicly available datasets permitted by laws and regulations, and the personal data underwent necessary anonymization or de-identification processing during use to ensure that it could not identify a specific individual and that the information was irreversible; ③ The use of personal data was limited to the tamper detection and model training purposes described in this application, and strict technical and management measures were taken to protect data security and prevent information leakage, abuse, and unauthorized access.

[0036] Furthermore, after acquiring CT medical images (CT medical images to be detected, unaltered samples, or tampered samples), this application can also perform preprocessing operations on the acquired CT medical images. Preprocessing includes grayscale normalization and Gaussian filtering, the purpose of which is to suppress irrelevant noise interference while preserving important structural information of the CT medical image. Grayscale normalization scales pixel values ​​to a uniform range (usually [0,1] or [-1,1]) to eliminate dimensional differences. This application does not specifically limit the normalization method; for example, Z-Score normalization or Min-Max normalization can be used. Gaussian filtering is used to smooth the image to suppress high-frequency noise interference. The preprocessed CT medical image can retain its true structural features and serve as the original input for the feature extraction module.

[0037] Step S110: Input the CT medical image to be detected into the multi-scale edge feature extraction module and the multi-scale noise feature extraction module in the pre-trained feature extraction model, respectively, and output the edge feature map (hereinafter referred to as multi-scale edge feature map) and the noise feature map (hereinafter referred to as multi-scale noise feature map) at multiple scales.

[0038] In unaltered CT medical images, normal complex structural regions may exhibit strong variations at the edge level, but show no significant anomalies at the noise statistics level; similarly, regions with local noise fluctuations may show statistical disturbances, but their boundary structures remain normal. However, genuinely tampered CT medical images typically exhibit anomalies in both the edge and noise dimensions simultaneously. Therefore, this application constructs edge anomaly and noise anomaly branches within the overall tamper detection framework, enabling the separate modeling of two key types of tampering traces in CT medical images.

[0039] The edge anomaly branch is used to extract tampering traces from CT medical images at the structural level. The extracted features may include structural anomalies, making the edge anomaly branch particularly suitable for detecting boundary distribution anomalies caused by geometric adjustments, morphological changes, and local content replacement. The noise anomaly branch mainly extracts tampering traces at the statistical level, and the extracted features may include both structural and statistical anomalies. Through this division of labor, the present invention can break down tampering traces in CT medical images into two more easily learned sub-problems, thereby improving overall detection performance.

[0040] More specifically, when CT medical images are tampered with, especially when lesions are inserted, erased, local tissues are copied and pasted, or other image areas are spliced ​​together, unnatural changes can easily appear in the local boundaries. These changes are not necessarily obvious differences in black and white; sometimes they simply manifest as sudden discontinuities in organ outlines, uneven lesion edges, abrupt changes in tissue boundary direction, or local geometric structures that visually fail to connect with surrounding tissues. Therefore, the main purpose of constructing the edge anomaly branch in this application is to identify structural anomalies such as edge breaks or geometric deformations, thereby using edge information to discover tampering clues such as edge breaks, outline discontinuities, local geometric deformations, or abnormal boundary distributions.

[0041] In the edge anomaly branch, a multi-scale edge feature extraction module and a multi-scale fusion module can be used to extract edge anomaly features, enabling the classification model to more sensitively detect the damage to boundary continuity and local geometric structure caused by tampering operations. Specifically, the multi-scale edge feature extraction module can identify edge breaks or geometric deformations, achieve edge segmentation, and extract multi-level edge features, extracting high-level structural information while preserving shallow edge details. The multi-scale edge features obtained by the multi-scale edge feature extraction module can be further input into the multi-scale fusion module for multi-scale fusion to obtain an edge anomaly feature representation suitable for tampering detection tasks. The multi-scale edge feature extraction module mentioned in this application can be constructed using structures such as U-Net, ResUNet, DeepLab, HRNet (High-Resolution Network), or Swin-UNet. This invention is not limited to these; any module that can utilize its encoding and decoding structure to complete multi-level edge information extraction is acceptable.

[0042] The multi-scale edge feature extraction module can extract and restore edge information of CT medical images layer by layer through encoding and decoding processes, and retain shallow edge details through skip connections, thus better balancing local edge information and overall structural information. Therefore, the edge anomaly branch proposed in this application can obtain a more stable boundary anomaly response when dealing with scenarios such as small-scale lesion tampering, edge blurring and replacement, and local geometric adjustments, and has a higher structural awareness capability compared with existing edge detection schemes that rely solely on single-scale edge information.

[0043] The multi-scale nature of edge feature maps is mainly reflected at the spatial structure level, with multiple edge feature maps extracted from edge anomaly branches having different scales. Taking the U-Net model as an example, high-level semantics can be extracted through downsampling, details can be recovered through upsampling, and shallow edge information can be preserved through skip connections, so that the extracted features contain structural boundaries and region contour information.

[0044] In some embodiments of the present invention, before inputting the CT medical image to be detected into the multi-scale edge feature extraction module, the tamper detection method proposed in this application further includes: performing edge enhancement processing on the CT medical image to be detected through gradient response. For example, Sobel edge enhancement processing can be performed on the image with abnormal edge branches to highlight structural boundaries. The Sobel operator calculates gray-level changes through convolution kernels in both horizontal and vertical directions, which can directly highlight the gradient changes of organ contours, lesion edges, and tissue junctions in CT medical images, and has the advantages of low computational cost and simple implementation. The above-mentioned edge enhancement method using the Sobel operator is only an example. The edge enhancement method designed in this application can adopt gradient response, such as the Prewitt operator, Laplacian operator, Scharr operator, or learnable edge convolution.

[0045] Furthermore, the noise anomaly branch extracts not ordinary random noise, but residual information left in the high-frequency components after image manipulation. For example, after manipulation operations such as copying and pasting, splicing, local repair, resampling, or interpolation, the image may appear visually natural, but it usually alters local texture, noise distribution, and high-frequency response. Since CT medical images inherently contain acquisition and reconstruction noise, if a region in the image has been manually edited, its noise pattern may be inconsistent with the surrounding real tissue. The multi-scale noise feature extraction module in the noise anomaly branch constructed in this application can amplify and extract high-frequency noise residuals that cannot be directly represented visually, thereby using a classification model to determine whether there are editing traces in a local area.

[0046] In the noise anomaly branch, a multi-scale noise feature extraction module and a multi-scale fusion module can be used to extract local statistical anomaly information (local statistical anomaly information is the noise feature map containing noise anomaly features). Even if the tampered area is visually similar to the surrounding area, it will still leave inconsistencies in high-frequency residuals, local texture distribution, and noise patterns. The multi-scale noise feature extraction module can model local noise patterns and residual information, decompose the image into low-frequency signal body and high-frequency sub-band in the multi-scale transform domain, and extract features that reflect local residuals, texture perturbations, and statistical inconsistencies in noise from the high-frequency sub-band. Therefore, the noise anomaly branch constructed in this invention can more effectively identify highly concealed tampering behaviors such as lesion insertion, local repair, lesion erasure, and content splicing. Compared with the method of performing uniform detection directly on the original grayscale image, this invention is more targeted in terms of statistical anomaly modeling.

[0047] The multi-scale noise feature extraction module can perform wavelet transform to further decompose and highlight noise anomalies in the CT medical image to be detected from multiple frequency perspectives. That is, the multi-scale nature of wavelet transform is mainly reflected in the frequency level, which can decompose the CT medical image into different frequency bands, making it convenient to observe high-frequency residuals, texture perturbations, and noise statistical changes. Wavelet transform can also be replaced by discrete cosine transform (DCT), Fourier transform (FFT), local time-frequency analysis (Gabor), Laplace pyramid, or learnable high-pass filtering. This application does not specifically limit the way the multi-scale noise feature extraction module extracts multi-scale noise feature maps. It is sufficient to utilize the prior knowledge that "noise is biased towards high frequencies" to separate noise from the signal body in the transform domain (multi-scale domain) and extract features from high-frequency components.

[0048] In some embodiments of the present invention, before inputting the CT medical image to be detected into the multi-scale noise feature extraction module, the tampering detection method proposed in this application further includes: performing noise residual enhancement processing on the CT medical image to be detected. Noise residual enhancement processing is used to enhance information on local high-frequency residuals, texture anomalies, and inconsistent noise distribution in the image. The CT medical image to be detected can be input into a Spatial Rich Model (SRM) filter for noise residual enhancement processing. The SRM can convolve the image through a set of high-pass filters (differential filters), thereby suppressing the main content of the image and amplifying and exposing noise / distortion components. The present invention is not limited thereto.

[0049] Taking the extraction and enhancement of local statistical anomalies in an image through a spatial rich model filter and wavelet transform as an example, the CT medical image input to the noise anomaly branch mainly undergoes the following processing steps: the input image is enhanced with a high-pass residual using a spatial rich model filter (i.e., enhancing high-frequency noise residuals) to highlight the inconsistent information of local noise introduced by editing, interpolation, splicing, or repair in the image; further, the image is decomposed into multiple frequencies using wavelet transform to enhance the expression of noise differences in different frequency bands.

[0050] The reason for introducing multi-scale features in step S110 of this application is that tampered areas in CT medical images may manifest as minor lesion-level modifications or as large-scale tissue splicing or content replacement. If the feature model cannot simultaneously integrate shallow detail information and deep semantic information, it is difficult to balance fine-grained tamper location and large-area anomaly recognition, resulting in coarse location boundaries or incomplete detection. Based on this, this application introduces multi-scale feature maps to improve the adaptability to tampered areas at different scales.

[0051] Step S120: After extracting multi-scale edge feature maps containing edge structure anomalies such as edge discontinuities, contour distortion, and local geometric anomalies caused by tampering, and multi-scale noise feature maps containing noise statistical anomalies such as local noise statistical inconsistencies, texture residual anomalies, and high-frequency feature perturbations caused by tampering, the edge feature maps and noise feature maps can be input into the multi-scale fusion module in the edge anomaly branch and the noise anomaly branch respectively, and fused to obtain the edge fusion feature map and the noise fusion feature map. That is, the multi-scale fusion module of edge feature maps of multiple scales outputs the edge fusion feature map; the multi-scale fusion module of noise feature maps of multiple scales outputs the noise fusion feature map.

[0052] In some embodiments of the present invention, such as Figure 3 As shown, this application designs a multi-scale fusion module that enhances the representation of abnormal features through contextual fusion. The main operations are as follows: The multi-scale feature map of the input multi-scale fusion module is downsampled through a bottom-up feature encoding path to obtain a first-scale sampled feature map; the first-scale sampled feature map is upsampled through a top-down feature encoding path to obtain a second-scale sampled feature map; the first and second sampled feature maps are aligned and weighted fusion is performed to obtain the fused feature map corresponding to the multi-scale feature map of the input multi-scale fusion module, thus integrating multi-scale contextual information. The bottom-up approach can simultaneously preserve shallow local texture information and deep abstract semantic information, while the top-down approach, through semantic propagation, upsampling, and shallow feature fusion, re-injects deep semantic information into high-resolution features, thereby enhancing the model's ability to perceive local detail anomalies.

[0053] More specifically, a bottom-up feature encoding path is used to downsample edge feature maps at multiple scales, and a top-down feature encoding path is used to upsample the downsampled feature maps. Scaling is applied to ensure that the downsampled and upsampled feature maps have the same spatial size and channel dimension (the spatial size and channel dimension can be designed independently; this invention does not limit the specific size). Weighted fusion (e.g., weighted summation of feature values ​​at the same pixel location) is then applied to the scaled feature maps to obtain an edge fusion feature map. Similarly, a bottom-up feature encoding path is used to downsample noise feature maps at multiple scales, and a top-down feature encoding path is used to upsample the downsampled feature maps. Scaling is applied to ensure that the downsampled and upsampled feature maps have the same spatial size and channel dimension (again, this invention does not limit the size of the spatial size and channel dimension). Weighted fusion (e.g., weighted summation) is then applied to the scaled feature maps to obtain a noise fusion feature map. For the edge anomaly branch, multi-scale feature fusion can enhance the edge anomaly response at different scales; for the noise anomaly branch, multi-scale feature fusion can enhance the statistical differences in noise at different scales.

[0054] The multi-scale fusion module's multi-scale aspect primarily lies in feature fusion. It aligns and fuses features of different resolutions and semantic levels, enabling the classification model to focus on both small-scale tampering and contextual anomalies in larger regions. Specifically, the multi-scale fusion module integrates edge or noise features at different scales to obtain a more complete representation of edge and noise features. It can perform multi-scale enhancement and spatial correction on the multi-scale intermediate features extracted from edge and noise anomaly branches, making it a crucial component in improving the detection accuracy of the tampering detection method in this application.

[0055] The multi-scale fusion module can employ models or networks capable of achieving feature alignment, semantic propagation, and multi-scale feature fusion at different levels, such as Feature Pyramid Network (FPN), Path Aggregation Network (PANet), Bidirectional Feature Pyramid Network (BiFPN), or Spatial Pyramid Pooling with Holes (ASPP), etc., but this invention is not limited to these.

[0056] As an example, during upsampling and downsampling, layer-by-layer convolution can be introduced to complete local feature extraction, hierarchical abstraction, and multi-scale representation construction. That is, multi-scale edge feature maps are input into a convolutional feature extraction network, as are multi-scale noise feature maps (the feature extraction model can also include a convolutional feature extraction network). Then, the edge and noise feature maps are upsampled or downsampled once. This is because feature maps at different levels may have different sizes and number of channels, and they cannot be directly added or multiplied. Therefore, it is necessary to first convert the feature maps into a size and dimension that can be fused through upsampling, downsampling, and convolutional mapping.

[0057] In addition, before inputting the edge feature map and noise feature map into the multi-scale fusion module respectively, feature preprocessing can be performed on the multi-scale edge feature map and multi-scale noise feature map obtained in step S110, such as normalization using the sigmoid function.

[0058] It should be noted that edge anomaly branches and noise anomaly branches can be integrated using the same multi-scale fusion module (in which case the tamper detection framework includes only one multi-scale fusion module), or they can be set up with separate multi-scale fusion modules (in which case the tamper detection framework includes two multi-scale fusion modules).

[0059] Step S130: Use the dual-feature fusion module to perform weighted fusion of the edge fusion feature map and the noise fusion feature map obtained in step S120 to obtain the fusion feature map corresponding to the CT medical image to be detected.

[0060] like Figure 1 As shown, the CT medical images to be detected are input into the edge anomaly branch and the noise anomaly branch, which undertake different tasks in the feature extraction part: the edge anomaly branch focuses more on boundary and contour changes and is suitable for detecting structural discontinuities and local morphological abnormalities, while the noise anomaly branch focuses more on local residuals and high-frequency disturbances and is suitable for detecting statistical traces with stronger concealment. Then, the extracted features can be jointly analyzed by the dual-branch feature fusion module, so that the features extracted by the two branches can work together to form the final tamper detection result.

[0061] The dual-branch feature fusion method designed in this application does not simply stitch together edge features and noise features. Instead, it introduces appropriate weights to enhance the response of regions that simultaneously exhibit edge and noise anomalies, thereby more accurately distinguishing between genuine tampered regions and ordinary structurally complex regions. This fusion method is particularly important for CT medical images because many normal tissue boundaries themselves have strong structural variations. Relying solely on edge information can easily lead to false positives, while relying solely on noise information may miss tampered regions with obvious boundaries but weak noise variations. Through dual-branch joint modeling, a crucial foundation can be laid for the high-precision discrimination capability of the tamper detection framework, thus more stably identifying genuine tampering traces. Specifically, this application integrates the fusion results at different scales by weighted summation of multi-scale features to form the final fused feature, thereby enhancing the discriminative power of the final feature representation.

[0062] As an example, this application can achieve bi-branch feature fusion based on a joint attention mechanism. The specific process is as follows: Align the edge fusion feature map and the noise fusion feature map to ensure they are of the same size (all sizes mentioned in this application refer to spatial dimensions and channel dimensions). Then, use the attention mechanism to generate a first attention map corresponding to the edge fusion feature map and a second attention map corresponding to the noise fusion feature map, respectively. Combine the first and second attention maps element-wise to form a joint attention map, and use this joint attention map as the joint weight. Finally, perform weighted fusion processing on the edge fusion feature map and the noise fusion feature map using the joint weight to obtain the fusion feature map corresponding to the CT medical image to be detected. The joint attention map has a stronger response to regions where both edge and noise anomalies are significant, enabling the fusion feature map to emphasize locations where both edge and noise anomalies are simultaneously significant.

[0063] After extracting the corresponding edge anomaly features and noise anomaly features from the edge anomaly branch and noise anomaly branch respectively, the two can be jointly modeled by the dual-branch feature fusion module to obtain fused features, thereby enabling the joint determination of image authenticity and identification of tampered regions. In addition to the joint attention mechanism, the dual-branch feature fusion module can also use features concatenated and then convolved, gated fusion, channel attention, spatial attention, Transformer cross attention, or bilinear fusion to fuse dual-branch features.

[0064] Not all areas in CT medical images have equal diagnostic significance. Without spatial prior constraints, detection models are prone to false alarms at normal organ boundaries, areas with significant density changes, or areas with fluctuating imaging noise. Therefore, this application also introduces an anatomical prior guidance mechanism in the tampering detection process to achieve high-precision identification and localization of CT medical image tampering behavior. The anatomical prior guidance mechanism plays two main roles: first, it provides medical structural constraints for spatial attention, making the model focus more on clinically relevant areas; second, it suppresses false alarms from anatomically unreasonable locations, improving the medical interpretability of the detection results.

[0065] In some embodiments of the present invention, after step S120, the tampering detection method proposed in this application further includes step S150: taking the prior knowledge graph and the specific feature map as input to the attention mechanism, outputting attention weights, and adjusting the specific feature map element by element using the attention weights; wherein, the specific feature map may be an edge fusion feature map, a noise fusion feature map, or a fusion feature map corresponding to the CT medical image to be detected; the prior knowledge graph is generated based on the untampered CT medical image and is used to represent the position of organ regions, tissue structure regions, or lesion boundary regions in the untampered CT medical image, and the untampered CT medical image and the CT medical image to be detected belong to the same anatomical region. The prior knowledge graph and the specific feature map input to the attention mechanism have the same spatial size, therefore, before inputting the prior knowledge graph into the attention mechanism, it is also necessary to adjust the prior knowledge graph to the same size as the specific feature map.

[0066] Taking a specific feature map as the noise fusion feature map as an example, untampered CT medical images inherently contain normal organ boundaries and grayscale transition regions. Furthermore, noise branches focus on abstract residuals and high-frequency statistical information, which are more susceptible to interference from irrelevant noise (such as CT imaging noise, reconstruction artifacts, or local texture fluctuations), making noise anomalous branches more easily affected. Without spatial constraints, noise anomalous branches might mistake some noise fluctuations or normal structural changes unrelated to tampering as suspicious tampering areas. For example... Figure 3 As shown, the specific process of step S150 is as follows: the prior knowledge graph and the noise fusion feature map are jointly input into the attention mechanism. Since the prior knowledge graph can spatially constrain the spatial attention distribution corresponding to the noise fusion features, it can generate an attention weight map by combining medically significant spatial regions such as organ regions, lesion boundaries, and tissue structure locations, making tamper detection more consistent with the structural rules of the medical image itself. The optimized feature map can be obtained by adjusting the noise fusion feature map element by element using the generated attention weight map. This application introduces the anatomical prior map (i.e., the prior knowledge graph) into the attention correction process in the noise anomaly branch, enabling the classification model to prioritize organ regions, lesion boundaries, and other medically significant regions, and suppress abnormal responses from anatomically unreasonable locations.

[0067] The location for introducing the anatomical prior map does not need to be strictly limited to the noisy anomalous branch, as long as it is used to constrain the model to focus on medically relevant regions and suppress anatomically unreasonable anomalous responses. It is worth mentioning that if the specific feature map is an edge fusion feature map or a noise fusion feature map, then step S150 needs to be performed before step S130; if the specific feature map is a fusion feature map corresponding to the CT medical image to be detected, then step S150 needs to be performed before step S140.

[0068] As an example, the tamper-proof CT medical images used to generate the prior knowledge graph can come from different users, and the users' physical states can be healthy or ill. This invention is not limited to this; the prior knowledge graph only needs to represent clinically relevant regions such as organ regions, lesion boundaries, or medical structure boundaries. Furthermore, the prior knowledge graph can be a CT image or a spatial feature of a CT image. This invention is not limited to this; for example, the prior knowledge graph can be generated by a pre-trained medical image segmentation model.

[0069] The purpose of introducing a prior knowledge graph into the tamper detection process in this application is to achieve feature-guided adjustment through medical structural prior constraints, so that the detection process can conform to the structural rules of medical images. That is, the anatomical prior guidance mechanism can realize the medical structural correction of features, making the tamper detection method proposed in this invention adaptable to medical scenarios. It can better adapt to the actual scenario of "complex structure but subtle tampering" in CT medical images, thereby making the tamper detection results more consistent with the structural rules of medical images and reducing false alarms.

[0070] Step S140: Input the fused feature map obtained in step S130 into the pre-trained tampering detection classification model, and output the tampering detection result corresponding to the CT medical image to be detected. The tampering detection result corresponding to the CT medical image to be detected may include a judgment result indicating whether the CT medical image to be detected has been tampered with (belonging to image-level true / false classification results), and may also include a tampering probability map representing the spatial distribution of the tampered area (i.e., pixel-level or region-level tampering location results). Compared with detection methods that only output true / false labels, this output method enables the present invention to meet the practical application needs of medical image integrity analysis, while simultaneously satisfying both automatic judgment and manual verification application needs, thus having higher practical value.

[0071] As an example, when the tampering detection result includes a tampering probability map, the process of determining whether an image is tampered can be based on the tampering probability map as follows: the tampering probability map is segmented according to a pixel-level probability threshold to obtain a mask of the suspected tampered region; the obtained mask is statistically analyzed as a whole (such as maximum value, average value, Top-k average value, or connected region area); if the overall abnormality exceeds the preset image judgment threshold, the CT medical image to be detected is judged to be suspected of being tampered, and the authenticity judgment result is obtained.

[0072] If the size of the fusion feature map corresponding to the CT medical image to be detected is smaller than that of the input image, it is necessary to upsample or deconvolve it to restore it to the same size as the original image. Figure 1 The probability map is a graph that represents the probability of tampering with a specific pixel or region in the CT medical image being inspected. A higher probability value for a pixel or region in the probability map indicates that the location is more likely to have been tampered with.

[0073] The key to this invention's tamper detection lies in separating and modeling structural and statistical anomalies in CT medical images, and combining these with prior constraints on medical structures to achieve a two-branch collaborative tamper detection and localization. The edge anomaly branch extracts the first type of structural anomalies, the noise anomaly branch extracts the second type of statistical anomalies, and the anatomical structure prior constrains the classification model to focus on more reasonable spatial locations. The two-branch feature fusion module combines the two types of features to output image-level authenticity judgments and region-level tamper location information.

[0074] This application also proposes a training method for a tampering detection model to generate a pre-trained tampering detection model. The tampering detection model may include a feature extraction model and a tampering detection classification model. Therefore, during the training phase, this application designs that the feature extraction model and the tampering detection classification model can be trained jointly or separately.

[0075] In some embodiments of the present invention, such as Figure 4 As shown, the training method for the tamper detection model includes steps S210 to S240, as detailed below: Step 210: Obtain a training sample set containing multiple training samples and the corresponding labels for each training sample. All training samples are CT medical images. Furthermore, this application does not limit the anatomical region to which the training samples belong; that is, all training samples in the obtained training sample set may belong to only the same anatomical region, or they may be CT medical images from multiple different anatomical regions.

[0076] The acquired training samples may include unaltered samples and altered samples obtained by adding edge perturbations and noise perturbations to the unaltered samples. Therefore, the acquired training sample set is regarded as a training sample pair formed by unaltered samples and their corresponding altered samples.

[0077] Step 220: Input the untampered sample and the tampered sample into the tamper detection model to be trained, and output the tamper detection results corresponding to the untampered sample and the tamper detection results corresponding to the tampered sample.

[0078] More specifically, by inputting an untampered sample into the tamper detection model to be trained, the model can obtain multi-scale edge feature maps, multi-scale noise feature maps, edge fusion feature maps, noise fusion feature maps, and fusion feature maps corresponding to the untampered sample, output by the multi-scale edge feature extraction module, multi-scale noise feature extraction module, multi-scale fusion module, and dual-feature fusion module. It can also obtain the tamper detection result corresponding to the untampered sample output by the tamper detection classification model. Conversely, by inputting a tampered sample into the tamper detection model to be trained, the model can obtain multi-scale edge feature maps, multi-scale noise feature maps, edge fusion feature maps, noise fusion feature maps, and fusion feature maps corresponding to the tampered sample, output by the multi-scale edge feature extraction module, multi-scale noise feature extraction module, multi-scale fusion module, and dual-feature fusion module. It can also obtain the tamper detection result corresponding to the tampered sample output by the tamper detection classification model.

[0079] Step 230: Based on the feature maps output by the multi-scale edge feature extraction module, multi-scale noise feature extraction module, multi-scale fusion module, and dual feature fusion module for unaltered and tampered samples, the training loss of the feature extraction model is formed; based on the tampering detection results and corresponding labels of the unaltered and tampered samples, the training loss of the tampering detection classification model is formed.

[0080] More specifically, throughout the training process of the tamper detection model, the training loss of the feature extraction model is based on edge loss. Noise loss and fusion loss This is a collective structure used to constrain the joint learning of various modules in the feature extraction model (e.g., iteratively updating the parameters of each module in the feature extraction model through gradient descent based on the training loss of the feature extraction model). The training loss of the feature extraction model can be expressed as: ; in, , and These are the corresponding weights, used to control the importance of different loss terms. Edge loss is determined based on the differences between the multi-scale edge feature maps corresponding to unaltered and tampered samples, and the differences between the edge fusion feature maps corresponding to unaltered and tampered samples. Noise loss is determined based on the differences between the multi-scale noise feature maps corresponding to unaltered and tampered samples, and the differences between the noise fusion feature maps corresponding to unaltered and tampered samples. Fusion loss is determined based on the differences between the fusion feature maps corresponding to unaltered and tampered samples. This application does not specifically limit the method for determining the differences in feature maps; for example, L2 norm calculation, cross-entropy, or Dice loss can be used.

[0081] Edge loss is used to constrain the edge anomaly branch to learn differences in edge breaks, contour anomalies, and geometric deformations caused by tampering. Noise loss is used to constrain the noise anomaly branch to learn high-frequency noise residuals and local statistical inconsistencies. Fusion loss is used to constrain the result of fusing the edge anomaly branch and the noise anomaly branch to be closer to the actual tampered area.

[0082] Furthermore, the differences between the tamper detection results and the corresponding labels of untampered samples can be determined, as well as the differences between the tamper detection results and the corresponding labels of tampered samples. The training loss of the tamper detection classification model is then formed based on the differences between untampered and tampered samples. (Training loss of the tamper detection classification model) The calculation formula can be: ; in, and These are the corresponding weights. The difference between the tampering detection results for untampered samples and the labels corresponding to untampered samples. This refers to the difference between the tampering detection result corresponding to the tampered sample and the label corresponding to the tampered sample. It is mainly used to constrain the tampering detection classification model, enabling it to distinguish between tampered and untampered images. Additionally, adversarial loss and other methods can be used to determine the difference between the tampering detection result and the label; however, this invention is not limited to these methods.

[0083] Step 240: Iteratively train the tamper detection model using the training loss of the feature extraction model and the training loss of the tamper detection classification model to obtain the trained tamper detection model.

[0084] In some embodiments of the present invention, the tampered sample and its corresponding label can be obtained by inputting an untampered sample into a pre-trained generative model. In this case, the tampered sample can be utilized. This constrains the tamper detection classification model and the generation model, making the generated tampered samples closer to the real tampering situation while distinguishing between untampered and tampered images.

[0085] If the pre-trained generative model is not used to generate tampered samples during the training phase, the feature extraction model and the tamper detection classification model can be selectively trained jointly or separately. If the pre-trained generative model is used to generate tampered samples during the training phase, the generative model, feature extraction model, and tamper detection classification model can be trained jointly, or the training loss of the tamper detection classification model can be used before training the feature extraction model. The generative model and the tamper detection and classification model are trained together.

[0086] As an example, such as Figure 5As shown, this application can employ a joint optimization approach during the training phase, combining a generative model, a feature extraction model, and a tampering detection classification model. This allows for the joint training of the generative model, edge anomaly branch, noise anomaly branch, dual-feature fusion module, and tampering detection classification model. The specific training process is as follows: the input training samples are normalized and processed using Gaussian filtering (e.g., using...). Figure 5 The model uses fully connected layers and ReLU activation functions to generate tampered samples. The tampered and untampered samples are then input into edge anomaly and noise anomaly branches, respectively, to extract corresponding edge and noise features. A dual-feature fusion module fuses these features, and a classification model is then used for joint discrimination. and Iteratively update the parameters of the generation model, feature extraction model, and tamper detection classification model.

[0087] The generative model can enhance the complexity of tampered samples based on generative perturbations, obtaining tampered CT medical image samples with edge perturbations and noise perturbations, thereby enhancing the tampering detection model's learning ability in complex tampering scenarios. This application does not specifically limit the type of generative model, such as building it based on ResNet, Generative Adversarial Network (GAN), Variational Autoencoder (VAE), or diffusion model. Furthermore, the settings for the number of training epochs, batch size, optimizer, and weights during the training process can be designed independently; this invention does not impose specific limitations.

[0088] The introduction of generative models enables the tamper detection framework proposed in this application to access more diverse and realistic tamper samples during the training phase, thereby improving its generalization ability.

[0089] Compared with existing technologies, the beneficial effects of the method of this invention stem from the complete technical chain formed by edge anomaly branches, noise anomaly branches, medical anatomical prior constraints, a multi-scale fusion module that establishes an effective connection between shallow local details and deep abstract semantics, and dual-branch collaborative fusion. The synergistic effect of these modules enables this invention to achieve higher detection accuracy, stronger fine-grained recognition capability, wider adaptability to medical scenarios, and better interpretability of results in CT medical image tampering detection tasks. Specifically, the tampering detection method proposed in this invention has the following significant advantages: ① Because tampering can disrupt the structural continuity of organ boundaries, lesion outlines, and tissue junctions, and may also alter the texture residuals, high-frequency responses, and noise statistical distributions of local areas, it is often difficult to simultaneously address both structural and statistical anomalies if all abnormal features are learned uniformly through a single feature extraction path. This invention utilizes an edge anomaly branch to focus on extracting structural anomalies and a noise anomaly branch to focus on extracting statistical anomalies. This allows different types of tampering features to be strengthened separately and ultimately used synergistically, thereby improving the coverage of complex tampering scenarios.

[0090] ② This invention uses a dual-branch feature fusion module to jointly model edge anomaly information and noise anomaly information, which can improve the stability of the final discrimination result. A truly tampered region usually exhibits anomalies in both edge and noise dimensions simultaneously. Therefore, this invention strengthens the response of regions with anomalies in both edge and noise dimensions through a weighted fusion mechanism, enabling more accurate differentiation between real tampered regions and normal complex structure regions, reducing the risk of misjudgment.

[0091] ③ This invention can generate a prior knowledge graph based on untampered CT medical images and use it for feature attention correction, making tamper detection more focused on organ regions, lesion boundaries, and other clinically relevant areas, while suppressing positional responses that are clearly inconsistent with anatomical structures. The introduction of the anatomical prior mechanism not only helps improve the accuracy of tamper detection but also enhances the medical rationality and interpretability of the results.

[0092] ④ This invention not only outputs image-level authenticity determination results but also tampering probability maps, better meeting practical application needs. Simply providing a "tampered" classification result is insufficient for subsequent manual review and security auditing. However, simultaneously providing detection results for the location and extent of tampering makes it more suitable for practical use in medical image security scenarios. The tampering detection results output by this invention through dual-branch joint modeling can determine whether CT medical images have been tampered with and also indicate the spatial distribution of the tampered area, thereby improving the practical value of the method.

[0093] Corresponding to the above method, the present invention also provides a tamper detection system for CT medical images. The system includes a computer device, which includes a processor and a memory. The memory stores computer programs / instructions, and the processor is used to execute the computer programs / instructions stored in the memory. When the computer programs / instructions are executed by the processor, the system performs the steps of the method described above.

[0094] The tamper detection system for CT medical images proposed in this application may include a two-branch feature extraction module (mainly containing a feature extraction model) and a classification module (mainly containing a tamper detection classification model). Each module can be configured according to... Figure 1The process connections and synergistic effects shown work together to complete the tampering detection task of CT medical images, and obtain the tampering judgment result and the tampering area location result.

[0095] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0096] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0097] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting tampering in CT medical images, characterized in that, This method is based on a pre-trained feature extraction model, which includes a multi-scale edge feature extraction module, a multi-scale noise feature extraction module, a multi-scale fusion module, and a dual-feature fusion module. The method includes the following steps: The CT medical image to be detected is input into the multi-scale edge feature extraction module, which outputs edge feature maps at multiple scales. The CT medical image to be detected is also input into the multi-scale noise feature extraction module, which outputs noise feature maps at multiple scales. The edge feature map and the noise feature map are respectively input into the multi-scale fusion module to obtain an edge fusion feature map and a noise fusion feature map; The edge fusion feature map and the noise fusion feature map are weighted and fused using a dual feature fusion module to obtain the fusion feature map corresponding to the CT medical image to be detected. The fused feature map is input into a pre-trained tamper detection classification model, and the tamper detection result corresponding to the CT medical image to be detected is output.

2. The method according to claim 1, characterized in that, After feature fusion using a multi-scale fusion module, the method further includes: A prior knowledge graph is generated based on an unaltered CT medical image; wherein the unaltered CT medical image and the CT medical image to be detected belong to the same anatomical region; the prior knowledge graph is used to represent the location of organ regions, tissue structure regions, or lesion boundary regions in the unaltered CT medical image; The prior knowledge graph and the specific feature map are used as inputs to the attention mechanism, and the output is the attention weight. The attention weight is then used to adjust the specific feature map. The specific feature map is an edge fusion feature map, a noise fusion feature map, or a fusion feature map corresponding to the CT medical image to be detected.

3. The method according to claim 1, characterized in that, The multi-scale fusion module is used to perform the following operations: The multi-scale feature map of the input multi-scale fusion module is downsampled by a bottom-up feature encoding path to obtain the first sampled feature map of the multi-scale; The first multi-scale sampled feature map is upsampled by a top-down feature encoding path to obtain the second multi-scale sampled feature map. Align the first and second sampled feature maps, and obtain the fused feature map corresponding to the multi-scale feature map input to the multi-scale fusion module through weighted fusion.

4. The method according to claim 1, characterized in that, The fusion feature map corresponding to the CT medical image to be detected is obtained in the following way: Align the edge fusion feature map and the noise fusion feature map, and use an attention mechanism to generate corresponding first attention map and second attention map respectively; By combining the first and second attention maps through element-wise multiplication, and using the resulting joint attention map as the joint weight; The edge fusion feature map and the noise fusion feature map are weighted and fused using the joint weights to obtain the fusion feature map corresponding to the CT medical image to be detected.

5. The method according to claim 1, characterized in that, The CT medical image to be detected is obtained by performing grayscale normalization and Gaussian filtering on the acquired original CT medical image to be detected. The tampering detection results corresponding to the CT medical image to be detected include a judgment result indicating whether tampering has occurred and a tampering probability graph; The multi-scale edge feature extraction module is built based on the U-Net model, while the multi-scale noise feature extraction module is used to extract noise features in the multi-scale transform domain.

6. The method according to claim 1, characterized in that, Before inputting the CT medical image to be detected into the multi-scale edge feature extraction module, the method further includes: performing edge enhancement processing on the CT medical image to be detected through gradient response.

7. The method according to claim 1, characterized in that, Before inputting the CT medical image to be detected into the multi-scale noise feature extraction module, the method further includes: performing noise residual enhancement processing on the CT medical image to be detected.

8. A training method for a tamper detection model of CT medical images, characterized in that, The tampering detection model includes a feature extraction model and a tampering detection classification model. The trained tampering detection model is used to execute the tampering detection method as described in any one of claims 1-7; the method includes the following steps: Multiple training samples and their corresponding labels are obtained; wherein, the training samples include unaltered samples and altered samples obtained by adding edge perturbation and noise perturbation to the unaltered samples, and both the unaltered samples and the altered samples in the training samples are CT medical images; Input the unaltered sample and the tampered sample into the tamper detection model to be trained, and output the tamper detection results corresponding to the unaltered sample and the tamper detection results corresponding to the tampered sample. The training loss of the feature extraction model is formed by the feature maps output by the multi-scale edge feature extraction module, multi-scale noise feature extraction module, multi-scale fusion module, and dual feature fusion module for untampered samples and tampered samples, respectively. The training loss of the tamper detection classification model is formed by the tamper detection results and corresponding labels of the untampered samples and tampered samples. The tamper detection model is then iteratively trained using the training losses of the feature extraction model and the tamper detection classification model to obtain the trained tamper detection model.

9. The method according to claim 8, characterized in that, The tampered samples and their corresponding labels are obtained by inputting untampered samples into a pre-trained generative model; The generative model and the tamper detection classification model are trained together.

10. A tamper detection system for CT medical images, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 7.

Images