DR equipment examination site AI automatic identification system based on image quality correction

Abstract

The present application relates to the field of digital medical image processing, in particular to a DR equipment examination site AI automatic identification system based on image quality correction. The system obtains a to-be-identified DR image, generates a quality state graph and a candidate site semantic response graph, and constructs an anatomical discrimination protection graph; according to the quality state graph, the degradation area is located, and correction is performed under the constraint of the anatomical discrimination protection graph; according to the correction residual and the consistent relationship of the anatomical structure before and after correction, a correction credible graph is generated; the corresponding features before and after correction are extracted, and the site response graph is formed by position fusion according to the correction credible graph and the semantic consistent relationship; the anatomical discrimination protection graph is updated according to the coincidence relationship of the site response graph and the correction residual, the correction and feature fusion are re-executed, and the examination site identification result is output. The present application maintains the discriminative anatomical structure when compensating for uneven exposure, noise and blur degradation, reduces the interference of correction pseudo-structure on category calculation, and improves the stability of site identification of low-quality DR images.

Description

Technical Field

[0001] This invention relates to the field of digital medical image processing, specifically to an AI-based automatic identification system for examination sites in DR equipment based on image quality correction. Background Technology

[0002] After exposure, digital radiography equipment typically saves the X-ray images and examination descriptions in a medical digital imaging and communication format. The automatic identification process for the examination site reads the pixel matrix from the image file, unifies the grayscale direction based on photometric interpretation information, and then performs window width and level conversion, grayscale normalization, projection boundary cropping, and size adjustment to ensure that images output from different devices have a unified representation suitable for processing by the classification network. Verified and annotated images of the head, neck, chest, abdomen, pelvis, and limbs are used to train a convolutional neural network. The network extracts image features such as bony contours, organ projection ranges, and tissue textures through layer-by-layer convolution, and then the classification layer forms the category response for each examination site. During deployment, the images to be identified are input into the classification network after undergoing the same preprocessing. Examination sites with higher category responses are identified, and the recognition results are written into the examination record or image description field.

[0003] To mitigate the interference of exposure conditions, background scattering, and detector response differences on part classification, the processing approach used in this project adds image quality correction before convolutional classification. This processing typically involves contrast adjustment based on the overall image's grayscale statistics, noise suppression or edge enhancement in high-frequency areas, and compensation for uneven local brightness. The processed image is then scaled to the input size specified by the classification network. During the training phase, a sample set is established using the uniformly quality-processed image and its part labels. The sample is expanded through rotation, cropping, and grayscale perturbation, and the classification error is used to update the network. While some features are extracted simultaneously from both the pre-corrected and post-corrected images, these features are usually merged using a fixed stitching or overall weighting method, and the quality processing result is still passed as a predetermined input to the part classification stage.

[0004] In existing workflows, image quality correction and examination site identification are independent. Correction relies on quality differences at the grayscale or frequency levels, failing to identify whether the adjusted area contains bony edges, joint spaces, or organ contours that determine the examination site. When local underexposure, scattering, or blurring coincides with the discriminant structure, uniform enhancement may alter edge positions and connectivity. Convolutional classification networks then treat the pseudo-textures created by correction as anatomical evidence. If the classification process preserves the original image features, structures obscured by quality degradation cannot provide stable category responses. The technical problem with existing workflows is that quality correction cannot limit pixel changes based on examination site discrimination information and verify the structural reliability of the correction results. This leads to structurally distorted correction features being included in category calculations, resulting in misidentification of examination sites in low-quality digital radiographic images. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based automatic identification system for inspection areas in DR equipment based on image quality correction, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An AI-based automatic identification system for examination sites in DR equipment based on image quality correction includes an image processor, which performs the following actions: acquiring the DR image to be identified, generating a quality status map and a semantic response map of candidate sites, and constructing an anatomical discrimination and protection map from the semantic response map; The degraded region is located based on the quality status map, and the degraded region is corrected under the constraints of the anatomical discrimination protection map to obtain the corrected image and the correction residual characterizing the pixel difference before and after correction. A correction confidence map is generated based on the correction residual and the consistency relationship of the anatomical structure before and after correction. Extract the spatial correspondence features before and after correction, and fuse them position by position according to the semantic consistency relationship between the correction confidence map and the spatial correspondence features to generate the first part response map; The overlap relationship between the first part response map and the correction residual updates the anatomical discrimination protection map, and the map is corrected and fused again to output the part identification result.

[0007] Preferably, the image processor unifies the grayscale direction and separates the effective illumination area based on the DICOM photometric interpretation information of the DR image to be identified; Within the effective illumination area, a low-frequency gray-level deviation map is formed based on the regional gray-level distribution, and a high-frequency structure attenuation map is formed based on the local gradient continuity. The low-frequency gray-level deviation map and the high-frequency structure attenuation map are aligned according to pixel coordinates, and their joint response is encoded into a quality state map. Pixels outside the effective illumination area are encoded as uncorrected regions.

[0008] Preferably, the image processor extracts the category correlation and spatial response of each candidate part from the DR image to be identified, and performs positional correspondence synthesis of the spatial response of each candidate part according to the category correlation to form a semantic response map of the candidate part; Extract continuous regions along the edge direction from the DR image to be identified, encode pixels that maintain response across candidate parts and are located within continuous regions along the edge direction as non-displaceable regions, and encode candidate-specific response regions as restricted grayscale adjustment regions to form the anatomical discrimination protection map.

[0009] Preferably, the image processor maps the quality status map and the anatomical discrimination protection map to the same pixel coordinates, decomposes the pixels in the degraded region into gray-level base components and edge structure components, and forms gray-level base compensation components based on the quality status map. The grayscale base component after compensation in the non-displaceable region is combined with the edge structure component before correction; the grayscale base component after compensation in the restricted grayscale adjustment region is combined with the edge structure component selected according to the continuity relationship of the edge direction. The combined result is reconstructed into a corrected image according to the original pixel coordinates, and the pixel difference before and after reconstruction is recorded as the correction residual.

[0010] Preferably, the image processor resamples the effective illumination area at both the global and local scales, calculates a low-frequency grayscale deviation map and a high-frequency structure attenuation map at each scale, and then maps each map back to the pixel coordinates of the DR image to be identified. A coupled degradation state is established for locations where low-frequency and high-frequency responses coexist, and a single degradation state is established for locations where only one type of response exists. The quality state map is updated by the spatial set of the two degradation states.

[0011] Preferably, the image processor divides the spatial response of each candidate region into a response connected region, uses the response connected region as a node, and establishes a candidate region response map structure according to the edge direction continuity relationship between adjacent nodes. Extract the common sub-image corresponding to the position in the response map structure of different candidate parts, write the pixels covered by the common sub-image into the non-displaceable region, and write the candidate dedicated node connected to the common sub-image into the restricted grayscale adjustment region.

[0012] Preferably, the image processor generates a corrected quality state map from the corrected image, and compares the quality state map with the corrected quality state map within the spatial support region of the correction residual to form a quality compensation consistency map; Within the area covered by the anatomical discrimination protection map, the edge directions and connectivity before and after correction are compared to form a structure preservation map; The quality compensation consistency map and the structure retention map are intersected by spatial location, and a correction reliability map is generated based on the corresponding range of the intersection and the correction residual.

[0013] Preferably, the image processor forms the original domain features of the DR image to be identified and the correction domain features of the corrected image through a shared feature extraction network, maps the corrected confidence map to the corresponding levels of the two domain features, and generates the response order of candidate parts according to the two domain features respectively. At the locations covered by the corrected confidence map and where the response order of the two candidate regions is consistent, the corrected domain features are selected. At the remaining locations, the original domain features are selected, and the selected features are spliced ​​together according to the feature continuity relationship of adjacent locations to form fused features.

[0014] Preferably, the image processor generates a first part response map based on the fusion features, and maps the first part response map and the correction residual to the same spatial scale; The spatial support region of the correction residual is intersected with the connected edge of the response map of the first part, and the positions where the response order of the two candidate parts of the intersection is different are extracted as correction conflict regions. The corrected conflict region is written into the anatomical discriminant protection map, and pixel correction is re-performed only on the corrected conflict region and its continuous boundaries to generate an updated corrected image and an updated correction residual.

[0015] Preferably, the image processor performs positional correspondence calculations on the updated correction residuals and the updated anatomical discriminant protection map, retains the updated correction residuals where the support region is located in the degraded region and does not intersect with the boundary of the immovable region as a quality compensation component, and marks the updated correction residuals where the support region intersects with the boundary of the immovable region as a structural conflict component. The quality compensation component is superimposed on the DR image to be identified to form a recorrected image. The original domain features are used at the location covered by the structural conflict component to re-form the fusion features and output the location identification result.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By generating a quality status map and a semantic response map of candidate sites from the image to be identified, the system can establish a correspondence between quality degradation locations and site discrimination structures on pixel coordinates, and form a discrimination protection map by continuous anatomical regions that maintain response across candidates. Quality correction locates degradation regions according to the quality status map, and is constrained by the discrimination protection map. Within the protected region, the compensated grayscale base and the original edge structure are combined, while in other degradation regions, edge structures filtered by continuity are used. The changes in quality status before and after correction and the consistency with anatomical structures together form a correction credibility map, allowing credible locations to use correction domain features, while locations with structural deviations retain the original domain features; the overlapping areas of site response and correction residuals are then written back into the discrimination protection map. The resulting closed-loop processing reduces the loss of anatomical evidence caused by quality degradation and suppresses the entry of correction pseudostructures into the examination site classification.

[0017] 2. Gray-level deviation and structural attenuation information at both global and local scales are mapped back to the original image coordinates, distinguishing between coupled degradation states and single degradation states at the same location. Local quality compensation can be performed according to the corresponding state. The spatial response of candidate regions is converted into a connected graph structure. The common sub-image corresponding to the location in different candidates is written into the immovable region, and the candidate-specific nodes connected to the common sub-image are written into the restricted gray-level adjustment region, so that the common anatomical structure is preserved even when the initial category is not yet determined. The corrected quality state and the original quality state form a quality compensation consistency map, and the edge relationship before and after correction forms a structure preservation map. The spatial intersection of the two maps limits the selection range of the correction domain features. The location where the correction residual intersects with the connected edge of the region response and the category order changes is identified as the correction conflict region. Only the conflict region and its continuous boundary are corrected again, reducing the rewriting of duplicate pixels in other regions. Attached Figure Description

[0018] Figure 1 Flowchart for automatic identification of inspection areas in a closed-loop process; Figure 2 Flowchart for generating credibility graphs for quality status analysis and correction; Figure 3 A graph showing the relationship between the overall strength of the quality status and the contributions of low and high frequencies; Figure 4 To correct the relationship between the confidence strength and the structural retention state; Figure 5 The graph shows the variation of the dual-domain feature fusion response with trust tendency. Detailed Implementation

[0019] Please refer to Figure 1 In one embodiment, an AI-based automatic identification system for examination sites in DR equipment based on image quality correction is provided. The system includes an image processor that performs the following processing: acquiring a DR image to be identified, generating a quality status map and a semantic response map of candidate sites, and constructing an anatomical discrimination protection map from the semantic response map; locating degraded areas based on the quality status map, correcting the degraded areas under the constraints of the anatomical discrimination protection map, obtaining a corrected image and a correction residual characterizing the pixel differences before and after correction, and generating a correction confidence map based on the consistency relationship between the correction residual and the anatomical structures before and after correction; extracting spatial correspondence features before and after correction, fusing them position by position according to the semantic consistency relationship between the correction confidence map and the spatial correspondence features, and generating a first site response map; updating the anatomical discrimination protection map based on the overlap relationship between the first site response map and the correction residual, correcting and fusing again, and outputting the site identification result. All of the aforementioned processing objects use a spatial index that can be mapped to the pixel coordinates of the DR image to be identified, so that the quality degradation location, anatomical discrimination location, correction change location, and classification response location can be associated position by position.

[0020] In this embodiment, the image processor uses the original pixel matrix of the image to be identified as a common data source and establishes the coordinate relationship between the original image object, the spatial map object, and the feature object in memory. The original image object stores the grayscale values ​​that have not been changed by quality correction. The spatial map object stores the quality status, semantic response, protection type, and correction confidence status according to the original pixel position. The feature object stores the original image coverage area corresponding to the position of each convolution feature layer. When the image is cropped, filled, scaled, or feature downsampled, the forward coordinate transformation and the inverse coordinate transformation are recorded simultaneously. Any spatial map is aggregated according to the forward transformation before entering the feature layer, and any category response is unfolded according to the inverse transformation before returning to the original image. The correction process always forms a new image copy in the original image coordinates without covering the original pixels, thereby giving the subsequent residual calculation a definite reference object and enabling each fused feature to be traced back to the corresponding pixel area before and after correction.

[0021] Specifically, the image processor reads pixel representations, photometric interpretations, grayscale transformation relationships, image dimensions, and examination descriptions from medical digital imaging and communication files. It converts stored pixels into processed grayscale values ​​according to the grayscale transformation relationships. When the photometric interpretation indicates that a smaller grayscale value corresponds to a brighter display, the grayscale order is reversed according to the available grayscale range in the file. When the photometric interpretation indicates that a larger grayscale value corresponds to a brighter display, the grayscale order is preserved. The conversion process preserves the mapping between stored values ​​and processed grayscale values. The display window is only used to form an image for the operator to observe and does not replace quality analysis pixels with pixels truncated by the display window. For images carrying image orientation information, an orientation transformation record is first established, and then the classification response and quality space map are unified to the coordinate orientation of the processed grayscale matrix to prevent mismatch between semantic response and correction position caused by image flipping or rotation.

[0022] Furthermore, the image processor tracks the gray-level stable band from the periphery of the image inwards, and combines the texture changes and connectivity within and outside the stable band to form collimation boundary candidates. Continuous tracking is performed on the boundary candidates, and the side that continuously contains the texture of the human projection is determined as the effective side. Then, connected regions belonging to the same human projection are merged and internal holes formed by local occlusion are filled. Blank areas at the outer edge of the image, collimation black edges, character areas not connected to the human projection, and filled areas are excluded to obtain an effective projection area mask. For non-rectangular projection areas, the gray-level sequences on both sides are compared along the local normal of the boundary, and the side that still retains the directional continuous edge after the gray-level abrupt change is included in the mask. For lead type markings close to the human contour, a marking mask is formed based on the relationship between the high gray-level closed contour and the lack of continuous connection with the surrounding anatomical edges, and the corresponding position is set to a non-participation state during quality state analysis and candidate semantic response aggregation.

[0023] During basic grayscale processing, the image processor establishes a sorting mapping for pixels in the effective illumination area. The main body of the sorted distribution, continuously occupying the grayscale range, is used as the mapping benchmark. The relative order of each pixel in the main distribution is transformed to a unified processing numerical domain, and low-end and high-end truncation markers are saved separately. When continuous saturated regions exist at the low or high end, these regions are not used as neighborhood grayscale references, but rather the grayscale trends of adjacent unsaturated regions are written into the quality status analysis input. For images with significant differences in grayscale range within the same inspection, the image processor only unifies the numerical expression range without changing the relative order of grayscale within each image. When cropping or padding to form classification input, the padding pixels carry independent mask values. The feature extraction network ignores the padding pixels during convolution response aggregation, and the inverse coordinate mapping does not expand the padding region to the original image, maintaining the coordinate closure of pixel correction, semantic response, and residual calculation.

[0024] The local reference used in quality analysis is formed by the adjacent non-degraded regions of the image to be identified and the offline site-independent quality distribution. The offline quality distribution only records the range of change of continuous gray-scale base in the X-ray image, the response relationship of continuous edges in the direction, and the statistical morphology of random discrete details, without recording the patient's identity or the shape of the specific examination site. The image processor first determines the direction of degradation change based on the adjacent regions of the image itself. When the adjacent regions are insufficient or truncated by collimation boundaries, the offline quality distribution is called to supplement the response range. Then, the reference propagation is limited by the edge connectivity of the original image. The reference value is transmitted along the connected regions of the same tissue and stops when it encounters bony edges, organ projection boundaries, or protective map preparation boundaries. This ensures that the natural gray-scale transition caused by the thickness of the human body is not rewritten by cross-regional references, and that saturated or blurred regions that cannot be directly judged from a single image still have an executable quality status coding basis.

[0025] Please refer to Figure 2 When generating the quality state map, the image processor constructs a low-frequency grayscale reference plane that preserves the main anatomical edges within the effective illumination area. A low-frequency grayscale deviation map is formed by the deviation of local grayscale statistics relative to the reference plane. At the same time, gradient magnitude, gradient direction, and directional continuity are extracted from the detail components after subtracting the low-frequency grayscale reference plane. A high-frequency structural attenuation map is formed by the attenuation of local directional response relative to adjacent continuous edges. The low-frequency grayscale deviation map records slowly changing exposure deviation and scattering background, while the high-frequency structural attenuation map records blurred, missing details, and discrete responses that cannot form continuous edges. The image processor maps the two images to the same pixel coordinates, encodes the low-frequency and high-frequency responses at the same location separately, and combines their spatial coexistence relationship to form the quality state map. Each element in the quality state map contains at least the degradation source, degradation direction, and degradation spatial support region.

[0026] When forming a low-frequency grayscale reference surface, the image processor uses the response connected regions in the effective illumination area as statistical boundaries. Within each statistical boundary, a local distribution is established according to the grayscale order, and adjacent regions with the same edge enclosure relationship are used as reference regions. If the grayscale order between the region to be analyzed and the reference region is maintained but the overall position is shifted, the shift is recorded as a base deviation. If the grayscale order is compressed at the same time, the change in the distribution span is recorded as a dynamic range deviation. For the inherent grayscale difference caused by the overlap of bone and soft tissue, the image processor checks whether the difference boundary coincides with the continuous bony edge. The overlapping part is retained in the tissue base, and the slow shift that does not coincide and crosses multiple adjacent statistical boundaries is entered into the low-frequency grayscale deviation map. The local statistical results are also compared with the grayscale order relationship of the entire effective illumination area, so that the large-scale scattering background of the chest and the local exposure shift of the small illumination range of the limbs adopt their own spatial references instead of directly mixing the absolute grayscale values.

[0027] The semantic response map of the candidate site is generated by the initial site recognition network, which adopts a multi-scale convolutional coding structure. The feature layer near the input side preserves the edge of the bone cortex, the boundary of the joint space, and the projection boundary. The middle feature layer expresses the spatial relationship between adjacent bony structures and soft tissue projections. The feature layer near the output side expresses the overall contour and anatomical structure combination of the examination site. The classification mapping layer forms the category association degree of each examination site based on the output side features, and projects the channel contribution corresponding to each category to the middle feature layer and the input pixel coordinates layer by layer to form the spatial response of each candidate site. The image processor retains the site responses whose category association degree is within the candidate set and has a competitive relationship. It performs positional correspondence synthesis according to the normalization result of the category association degree to obtain the semantic response map of the candidate site that simultaneously represents multiple reasonable site hypotheses when the category is not yet determined.

[0028] In this embodiment, the semantic response map of the candidate region does not directly cover the other candidate responses with the highest category response. The image processor saves the response channel of each candidate region separately, and then checks whether the response has anatomical edge support on the gradient direction field of the original image. When the coarse-grained response in the classification feature covers multiple unconnected regions after being mapped back to the original image, the continuous region is expanded along the gradient direction starting from the response center, and only the response part that can be connected to the human body projection body is retained. When the response region is located near the collimation boundary or character mark, the non-anatomical part is cut off according to the effective projection area mask and the mark mask. Each candidate response after processing saves the category source, response strength and connectivity according to the original pixel coordinates, so that the subsequent protection map can identify the structure that different candidates depend on, as well as the differential structure used only by some candidates.

[0029] In a preferred embodiment, the candidate set is determined by the relative order of category responses and the response interval. The image processor first converts the output of the classification mapping layer into a category correlation degree with a fixed sum, and then arranges them from high to low correlation degree. Categories that can still form independent response connected regions on the original image are retained, while categories whose responses fall entirely into higher-order categories and lack independent edge support are merged. When there is competition between adjacent categories between the chest and thoracic vertebrae, pelvis and lumbar vertebrae, or forearm and wrist, the candidate set simultaneously saves the adjacent categories and generates spatial response channels respectively, without prematurely excluding candidates based on checking the description field or the highest single response. The category correlation degree only controls the relative contribution of candidate responses to location synthesis. The anatomical discrimination protection type is still determined by the cross-candidate common response and the edge continuity relationship of the original image. Therefore, the incorrect initial first category will not independently determine the protected pixel range.

[0030] Furthermore, when the semantic response returns to pixel coordinates from feature coordinates, the image processor reads the original image coverage area corresponding to each layer feature position, allocates the deep category contribution to the connected region of the middle layer response, and then uses the shallow gradient direction to limit the middle layer response to both sides of the actual edge. In the case where a deep position covers multiple shallow edges, the category contribution is allocated according to the connected path between each shallow edge and the adjacent deep response center. In the case where the same shallow edge receives multiple candidate category contributions at the same time, the independent channels of each category are retained and the contribution source is recorded in the composite image. After inverse mapping, connected merging is performed on adjacent pixels. Pixels located in the same edge chain and with the same category source form a semantic response fragment. The mapped part that crosses the collimation boundary or lacks the original image gradient support is discarded, so that the spatial accuracy of the candidate semantic response is determined by the original image structure and is not limited by the deep feature resolution.

[0031] In a preferred embodiment, the degree of quality degradation is determined according to the following formula: ; Where x and y represent the pixel coordinates in the image to be identified. This represents the deviation of low-frequency grayscale values ​​after normalization. This represents the normalized high-frequency structure attenuation. This represents the fusion coefficient obtained based on the continuity of low-frequency responses in adjacent regions, with values ​​ranging from 0 to 1. The plus sign indicates a weighted combination of low-frequency and high-frequency terms, and the multiplication sign indicates the combination of the corresponding response and the fusion coefficient. The complementary coefficients of high-frequency terms are represented. This indicates the overall degradation level at the corresponding location. For example, if the low-frequency grayscale deviation at a certain location is 0.8, the high-frequency structural attenuation is 0.4, and the fusion coefficient is 0.75, the overall degradation level is 0.7. The corresponding location is preferentially included in the grayscale base compensation area relative to the surrounding low-degradation locations, while the structural attenuation information expressed by 0.4 is retained for edge structural component filtering.

[0032] Please refer to Figure 3 , Figure 3 The horizontal axis represents the sampling phase, arranged according to the continuous windows within the effective illumination area from the beginning, front, middle, and end; the vertical axis represents the overall intensity, ranging from approximately zero to 0.7. The solid black line indicates... The corresponding overall strength is represented by the long dashed dot, indicating the stress. The modulated low-frequency contribution is represented by a dotted-line square. The relevant high-frequency contributions are represented by dotted triangles and long and short rhombuses, which are the first and second thresholds, respectively. The curves show... The overall intensity rises as sampling progresses, approaching the second threshold in the later stages, indicating that grayscale base shift and structural attenuation are enhanced together in this region. The low-frequency contribution remains consistently low, suggesting that simple grayscale changes are insufficient to form a high-risk judgment; the high-frequency contribution is closer to the overall intensity trend, indicating that the decline in edge continuity plays a more significant role in the overall judgment.

[0033] When constructing the anatomical discrimination and protection map, the image processor divides the spatial response of each candidate site into a connected response region. This connected region serves as a node, and connecting edges are established based on the presence of directionally continuous original image edges between adjacent nodes. Each node records the response coverage, candidate category, and internal edge direction distribution, while the connecting edges record the pixel paths traversed and their directional changes. The image processor performs positional matching between the response map structures of different candidate sites. Starting from nodes with intersecting coverage and compatible internal edge directions, it synchronously expands along the connecting edges, identifying graph structures among multiple candidates that maintain the same connection order and spatial orientation as common subgraphs. Pixels covered by the common subgraph are written into immovable regions, nodes connected to the common subgraph but belonging only to some candidates are written into restricted grayscale adjustment regions, and the remaining regions marked by the quality status map are written into correctable regions. These three types of regions are combined according to their original pixel coordinates to form the anatomical discrimination and protection map. Continuous edge-direction regions are extracted from the DR image to be identified, and pixels that maintain response across candidate sites and are located within continuous edge-direction regions are encoded as immovable regions. During pixel correction, the image processor uses edge-preserving grayscale decomposition to divide degraded region pixels into grayscale base components and edge structure components. The grayscale base component records the slow grayscale changes and low-frequency deviations formed by tissue projection, while the edge structure component records the directional changes left after subtracting the grayscale base from the original image. The image processor forms a grayscale base compensation component based on the low-frequency deviation direction, the coverage area of ​​the degraded region, and the grayscale continuity relationship of adjacent non-degraded regions. In non-displaceable regions, the compensated grayscale base component is combined with the original edge structure component. In restricted grayscale adjustment regions, the compensated grayscale base component is combined with the edge structure component selected according to the continuity relationship of edge direction. In correctable regions, discrete components that cannot form a connection with the response of adjacent directions are removed or structural components that continue along the existing edge path are added based on the high-frequency structural attenuation state. The combined result is reconstructed into a corrected image according to the original pixel coordinates.

[0034] After the corrected image is formed, the image processor calculates the signed difference between the corrected image and the image to be identified pixel by pixel and records it as the correction residual. The positive and negative directions represent the grayscale compensation directions, respectively. Residuals that are continuous and have the same direction of change are merged into the residual spatial support region. The image processor uses the same effective projection area mask and quality resolution process as the image to be identified to generate a corrected quality state map. In the spatial support region of the correction residual, the degradation direction and degradation connectivity range before and after correction are compared to form quality compensation consistency information. At the same time, in the anatomical discriminant protection map coverage area, the edge direction and node connection relationship before and after correction are compared to form structural consistency information. The positions where both quality compensation and structural consistency are valid are encoded as reliable correction positions. The positions where the residual crosses the boundary of the immovable region or changes the connection relationship of the common subgraph are encoded as structural conflict positions. The remaining positions are retained as transition positions according to the state of the adjacent region, thus forming a correction reliability map.

[0035] When extracting spatial correspondence features before and after correction, the image processor inputs the image to be identified and the corrected image into a feature extraction network with shared parameters. The two input branches use the same convolution transformation, the same channel projection, and the same scale transformation. Therefore, each original domain feature position has a corresponding corrected domain feature position. The image processor maps the corrected confidence map to each feature layer and generates the candidate site response order from the original domain features and the corrected domain features respectively. Corrected domain features are selected at positions covered by the corrected confidence map and where the original domain and the corrected domain candidate site response order are consistent. Original domain features are selected at positions with structural conflicts or different candidate orders. At transitional positions, the two domain features are combined according to the continuity relationship of adjacent features. The features selected at each position are then spliced ​​together into a fusion feature according to the continuous direction of the anatomical edge. The classification mapping layer generates the first site response map based on the fusion feature.

[0036] During feedback processing, the image processor maps the response map of the first part and the correction residual to the same spatial scale, extracts the continuous response connectivity edges along the gradient direction of the original image in the response map of the first part, intersects the spatial support region of the correction residual with the response connectivity edges, and then extracts the positions where the response order of the candidate parts in the original domain and the correction domain is different from the intersection as correction conflict regions. The correction conflict regions are written into the immovable region of the anatomical discrimination protection map, and the continuous boundary adjacent to them is written into the restricted grayscale adjustment region. The image processor only re-extracts the grayscale base component and edge structure component for the correction conflict regions and continuous boundaries, and re-corrects them using the unmodified correction image pixels outside the regions as boundary conditions to obtain the updated correction image and updated correction residual, while the correction pixels at other positions remain unchanged.

[0037] During recalibration, the image processor performs positional correspondence calculations between the updated calibration residuals and the updated anatomical discriminant preservation map. The updated calibration residuals where the support region is located in the degraded region and does not intersect with the boundary of the immovable region are retained as quality compensation components. The updated calibration residuals where the support region intersects with the boundary of the immovable region are marked as structural conflict components. The quality compensation components are superimposed onto the image to be identified according to the original pixel coordinates to form a recalibrated image. The structural conflict component coverage position maintains the pixels of the image to be identified and uses the original domain features. The remaining reliable positions in the recalibrated image use recalibrated domain features. The features at each position are re-fused according to the continuity relationship of adjacent features to form the part identification result. Feedback ends when the update processing does not form a new calibration conflict region and the candidate part response order is maintained. If there is still an order change caused by the updated calibration residuals, the corresponding region continues to be processed. The original image remains unchanged throughout the process.

[0038] In a preferred embodiment, the image processor performs multi-scale analysis on the low-frequency grayscale distribution of the effective projection area. The global scale preserves the human projection contour and cross-regional grayscale changes, while the local scale preserves the grayscale transition of tissues around the bony edges. Each scale uses edge-preserving smoothing operations to form a grayscale reference plane, and then writes the deviation direction and continuous range of the local median grayscale relative to the grayscale reference plane into the low-frequency grayscale deviation map. For grayscale transitions that change slowly with the thickness of the human body and do not cross anatomical boundaries, they are preserved as tissue bases according to the edge direction of adjacent regions. For slow changes that cross edges and are inconsistent with the grayscale trend of the same connected tissue region, they are written into the degradation response. The results of each scale are mapped back to the original image coordinates using the pixel center correspondence before and after resampling. The positions covered by multiple scales are merged according to spatial continuity, rather than directly determining the correction range based on a single local extremum.

[0039] Preferably, the high-frequency structural attenuation map is generated from the detailed components after removing the grayscale reference plane. The image processor calculates the grayscale changes at each location along multiple directions and forms directional responses. For responses at adjacent locations that are close in direction and connected end to end, a chain connection is performed. The continuity of the connection path and the response changes describe the bone cortex edge, joint space edge, and organ projection boundary. When the local directional response decreases but there are still continuous edges in the same direction in its front and rear neighborhoods, the corresponding location is written into the structural attenuation region. When the high-frequency response increases but the directional distribution is discrete and cannot be connected to the existing edge chain, the corresponding location is written into the noise interference region. The image processor saves the structural attenuation region and the noise interference region separately. When forming edge structural components, only the former is supplemented with structural responses along the existing edge path, and the latter is deleting discrete components that lack directional continuity.

[0040] Furthermore, the image processor maps the low-frequency grayscale deviation map and high-frequency structural attenuation map at each scale back to the coordinates of the image to be identified. For cases where low-frequency and high-frequency responses coexist in the same location or the same continuous anatomical region, a coupled degradation state is established; for cases where only one type of response exists, a single degradation state is established. Grayscale floor compensation and edge structure processing in the coupled degradation state are performed according to the same spatial support region. The floor compensation result serves as the grayscale reference for structural response screening. The single low-frequency degradation state only forms grayscale floor compensation, and the single high-frequency degradation state only processes the edge structure component. When scale mapping causes multiple sampling positions to fall into the same original pixel, degradation descriptions that have spatial intersection with the anatomical discriminant protection map are preferentially retained, and the remaining descriptions are saved as neighborhood continuous information, so that the quality state map can distinguish different degradation components under the same comprehensive degradation degree.

[0041] In the correction process of coupled degradation regions, the image processor first forms grayscale base compensation candidates according to the low-frequency deviation direction, and then recalculates the directional responses on both sides of the edge in the local grayscale environment formed by the compensation candidates. If the original high-frequency structure attenuation position can be connected with the preceding and following edge segments, the edge structure components are reconstructed along the connection path. If the compensated grayscale environment causes the original edge segment to lose its directional continuity with the neighborhood, the edge structure components before correction are retained and the corresponding positions are recorded as the correction positions to be verified. In a single high-frequency degradation region, the spatial trend of the grayscale base is not changed, and the structural components are processed only according to the original image edge path. In a single low-frequency degradation region, no new edge response is generated, and the grayscale base compensation amount is combined with the original structural components. The coupling state thus expresses the processing order and mutual constraints of the two degradation factors in the same anatomical region, rather than simply adding the two quality images together and then uniformly enhancing them.

[0042] In a preferred embodiment, the confidence level of the correction is calculated according to the following formula: ; in, Represents pixel coordinates The reliability of the correction at that point This indicates the consistency of quality compensation between the changes in quality status before and after correction and the original direction of degradation. This indicates the structural preservation degree of edge direction and connectivity before and after correction. Both quantities are normalized to 0 to 1, and the multiplication sign indicates the intersection constraint of the two types of conditions. For example, when the quality compensation consistency degree is 0.9 and the structural preservation degree is 0.8, the correction confidence degree is 0.72, and the corresponding position can enter the candidate range of correction domain features. However, when the quality compensation consistency degree is still 0.9 but the structural preservation degree is 0.2, the correction confidence degree drops to 0.18, and the corresponding position retains the original domain features. The above value relationship ensures that the position where the grayscale quality is compensated but the anatomical structure deviates will not dominate the part classification simply because of the improved quality state.

[0043] Please refer to Figure 4 , Figure 4 The key focus is on presenting the threshold discrimination relationship for the calibration confidence strength. The horizontal axis represents the degree of consistency, expanded in the order of very low, low, medium, high, and very high; the vertical axis represents the confidence strength, with values ​​ranging from [0,1]. Solid black lines correspond to a stable structural state, long dashed dots correspond to a weaker structural state, dotted-dash squares correspond to a medium structural state, dotted-line triangles represent transition thresholds, and long and short dashed diamonds represent adoption thresholds. It can be seen that in... When kept at a high level, Will follow The improvement allows for a rapid crossing of the threshold, enabling the position to enter the feature candidate range of the correction domain; when... When insufficient, although If the curve continues to increase but remains below the threshold for an extended period, it indicates that the direction of quality compensation and the maintenance of the anatomical edge must both be valid.

[0044] In this embodiment, the quality compensation consistency is obtained by comparing the quality state maps before and after correction position by position within the support area of ​​the correction residual space. The image processor checks whether the original low-frequency deviation converges in the opposite direction, whether the original high-frequency structure attenuation recovers along the existing edge path, and whether the degradation connectivity range remains within the original degradation area. Positions that meet the above correspondence are written into the quality compensation consistency map. The structure retention is obtained by comparing the edge structures in the coverage area of ​​the dissection discrimination protection map. The image processor finds the corresponding path after correction for each edge path before correction, compares the path start and end positions, direction change order, and node connection relationship, and writes the positions that are still within the original edge coverage area and maintain the connection order into the structure retention map. Positions that are broken at the edge, cross other nodes, or have undergone position migration within the common subgraph are written into the structure conflict area. The spatial intersection of the quality compensation consistency map and the structure retention map forms the correction reliability map.

[0045] Specifically, the quality compensation consistency map uses directed comparison rather than simply comparing the difference before and after correction. When the low-frequency gray-level deviation is too high, the corrected local gray-level distribution should approach the reference distribution of the same connected tissue region along the decreasing direction. When the low-frequency gray-level deviation is too low, it should approach the reference distribution along the increasing direction. The correction response of the high-frequency structural attenuation position should extend along the direction of the edge before correction and establish a continuous connection between adjacent unattenuated segments. If the quality value changes but the direction of change is the same as the original degradation direction, or the range of change exceeds the original degradation connected region, the corresponding position is not written into the quality compensation consistency region. Different consistency states are given for different scales. The image processor checks the spatial continuity along the edge direction of the preservation map, retains the state that matches the continuous degradation region, and records the scale conflict position as the transition position.

[0046] The structure-preserving map is jointly determined by the edge path and the candidate response map structure. The image processor sets sampling positions in spatial order on the edge path before correction, and then finds corresponding positions with compatible directions and overlapping coverage in the corrected image, and connects them in the original order to form the corrected path. When the corresponding position exists but the order is changed, the path is recorded as a reversal. When the corresponding position is missing, the path is recorded as an interruption. When a new edge appears after correction that cannot be mapped to the original path and is connected to the candidate response in a common sub-map, the path is recorded as a bridging. Positions where the path maintains the original order and the node connection belongs to the structure-preserving region are included in the structure-preserving region. Reversal, interruption and bridging positions are included in the structure conflict region. Positions in the restricted grayscale adjustment region where only the response intensity changes and the path order remains unchanged are still included in the structure-preserving region, so that the topological relationship on which the structure preservation judgment and part recognition depend is consistent.

[0047] In another preferred embodiment, the candidate site response map structure is established according to connected regions. Nodes store the response source, coverage area, and internal edge direction distribution. Connecting edges are generated based on the continuous edge direction relationship between adjacent nodes. The graph structures of different candidates are matched with corresponding nodes through the original image coordinates and synchronously expand along the connecting edges starting from the common response area. When the expansion path maintains the same spatial orientation and connection order in each candidate, a common subgraph is formed. When a node appears only in some candidates but is directly connected to the common subgraph, a candidate-specific node is formed. Pixels covered by the common subgraph are written into the non-displaceable region, and pixels covered by the candidate-specific node are written into the restricted grayscale adjustment region. The former uses the edge structure components before correction during correction, while the latter allows changes in edge response strength but restricts edge migration through the coverage range of the original connecting edges. When there is competition in the initial site category, the anatomical topology that multiple candidates commonly depend on is maintained.

[0048] Specifically, the shared feature extraction network employs hierarchical encoding and cross-layer feature mapping. The image to be identified and the corrected image share the same convolutional kernel, channel projection relationship, and classification mapping relationship in two forward processing steps, thus the same channel represents the same category of related structures. Each feature location is associated with the original image coverage area through coordinate records. When the corrected credibility map is mapped to the feature layer, a correction domain selection state is formed for locations where the entire coverage area is located in the quality compensation consistent area and does not cross the structure preservation boundary. An original domain selection state is formed for locations where the coverage area contains structural conflict areas. For locations that simultaneously contain credibility areas and transition areas, credibility information is aggregated after being separated along the edge direction of the original image. The original domain and the corrected domain are respectively processed by the same category response mapping layer to generate the candidate part response order. The order comparison is limited to the same location, the same scale, and the same candidate set, and the case of equal response is handled by class assignment of adjacent response connected regions.

[0049] Furthermore, when the calibration confidence map is mapped from pixel coordinates to feature coordinates, the image processor reads the confidence state covered by the receptive region of each feature location and maintains the priority of non-displaceable regions. When the entire receptive region falls into the intersection of the quality compensation consistency map and the structure preservation map, the corresponding feature location is marked as a calibration domain candidate. When the receptive region intersects with the structural conflict region, the corresponding feature location is marked as an original domain candidate. When the receptive region spans two states but the structural conflict part does not coincide with the candidate semantic response, the corresponding feature location is marked as a transitional candidate. In deeper feature layers, the original image range covered by a location expands. The image processor downsamples the connectivity of the common subgraph of the preservation map along with the coordinate record and aggregates the confidence state only when the covered area belongs to the same common subgraph to avoid non-adjacent anatomical structures sharing the same calibration confidence value due to reduced feature resolution.

[0050] In a preferred embodiment, the fusion feature is formed according to the following formula: ; Where u and v represent the spatial coordinates of the current feature layer, and k represents the feature channel index. Indicates the correction domain in coordinates and the eigenvalue at channel k, Represents the corresponding eigenvalues ​​of the original domain. This indicates the degree of correction confidence obtained after mapping the correction confidence map from pixel coordinates to the current feature layer. The plus sign indicates the addition of corresponding features from two domains, and the multiplication sign indicates the modulation of the corresponding domain features by the correction confidence level. This represents the amount of complementary retention of the original domain. For example, when the correction confidence level is 0.8, the correction domain feature value is 0.7, and the original domain feature value is 0.5, the fused feature value is 0.66, and the correction domain information accounts for the majority. However, when the correction confidence level is 0.1 and the other values ​​remain unchanged, the fused feature value is 0.52, and the fusion result is close to the original domain feature, so that the feature source changes continuously with the correction confidence state at the same position.

[0051] Please refer to Figure 5 , combined Figure 5 Observe the transition process of dual-domain features. The horizontal axis represents the confidence tendency, gradually changing from the original side, partially original, balanced, partially corrected, to the corrected side; the vertical axis represents the fusion strength, ranging from approximately 0.3 to 0.9. The black solid line, long dashed dotted circle, and dotted square correspond to three sets of feature combinations, used to describe different channels. Follow The response to changes. Combination 1 shows a steady increase, indicating that the contribution of the correction domain is continuously amplified as the confidence tendency increases; Combination 2 shows a rapid increase from a low level, indicating... Clearly stronger than When the high-confidence region provides a clearer usable structure, the combination of three regions decreases as the confidence tendency increases, indicating that when the original domain information is more complete, the fusion result will not be dominated by the correction domain alone.

[0052] In this embodiment, discrete selection and continuous fusion methods can be used in combination. Locations covered by the calibration confidence map and where the response order of candidate parts in both domains is consistent are directly selected from the calibration domain features. Locations with structural conflicts are directly selected from the original domain features. Transitional regions at the boundary between two types of locations form fusion values ​​according to the aforementioned formula. The image processor establishes a continuous feature chain based on the edge direction, response connectivity, and candidate part order of adjacent locations. At the boundary between the original and calibration domain selection results, the channel vector at the center location is concatenated with the adjacent channel vectors on the continuous feature chain, and then the original channel arrangement is restored through shared channel mapping. Fusion features at different scales are mapped layer by layer according to coordinate records. Overall contour semantics and local edge semantics are merged by channel in the same original image region. The classification mapping layer then forms the spatial and category responses of candidate parts accordingly.

[0053] In the boundary region between the original domain features and the correction domain features, the image processor does not directly average the two domain features along the rectangular window. Instead, it reads the edge path passing through the boundary position from the anatomical discriminant preservation map and assigns the features on both sides of the same edge path to the corresponding tissue regions. Features that are adjacent along the edge tangentially and have the same source retain the original selection results. Features that cross the edge normal only participate in continuous fusion when the structure preservation map shows that the edge position has not changed. The transition value near the structural conflict area changes positionally from the original domain side to the quality compensation consistent area. After channel stitching, the shared channel mapping merges the same type of structural response in the original domain and the correction domain according to the channel semantics saved during the training phase. The exclusive channels of different candidate categories remain separated. Therefore, the fusion boundary will not propagate the local gray-level changes in the correction domain to the original domain anatomical structure that is not connected to it, nor will it cut off the category evidence continuously distributed along the same bony edge.

[0054] When generating the response map of the first part, the image processor merges the spatial responses that belong to the same candidate part and are continuous along the anatomical edge into response connected regions. The category of each connected region is calibrated by fusing the category response, and the peripheral edge of the response, the internal response ridge line, and the boundary edge between adjacent connected regions are recorded. The response connected edge is composed of the peripheral edge and the boundary edge that are continuous with the gradient direction of the original image. The spatial support region of the correction residual and the response connected edge are transformed to the same scale and then intersected. Within the intersection, the image processor reads the response order of the two candidate parts, determines the position where the first candidate is exchanged or the response connected attribute changes as the position of order difference, and then performs a connectivity search along the residual direction of the corresponding position as the starting point. The search results form the correction conflict region, while the residual region that does not touch the response connected edge is retained as the quality compensation candidate region.

[0055] When determining the response order of candidate regions, the category response mapping layer outputs a response vector arranged according to the candidate set for each feature location. The image processor forms a positional order record based on the relative order of each element in the response vector and merges the positional orders belonging to the same response connected region according to the connectivity relationship. If the first candidate of the original domain and the correction domain are the same but adjacent candidates are swapped, the image processor checks whether the swapped position coincides with the spatial support region of the correction residual. If they coincide, they are retained as evidence of potential conflict; if they do not coincide, they are regarded as local fluctuations caused by feature propagation. If the first candidate is swapped, the process directly proceeds to the response connected edge check. The order difference is only extended into a correction conflict region when the three spatial conditions of residual support region, response connected edge and the same candidate set are continuously corresponding, thus avoiding writing category competition far from the correction position into the protection map.

[0056] Preferably, after correcting the conflict region and writing it back, the image processor sets it as an immovable region, sets the continuous edge region adjacent to it as a restricted grayscale adjustment region, and re-obtains the grayscale base component and edge structure component from the original image during local recorrection. The immovable region is reconstructed with the compensated grayscale base component and the edge structure component before correction. The restricted grayscale adjustment region adjusts the grayscale base according to the continuity of the surrounding quality state, and at the same time filters the structure component along the original edge direction. The unmodified correction pixels outside the region are used as boundary conditions to connect with the reconstructed pixels. The updated correction image and the image to be identified are subtracted pixel by pixel to form an updated correction residual. The part of the updated residual located in the degradation region and not crossing the boundary of the immovable region is written with the quality compensation component, and the part intersecting with the boundary of the immovable region is written with the structure conflict component.

[0057] In a preferred embodiment, the image processor superimposes the quality compensation component onto the image to be identified according to the original pixel coordinates to form a recalibrated image. The original pixels are maintained at the locations covered by the structural conflict component, and the boundaries are reconstructed according to the spatial continuity of the grayscale base at the locations where the quality compensation component meets the uncalibrated area. The recalibrated image is input into a shared feature extraction network to generate recalibrated domain features. The original domain features are directly called at the locations covered by the structural conflict component, while the recalibrated domain features are called at other locations when the calibration confidence state is established and the response order of the two candidate regions is consistent. The selected features are reassembled according to a continuous feature chain to form updated fusion features. The updated fusion features are then classified and mapped to form the region identification result. Simultaneously, the coordinate correspondence records of the final spatial response, the calibrated conflict region, the quality compensation component, and the structural conflict component are saved, enabling each category contribution to be associated with a specific region using either the original pixel or the calibrated pixel.

[0058] In the model building phase, a training set is formed using verified images of the examination sites. Each training image retains its original quality version and constructs quality change versions based on the same anatomical coordinates, including low-frequency grayscale deviation, high-frequency structural attenuation, and the coupling of the two. The candidate site recognition branch learns the category correlation and spatial response based on site annotations. The quality analysis branch learns the low-frequency grayscale deviation map and high-frequency structural attenuation map based on the difference between the quality change version and the original quality version. The correction branch learns the compensation component based on the correspondence between the grayscale base and the edge structure. The training constraints include site category consistency, corresponding edge directions of the protected area, correction residuals located in the degraded area, and maintaining connectivity between the common subgraphs of the original and correction domains. After training, the network parameters used for inference are fixed. During runtime, the descriptive information is checked and not used as the basis for site classification to avoid incorrect descriptions directly determining the semantic response of pixels.

[0059] In one engineering application, the image to be identified is a frontal digital radiograph of the chest. The quality status map records local grayscale deviations in the scapular region and scattering background in the lower lung field. The candidate site semantic response map simultaneously preserves the spatial responses corresponding to the chest, thoracic vertebrae, and ribs. The anatomical discrimination protection map writes common sub-images from the continuous edge of the clavicle, the inner edge of the scapula, the midline of the spine, and the contour of the diaphragm into the immovable region. Constrained correction compensates for the grayscale base of the lower lung field and preserves the pre-correction edge structure at the inner edge of the scapula. The shared feature extraction network uses correction domain features at locations in the lower lung field where quality compensation is consistent and structure is preserved, and uses original domain features at structural conflict locations at the junction of the clavicle and scapula. When the local candidate response in the initial fusion changes from chest-led to rib-led, the corresponding location and continuous residuals together form a correction conflict region and trigger local recorrection.

[0060] In the local recalibration of this application example, the image processor writes the inner edge of the scapula into an immovable region, compensates only for the grayscale base component corresponding to the exposure deviation, updates the correction residuals to include the portion that maintains the edge direction and is located in the original degraded region into the quality compensation component, and includes the portion that causes a change in candidate order or crosses the scapular boundary into the structural conflict component. The recalibrated domain features are then used at the location covered by the quality compensation component, and the original domain features are called at the location covered by the structural conflict component. The spatial response is merged along the continuous edges of the clavicle, scapula, and diaphragm. The classification mapping layer outputs the chest examination site and its spatial response. The original image, the recalibrated image, the calibration confidence map, and the conflict region maintain the same coordinate association, thus enabling the verification of the image region actually used in the process of forming the examination site category.

[0061] In another engineering application, the image to be identified is a radiographic image containing the junction of the pelvis and lumbar spine. The initial category response is close between the pelvis and lumbar spine. The quality state map records high-frequency structural attenuation at the outer edge of the iliac bone and low-frequency gray-level deviation in the abdominal soft tissue projection area. The candidate response common sub-map covers the continuous edge of the sacroiliac joint and pelvic ring and is written into the immovable region. The lumbar spine candidate specific nodes are written into the restricted gray-level adjustment region. The correction process combines the compensated gray-level base with the original edge structure at the sacroiliac joint. At the lumbar spine nodes, only the continuous structural components along the original vertebral body edge direction are retained. If the correction residual does not cross the response connectivity edge of the pelvic ring and the candidate order of the two domains is consistent, the corresponding correction domain features are fused. If the residual causes the pelvic ring to break and changes the candidate order, the conflict location adopts the original domain features and enters local recorrection. The final site response is formed by the pelvic ring that remains connected and the soft tissue region that has been reliably compensated.

Claims

1. An AI-based automatic identification system for inspection areas in DR equipment based on image quality correction, characterized in that, The image processor includes an image processor that performs the following actions: acquiring a DR image to be identified, generating a quality status map and a semantic response map of candidate parts, and constructing an anatomical discriminant protection map from the semantic response map; The degraded region is located based on the quality status map, and the degraded region is corrected under the constraints of the anatomical discrimination protection map to obtain the corrected image and the correction residual characterizing the pixel difference before and after correction. A correction confidence map is generated based on the correction residual and the consistency relationship of the anatomical structure before and after correction. Extract the spatial correspondence features before and after correction, and fuse them position by position according to the semantic consistency relationship between the correction confidence map and the spatial correspondence features to generate the first part response map; The overlap relationship between the first part response map and the correction residual updates the anatomical discrimination protection map, and the map is corrected and fused again to output the part identification result.

2. The AI-based automatic identification system for DR equipment inspection areas based on image quality correction according to claim 1, characterized in that, The image processor unifies the grayscale direction and separates the effective illumination area based on the DICOM photometric interpretation information of the DR image to be identified; Within the effective illumination area, a low-frequency gray-level deviation map is formed based on the regional gray-level distribution, and a high-frequency structure attenuation map is formed based on the local gradient continuity. The low-frequency gray-level deviation map and the high-frequency structure attenuation map are aligned according to pixel coordinates, and their joint response is encoded into a quality state map. Pixels outside the effective illumination area are encoded as uncorrected regions.

3. The AI-based automatic identification system for DR equipment inspection areas based on image quality correction according to claim 1, characterized in that, The image processor extracts the category correlation and spatial response of each candidate part from the DR image to be identified, and performs positional correspondence synthesis of the spatial response of each candidate part according to the category correlation to form a semantic response map of the candidate part. Extract continuous regions along the edge direction from the DR image to be identified, encode pixels that remain responsive across candidate parts and are located within continuous regions along the edge direction as non-displaceable regions, and encode candidate-specific response regions as restricted grayscale adjustment regions to form the anatomical discrimination protection map.

4. The AI-based automatic identification system for DR equipment inspection areas based on image quality correction according to claim 1, characterized in that, The image processor maps the quality status map and the anatomy discrimination protection map to the same pixel coordinates, decomposes the pixels in the degraded region into gray-level base components and edge structure components, and forms gray-level base compensation components based on the quality status map. The grayscale base component after compensation in the non-displaceable region is combined with the edge structure component before correction; the grayscale base component after compensation in the restricted grayscale adjustment region is combined with the edge structure component selected according to the continuity relationship of the edge direction. The combined result is reconstructed into a corrected image according to the original pixel coordinates, and the pixel difference before and after reconstruction is recorded as the correction residual.

5. The AI-based automatic identification system for DR equipment inspection areas based on image quality correction according to claim 2, characterized in that, The image processor resamples the effective illumination area at both the global and local scales, calculates low-frequency grayscale deviation maps and high-frequency structure attenuation maps at each scale, and then maps each map back to the pixel coordinates of the DR image to be identified. A coupled degradation state is established for locations where low-frequency and high-frequency responses coexist, and a single degradation state is established for locations where only one type of response exists. The quality state map is updated by the spatial set of the two degradation states.

6. The AI-based automatic identification system for DR equipment inspection areas based on image quality correction according to claim 3, characterized in that, The image processor divides the spatial response of each candidate region into a response connected region, uses the response connected region as a node, and establishes a candidate region response map structure according to the edge direction continuity relationship between adjacent nodes. Extract the common sub-image corresponding to the position in the response map structure of different candidate parts, write the pixels covered by the common sub-image into the non-displaceable region, and write the candidate dedicated node connected to the common sub-image into the restricted grayscale adjustment region.

7. The AI-based automatic identification system for DR equipment inspection areas based on image quality correction according to claim 4, characterized in that, The image processor generates a corrected quality state map from the corrected image, and compares the quality state map with the corrected quality state map within the spatial support region of the correction residual to form a quality compensation consistency map. Within the area covered by the anatomical discrimination protection map, the edge directions and connectivity before and after correction are compared to form a structure preservation map; The quality compensation consistency map and the structure retention map are intersected by spatial location, and a correction reliability map is generated based on the corresponding range of the intersection and the correction residual.

8. The AI-based automatic identification system for inspection areas of DR equipment based on image quality correction according to claim 7, characterized in that, The image processor forms the original domain features of the DR image to be identified and the correction domain features of the corrected image through a shared feature extraction network, maps the corrected confidence map to the corresponding level of the two domain features, and generates the response order of candidate parts according to the two domain features respectively. At the locations covered by the corrected confidence map and where the response order of the two candidate regions is consistent, the corrected domain features are selected. At the remaining locations, the original domain features are selected, and the selected features are spliced ​​together according to the feature continuity relationship of adjacent locations to form fused features.

9. The AI-based automatic identification system for inspection areas of DR equipment based on image quality correction according to claim 7, characterized in that, The image processor generates a first part response map based on the fusion features, and maps the first part response map and the correction residual to the same spatial scale. The spatial support region of the correction residual is intersected with the connected edge of the response map of the first part, and the positions where the response order of the two candidate parts of the intersection is different are extracted as the correction conflict region. The corrected conflict region is written into the anatomical discriminant protection map, and pixel correction is re-performed only on the corrected conflict region and its continuous boundaries to generate an updated corrected image and an updated correction residual.

10. The AI-based automatic identification system for inspection areas of DR equipment based on image quality correction according to claim 9, characterized in that, The image processor performs position correspondence calculations between the updated correction residual and the updated anatomical discriminant protection map, retains the updated correction residual where the support region is located in the degraded region and does not intersect with the boundary of the immovable region as a quality compensation component, and marks the updated correction residual where the support region intersects with the boundary of the immovable region as a structural conflict component. The quality compensation component is superimposed on the DR image to be identified to form a recorrected image. The original domain features are used at the location covered by the structural conflict component to re-form the fusion features and output the location identification result.

Images