A discriminative and certainty-enhancing method and system for complex background medical image segmentation

Abstract

The application discloses a discriminative and deterministic enhancement method and system for complex background medical image segmentation, and relates to the technical field of intelligent auxiliary diagnosis and treatment. The application comprises the following steps: a medical image segmentation model obtains a multi-scale feature map of a medical image through a feature extraction network; a discriminative query network generates a global semantic label according to the multi-scale feature map, initializes a foreground query vector and a background query vector in combination with foreground semantic information and background semantic information, and fuses and updates query responses obtained at different scales to obtain an enhanced multi-scale feature map; the enhanced multi-scale feature map is subjected to a decoder to generate an initial segmentation probability map, a class-aware feature propagation module is used to correct the features of the initial segmentation probability map, and image segmentation is completed. The application is used for realizing enhancement on a real target region, suppression on a similar background region and fine correction on an uncertain region.

Description

Technical Field

[0001] This invention relates to the field of intelligent assisted diagnosis and treatment technology, and more specifically to a discriminative and deterministic enhancement method and system for medical image segmentation with complex backgrounds. Background Technology

[0002] Medical image segmentation is a key technology in medical image analysis, widely used in lesion detection, tissue structure analysis, surgical planning, assisted diagnosis, and efficacy evaluation. With the development of deep learning technology, medical image segmentation methods based on convolutional neural networks, encoder-decoder structures, attention mechanisms, and Transformer structures have been extensively researched and applied.

[0003] Existing technologies typically employ U-Net and its improved networks, such as U-Net++, ResUNet, and Attention U-Net. These methods usually use an encoder to extract multi-scale features and a decoder to progressively restore spatial resolution, achieving pixel-level segmentation results. To improve the segmentation accuracy of complex lesions, subsequent research has further introduced attention mechanisms, multi-scale feature fusion, boundary enhancement, and contextual modeling, enabling the model to focus more on the target region and reduce interference from the background region.

[0004] Furthermore, to address common issues in medical images such as blurred boundaries, low contrast, and significant morphological variations, existing technologies have proposed methods such as boundary-aware learning, uncertainty modeling, Bayesian inference, probabilistic segmentation networks, contrastive learning, and semantic decoupling. These methods primarily improve the segmentation reliability of the model by estimating prediction confidence, enhancing boundary features, modeling foreground-background differences, or utilizing auxiliary tasks.

[0005] In recent years, some methods have attempted to combine CNN and Transformer structures to improve medical image segmentation performance by leveraging local texture modeling and global context modeling capabilities. Other methods have introduced contrastive learning or uncertainty constraints to enhance the model's ability to distinguish foreground, background, and uncertain regions. These existing techniques have achieved certain results in routine medical image segmentation tasks and have a basis for application in tasks such as polyp segmentation, skin lesion segmentation, gland segmentation, and tumor region segmentation.

[0006] Despite the significant progress made by the aforementioned technologies in the field of medical image segmentation, the following problems and shortcomings still exist in complex and realistic clinical image environments.

[0007] First, existing methods typically attribute segmentation errors primarily to blurred or uncertain target boundaries. Therefore, most methods focus on boundary enhancement, edge refinement, or modeling of prediction uncertainty. However, in real medical images, segmentation errors do not always originate from boundary regions. When background tissue is highly similar to the actual lesion in texture, grayscale, color, morphology, or structure, the model is prone to producing a higher response in non-lesion areas, misclassifying similar background as lesions or uncertain regions, thus creating pseudo-uncertainty. This type of pseudo-uncertainty can lead to problems such as false detections, boundary shifts, region adhesion, segmentation discontinuities, or false positive background enhancement.

[0008] Secondly, existing attention mechanisms or feature enhancement methods mostly rely on fixed convolutional receptive fields, static weight allocation, or global attention calculations. They can usually only passively enhance salient regions and lack the ability to actively discover truly discriminative regions. When lesions are similar in appearance to surrounding tissues, the model struggles to accurately distinguish between the features of the real lesion and similar background features, easily leading to insufficient foreground enhancement or inadequate background suppression.

[0009] Furthermore, while existing uncertainty modeling methods can identify low-confidence regions, they often fail to further distinguish between "true uncertainty regions" and "pseudo-uncertainty regions caused by similar backgrounds." Therefore, during subsequent feature propagation or prediction correction, the model may retain or even amplify erroneous responses from background regions, leading to error propagation and category confusion.

[0010] Furthermore, existing medical image segmentation models often suffer from insufficient generalization ability across different imaging devices, hospital data, lighting conditions, imaging modalities, or patient groups. Due to the significant cross-domain distribution differences in medical images, fixed feature extraction methods and ordinary supervised learning strategies are insufficient to ensure stable performance of the model in multi-center, multi-device, and multi-modal scenarios, easily leading to prediction fluctuations, increased false positives, or decreased segmentation accuracy.

[0011] Furthermore, existing methods typically lack a unified mechanism for foreground enhancement, background suppression, and uncertain region correction when dealing with boundary regions, low-contrast regions, and similar background regions. Their feature learning process does not explicitly construct high-confidence semantic anchors, nor does it fully utilize high-confidence regions to guide class consistency in low-confidence regions, resulting in insufficient feature discriminativeness, boundary continuity, and prediction stability in complex backgrounds.

[0012] Therefore, the existing technology has at least the following technical problems: First, it is difficult to suppress spurious responses generated by background areas that resemble the appearance of the lesion; Second, it is difficult to distinguish between genuine uncertainty and pseudo-uncertainty caused by similar backgrounds; Third, there is a lack of mechanisms for actively extracting discriminative regions and separating foreground / background features; Fourth, the correction for low-confidence regions is not stable enough and is prone to cross-class error propagation. Fifth, it lacks robustness and generalization ability under cross-device, cross-center, cross-modal, or complex imaging conditions.

[0013] The purpose of this invention is to address the aforementioned shortcomings by providing a method, system, device, and storage medium for enhancing the discriminative and deterministic capabilities of medical image segmentation with complex backgrounds. This is achieved by constructing a discriminative anchor learning mechanism to actively identify foreground / background semantic regions with discriminative capabilities; by using a category-aware feature propagation mechanism to transfer reliable information from high-confidence regions to uncertain regions; and by enhancing the stability of the model under varying distribution conditions through foreground and background feature separation constraints. This suppresses pseudo-uncertainty caused by similar backgrounds and improves the accuracy, robustness, and generalization ability of medical image segmentation. Summary of the Invention

[0014] In view of this, the present invention provides a discriminative and deterministic enhancement method and system for medical image segmentation with complex backgrounds. Addressing the problem of pseudo-uncertainty arising when lesion regions and background tissues in medical images are highly similar in color, texture, grayscale, and morphology, the invention utilizes a discriminative query network, a category-aware feature propagation module, and discriminative optimization constraints to enhance the true target region, suppress similar background regions, and refine uncertain regions. The overall structure of DCE-Net comprises three core components: DQN, CFPM, and ADL. DQN extracts discriminative semantic anchors from multi-scale features; CFPM propagates information from high-confidence regions to uncertain regions under spatial-semantic constraints; and ADL enhances the feature separation capability between foreground and background.

[0015] To achieve the above objectives, the present invention adopts the following technical solution: A discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds includes the following steps: Medical image segmentation models acquire multi-scale feature maps of medical images through feature extraction networks; The discriminative query network generates global semantic labels based on multi-scale feature maps, initializes foreground and background query vectors by combining foreground and background semantic information, and fuses and updates query responses obtained at different scales to obtain enhanced multi-scale feature maps. The enhanced multi-scale feature map is used by the decoder to generate an initial segmentation probability map. The features of the initial segmentation probability map are then corrected using a category-aware feature propagation module to complete image segmentation.

[0016] Optionally, the medical images include endoscopic images, dermatoscope images, pathological images, CT images, MRI images, ultrasound images, or other medical images; after preprocessing, the medical images are sent to a feature extraction network; the preprocessing includes normalization and pixel normalization.

[0017] Optionally, the foreground query vector is used to actively find the most discriminative region in the lesion or target tissue, while the background query vector is used to model interference regions that are similar to the lesion but actually belong to the background.

[0018] Optionally, the foreground query vector and background query vector predict reference positions and local offset positions on feature maps of different scales, and actively collect the most relevant local region features to the query through deformable sampling. This allows the medical segmentation model to adaptively select the region of interest based on the shape, position and structural changes of the target, without relying on fixed convolutional windows or static attention weights.

[0019] Optionally, the updated foreground query vector is used to enhance the response of the lesion region or target tissue region, and the updated background query vector is used to suppress the response of similar backgrounds, pseudo-lesion regions or other non-target regions.

[0020] Optionally, the initial segmentation probability map features can be corrected using a category-aware feature propagation module. Specifically, the pixels are divided into regions based on the initial segmentation probability map, and spatial-semantic associations are established between uncertain pixels and high-confidence pixels. During feature propagation, uncertain background regions receive information from high-confidence background regions. The uncertain prediction results are then updated with weights based on the propagation weights.

[0021] Optionally, the training may also include optimizing the medical image segmentation model, specifically: the training loss includes basic segmentation loss, query diversity constraint, category-aware propagation loss, and category protection loss; wherein, the basic segmentation loss is used to ensure that the predicted segmentation result is consistent with the manually labeled result; the query diversity constraint is used to avoid multiple queries learning the same attention region, so that different queries can cover different discriminative structures; the category-aware propagation loss is used to guide uncertain regions to move closer to high-confidence regions of the same category; and the category protection loss is used to prevent errors from propagating across categories.

[0022] A discriminative and deterministic enhancement system for medical image segmentation with complex backgrounds, comprising: Image input and preprocessing module: used to receive medical image data and preprocess the medical image data; Multi-scale feature extraction module: used to extract multi-level features from the input image; Discriminative Query Network Module: Used to proactively discover semantic anchors with discriminative capabilities, including query initialization, deformable sampling, query updating, and feature enhancement; Multi-scale fusion and initial segmentation module: used to upsample and fuse the enhanced features output by DQN to obtain an initial segmentation probability map corresponding to the size of the input image; Category-aware feature propagation module (CFPM): used to correct uncertain regions in the initial segmentation probability map; Discriminative optimization module: used to constrain the model to learn more stable foreground and background representations; Segmentation Result Output Module: Used to output the final segmentation result, which can be in the form of a segmentation probability map, a binary mask, a target boundary, a color overlay map, a confidence map, or a structured measurement result.

[0023] As can be seen from the above technical solution, compared with the prior art, the present invention provides a discriminative and deterministic enhancement method and system for medical image segmentation with complex backgrounds, which has the following beneficial effects: 1. It can effectively suppress pseudo-uncertainty caused by similar backgrounds.

[0024] To address the issue of background tissues and lesion areas being highly similar in texture, color, grayscale, or morphology in medical images, this invention models the target region and interference region respectively through foreground and background queries. This enables the model to actively identify truly discriminative lesion areas and suppress erroneous high responses in similar backgrounds, thereby reducing false detections and false positive segmentation.

[0025] 2. Improve the accuracy of target segmentation in complex backgrounds.

[0026] This invention no longer relies solely on ordinary convolution or static attention mechanisms. Instead, it actively locates key semantic anchors through a discriminative query network and enhances them by combining multi-scale features. This enables the model to extract lesion, tissue, or organ boundaries more accurately, improving the consistency between the segmentation results and the real annotations.

[0027] 3. Improve the segmentation stability of regions with ambiguous boundaries and low confidence.

[0028] This invention uses a category-aware feature propagation module to use high-confidence foreground or background regions as reliable semantic references, and performs same-category feature propagation and prediction correction on uncertain regions, which can improve problems such as boundary breakage, contour offset, local missing and region adhesion.

[0029] 4. Reduce the risk of cross-category error propagation.

[0030] Existing uncertainty correction methods are prone to mispropagating foreground information to background regions or vice versa. This invention incorporates category constraints during feature propagation, ensuring that uncertain foregrounds are primarily guided by high-confidence foreground regions, and uncertain backgrounds are primarily guided by high-confidence background regions, thereby improving semantic consistency and reducing category confusion.

[0031] 5. Enhance the separability of foreground and background features.

[0032] This invention uses discriminative optimization constraints to enable foreground and background queries to learn more separate and stable feature representations, avoiding different queries from focusing on the same region or causing semantic overlap, thereby improving the model's ability to distinguish between real lesion regions and similar background regions.

[0033] 6. Improve the robustness and generalization ability of the model under different imaging conditions.

[0034] Because medical images are often affected by differences in equipment, lighting, imaging modality, hospital origin, and patient, existing models are prone to performance fluctuations. This invention improves the stability of the model in cross-device, cross-center, cross-dataset, and cross-modal scenarios through discriminative anchor learning and foreground / background separation mechanisms. Experimental results also show that this method achieves stable improvements on multiple medical image segmentation datasets and can better suppress false positive responses.

[0035] 7. Improve the reliability of clinical auxiliary analysis results.

[0036] The segmentation probability map, binary mask, or target contour output by this invention is more accurate and can be further used for lesion localization, area measurement, boundary analysis, follow-up comparison, surgical planning, and auxiliary diagnosis, reducing the burden of manual delineation for doctors and improving the efficiency of medical image analysis.

[0037] 8. It has good scalability and engineering deployment value.

[0038] This invention can be used as a standalone medical image segmentation algorithm, or integrated into endoscopy systems, dermatoscopy systems, pathology image analysis systems, medical imaging workstations, cloud-based assisted diagnostic platforms, or edge computing devices. Its technical concept can also be extended to other image segmentation tasks with similar background interference, such as industrial defect detection, remote sensing target extraction, and microscopic image analysis. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0040] Figure 1 This is a diagram showing the overall network structure of the method of the present invention; Figure 2 A structural diagram of the Discriminative Query Network (DQN); Figure 3 This is a structural diagram of the Category Aware Feature Propagation Module (CFPM). Figure 4 This is a flowchart of the training and optimization process of the present invention; Figure 5 This is a schematic diagram illustrating the segmentation effect of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] This invention improves the network structure, feature processing flow, and training optimization methods of medical image segmentation systems, addressing the shortcomings of existing methods in areas such as similar background interference, pseudo-uncertainty, boundary blurring, and insufficient cross-domain generalization. Specifically, this invention achieves synergistic optimization of foreground enhancement, background suppression, and uncertain region correction through a Discriminant Query Network (DQN), a Category Aware Feature Propagation Module (CFPM), and discriminant optimization constraints. DQN is used to extract semantic anchors, CFPM is used to correct pseudo-uncertain regions under spatial-semantic constraints, and ADL is used to enhance the feature separation capability between foreground and background.

[0043] This invention discloses a discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds, comprising the following steps: Medical image segmentation models acquire multi-scale feature maps of medical images through feature extraction networks; The discriminative query network generates global semantic labels based on multi-scale feature maps, initializes foreground and background query vectors by combining foreground and background semantic information, and fuses and updates query responses obtained at different scales to obtain enhanced multi-scale feature maps. The enhanced multi-scale feature map is used by the decoder to generate an initial segmentation probability map. The features of the initial segmentation probability map are then corrected using a category-aware feature propagation module to complete image segmentation.

[0044] Optionally, the medical images include endoscopic images, dermatoscope images, pathological images, CT images, MRI images, ultrasound images, or other medical images; after preprocessing, the medical images are sent to a feature extraction network; the preprocessing includes normalization and pixel normalization.

[0045] Specifically, such as Figure 1 As shown, the medical image segmentation method of the present invention mainly includes steps such as image input and preprocessing, multi-scale feature extraction, discriminative query enhancement, initial segmentation prediction, category-aware feature propagation, joint optimization training, and segmentation result output.

[0046] First, the medical image to be segmented is input into the model. The medical image can be an endoscopy image, dermatoscopy image, pathology image, CT image, MRI image, ultrasound image, or other medical images. After preprocessing such as size normalization and pixel normalization, the input image is fed into a feature extraction network. The feature extraction network can employ a convolutional neural network, a Transformer network, or a combination of both encoder structures to obtain multi-scale feature maps with different resolutions and semantic levels.

[0047] After obtaining multi-scale features, such as Figure 2 As shown, this invention employs a Discriminative Query Network (DQN) to actively model features for discrimination. Instead of simply applying uniform attention weights to the entire image, DQN constructs separate foreground and background queries. The foreground query actively identifies the most discriminative regions of lesions or target tissues, while the background query models interference regions that are similar to lesions but actually belong to the background. In this way, the model can actively distinguish between the true foreground and similar background in the feature space.

[0048] Specifically, DQN first generates a global semantic representation based on multi-scale features, and initializes the foreground query vector and background query vector by combining foreground and background semantic information, respectively. Then, each query vector predicts a reference position and a local offset position on feature maps at different scales, and actively collects the local region features most relevant to the query through deformable sampling. This process enables the model to adaptively select regions of interest based on changes in the target's shape, position, and structure, rather than relying on fixed convolutional windows or static attention weights.

[0049] Next, DQN fuses and updates the query responses obtained at different scales, ensuring that each query vector simultaneously possesses both local detail information and global semantic information. The updated foreground query is used to enhance the response to the lesion region or target tissue region, while the updated background query is used to suppress the response to similar backgrounds, pseudo-lesion regions, or other non-target regions. After processing by this module, the original multi-scale features are transformed into more discriminative enhancement features.

[0050] Then, the enhanced multi-scale features are upsampled and fused by the decoder to generate an initial segmentation probability map. Each pixel value in this probability map represents the probability that the pixel belongs to the target region. Since medical images often contain blurred boundaries, low contrast, and similar background interference, relying solely on the initial probability map may still result in uncertain areas or erroneous responses. Therefore, this invention further introduces a category-aware feature propagation module (CFPM) for correction.

[0051] like Figure 3 As shown, CFPM first divides pixels into regions based on the initial segmentation probability map. Taking binary lesion segmentation as an example, when the predicted probability of a pixel is higher than a preset foreground confidence threshold, it is classified as a high-confidence foreground region; when the predicted probability is lower than a preset background confidence threshold, it is classified as a high-confidence background region; and when the predicted probability is in the middle range between the foreground and background, it is classified as an uncertain region. Preferably, pixels with a probability greater than or equal to 0.7 can be designated as high-confidence foreground regions, pixels with a probability less than 0.3 as high-confidence background regions, and pixels between 0.3 and 0.7 as uncertain regions; these thresholds can be adjusted according to the specific task, image modality, or data distribution.

[0052] After region segmentation, CFPM establishes spatial-semantic associations between uncertain pixels and high-confidence pixels. For any uncertain pixel, CFPM calculates its feature similarity and spatial distance similarity with high-confidence pixels of the same category. Feature similarity measures the semantic proximity of two pixels, while spatial distance similarity measures their proximity in the image. Together, they constitute the propagation weights, enabling reliable high-confidence regions to guide adjacent and semantically similar uncertain regions.

[0053] During feature propagation, this invention restricts information to propagate only within the same category. That is, uncertain foreground regions primarily receive information from high-confidence foreground regions, and uncertain background regions primarily receive information from high-confidence background regions. This category constraint prevents foreground information from being mistakenly propagated to the background region, or vice versa, thereby reducing cross-category interference and semantic drift.

[0054] Subsequently, CFPM updates the prediction results for uncertain regions based on propagation weights. For uncertain foreground pixels, if they are highly consistent with high-confidence foreground regions in terms of features and spatiality, their foreground probability is gradually increased; for uncertain background pixels, if they are more consistent with high-confidence background regions, their foreground probability is gradually decreased. To avoid amplifying errors due to one-time corrections, this invention employs a progressive semantic update strategy, gradually increasing the update intensity during training to keep the model stable in the early stages and gradually enhance its ability to correct uncertain regions in the later stages. CFPM improves boundary consistency and reduces cross-class interference by distinguishing between high-confidence and uncertain regions and correcting uncertain regions based on spatial-semantic consistency.

[0055] During the training phase, such as Figure 4 As shown, this invention employs a joint optimization approach to train the entire network. The training loss can include basic segmentation loss, query diversity constraint, category-aware propagation loss, and category protection loss. Specifically, the basic segmentation loss ensures consistency between the predicted segmentation result and the manually labeled result; the query diversity constraint prevents multiple queries from learning duplicate attention regions, enabling different queries to cover different discriminative structures; the category-aware propagation loss guides uncertain regions toward higher-confidence regions of the same category; and the category protection loss prevents errors from propagating across categories. Through this joint optimization, DQN provides stable discriminative semantic anchors in the feature space, while CFPM improves the category consistency of uncertain regions in the prediction space. Together, they enhance the accuracy and robustness of the segmentation results.

[0056] During the inference phase, after receiving the medical image to be processed, the trained model sequentially performs multi-scale feature extraction, DQN discriminative enhancement, multi-scale decoding fusion, and CFPM uncertainty region correction, ultimately outputting a segmentation probability map of the target region. The segmentation probability map can be further converted into a binary segmentation mask through thresholding operations. Alternatively, it can output target contours, heatmaps, confidence maps, or structured measurement results, as shown in the image. Figure 5 As shown. For example, lesion area, maximum diameter, boundary length, and location coordinates, etc.

[0057] Figure 5 The comparison between the input image, manually annotated results, segmentation results from traditional methods, and segmentation results from the present invention is presented to demonstrate the effectiveness of the present invention in terms of boundary integrity, background suppression, and reduction of pseudo-uncertainty.

[0058] This embodiment also discloses a discriminative and deterministic enhancement system for medical image segmentation with complex backgrounds, including: Image input and preprocessing module: used to receive medical image data and preprocess the medical image data; Multi-scale feature extraction module: used to extract multi-level features from the input image; Discriminative Query Network Module: Used to proactively discover semantic anchors with discriminative capabilities, including query initialization, deformable sampling, query updating, and feature enhancement; Multi-scale fusion and initial segmentation module: used to upsample and fuse the enhanced features output by DQN to obtain an initial segmentation probability map corresponding to the size of the input image; Category-aware feature propagation module (CFPM): used to correct uncertain regions in the initial segmentation probability map; Discriminative optimization module: used to constrain the model to learn more stable foreground and background representations; Segmentation Result Output Module: Used to output the final segmentation result, which can be in the form of a segmentation probability map, a binary mask, a target boundary, a color overlay map, a confidence map, or a structured measurement result.

[0059] Among them, the image input and preprocessing module

[0060] This module receives medical image data and converts it into a data format that the model can process. Preprocessing operations can include image scaling, cropping, normalization, and grayscale or color channel conversion. For the training process, data augmentation techniques such as random rotation, horizontal flipping, and vertical flipping can be used to improve the model's adaptability to different imaging angles, lighting conditions, and data from different devices.

[0061] Multi-scale feature extraction module: This module is used to extract multi-level features from the input image. Shallow features contain rich edge, texture, and detail information, while deep features contain strong semantic information. By simultaneously preserving features at multiple scales, this invention can take into account small objects, fine boundaries, complex shapes, and global contextual information.

[0062] Discriminative Query Network (DQN): This module is one of the core modules of this invention, used to proactively discover semantic anchors with discriminative capabilities. DQN includes four processes: query initialization, deformable sampling, query update, and feature enhancement.

[0063] During query initialization, the model generates foreground and background queries based on global context information. The foreground query learns the discriminative features of the target region, while the background query learns the structural features of similar backgrounds or interfering tissues.

[0064] In deformable sampling, each query predicts a reference point and offset point based on its own semantic information, and collects relevant regional features from multi-scale feature maps. This mechanism enables the model to adapt to changes in lesion morphology, scale, and location.

[0065] During the query update process, the model integrates sampled responses at different scales, enabling the query to possess both local structural information and global semantic information.

[0066] During feature enhancement, foreground queries enhance the target region, while background queries suppress non-target regions, ultimately yielding enhanced features more suitable for segmentation tasks. DQN actively samples and aggregates target-related regions through semantic queries, achieving discriminative structure modeling and feature decoupling.

[0067] Multi-scale fusion and initial segmentation module: This module upsamples and fuses the enhanced features output by DQN to obtain an initial segmentation probability map corresponding to the input image size. This probability map provides the initial confidence basis for the subsequent CFPM module.

[0068] The Category-Aware Feature Propagation Module (CFPM) corrects uncertain regions in the initial segmentation probability map. CFPM first divides regions into high-confidence and uncertain regions based on predicted probabilities. Then, it calculates propagation weights between uncertain and high-confidence pixels based on same-class constraints. These propagation weights consider both semantic feature similarity and spatial proximity. Finally, CFPM uses high-confidence regions of the same class to weightedly correct the uncertain regions, gradually bringing them closer to the correct class.

[0069] This module effectively addresses prediction instability caused by blurred boundaries, low contrast, and similar backgrounds, reducing the impact of pseudo-uncertain regions on segmentation results.

[0070] Discriminative Optimization Module: This module primarily functions during the training phase, constraining the model to learn more stable foreground and background representations. Through segmentation supervision, query diversity constraints, same-class propagation constraints, and class protection constraints, this module enables clearer feature separation between foreground and background queries and makes the correction process for uncertain regions more stable. After training, the constraints of this module are already reflected in the model parameters, allowing segmentation to be completed without manual intervention during the inference phase.

[0071] Segmentation Result Output Module: This module outputs the final segmentation result. The output format can be a segmentation probability map, a binary mask, target boundaries, a color overlay image, a confidence map, or structured measurement results. For clinical auxiliary diagnostic scenarios, the segmentation result can be overlaid on the original medical image to allow doctors to quickly observe the location and boundaries of lesions.

[0072] This invention relates to the fields of artificial intelligence, computer vision, medical image processing, and intelligent assisted diagnosis and treatment, and is particularly applicable to target segmentation, lesion region identification, tissue structure extraction, and refined discrimination of uncertain regions in medical images under complex backgrounds. Its known and potential applications and methods include, but are not limited to, the following: 1. Field of digestive endoscopy image analysis This invention can be applied to the automatic segmentation of polyps, tumors, ulcers, inflammation or other abnormal tissue areas in endoscopic images such as gastroscopy, colonoscopy, and capsule endoscopy.

[0073] In this application, images acquired by endoscopy are input into a trained segmentation model. Through discriminative anchor learning, category-aware feature propagation, and foreground / background feature separation mechanisms, segmentation masks, contour boundaries, or probability maps of lesion areas are automatically generated to assist doctors in locating lesion areas, reducing the false negative rate, and minimizing missegmentation caused by similarity between mucosal texture, color, shape, and lesion.

[0074] 2. Field of image analysis of skin lesions

[0075] This invention can be applied to the segmentation of lesions such as melanoma, pigmented nevi, skin tumors, and areas of skin inflammation in images such as dermoscopy images and clinical skin photographs.

[0076] In this application, the present invention can extract the lesion area and suppress the spurious response of the surrounding normal skin under complex conditions such as skin texture, hair occlusion, light changes, and blurred lesion boundaries, providing an image basis for skin lesion area measurement, boundary analysis, follow-up comparison and auxiliary diagnosis.

[0077] 3. Field of Pathological Tissue Image Analysis

[0078] This invention can be applied to the segmentation of tissue structures such as glands, cell nuclei, tumor areas, necrotic areas, and inflammatory areas in digital pathological slide images.

[0079] In this application, the present invention uses high-confidence tissue regions as semantic anchors to perform feature propagation and category consistency correction on low-confidence or ambiguous boundary regions, thereby improving the segmentation continuity and boundary integrity under complex tissue morphology. It can be used in scenarios such as pathological quantitative analysis, tumor infiltration area assessment, and tissue morphology analysis.

[0080] 4. Multimodal medical image segmentation in CT, MRI, ultrasound, and other fields

[0081] This invention can be extended to the segmentation of organ, lesion, blood vessel, mass, nodule or tissue boundary in medical imaging such as CT, MRI, ultrasound, PET, and OCT.

[0082] In this application, two-dimensional or three-dimensional medical image slices are input into the model. The target region is identified through cross-scale feature extraction and discriminative query mechanism. The uncertain prediction in low-contrast, noise interference, or tissue similarity regions is corrected through the category-aware propagation module. It is suitable for medical image analysis of liver, lung, brain, breast, prostate, thyroid and other parts of the body.

[0083] 5. Medical Auxiliary Diagnostic Software Field

[0084] This invention can be used as a core segmentation module in medical image-assisted diagnostic software.

[0085] In this application, the software system receives image data from endoscopic equipment, image archiving systems, pathology scanners, or other medical imaging equipment, calls the model of this invention to output the target region segmentation results, and further calculates quantitative indicators such as lesion area, diameter, shape, boundary irregularity, number, and location, providing doctors with visual references and structured analysis results.

[0086] 6. Surgical planning and intraoperative navigation

[0087] This invention can be applied to preoperative surgical planning, intraoperative navigation, and minimally invasive surgical assistance systems.

[0088] In this application, the present invention can automatically segment lesions, organ boundaries, dangerous tissues or surgical target areas to form visual contours or three-dimensional reconstruction inputs, which can be used to assist doctors in determining the resection range, puncture path, ablation area or surgical boundary, and improve the accuracy of target localization in complex backgrounds.

[0089] 7. Radiotherapy, interventional therapy, and efficacy evaluation

[0090] This invention can be applied to tumor radiotherapy target delineation, interventional therapy localization, and pre- and post-treatment image comparison analysis.

[0091] In this application, the present invention can automatically or semi-automatically segment tumor areas or organ danger zones, assisting doctors in target area contour drawing, dose planning, lesion volume measurement, and efficacy follow-up evaluation, reducing the burden of manual delineation and improving repeatability.

[0092] 8. Embedded intelligent analysis in medical devices

[0093] This invention can be embedded in endoscopy systems, ultrasound equipment, dermatoscopes, pathology scanning equipment, imaging workstations, or edge computing terminals.

[0094] In this application, the model can run in real-time or near real-time on the device or server side, automatically segment the acquired images, and display the target region on the device interface in the form of color overlay, outline, heat map, probability map or confidence level prompt, for real-time auxiliary observation during clinical examination.

[0095] 9. Medical imaging cloud platforms and telemedicine

[0096] This invention can be applied to cloud-based medical image analysis platforms, remote consultation platforms, and regional medical data centers.

[0097] In this application, after hospitals or primary healthcare institutions upload medical images, the model of this invention, deployed in the cloud, automatically completes target segmentation and quantitative analysis, and generates structured results for remote doctors to review, thereby improving the efficiency of medical image processing and the auxiliary diagnostic capabilities of primary healthcare.

[0098] 10. Medical data annotation and scientific data processing

[0099] This invention can be applied to the construction of medical image datasets, data pre-annotation before training artificial intelligence models, and batch analysis of scientific research images.

[0100] In this application, the present invention can automatically generate an initial segmentation mask, which can be corrected by doctors or annotators, thereby reducing the workload of manual annotation and improving annotation consistency; at the same time, it can be used for lesion statistics, morphological analysis and model performance evaluation in large-scale medical image data.

[0101] 11. Cross-domain medical image segmentation and generalization applications

[0102] This invention is particularly applicable to medical image segmentation tasks where there are distributional differences between different hospitals, different equipment, different imaging parameters, different lighting conditions, or different patient groups.

[0103] In this application, the present invention improves the stability and generalization ability of the model in cross-device, cross-center, and cross-dataset environments by enhancing the separation of foreground and background features and suppressing pseudo-uncertainty of similar backgrounds. It is suitable for deployment in multi-center clinical systems or general medical image analysis platforms.

[0104] 12. Segmentation of targets with similar non-medical backgrounds

[0105] In addition to medical images, the technical concept of this invention can also be extended to general image segmentation tasks with similar background interference, such as industrial defect detection, remote sensing target extraction, crop lesion identification, microscopic image analysis, and material surface defect detection.

[0106] In this application, the present invention can be used to distinguish target regions that are highly similar to the background texture, color or shape, reduce false detection and false negative detection, and improve the reliability of target segmentation in complex scenes.

[0107] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0108] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds, characterized in that, Includes the following steps: Medical image segmentation models acquire multi-scale feature maps of medical images through feature extraction networks; The discriminative query network generates global semantic labels based on multi-scale feature maps, initializes foreground and background query vectors by combining foreground and background semantic information, and fuses and updates query responses obtained at different scales to obtain enhanced multi-scale feature maps. The enhanced multi-scale feature map is used by the decoder to generate an initial segmentation probability map. The features of the initial segmentation probability map are then corrected using a category-aware feature propagation module to complete image segmentation.

2. The discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds according to claim 1, characterized in that, The medical images include endoscopic images, dermoscopic images, pathological images, CT images, MRI images, ultrasound images, or other medical images; after preprocessing, the medical images are fed into a feature extraction network; Preprocessing includes normalization, specifically pixel normalization.

3. The discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds according to claim 1, characterized in that, Foreground query vectors are used to actively find the most discriminative regions in lesions or target tissues, while background query vectors are used to model interference regions that are similar to lesions but actually belong to the background.

4. The discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds according to claim 1, characterized in that, The foreground query vector and background query vector predict reference positions and local offset positions on feature maps of different scales. They also actively collect the most relevant local region features to the query through deformable sampling. This allows the medical segmentation model to adaptively select regions of interest based on the shape, position, and structural changes of the target, without relying on fixed convolutional windows or static attention weights.

5. The discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds according to claim 1, characterized in that, The updated foreground query vector is used to enhance the response of the lesion area or target tissue area, while the updated background query vector is used to suppress the response of similar backgrounds, pseudo-lesion areas, or other non-target areas.

6. The discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds according to claim 1, characterized in that, The initial segmentation probability map features are corrected using a category-aware feature propagation module. Specifically, the pixels are divided into regions based on the initial segmentation probability map, and spatial-semantic associations are established between uncertain pixels and high-confidence pixels. During feature propagation, uncertain background regions receive information from high-confidence background regions; the uncertain prediction results are then updated with weights based on the propagation weights.

7. The discriminative and deterministic enhancement method for medical image segmentation with complex backgrounds according to claim 1, characterized in that, It also includes optimizing the training of the medical image segmentation model, specifically: the training loss includes basic segmentation loss, query diversity constraint, category-aware propagation loss, and category protection loss; among which, the basic segmentation loss is used to ensure that the predicted segmentation result is consistent with the manually labeled result; the query diversity constraint is used to avoid multiple queries learning the same attention region, so that different queries can cover different discriminative structures; the category-aware propagation loss is used to guide uncertain regions to move closer to high-confidence regions of the same category; and the category protection loss is used to prevent errors from propagating across categories.

8. A discriminative and deterministic enhancement system for medical image segmentation with complex backgrounds, characterized in that, include: Image input and preprocessing module: used to receive medical image data and preprocess the medical image data; Multi-scale feature extraction module: used to extract multi-level features from the input image; Discriminative Query Network Module: Used to proactively discover semantic anchors with discriminative capabilities, including query initialization, deformable sampling, query updating, and feature enhancement; Multi-scale fusion and initial segmentation module: used to upsample and fuse the enhanced features output by DQN to obtain an initial segmentation probability map corresponding to the size of the input image; Category-aware feature propagation module (CFPM): used to correct uncertain regions in the initial segmentation probability map; Discriminative optimization module: used to constrain the model to learn more stable foreground and background representations; Segmentation Result Output Module: Used to output the final segmentation result, which can be in the form of a segmentation probability map, a binary mask, a target boundary, a color overlay map, a confidence map, or a structured measurement result.

Images