A deep learning analysis method based on MA-net plain CT and enhanced CT isotope contrast

Abstract

The application relates to the technical field of medical image analysis, in particular to a deep learning analysis method based on MA-net plain CT and enhanced CT isotope contrast, which comprises the following steps: constructing a structure anchor representation for representing a structure in a plain CT image that can be enhanced and responded to, the structure anchor representation being used to limit an enhanced attention source area; inputting the plain CT image, the enhanced CT image and the structure anchor representation into an MA-net network comprising a double-branch structure, introducing an attention source tracing constraint mechanism, and generating an enhanced feature representation subjected to the source tracing constraint; and performing lesion area discrimination to output a lesion identification result. According to the application, a source-tracing reliable enhanced feature representation is constructed through structure-guided feature fusion, the stability and reliability of lesion identification are enhanced, and the application is particularly suitable for a low-contrast or atypical lesion identification task.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to a deep learning analysis method based on the isotopic comparison of MA-net plain CT and enhanced CT scans. Background Technology

[0002] Computer-aided diagnostic technology has been increasingly widely applied in medical image analysis, particularly in the automatic detection and segmentation of lesion areas in CT images using deep learning models. In clinical practice, enhanced CT images, due to their higher soft tissue contrast, are widely used for the identification of tumors, hemorrhages, and inflammatory lesions; while plain CT images provide stable, uncontrast-free information on the original anatomical structures. Given the differences in their presentation and imaging mechanisms, how to synergistically utilize these two types of image information to improve the model's discrimination accuracy and interpretability has become an important research direction.

[0003] Most existing deep learning-based medical image analysis methods typically rely on a single type of image as input, or simply stitch together or fuse features from multimodal images before feeding them into a neural network for training. However, while density changes in lesion areas in enhanced images help highlight them, they can also introduce high-response artifacts due to factors such as contrast agent leakage, local enhancement delay, and vascular signal diffusion. This can cause the model to form false attention at locations without structural basis, affecting the accuracy of discrimination. Traditional fusion methods often fail to establish an effective mapping relationship between enhancement response and anatomical structure; the model may rely solely on differences in image surface intensity for judgment, lacking structural consistency supervision. Although plain and enhanced images are usually acquired consecutively from the same patient and the same site in clinical practice, most models fail to effectively align image information and fail to uncover the potential clinical judgment logic of whether the enhancement signal has a structural basis. Summary of the Invention

[0004] This invention provides a deep learning analysis method based on the isotopic comparison of MA-net plain CT and enhanced CT scans.

[0005] A deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison includes the following steps: S1, based on the plain CT image and enhanced CT image of the same patient, perform co-positional region alignment on the plain CT image and enhanced CT image, and extract stable anatomical structure response in the plain CT image to construct a structural anchoring representation for characterizing the structure that can be attached to the enhancement response in the plain CT image. The structural anchoring representation is used to define the source region of enhancement attention. S2, the plain CT image, enhanced CT image, and structural anchoring representation are input into the MA-net network including a dual-branch structure, wherein: Plain CT images are input into the plain CT branch to extract latent structural features; Enhanced CT images are input into the enhanced CT branch to extract contrast enhancement features; An attention source constraint mechanism is introduced into the MA-net network, which requires that the enhanced attention response generated in the enhanced CT branch must be mapped back to the corresponding structural anchoring region in the plain CT branch. Only when the enhanced attention response can establish an effective correspondence in the structural anchoring representation of the plain CT branch is it allowed to participate in feature modeling and generate enhanced feature representations subject to source constraints. S3, based on the enhanced feature representation subject to source tracing constraints, perform lesion region discrimination, output lesion identification results and corresponding attention source tracing consistency map, which are used to characterize the correspondence reliability between enhanced response and plain scan structure.

[0006] Optionally, the alignment of the co-position region adopts an image registration algorithm based on deformation field learning, which maps the spatial coordinates of the enhanced CT image to the spatial coordinates of the plain CT image, so as to achieve pixel-level anatomical position alignment between the two sets of images.

[0007] Optionally, the extraction of the stable anatomical structure response includes using a pre-trained structural segmentation network with plain CT images as input to extract the contour or region probability map of stable anatomical structures, including major blood vessels, organ boundaries, and bones, from the plain CT images; binarizing or preserving the probability values ​​of the contour or region probability map of the stable anatomical structure to generate a matrix or feature map of the same size as the plain CT image, which serves as the structural anchoring representation.

[0008] Optionally, the structural anchoring representation serves as a spatial mask or attention guide map, constraining any attention weight map derived from enhanced CT branches to be activated or have non-zero values ​​only within the anatomical region identified by the structural anchoring representation.

[0009] Optionally, both the plain CT branch and the enhanced CT branch include a feature encoder composed of multiple convolutional layers, normalization layers, and nonlinear activation layers connected in sequence; wherein: The feature encoder of the plain CT branch takes the plain CT image as input, extracts and outputs multi-scale structural latent feature maps; The feature encoder of the enhanced CT branch takes the enhanced CT image as input and extracts and outputs multi-scale initial enhanced feature maps.

[0010] Optionally, the attention source constraint mechanism introduced in the MA-net network is implemented by the attention source constraint module. The input of the attention source constraint module includes: the initial enhancement feature map from the enhanced CT branch, the structural latent feature map from the plain CT branch, and the structural anchoring representation.

[0011] Optionally, the attention source constraint module first calculates the spatial attention weight map based on the initial enhanced feature map, and then multiplies the spatial attention weight map element-wise with the structural anchoring representation to obtain the attention weight map after source constraint, which forces the enhanced attention response generated in the enhanced CT branch to be restricted to the anatomical structure region defined by the structural anchoring representation, that is, it must be mapped back to the corresponding structural anchoring region in the plain CT branch.

[0012] Optionally, the attention weight map after source tracing constraint is then used to weight and modulate the initial enhanced feature map to generate intermediate enhanced features. The intermediate enhanced features are then fused with the latent structural feature map of the same scale from the plain CT branch to generate the enhanced feature representation subject to source tracing constraint. The participation feature modeling is achieved through the element-wise multiplication operation. The response intensity can be retained and transmitted if and only if the response position in the spatial attention weight map has a non-zero value or a high probability value in the structural anchoring representation.

[0013] Optionally, the lesion region discrimination is implemented through a lesion discrimination decoder. The lesion discrimination decoder takes the enhanced feature representation subject to source tracing constraints as input, processes it through at least one decoding layer including convolution and upsampling operations, and finally outputs a lesion probability map through an activation function, where the value of each pixel represents the probability that the corresponding position belongs to a lesion. The lesion probability map is binarized by applying a predetermined probability threshold to obtain a binarized lesion segmentation mask, and the lesion recognition result includes the lesion segmentation mask.

[0014] Optionally, the generation of the attention source consistency map includes: for each spatial location in the attention weight map after source constraint, calculating the product or correlation coefficient of its weight value and the corresponding position value in the structural anchoring representation, and generating a consistency map of the same size as the image with a value range between [0,1]; the value of each pixel in the consistency map represents the reliability of the correspondence between the contrast enhancement response observed by enhanced CT at that location and the anatomical basis defined by plain CT.

[0015] The beneficial effects of this invention are: This invention, for the first time, introduces a priori constraint mechanism for structurally traceable regions in enhanced CT analysis by constructing stable anatomical structural responses in plain CT images and generating structural anchoring representations. This structural anchoring representation not only originates from anatomical probability maps extracted by a pre-trained structural segmentation network but also ensures accurate spatial alignment with the enhanced image through registration. By using this representation as a spatial mask for subsequent attention modulation, the model significantly reduces false activation issues in regions with contrast agent interference, high-contrast artifacts, and abnormal tissue density, thereby improving the structural rationality and medical interpretability of the model's discrimination regions.

[0016] This invention proposes an attention source constraint module in deep networks. It performs point-by-point multiplication constraint on the spatial attention map automatically generated in the enhanced CT branch and the structural anchoring representation in the plain scan branch, ensuring that only those locations with anatomical support are allowed to form enhanced attention responses. This enhanced attention must have a structural source mechanism, breaking the conventional design of free drift in traditional attention mechanisms. It establishes a pixel-level mapping relationship between enhanced response and structural basis, and further constructs a source-based and reliable enhanced feature representation through structure-guided feature fusion, enhancing the stability and reliability of lesion identification. It is particularly suitable for the identification of low-contrast or atypical lesions.

[0017] This invention not only outputs lesion probability maps and segmentation masks, but also introduces an attention source consistency map as a visual representation of the discrimination criterion, quantitatively describing the spatial correspondence between enhanced response and structural anchoring. This map provides doctors with a reliability indicator of the model's attention source, improving the transparency and clinical trust in AI interpretation. Furthermore, this invention links lesion masks with enhanced CT images for analysis, supporting the output of various quantitative parameters such as lesion spatial location, area, and mean density, providing structured support information for subsequent clinical decision-making. It possesses strong practical application potential and industrial transformation value. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the analysis method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the execution of MA-net feature modeling based on attention source tracing constraints in an embodiment of the present invention; Figure 3 This is a schematic diagram of the lesion discrimination output based on source consistency in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] like Figures 1-3 As shown, a deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison includes the following steps: S1, Constructing a planar scan-enhanced isotopic anchoring representation: Based on the plain CT images and enhanced CT images of the same patient, the co-position regions of the plain CT images and enhanced CT images are aligned, and stable anatomical structure responses are extracted from the plain CT images. A structural anchoring representation is constructed to characterize the structural anchoring representation in the plain CT images that can be attached to the enhanced response. The structural anchoring representation is used to define the source region of enhanced attention.

[0022] S11, perform co-location alignment between the plain CT image and the enhanced CT image. Using an image registration algorithm, map the spatial coordinates of the enhanced CT image to the spatial coordinates of the plain CT image, achieving pixel-level anatomical alignment between the two sets of images. In this invention, to achieve precise pixel-level alignment between the plain CT image and the enhanced CT image, cross-modal image registration is required. Since the plain and enhanced CT images differ significantly in grayscale distribution and boundary representation, traditional rigid or affine registration methods based on pixel similarity have limited accuracy. Therefore, an image registration algorithm based on deformation field learning is used to generate a high-resolution, spatially continuous non-rigid deformation field to spatially align the enhanced CT image to the coordinate system of the plain CT image. Specifically: Input plain CT images Enhanced CT images ; The registration network uses a variant of UNet, namely the deformation field generation network, whose structural input is the stitching of two images or feature fusion: Network input: concat ; Network output: That is, the two-dimensional deformation vector field corresponding to each pixel position.

[0023] The generated deformation field Used to enhance images Perform spatial transformation to obtain the registered enhanced image: ;in, This represents image transformation operations based on deformation fields, typically implemented using bilinear interpolation.

[0024] To ensure accurate registration, the deformation field learning process incorporates a loss term: Registration similarity loss Used for measurement and Structural similarity is expressed as: ; Smoothing loss To prevent drastic and discontinuous changes in the deformation field and maintain the smoothness of spatial transformation, it is represented as: ; Total loss: ;in These are the weighting coefficients.

[0025] After training, input any pair of plain CT and enhanced CT images to obtain the registration deformation field. and the registered image Subsequent operations in this invention, such as structural anchoring representation extraction and attention source constraint, are all performed on the registration results. and Completed under the same coordinate system, thus ensuring that all regions of interest have a consistent anatomical semantic basis.

[0026] S12 extracts the stable anatomical structure response from plain CT images. In extracting anatomical regions with high structural stability and diagnostic reference value from plain CT images, such as large blood vessels, bone boundaries, and organ margins, density artifacts or inconsistent signal enhancement may occur in contrast-enhanced CT images at different time windows due to the influence of contrast agent distribution. Therefore, plain CT images are used as a source of stable structures, and a segmentation network outputs region probability maps or contour maps to construct the structural anchoring representation in the subsequent attention mechanism. A pre-trained structural segmentation network is used. The network uses plain CT images As input, output a contour map or region probability map of stable anatomical structures in the image. Stable anatomical structures include major blood vessels, organ boundaries, and skeletal regions; represented as: ;in, This represents a plain CT scan image. Represents a structurally segmented network. A probability map representing regions with stable anatomical structures.

[0027] The specific structure segmentation network is as follows: S121. Network Input and Output: Input: A single 2D or 3D plain CT image slice or volume data ; Output: Probability map of structural regions of the same size as the input image. The value of each pixel or voxel represents the probability that it belongs to a stable anatomical structure, ranging from 0 to 1. S122. Network Structure: A structurally segmented network consists of the following parts: a) Encoder part, feature extraction: contains 4~5 convolutional modules, each layer contains two convolutions (3×3 convolutional kernel, ReLU activation) and one max pooling operation. This part gradually extracts features from low-level edge texture to high-level semantic structure. The number of feature channels increases layer by layer, configured as: 64→128→256→512→1024.

[0028] b) Decoder section, upsampling and localization: Each layer uses transposed convolution or upsampling followed by concatenation of features from the same layer of the encoder, i.e., skip connections in U-Net. Each decoding module further refines the concatenated features through two convolution and activation operations, and the final output feature map is mapped to a 1-channel probability map via a 1×1 convolution. .

[0029] c) Add an attention gating module before each pair of skip connections. This module learns which regions in the feature map are related to structural anatomy, suppresses the interference of unstructured information on the segmentation results, and makes the output more focused on stable structural regions such as the spine, ribs, pericardial outline, and large blood vessels.

[0030] S123. Use publicly available CT structural anatomy segmentation datasets or self-built datasets, labeled with structures including bones, lung tissue, large blood vessels, liver, etc. Each structural category can be modeled independently or merged into a stable structural label. Use a hybrid form of cross-entropy loss and Dice coefficient to balance the foreground-background imbalance problem. To enhance the network's adaptability to CT scale, noise, and positional changes, data augmentation is added during training, including translation, rotation, affine deformation, and artifact simulation.

[0031] In actual operation, the structural segmentation network executes as follows: A plain CT image is input into a trained structural segmentation network; The network outputs a probability graph. Each pixel value represents the probability that the location belongs to a stable structure; You can choose to directly retain this probability map as a continuous structure anchoring representation, or you can perform threshold binarization to extract significant structural contour regions. The output results serve as the core input for structural anchoring representation in subsequent steps, limiting the area of ​​influence of the attention mechanism in the enhanced CT branch.

[0032] S13, Constructing a structural anchoring representation: The purpose is to transform the structural region probability map output by the structural segmentation network in the previous step, i.e., the stable anatomical structural response, into a clear, controllable structural anchoring representation that has a practical limiting effect on the subsequent attention mechanism. This ensures that attention enhancement is performed only in anatomically reliable regions during the subsequent enhancement feature modeling process, thereby avoiding irrelevant regions from being incorrectly highlighted or misidentified.

[0033] Will Processing is performed to generate images similar to plain CT scans. Anchoring feature diagram of the same size In order to convert this probability graph into a directly usable structured anchoring representation... The present invention can provide two processing methods, including... Binarize or retain the probability value: 1. Retain original probability values ​​(soft mask): Under certain model settings, retain... The original continuous probability value as These values ​​serve directly as guide maps for subsequent attention maps. The subsequent attention mechanism can then flexibly determine which regions are more strongly activated and which are less so through these continuous values. The advantages are a natural transition and preservation of detail; however, it is more sensitive to model training and can easily amplify probability prediction errors.

[0034] 2. Perform binarization (hard masking): Another approach is to set a structural saliency threshold. This transforms the probability map into a 0-1 binary map, activating attention only in high-confidence regions. This is represented as: ;or ; in, This indicates structural anchoring; a value of 1 means that enhanced attention is allowed to activate in this region, while a value of 0 means that the region is unstructured and attention activation should be blocked. The threshold representing structural saliency is used in most structural anatomy segmentation models. Regions with a predicted probability higher than 0.6 tend to have a high IOU and GT matching degree, representing a high-confidence output of the model. Therefore, this invention... Take 0.6, It represents the spatial coordinates of the image. Its advantages are stronger control and decision-making, and easier model convergence; however, the boundaries may not be smooth enough, resulting in significant information loss.

[0035] Regardless of the form used, the final result is All of these will serve as a spatial mask or attention guide map, which will be multiplied with the attention weight map generated in the enhanced CT branch in subsequent steps, allowing attention responses to be retained only in structurally anchored regions. This avoids untrusted regions being over-attented or misjudged by the model, emphasizes the importance and credibility of known anatomical regions, and provides structural boundary conditions for subsequent attention source tracing constraints.

[0036] S14 is a constraint on the attention response generated in the enhanced CT image branch during feature modeling in the deep network (MA-net), which is to ensure that it can only be activated in areas with credible structures. In other words, not all areas in the enhanced image are allowed to participate in modeling. Only those areas that are confirmed to have a clear anatomical basis in the plain scan image are given attention and further processing. This avoids the contrast agent-induced bright areas in the enhanced image from misleading the model and reduces the model's dependence on non-structural artifacts or noise. After the enhanced CT image is processed by convolution of its independent branches, an attention response map is generated. Each pixel value in this image represents the intensity of the enhancement feature that the model perceives at that location. The problem is that contrast-enhanced CT scans may contain contrast agent exudation, high-density lesions, and organ motion artifacts. These high-response areas do not necessarily correspond to the actual structures, easily introducing false positives. Therefore, a dot-multiplication operation (pixel-by-pixel multiplication) is used to transform the attention map generated by the enhanced CT branch. With structural anchoring diagram Multiplying them together yields a new restricted attention map. The underlying logic is that only those locations that have values ​​in the structural anchoring graph, i.e., are considered reliable structural regions, will have their enhanced attention retained; otherwise, the attention response at that location will be forcibly reset to zero.

[0037] In layman's terms, this refers to the structure-anchored representation in subsequent MA-net networks. It is used as a spatial mask or attention guide map to constrain the activation region of the attention weight map generated in the enhanced CT branch.

[0038] This masking mechanism is highly efficient, requiring no additional network branches or parameters. It can be embedded in the network structure using pixel-wise multiplication and supports end-to-end training. If the structure anchor map is probabilistic (non-binarized), the masking operation will create weighted attention, allowing partial activation. If the structure anchor map is a binary mask, the masking operation will create hard constraints, completely deactivating unstructured regions.

[0039] S2, Perform MA-net feature modeling based on attention source tracing constraints: Plain CT images, enhanced CT images, and structural anchoring representations are input into the MA-net network, which includes a two-branch structure. Plain CT images are input into the plain CT branch to extract latent structural features; enhanced CT images are input into the enhanced CT branch to extract contrast enhancement features. An attention source constraint mechanism is introduced into the MA-net network, which requires that the enhanced attention response generated in the enhanced CT branch must be mapped back to the corresponding structural anchoring region in the plain CT branch. Only when the enhanced attention response can establish an effective correspondence in the structural anchoring representation of the plain CT branch is it allowed to participate in feature modeling and generate enhanced feature representations subject to source constraints.

[0040] In the MA-net structure of this invention, plain CT images and enhanced CT images are respectively fed into two parallel network branches. The core components of these two branches are typical feature encoders, whose task is to extract layer-by-layer abstract feature maps from the original images for subsequent attention mechanisms, fusion processing, and lesion discrimination.

[0041] Each encoder is composed of the following basic modules stacked together: Convolutional layers: used to extract spatial features such as local texture, edges, and intensity variations; Normalization layer: Standardizes the feature distribution to avoid gradient explosion or vanishing. Nonlinear ReLU activation function: Introduces nonlinearity to enhance the network's ability to express complex patterns; Max pooling downsampling operation: reduces the feature map size, gradually expands the receptive field, and obtains higher-level abstract information.

[0042] The above steps are repeatedly stacked to form a multi-level structure (L layers). Each layer outputs a feature map, ultimately resulting in a multi-scale feature map set. S2 is detailed below: S21, both the plain CT branch and the enhanced CT branch contain a feature encoder composed of multiple convolutional layers, normalized layers, and nonlinear activation layers connected in sequence: The feature encoder of the plain CT branch is used to analyze plain CT images. As input, extract and output multi-scale structural latent feature maps, denoted as: This includes stable structures such as skeletal boundaries, organ margins, and major blood vessels. The execution process is as follows: Layer 1: Uses convolution to extract low-level features such as basic edges and density texture, i.e., output feature map. ; Layer 2: Continue convolution to extract larger structural regions and anatomical contours, i.e., output feature maps.

[0043] … Layer L: Extracts the global structural latent features of the entire anatomical region, i.e., outputs the highest-level feature map. ; The output size of each layer gradually decreases while the number of channels increases, together forming a feature pyramid structure that contains low-level details and high-level semantics.

[0044] Enhanced CT branch feature encoder to enhance CT images As input, extract and output multi-scale initial enhanced feature maps, denoted as: This includes information on contrast-enhanced areas and lesion contrast enhancement; the execution process is almost entirely symmetrical to the plain scan branch. Layer 1: Extracting local enhancement signals and contrast agent density abrupt changes, i.e., outputting feature maps. ; Layer 2: Extracting mid-layer features related to changes in organ contour after enhancement, i.e., output ; … Layer L: Identifies high-level features such as lesion areas with strong contrast changes, i.e., output ; Because the density features of enhanced images differ from those of plain scans, this branch helps to uncover the diagnostic contribution of enhancement features to lesions.

[0045] in, Indicates the number of feature levels. Indicates the branch of plain CT scan in the first The structural latent feature map output by the layer. Indicates the enhanced CT branch in the first The initial enhanced feature map output by the layer.

[0046] In the above multi-scale extraction: Low-level features (shallow layer): capture detailed edges and local textures of the image, such as small blood vessels and subtle density variations; Mid-level features: The learned structural morphology begins to possess semantic information, such as organ outlines and partition boundaries; Deep features: They possess global anatomical context and abstract lesion patterns, which helps distinguish abnormal and normal areas.

[0047] The acquisition of multi-scale features enables the model to consider both local information and overall structure, which is the foundation for achieving high-accuracy recognition.

[0048] The differences and complementarities between plain CT and contrast-enhanced CT branches: Plain CT images have stable structures but lack contrast enhancement information; contrast-enhanced CT images may have structural details obscured, but have strong lesion identification capabilities; by encoding each branch separately and then guiding fusion through an attention mechanism, the advantages of the two types of information can be complemented.

[0049] S22, The MA-net network introduces an attention source constraint mechanism, implemented by an attention source constraint module. This module is a key component of the proposed MA-net network structure, and its core task is to ensure that the spatial attention response map generated in the enhanced CT image branch is only allowed to be effective within the anatomical structure regions already identified in the plain CT image. This guarantees that the enhanced response has structural basis and medical interpretability. This addresses the problem of attention misalignment or false activation caused by high-contrast artifacts and signal drift in existing methods.

[0050] The attention source tracing constraint module receives three types of input: Initial enhancement feature map generated by enhanced CT branch Used to generate the original attention response map, which includes the enhancement signal; Structural latent feature maps at the same scale extracted from plain CT branches Provides structural semantic information for subsequent fusion; Structural anchoring representation (i.e., the trustworthy structure mask marked in the flat scan image): serves as a hard constraint condition for the trustworthy region.

[0051] S23, the execution process of the attention source tracing constraint module is as follows: S231, Based on the initial enhanced feature map Generate spatial attention weight map This is obtained through a fully connected attention gate or a combination of convolution-activation-normalization. Specifically, in the intermediate layer of the enhanced CT branch... The layer yielded a feature map. Its channels contain various texture, density gradient, and contrast enhancement signals, and this feature map is used to generate a spatial attention map. That is, the model's assessment of the degree of attention given to each spatial location. First, a 1×1 convolution is performed to extract inter-channel interactions. Then, ReLU activation is applied, followed by mapping to the [0,1] interval using the Sigmoid function. It is a two-dimensional matrix with the same spatial dimensions as the input feature map, where each value represents the intensity of the augmented response at that location. It's important to note that this attention map is entirely computed spontaneously from the augmentation map and has not been aligned with any structural information, potentially leading to misalignment or false activations.

[0052] S232, the above attention map Structure anchoring representation extracted from flat scan images Perform element-wise multiplication to obtain the attention weight graph with source tracing constraints. : ;in: 1. If a binarization mask is used as described above, then: In this case, the location is marked as a structurally reliable region in the flat scan image, which allows enhanced attention to participate in feature modeling, and the attention value can be preserved; At that time, the location did not belong to the structurally reliable region, and no structural support was found in the flat scan image, so the enhanced attention was forcibly set to zero; 2. If the above uses a probabilistic mask (between 0 and 1): At this point, the position allows enhanced attention to participate in feature modeling and to suppress or modulate the attention response proportionally as a confidence weight.

[0053] In terms of effect, it's like covering the attention enhancement map with a structural transparent sheet. Without structural support, even if the brightness in the enhancement map is high, the area will not be allowed to pass through.

[0054] S233, Attention weight graph after applying source tracing constraints This is used to reweight each channel in the initial enhanced feature map, that is, to perform weighted modulation on the initial enhanced feature map to obtain the intermediate enhanced feature map. : ;in, This operation indexes the channels; it selectively activates enhanced channel features based on the structure anchoring region in the spatial dimension. The feature vector at each spatial location is modulated by the credible attention value at that location; if a location has no corresponding structure in the structure anchoring map (i.e., 0), the enhanced features of all channels at that location will be cleared; if there is structural support, the enhanced features will be retained or even strengthened. In other words, the entire enhanced feature map is filtered by structural credibility in the spatial dimension, achieving precise selection and cleaning of local enhanced features.

[0055] S234, enhance the intermediate feature map Latent feature map of corresponding structures in plain CT branch Perform feature fusion to generate enhanced feature representations subject to source tracing constraints. The fusion operations include channel splicing and convolutional fusion methods: The final output fused features It has the following attributes: It possesses high-contrast lesion information in enhanced images; Maintain a stable, reliable, and structurally continuous spatial basis in the plain scan image; Achieve a balance between structural consistency and response sensitivity.

[0056] In summary, by introducing structural constraints as a hard filter for attention generation through the attention source constraint module of S23, traditional attention mechanisms typically rely solely on the features themselves to generate attention weights, lacking verification of the rationality of medical structures. This invention introduces structural anchoring representation into the attention generation path and uses element-wise multiplication to forcibly control the spatial activation range of the attention response, realizing a structural source tracing mechanism at the attention level. This effectively overcomes the fundamental problems of attention misalignment, drift, or false highlighting in enhanced images. Furthermore, enhanced CT images often suffer from incorrect network attention due to factors such as contrast agent leakage, incomplete vascular perfusion, and unstable lesion density. This invention improves the interpretability and reliability of the attention response by requiring enhanced attention to find corresponding support in the plain scan at the structural level. This invention is no longer a simple feature fusion but introduces an attention-level coupling step before fusion, essentially forming a product spatial modulation mechanism of enhanced signal-structural trust, balancing high response and structural trust.

[0057] S3, Lesion discrimination output based on consistency of source tracing: Based on the enhanced feature representation constrained by the source, lesion region discrimination is performed, and the lesion identification results and the corresponding attention-source consistency map are output to characterize the correspondence reliability between the enhanced response and the plain scan structure.

[0058] S31, lesion region discrimination is performed through a lesion discrimination decoder, which represents enhanced features subject to source tracing constraints. The input is passed sequentially through at least one decoding layer, which includes convolution and upsampling operations to progressively restore spatial resolution and integrate multi-scale semantic information.

[0059] Finally, a lesion probability map is output through a Sigmoid activation function. This image has the same spatial dimensions as the input image, where the value of each pixel represents the probability that it belongs to the lesion region. ;in, This represents the feature decoding function of the lesion discrimination decoder. It is the Sigmoid activation function. Represents pixels The probability value that it belongs to the lesion, and The input to the feature decoding function is the fused feature map. Its characteristics include low resolution, high number of channels, and feature vectors at each location that combine the response of enhanced CT with the structural anchoring constraints of plain CT. It also suffers from significant spatial information loss but high semantic condensation. Therefore, the main task of the decoder is spatial recovery and semantic preservation. The feature decoding function consists of multiple stacked decoding layers, each including: Upsampling operation: Bilinear interpolation is used to restore the spatial size of the feature map to the previous level, for example, from 32×32 to 64×64; Convolution operation: restore details and refine spatial patterns by performing two consecutive 3×3 convolutions, each followed by ReLU activation; Normalization operation: used for stable training; The above process will be repeated multiple times until the spatial resolution of the final feature map is restored to match that of the original image. After restoring the full spatial size, the final projection layer will be entered, and the following operations will be performed: A 1×1 convolution is used to compress a multi-channel feature map into a 1-channel feature map; Next, the Sigmoid activation function is used to map the output to the range [0,1] to form a lesion probability map.

[0060] Represented as: ; This indicates a convolution operation compressed to a single channel. It is the final feature map output by the decoder.

[0061] S32, Output lesion identification results: lesion probability map Apply a predetermined probability threshold Binarization is performed to obtain a binarized lesion segmentation mask. : .

[0062] The probability threshold is set to a neutral probability boundary of 0.5, which means that lesions and non-lesions are equally likely. Based on this mask, the spatial location and morphological attributes of the lesions can be further calculated, which can be used as part of the lesion identification result or as a derivative output.

[0063] 1. Spatial location index is represented by the centroid coordinates of the lesion: Calculate the average spatial coordinates of all pixels in the mask region: ;in, It is the set of all pixels with a mask value of 1. It represents the number of pixels in the lesion area; 2. Morphological and structural attributes are characterized as follows: Area: The total number of pixels with a value of 1 in the mask, representing the projected area of ​​the lesion region.

[0064] Boundary contour and perimeter: The edge of the lesion can be obtained by using a contour extraction algorithm, and then its length can be calculated to represent the complexity of the lesion's shape.

[0065] Boundary irregularity: The roundness factor is used to measure the regularity of lesion boundaries. The closer the value is to 1, the more regular the shape; the lower the value, the more complex the edges.

[0066] S33, output the corresponding attention source consistency graph: Attention Source Consistency Graph It is achieved by calculating the attention weight graph after source tracing constraints. Representation of structural anchoring It is generated by corresponding metrics in spatial location. Metric methods include: Dot product method: This method directly reflects the actual degree of overlap of the enhanced attention response within the structural anchoring region.

[0067] Alternatively, the correlation coefficient method can be used: within a local window Calculate the Pearson correlation within this region: ;in, Indicates The small window centered on Corr represents the Pearson correlation function.

[0068] The final generated The consistency map is the same size as the input image, and each pixel value represents the reliability level of whether the enhanced attention has structural support at that location.

[0069] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0070] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning analysis method based on the isotopic comparison of plain CT and enhanced CT scans using MA-net, characterized in that, Includes the following steps: Based on the plain CT images and enhanced CT images of the same patient, the co-position regions of the plain CT images and enhanced CT images are aligned, and stable anatomical structure responses are extracted from the plain CT images. A structural anchoring representation is constructed to characterize the structural anchoring representation in the plain CT images that can be attached to the enhancement response. The structural anchoring representation is used to define the source region of enhancement attention. The plain CT images, enhanced CT images, and structural anchoring representations are input into an MA-net network including a dual-branch structure, wherein: Plain CT images are input into the plain CT branch to extract latent structural features; Enhanced CT images are input into the enhanced CT branch to extract contrast enhancement features; An attention source constraint mechanism is introduced into the MA-net network, which requires that the enhanced attention response generated in the enhanced CT branch must be mapped back to the corresponding structural anchoring region in the plain CT branch. Only when the enhanced attention response can establish an effective correspondence in the structural anchoring representation of the plain CT branch is it allowed to participate in feature modeling and generate enhanced feature representations subject to source constraints. Based on the enhanced feature representation constrained by the source, lesion region discrimination is performed, and the lesion identification result and the corresponding attention source consistency map are output to characterize the correspondence reliability between the enhanced response and the plain scan structure.

2. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison as described in claim 1, characterized in that, The alignment of the co-position region adopts an image registration algorithm based on deformation field learning, which maps the spatial coordinates of the enhanced CT image to the spatial coordinates of the plain CT image, so as to achieve pixel-level alignment of the two sets of images at the anatomical position.

3. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison according to claim 2, characterized in that, The extraction of the stable anatomical structure response includes using a pre-trained structural segmentation network with plain CT images as input to extract the contour or region probability map of stable anatomical structures, including major blood vessels, organ boundaries, and bones, from the plain CT images; binarizing or retaining the probability values ​​of the contour or region probability map of the stable anatomical structure to generate a matrix or feature map of the same size as the plain CT image, which serves as the anchoring representation of the structure.

4. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison according to claim 1, characterized in that, The structural anchoring representation, as a spatial mask or attention guide map, constrains any attention weight map originating from enhanced CT branches to be activated or have non-zero values ​​only within the anatomical region identified by the structural anchoring representation.

5. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison according to claim 1, characterized in that, Both the plain CT branch and the enhanced CT branch include a feature encoder composed of multiple convolutional layers, normalized layers, and nonlinear activation layers connected in sequence; wherein: The feature encoder of the plain CT branch takes the plain CT image as input, extracts and outputs multi-scale structural latent feature maps; The feature encoder of the enhanced CT branch takes the enhanced CT image as input and extracts and outputs multi-scale initial enhanced feature maps.

6. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison according to claim 5, characterized in that, The attention source constraint mechanism introduced in the MA-net network is implemented by the attention source constraint module. The input of the attention source constraint module includes: the initial enhancement feature map from the enhanced CT branch, the structural latent feature map from the plain CT branch, and the structural anchoring representation.

7. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison according to claim 6, characterized in that, The attention source constraint module first calculates the spatial attention weight map based on the initial enhanced feature map, and then multiplies the spatial attention weight map element-wise with the structural anchoring representation to obtain the attention weight map after source constraint. This forces the enhanced attention response generated in the enhanced CT branch to be restricted to the anatomical structure region defined by the structural anchoring representation, that is, it must be mapped back to the corresponding structural anchoring region in the plain CT branch.

8. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison according to claim 7, characterized in that, The attention weight map after source tracing constraint is then used to weight and modulate the initial enhanced feature map to generate intermediate enhanced features. The intermediate enhanced features are then fused with the same-scale structural latent feature map from the plain CT branch to generate the enhanced feature representation subject to source tracing constraint. The participation feature modeling is achieved through the element-wise multiplication operation. The response intensity can be retained and transmitted only if the response position in the spatial attention weight map has a non-zero value or a high probability value in the structural anchoring representation.

9. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison according to claim 4, characterized in that, The lesion region discrimination is performed through a lesion discrimination decoder. The lesion discrimination decoder takes the enhanced feature representation subject to source tracing constraints as input, processes it through at least one decoding layer containing convolution and upsampling operations, and finally outputs a lesion probability map through an activation function, where the value of each pixel represents the probability that the corresponding position belongs to a lesion. The lesion probability map is binarized by applying a predetermined probability threshold to obtain a binarized lesion segmentation mask, and the lesion identification result includes the lesion segmentation mask.

10. The deep learning analysis method based on MA-net plain CT and enhanced CT isotopic comparison according to claim 9, characterized in that, The generation of the attention source consistency map includes: for each spatial location in the attention weight map after source constraint, calculating the product or correlation coefficient of its weight value and the corresponding position value in the structural anchoring representation, and generating a consistency map of the same size as the image with a value range between [0,1]; the value of each pixel in the consistency map represents the reliability of the correspondence between the contrast enhancement response observed by enhanced CT at that location and the anatomical basis defined by plain CT.

Images