A medical image quality evaluation method and system based on deep learning

Abstract

The application belongs to the technical field of image quality evaluation, and discloses a medical image quality evaluation method and system based on deep learning, which comprises the following steps: extracting medical image features; performing global feature correlation analysis on the medical image features to determine the quality defect type of the medical image to be evaluated and to determine the severity corresponding to the quality defect type; performing quantitative analysis on the quality defect type of the medical image to be evaluated and the severity corresponding thereto to identify the key factors affecting the imaging quality of the medical image to be evaluated; and based on the key factors, comprehensively analyzing and judging by combining preset medical image quality evaluation indexes to output the final quality evaluation result of the medical image to be evaluated. The application can accurately identify the quality defect type of the medical image and determine the severity thereof, and through quantitative analysis, the key factors affecting the imaging quality are located, so that the overall efficiency and determination accuracy of the medical image quality evaluation are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of image quality assessment technology, and in particular to a method and system for medical image quality assessment based on deep learning. Background Technology

[0002] Medical imaging is the technology and processing of acquiring images of internal human tissues non-invasively. It encompasses categories such as radiology, endoscopy, medical thermal imaging, medical photography, and microscopy, and falls under the category of bioimaging. It serves as a crucial basis for clinical disease diagnosis and treatment planning, providing medical professionals with clear and accurate reference information. Furthermore, standardized image quality is essential for sharing medical examination data, mutual recognition of results, and steadily promoting the homogenization of medical quality, playing an indispensable role in modern clinical medicine.

[0003] However, current medical image quality assessments largely rely on manual judgment, which is inefficient and easily influenced by subjective factors, making it difficult to standardize evaluation criteria. It is challenging to accurately distinguish defect types, determine defect severity, and effectively locate key factors affecting image quality. Overall, the accuracy of the assessment is insufficient, making it difficult to generate standardized and objective quantitative evaluation results.

[0004] Therefore, how to provide a method and system for evaluating the quality of medical images based on deep learning is an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a method and system for evaluating the quality of medical images based on deep learning, in order to solve the aforementioned technical problems in the prior art.

[0006] According to a first aspect of the present invention, a method for evaluating the quality of medical images based on deep learning is provided.

[0007] In one embodiment, a deep learning-based medical image quality assessment method includes: The pre-acquired medical images to be evaluated are preprocessed, and medical image features are extracted. Global feature association analysis is performed on medical image features to determine the type of quality defect in the medical image to be evaluated and to determine the severity of the corresponding quality defect type. A quantitative analysis of the types of quality defects in the medical images to be evaluated and their corresponding severity was conducted to identify the key factors affecting the imaging quality of the medical images to be evaluated. Based on key factors and combined with preset medical image quality evaluation indicators, a comprehensive assessment is conducted to output the final quality evaluation result of the medical image to be evaluated.

[0008] In one embodiment, preprocessing the pre-acquired medical images to be evaluated and extracting medical image features includes: The medical images to be evaluated are denoised, and the brightness and grayscale deviations of the denoised medical images are corrected. The corrected medical image to be evaluated is enhanced to obtain a medical image with enhanced contrast and clarity. The enhanced medical image to be evaluated is converted to grayscale to obtain a grayscale image, and the edge gradient of each pixel in the grayscale image is calculated to generate a gradient image. Based on gradient images, quality defect feature regions are delineated, and the area, morphological parameters, and gradient intensity statistics of these quality defect feature regions are calculated to construct defect quantification features. The quality defect feature region is binarized, edge information is extracted and defect contours are generated, and the defect contours are fused with the defect quantification features to obtain medical image features.

[0009] In one embodiment, global feature association analysis is performed on medical image features to determine the type of quality defect in the medical image to be evaluated, and the severity of the corresponding quality defect type is determined, including: A baseline feature region is selected from medical image features as the initial analysis object and included in the set to be analyzed. Check if there are any valid analysis objects in the set to be analyzed. If there are no valid analysis objects, terminate the feature analysis; if there are valid analysis objects, perform subsequent operations. Select the region with the best comprehensive feature parameters from the set to be analyzed as the current analysis region, and then determine and process the surrounding related feature regions of the current analysis region one by one; Move the current analysis area into the completed analysis set and check whether the target defect feature area is included in the set to be analyzed; if not included, continue feature analysis; if included, terminate feature retrieval. Starting from the target defect feature area, the system traces back to the initial analysis object according to the reference benchmark corresponding to each area to determine the quality defect type of the medical image to be evaluated and to determine the severity of the corresponding quality defect type.

[0010] In one embodiment, selecting the region with the optimal comprehensive feature parameters from the set to be analyzed as the current analysis region, and then determining and processing the surrounding related feature regions of the current analysis region one by one, includes: If the surrounding related feature regions have already been included in the completed analysis set, no operation will be performed; If the surrounding related feature regions are not included in the set to be analyzed, then the surrounding related feature regions will be included in the set to be analyzed, and the comprehensive feature parameters and basic feature parameters of the surrounding related feature regions will be calculated using the current analysis region as a reference. If the surrounding related feature regions are already within the set to be analyzed, and the basic feature parameters of the surrounding related feature regions are better than the original values ​​after using the current analysis region as a reference, then the reference is updated, and the comprehensive feature parameters and basic feature parameters of the surrounding related feature regions are recalculated.

[0011] In one embodiment, starting from the target defect feature region, the process traces backward to the initial analysis object according to the reference benchmark corresponding to each region to determine the type of quality defect in the medical image to be evaluated, and to determine the severity of the corresponding quality defect type, including: Extract all feature parameters of the target defect feature region and simultaneously retrieve the reference benchmark information corresponding to each region on the tracing path; By tracing back layer by layer based on the reference benchmark information, the complete feature association link from the initial analysis object to the target defect feature area is restored; All feature parameters within the feature association link are compared and matched with a preset standard defect sample library, and the type of quality defect corresponding to the medical image to be evaluated is determined based on the matching results. Based on the preset defect severity grading index and combined with the deviation of characteristic parameters from the standard value, the severity of the quality defect type is determined.

[0012] In one embodiment, all feature parameters within the feature association link are compared and matched with a preset standard defect sample library, and the type of quality defect corresponding to the medical image to be evaluated is determined based on the matching results, including: Extract all feature parameters within the feature association link, and perform data normalization and format unification processing on all feature parameters; Call the preset standard defect sample library and load the standard feature samples corresponding to various defects; After data normalization and format unification, all feature parameters are compared with standard feature samples one by one to obtain the feature fit of all feature parameters with standard feature samples. Feature classification is performed based on feature fit, and the type of quality defect corresponding to the medical image to be evaluated is determined.

[0013] In one embodiment, the severity of a quality defect type is determined based on a preset defect severity grading index and the magnitude of deviation of characteristic parameters from standard values, including: Retrieve pre-configured defect severity grading indicators and organize the standard range of characteristic parameters and grading rules corresponding to each severity level; Extract all feature parameters from the feature association link after data normalization and format unification. Using the standard feature parameters as a reference, calculate the offset of each feature parameter relative to the standard value to obtain the magnitude of the feature parameter's deviation from the standard value. The deviation of the characteristic parameters from the standard value is compared with the defect severity grading index one by one to determine the grade range to which the deviation belongs; By combining the grading rules with the grading range matching results, the grading is completed, and the severity of the quality defect type is determined.

[0014] In one embodiment, a quantitative analysis of the types of quality defects in the medical images to be evaluated and their corresponding severity is performed to identify key factors affecting the imaging quality of the medical images to be evaluated, including: A medical image feature quantitative analysis system was constructed. All feature parameters after data normalization and format unification were initialized to an unanalyzed state. At the same time, the matching marks of quality defects corresponding to each feature parameter were uniformly set to an unmatched state. Select initial baseline feature parameters from all feature parameters, complete the feature parameter grouping based on the preset feature matching degree threshold, and calculate the ratio of the representation contribution to the matching strength of each feature parameter group after removing redundant interference. Feature parameters with a matching degree greater than the preset feature matching degree threshold are classified into the same feature group. The currently selected baseline feature parameters are marked as analyzed. Among all unanalyzed feature parameters, the feature parameters with the smallest ratio of characterization contribution to matching strength and the highest defect abnormality tendency are selected as the baseline feature parameters for the next stage. The quality defect types of the medical images to be evaluated and their corresponding severity are simultaneously associated. The process of selecting baseline feature parameters and dividing feature groups is repeated iteratively until all feature parameters have been traversed and analyzed. By combining the quality defect type, defect severity and defect abnormality tendency corresponding to each feature group, the key factors affecting the imaging quality of the medical image to be evaluated are comprehensively identified.

[0015] In one embodiment, initial baseline feature parameters are selected from all feature parameters, feature parameter groups are divided based on a preset feature matching threshold, and the ratio of representation contribution to matching strength of each feature parameter group after removing redundant interference is calculated, including: Parameters with benchmark characterization properties are selected from all feature parameters and used as initial benchmark feature parameters; Based on the preset feature matching threshold, all other feature parameters are matched and verified with the initial benchmark feature parameters. Based on the matching results, all feature parameters are grouped and formed into multiple independent feature parameter groups. Redundancy and interference removal are performed on each of the divided feature parameter groups one by one to remove invalid, duplicate and interference parameter data within the group, and retain the effective feature parameters that can reflect image quality defects. Based on each feature parameter group after removing redundant interference, the ratio of overall representation contribution to matching strength for each feature parameter group is calculated.

[0016] According to a second aspect of the present invention, a deep learning-based medical image quality assessment system is provided.

[0017] In one embodiment, a deep learning-based medical image quality assessment system includes: The image processing module is used to preprocess the pre-acquired medical images to be evaluated and extract medical image features. The defect analysis module is used to perform global feature correlation analysis on medical image features, determine the type of quality defect in the medical image to be evaluated, and determine the severity of the corresponding quality defect type. The factor analysis module is used to quantitatively analyze the types of quality defects in the medical images to be evaluated and their corresponding severity, and to identify the key factors that affect the imaging quality of the medical images to be evaluated. The evaluation result output module is used to conduct a comprehensive assessment based on key factors and preset medical image quality evaluation indicators, and output the final quality evaluation result of the medical image to be evaluated.

[0018] According to a third aspect of the present invention, a computer device is provided.

[0019] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0020] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0021] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0022] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: 1) This invention can accurately identify the types of defects in medical image quality and determine their severity, and locate key factors affecting image quality through quantitative analysis. Combined with preset evaluation indicators, it conducts a comprehensive assessment, achieving fully automated evaluation throughout the entire process. The evaluation results are objective and consistent, effectively improving the overall efficiency and accuracy of medical image quality evaluation.

[0023] 2) This invention accurately locates defect types and quantifies severity through region-by-region correlation analysis and reverse tracing, avoiding redundant global calculations and improving evaluation efficiency. Combined with feature comparison and grading indicators, it achieves standardized and interpretable quality assessment, enhancing the objectivity and clinical credibility of medical image evaluation.

[0024] 3) This invention iteratively selects benchmark features and analyzes them in groups to quantify the representational contribution and matching strength of each feature parameter, identify key influencing factors, thereby effectively eliminating redundant interference, improving the accuracy and efficiency of key factor identification, and providing a clear and quantifiable basis for optimizing medical image quality.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0027] Figure 1 This is a flowchart illustrating a deep learning-based medical image quality assessment method according to an exemplary embodiment; Figure 2 This is a schematic diagram illustrating the principle of a deep learning-based medical image quality assessment system according to an exemplary embodiment. Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment. Detailed Implementation

[0028] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented 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.

[0029] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0030] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0031] Figure 1 An embodiment of a deep learning-based medical image quality assessment method of the present invention is shown.

[0032] In this optional embodiment, the deep learning-based medical image quality assessment method includes: S101. Preprocess the pre-acquired medical images to be evaluated and extract the medical image features; Specifically, medical image features include morphological region features (i.e., regional attribute features such as the outline, range, location, and regional correlation of each region of the image), grayscale / pixel features (i.e., basic pixel parameters such as pixel grayscale value, brightness, contrast, and uniformity), and texture detail features (i.e., detail features such as image texture distribution, detail clarity, and edge integrity).

[0033] S102. Perform global feature association analysis on medical image features to determine the type of quality defect in the medical image to be evaluated and determine the severity of the corresponding quality defect type. Specifically, quality defect types include sharpness defects, grayscale / brightness defects, noise defects, and deformation / misalignment defects. The severity levels corresponding to these quality defect types are categorized as minor, moderate, moderate, and severe.

[0034] S103. Quantitatively analyze the types of quality defects in the medical images to be evaluated and their corresponding severity to identify the key factors affecting the imaging quality of the medical images to be evaluated. Specifically, key factors affecting the imaging quality of the medical images to be evaluated include acquisition parameters, equipment hardware, the image itself, and operational and environmental factors.

[0035] S104. Based on key factors and combined with preset medical image quality evaluation indicators, a comprehensive assessment is conducted to output the final quality evaluation result of the medical image to be evaluated.

[0036] In this optional embodiment, the pre-acquired medical image to be evaluated is preprocessed and medical image features are extracted, including: denoising the medical image to be evaluated and correcting the brightness and grayscale deviation of the denoised medical image; enhancing the corrected medical image to obtain a medical image with enhanced contrast and clarity; performing grayscale conversion on the enhanced medical image to obtain a grayscale image, and calculating the edge gradient of each pixel in the grayscale image to generate a gradient image; delineating quality defect feature regions based on the gradient image, and calculating the area, morphological parameters, and gradient intensity statistics of the quality defect feature regions to construct defect quantification features; binarizing the quality defect feature regions, extracting edge information and generating defect contours, and fusing the defect contours with the defect quantification features to obtain medical image features.

[0037] It should be explained that the raw medical images acquired clinically are used as the objects to be processed. Noise reduction is performed using Gaussian filtering, which smooths the image and eliminates noise and artifacts through a weighted average of neighboring pixels. The filter kernel size is selected from 3×3 to 7×7 (5×5 is preferred), and noise is filtered out by traversing all pixels of the entire image. After noise reduction, brightness and grayscale correction are performed. Referring to a preset brightness standard (clinical medical image baseline brightness, grayscale level 120) and a grayscale baseline range [40, 220], the grayscale value of each pixel is compared pixel by pixel. Pixels exceeding the range are numerically corrected to unify the brightness and grayscale representation of the entire image.

[0038] The corrected image is enhanced and optimized by stretching pixel grayscale to improve contrast and sharpening edges to improve clarity. The contrast enhancement factor is set between 1.2 and 1.8 to amplify the grayscale difference between pixels. Clarity enhancement follows a predetermined rule: strengthen the grayscale response of pixels at the edges and weaken invalid details in flat areas, ultimately resulting in an enhanced image with satisfactory contrast and clarity.

[0039] Multi-channel color images are converted into single-channel grayscale images according to a preset grayscale mapping rule. The rule is a weighted sum of the three color channels, with weights of 0.299 for the red channel, 0.587 for the green channel, and 0.114 for the blue channel, thus unifying the data format. The Sobel operator is used to calculate edge gradients. This operator uses a 3×3 computation window to traverse pixels, detecting horizontal and vertical edge changes and outputting gradient values. Based on the gradient values ​​of all pixels, a corresponding gradient image is generated.

[0040] A gradient threshold range of 25-60 is set, and regions in the gradient image whose pixel gradient values ​​fall outside this range are defined as quality defect feature regions. For each defined region, three types of quantization parameters are calculated: the total number of pixels in the region represents the area; aspect ratio and contour roundness (roundness within the normal range of 0.3-1.0) are used as morphological parameters; and the average gradient, maximum gradient, and minimum gradient of the region are calculated as gradient intensity statistics. All parameters are integrated to form a set of defect quantification features.

[0041] Binarization segmentation is performed on the defect feature region, with a binarization threshold ranging from 100 to 160. Pixels with gray levels above the threshold are marked as foreground regions, and those below are marked as background regions. Edge pixels are extracted based on the binarization results to generate the defect contour. Integration is performed according to feature fusion rules: the coordinate information of the defect contour is bound to and synchronously stored with the aforementioned defect quantization parameters. The fused data is the final extracted medical image feature, which is input into the next stage for global feature association analysis.

[0042] In this optional embodiment, performing global feature association analysis on medical image features to determine the type of quality defect in the medical image to be evaluated and to determine the severity of the corresponding quality defect type includes: selecting a baseline feature region from the medical image features as the initial analysis object and including it in the analysis set; checking whether there is a valid analysis object in the analysis set; if there is no valid analysis object, terminating the feature analysis; if there is a valid analysis object, performing subsequent operations; selecting the region with the best comprehensive feature parameters from the analysis set as the current analysis region, and judging and processing the surrounding associated feature regions of the current analysis region one by one; moving the current analysis region into the completed analysis set and checking whether the target defect feature region is included in the analysis set; if not included, continuing feature analysis; if included, terminating feature retrieval; starting from the target defect feature region, tracing back to the initial analysis object according to the reference benchmark corresponding to each region to determine the type of quality defect in the medical image to be evaluated and to determine the severity of the corresponding quality defect type.

[0043] Specifically, based on medical image features, and following preset region selection rules—using gradient intensity and contour integrity as selection criteria—preferentially selecting feature regions corresponding to the image center and main tissue as benchmarks. These regions are the core observation areas of the image and possess representativeness for analysis. From these, a benchmark feature region is selected as the initial analysis object. This region is then uniformly categorized into the analysis set, which is used to centrally store all feature regions to be traversed and analyzed.

[0044] The system retrieves the contents of the set to be analyzed in real time and performs verification based on the valid object determination rules. Valid object determination rules: If the set contains feature regions with complete feature parameters and normal contours, a valid analysis object is determined to exist; if the set is empty, or if the internal region feature data is missing or the contour is abnormal, no valid analysis object is determined to exist. If no valid analysis object is determined, the current global feature analysis process is terminated directly; if a valid analysis object exists, subsequent analysis operations continue.

[0045] Based on the comprehensive feature parameter optimization rule: the comprehensive feature parameters are scored by combining regional gradient strength, morphological parameters, and feature completeness, with a score range of 0-100. The region with the highest score is selected as the optimal region for comprehensive feature parameters. The current analysis region is determined from the set to be analyzed. According to the neighborhood association determination rule: a range of 8-15 pixels extending outward from the boundary of the current analysis region is defined as the surrounding associated feature region. The surrounding regions are determined and processed one by one.

[0046] The currently analyzed region, once its assessment is complete, is transferred from the set to be analyzed to the set of regions that have been analyzed. This set specifically includes all feature regions that have been traversed and do not require repeated processing. The system then checks whether the target defect feature region has been included in the set to be analyzed. If not, it returns to the previous loop to continue feature analysis; if successfully included, the entire feature retrieval process is immediately terminated.

[0047] Starting from the target defect feature region, the system traces back layer by layer to the initially selected benchmark feature region based on the reference benchmarks retained for each feature region (i.e., the comparison standard feature parameters and regional relationships bound to each feature region during the analysis process, serving as the basis for tracing). This process establishes a complete feature association chain. All feature parameters within this chain are then matched against a pre-defined standard defect sample library (which includes standard feature data corresponding to various image defects such as blurring, artifacts, and grayscale anomalies, serving as a general reference sample for medical image defects in the industry). The specific quality defect type of the image is determined based on defect classification matching rules. Defect classification matching rules: Each feature to be analyzed is compared with the standard features of various defects in the library, and the overall similarity is calculated comprehensively. When the similarity reaches a pre-defined acceptable standard, the current image is classified into that defect category.

[0048] The severity grading index of defects is retrieved and divided into four levels: minor, moderate, severe, and serious. The range is divided according to the deviation of the characteristic parameter from the standard parameter. A small deviation is minor, a moderate deviation is moderate, a large deviation is severe, and an extremely large deviation is serious. The determination is made by combining the parameter deviation range, and finally the severity of the defect is obtained.

[0049] In this optional embodiment, selecting the region with the optimal comprehensive feature parameters from the set to be analyzed as the current analysis region, and determining and processing the surrounding related feature regions of the current analysis region one by one, includes: if the surrounding related feature regions have been included in the completed analysis set, no operation is performed; if the surrounding related feature regions have not been included in the set to be analyzed, the surrounding related feature regions are included in the set to be analyzed, and the comprehensive feature parameters and basic feature parameters of the surrounding related feature regions are calculated using the current analysis region as a reference benchmark; if the surrounding related feature regions are already in the set to be analyzed, and the basic feature parameters of the surrounding related feature regions are better than the original values ​​after using the current analysis region as a reference benchmark, the reference benchmark is updated, and the comprehensive feature parameters and basic feature parameters of the surrounding related feature regions are recalculated.

[0050] Specifically, based on the set to be analyzed, according to the comprehensive feature parameter optimization rule: the region morphology, gradient intensity, contour integrity and feature stability are used as evaluation dimensions for comprehensive evaluation. The region with the best comprehensive performance is the region with the best comprehensive feature parameters. This rule is used to uniquely determine the priority analysis object in this round, and select the feature region with the best comprehensive feature parameters as the current analysis region.

[0051] Centered on the current analysis area, the neighborhood range is defined according to the following criteria: using the outer edge of the current area as a reference, the surrounding related feature areas are delineated by extending a pixel range outward. The extended pixel range is 8-15 pixels, which is the standard neighborhood determination range for medical image feature association analysis. All surrounding related feature areas corresponding to the current analysis area are then identified, and each surrounding related feature area is subsequently classified.

[0052] Retrieve the status of the current surrounding related feature regions. If the region has been moved to the completed analysis set (which stores all feature regions that have been traversed and do not require further analysis or updating), skip that region without performing any operations and continue checking the next surrounding related feature region. This follows the idle processing rule: regions that have already been analyzed will not participate in repeated calculations and updates to avoid invalid loop processing.

[0053] If the surrounding related feature region is neither in the set to be analyzed nor in the set of completed analyses, a region migration operation is performed to uniformly include it in the set to be analyzed. The set to be analyzed is used to store all feature regions waiting to be traversed, analyzed, and have their parameters calculated. The current analysis region is set as a unified reference benchmark, which serves as the standard reference object for feature parameter comparison and calculation. Using the various features of the current analysis region as comparison benchmarks, and following the parameter calculation specifications: according to the preset feature statistical logic, the basic feature parameters and comprehensive feature parameters corresponding to the region are collected and calculated sequentially. The basic feature parameters include basic indicators such as regional gradient, shape, and grayscale, while the comprehensive feature parameters are the overall evaluation indicators obtained by fusing multiple basic parameters. The calculation and storage of the two types of parameters for the surrounding region are completed.

[0054] If the surrounding related feature region is already stored in the analysis set, the current analysis region is used as a new reference benchmark to re-verify the region's basic feature parameters. According to the parameter quality judgment rules: the basic feature parameters obtained after updating the reference benchmark are compared with the original parameters previously stored for that region; if the feature performance better matches the standard image features, the parameters are considered superior. If the newly obtained basic feature parameters are better than the original values, a benchmark update operation is performed, replacing the reference benchmark corresponding to that region with the current analysis region. Then, the comprehensive feature parameters and basic feature parameters of the surrounding related feature region are recalculated and updated according to the aforementioned parameter calculation specifications; if the new parameters are not better than the original values, no changes are made, and the next region is processed directly.

[0055] In this optional embodiment, starting from the target defect feature region, the process traces back to the initial analysis object according to the reference benchmarks corresponding to each region to determine the quality defect type of the medical image to be evaluated and to determine the severity of the quality defect type. This includes: extracting all feature parameters of the target defect feature region and simultaneously retrieving the reference benchmark information corresponding to each region on the tracing path; tracing back layer by layer according to the reference benchmark information to restore the complete feature association link from the initial analysis object to the target defect feature region; comparing and matching all feature parameters in the feature association link with a preset standard defect sample library, and determining the quality defect type of the medical image to be evaluated based on the matching results; and determining the severity of the quality defect type based on a preset defect severity grading index and the magnitude of the deviation of the feature parameters from the standard value.

[0056] It should be explained that, starting from the target defect feature area, all feature parameters of that area, including morphology, gradient, and contour, are completely extracted. At the same time, the reference benchmark information pre-bound to each feature area on the tracing path is retrieved. The reference benchmark information is the comparison standard features, correlation relationships, and parameter reference basis used in the analysis of each area. It is the core evidence for reverse backtracking and completes the synchronous collection of the two types of data.

[0057] Based on the retrieved reference baseline information, the process is reversed layer by layer from the target defect feature area backwards. Following the relationships between each area, each feature area is sequentially linked to fully reconstruct the entire feature association chain from the initially selected initial analysis object to the target defect feature area. The feature association chain is used to record the sequential analysis order and interrelationships of each feature area, intuitively presenting the propagation and distribution patterns of defect features, and ensuring the integrity of the chain nodes and the accuracy of the association relationships.

[0058] The feature parameters corresponding to all nodes in the feature association link are input into a pre-defined standard defect sample library. This library stores standard feature data for various common quality defects, such as image blurring, grayscale anomalies, noise artifacts, and edge loss. Each type of defect corresponds to a separate set of standard feature samples, which serve as a comparison reference for item-by-item comparison. Defect matching judgment rules are executed: the similarity between the parameter to be analyzed and the standard parameters in the library is comprehensively compared. When the overall similarity reaches the pre-defined judgment standard, it is matched as a defect of the corresponding category. Based on the final matching result, the specific quality defect type of the current image is determined.

[0059] The system invokes a preset defect severity grading index, which classifies defects into four levels: minor, moderate, severe, and critical. The grading is based on the deviation of the characteristic parameters from the standard parameters; a small deviation is classified as minor, a moderate deviation as moderate, a large deviation as severe, and an extremely large deviation as critical. By combining the overall deviation of all characteristic parameters within the link from the standard values ​​and comparing them against the grading index, the system completes the level determination, ultimately obtaining the severity result corresponding to this type of defect.

[0060] In this optional embodiment, comparing and matching all feature parameters within the feature association link with a preset standard defect sample library, and determining the quality defect type corresponding to the medical image to be evaluated based on the matching results includes: extracting all feature parameters within the feature association link and performing data normalization and format unification processing on all feature parameters; calling the preset standard defect sample library and loading standard feature samples corresponding to various defects; comparing all feature parameters after data normalization and format unification processing with the standard feature samples item by item to obtain the feature fit degree between all feature parameters and the standard feature samples; completing feature classification and determination based on the feature fit degree, and determining the quality defect type corresponding to the medical image to be evaluated.

[0061] It needs to be explained that the process involves extracting all feature parameters corresponding to all nodes in the feature association chain, and then standardizing them according to preset data organization and format unification rules: unifying the data storage structure, field names, data types, and arrangement order of the parameters; eliminating invalid and duplicate redundant data within the chain; and ensuring that all parameters have consistent formats and are valid to meet the requirements of subsequent batch comparisons. All feature parameters are then standardized to obtain a set of parameters to be compared with a unified format. Finally, a preset standard defect sample library is retrieved, and all standard feature samples in the library are loaded according to category, completing the data preparation work before comparison.

[0062] The normalized feature parameters to be compared are compared item by item with the loaded standard feature samples. Following the feature fit calculation rules: using standard sample parameters as a reference, the matching degree of individual parameters and the entire set of parameters is comprehensively evaluated, and the feature fit degree, representing the similarity between the two, is finally output. The higher the fit degree, the closer it is to the corresponding defect sample features. The feature fit degree between each set of parameters to be analyzed and the corresponding standard sample is calculated sequentially.

[0063] Referring to the preset feature classification and judgment rules: a unified matching degree judgment benchmark is set. When the feature matching degree corresponding to a certain type of standard sample meets the benchmark requirement, the current image feature is classified into this type of defect category. If the matching degree of multiple groups of samples meets the benchmark, the category with the highest matching degree is selected as the final judgment result. Combining the feature matching degree corresponding to each sample, the classification is completed, and the specific quality defect type of the medical image to be evaluated is finally determined.

[0064] In this optional embodiment, determining the severity of a quality defect type based on a preset defect severity grading index and the deviation of feature parameters from standard values ​​includes: retrieving a pre-configured defect severity grading index and organizing the standard range of feature parameters and grading rules corresponding to each severity level; extracting all feature parameters from the feature association link after data normalization and format unification processing, using standard feature parameters as a reference benchmark, calculating the offset of each feature parameter relative to the standard value, and obtaining the deviation of the feature parameter from the standard value; comparing the deviation of the feature parameter from the standard value with the defect severity grading index one by one to determine the grading range to which the deviation belongs; and combining the grading rules with the grading range matching results to complete the grading and determine the severity of the quality defect type.

[0065] Specifically, the pre-configured defect severity grading index is retrieved locally. This index is a dedicated grading system for medical image quality evaluation, divided into four levels: minor, moderate, moderate, and severe, serving as the unified basis for determining defect severity. Based on the grading index, the standard ranges of characteristic parameters corresponding to each severity level are identified (i.e., for various characteristic parameters such as gradient, morphology, and grayscale, a corresponding normal reference range is defined for each level, with different types of characteristic parameters each having their own dedicated range) and the corresponding grading rules: specifying the judgment logic for single parameters and multi-parameter combinations, using the overall performance of the mainstream characteristic parameters within the link as the core judgment criterion, while also considering the influence weight of individual abnormal parameters. The work of organizing the judgment criteria is completed.

[0066] Extract all feature parameters from the feature association chain that have completed data normalization and format standardization, and use standard feature parameters (i.e., the benchmark feature parameters corresponding to normal and qualified medical images, which serve as the unified reference benchmark for the entire process of offset calculation) as the comparison standard. Perform comparative analysis on each feature parameter in turn, calculate the offset of each parameter relative to the standard value, and finally obtain the overall magnitude of the deviation of the entire set of feature parameters from the standard value.

[0067] The calculated deviation of the feature parameters is compared and matched one by one with the compiled standard intervals for each level of feature parameters. According to the interval assignment logic: the intervals of each level are sorted in ascending order of deviation, and the intervals are independent and non-overlapping. Based on the deviation value, the interval to which it uniquely belongs is determined. If the deviation falls exactly at the boundary between two intervals, it is defaulted to the higher-level interval. The specific interval range corresponding to the current parameter deviation is determined, and the preliminary matching result is locked. Combining the previously compiled level determination rules and the level interval matching results, the final level classification is completed. Based on the comprehensive judgment result, the specific severity of the current medical image quality defect is determined, and the judgment conclusion is output.

[0068] In this optional embodiment, the quantitative analysis of the types of quality defects in the medical images to be evaluated and their corresponding severity, identifying key factors affecting the imaging quality of the medical images to be evaluated, includes: constructing a quantitative analysis system for medical image features; initializing all feature parameters after data normalization and format unification to an unanalyzed state, and simultaneously setting the matching markers of quality defects corresponding to each feature parameter to an unmatched state; selecting initial baseline feature parameters from all feature parameters, classifying feature parameter groups based on a preset feature matching degree threshold, and calculating the ratio of the representation contribution to the matching strength of each feature parameter group after removing redundant interference; and selecting feature parameters with a matching degree greater than the baseline feature parameter. Feature parameters with a preset feature matching threshold are grouped into the same feature group. The currently selected benchmark feature parameters are marked as analyzed. Among all unanalyzed feature parameters, the feature parameters with the smallest ratio of representation contribution to matching strength and the highest defect abnormality tendency are selected as the benchmark feature parameters for the next stage. The quality defect type and its corresponding severity of the medical image to be evaluated are simultaneously associated. The process of selecting benchmark feature parameters and dividing feature groups is repeated iteratively until all feature parameters have been traversed and analyzed. Combining the quality defect type, defect severity and defect abnormality tendency corresponding to each feature group, the key factors affecting the imaging quality of the medical image to be evaluated are comprehensively identified.

[0069] Specifically, a medical image feature quantitative analysis system is established (i.e., an analytical framework for grouping, statistically analyzing, scoring, and associating defects among various image feature parameters). All image feature parameters that have already been standardized and formatted are retrieved, and all parameters are uniformly marked as "unanalyzed," indicating that this group of parameters has not yet participated in benchmark selection, grouping, or other analytical operations; this is the initial default state. Simultaneously, the quality defect matching record corresponding to each feature parameter is uniformly set to "unmatched," indicating that the current parameter has not yet established an association with any defect type or severity, thus completing the state initialization work before analysis.

[0070] Initial baseline feature parameters are selected from all feature parameters. The parameters with the strongest representativeness from the core image region are chosen according to predetermined requirements and used as the reference standard for group comparison. A preset feature matching threshold is enabled. This threshold is the criterion for dividing feature groups, with a value range of 60-85. Parameters with a matching degree higher than this threshold are considered to have a strong correlation and are assigned to the same feature group. All parameters are grouped according to this standard.

[0071] After grouping, the data within each parameter group is cleaned up, and redundant interference is removed. The ratio of the representational contribution to the matching strength of each parameter group is calculated. The representational contribution reflects the extent to which the parameter group reflects the image quality and defect characteristics; the matching strength ratio reflects the degree of correlation between the parameter group and the defect features. Both data are important bases for judging the analysis results. Parameters whose similarity to the initial baseline parameters exceeds the set standard are grouped into the same feature group.

[0072] Change the initial baseline parameters used in this round from "unanalyzed" to "analyzed". From the remaining unanalyzed parameters, select the baseline parameters for the next round: prioritize parameters with the lowest ratio of characterization contribution to matching strength and the highest defect anomaly tendency. Defect anomaly tendency represents the degree to which the parameter deviates from the normal standard and exhibits defects; the higher the tendency, the more likely it is to cause imaging anomalies. After selecting the new baseline parameters, link them to the defect types and severity found in the current images, and maintain the corresponding relationship.

[0073] The entire process of selecting baseline parameters, dividing parameter groups, cleaning data, calculating indicators, updating status, and binding defect information is repeatedly executed. This process is continuously looped until all feature parameters are marked as analyzed, at which point the loop stops. The defect types, severity levels, and abnormal tendencies of each parameter group are summarized, and all analysis results are integrated to identify key factors affecting the quality of medical imaging.

[0074] In this optional embodiment, initial benchmark feature parameters are selected from all feature parameters. Feature parameter grouping is completed based on a preset feature matching threshold, and the ratio of representation contribution to matching strength after removing redundant interference for each feature parameter group is calculated. This includes: selecting parameters with benchmark representation characteristics from all feature parameters as initial benchmark feature parameters; matching and verifying all remaining feature parameters with the initial benchmark feature parameters based on the preset feature matching threshold, and grouping all feature parameters according to the matching results to form multiple independent feature parameter groups; performing redundant interference removal processing on each of the divided feature parameter groups, clearing invalid, duplicate, and interfering parameter data within the group, and retaining valid feature parameters that reflect image quality defects; and calculating the overall representation contribution to matching strength ratio for each feature parameter group after removing redundant interference.

[0075] Specifically, among all the image feature parameters that have been standardized and formatted, a single parameter is selected as the initial baseline feature parameter, serving as the unified reference for subsequent matching and grouping. A preset feature matching threshold is applied, which serves as the boundary standard for determining the degree of correlation between parameters, with a value range of 60-85. Parameters with a matching degree higher than this threshold are considered to have a strong correlation, while those lower are considered to have a weak correlation. All remaining feature parameters are then matched and verified against the initial baseline feature parameter one by one. Based on the matching results, grouping and aggregating are completed, forming multiple independent feature parameter groups with no overlapping data.

[0076] For each group of feature parameters that has been divided, a redundancy and interference removal operation is performed sequentially. Redundancy is determined according to the following criteria: invalid parameters are those with missing data or values ​​outside the normal range; duplicate parameters are those with completely identical content; and interfering parameters are those that deviate from the overall feature pattern and affect defect assessment. Data within each group is cleared, retaining only valid feature parameters that truly reflect defects in medical image quality.

[0077] For each feature parameter group after redundancy removal, two sets of core indicators are calculated. The overall representation contribution measures the comprehensive ability of the effective parameters in this group to reflect image quality and defect features; a higher value indicates greater reference value. The matching strength ratio measures the overall correlation between the parameters in this group and their corresponding defect features, reflecting the strength of the correspondence between features and defects. The indicator calculations for all feature parameter groups are then completed.

[0078] Figure 2 An embodiment of a deep learning-based medical image quality assessment system of the present invention is shown.

[0079] In this optional embodiment, the deep learning-based medical image quality assessment system includes: The image processing module 201 is used to preprocess the pre-acquired medical images to be evaluated and extract medical image features. The defect analysis module 202 is used to perform global feature correlation analysis on medical image features, determine the type of quality defect in the medical image to be evaluated, and determine the severity of the corresponding quality defect type. The factor analysis module 203 is used to quantitatively analyze the types of quality defects in the medical images to be evaluated and their corresponding severity, and to identify the key factors affecting the imaging quality of the medical images to be evaluated. The evaluation result output module 204 is used to conduct a comprehensive assessment based on key factors and preset medical image quality evaluation indicators, and output the final quality evaluation result of the medical image to be evaluated.

[0080] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0081] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0082] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0083] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0084] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0085] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A deep learning-based method for evaluating the quality of medical images, characterized in that, include: The pre-acquired medical images to be evaluated are preprocessed, and medical image features are extracted. Global feature association analysis is performed on medical image features to determine the type of quality defect in the medical image to be evaluated and to determine the severity of the corresponding quality defect type. A quantitative analysis of the types of quality defects in the medical images to be evaluated and their corresponding severity was conducted to identify the key factors affecting the imaging quality of the medical images to be evaluated. Based on key factors and combined with preset medical image quality evaluation indicators, a comprehensive assessment is conducted to output the final quality evaluation result of the medical image to be evaluated.

2. The deep learning-based medical image quality assessment method according to claim 1, characterized in that, The process of preprocessing the pre-acquired medical images to be evaluated and extracting medical image features includes: The medical images to be evaluated are denoised, and the brightness and grayscale deviations of the denoised medical images are corrected. The corrected medical image to be evaluated is enhanced to obtain a medical image with enhanced contrast and clarity. The enhanced medical image to be evaluated is converted to grayscale to obtain a grayscale image, and the edge gradient of each pixel in the grayscale image is calculated to generate a gradient image. Based on gradient images, quality defect feature regions are delineated, and the area, morphological parameters, and gradient intensity statistics of these quality defect feature regions are calculated to construct defect quantification features. The quality defect feature region is binarized, edge information is extracted and defect contours are generated, and the defect contours are fused with the defect quantification features to obtain medical image features.

3. The deep learning-based medical image quality assessment method according to claim 1, characterized in that, The step of performing global feature correlation analysis on medical image features to determine the type of quality defect in the medical image to be evaluated and to determine the severity of the corresponding quality defect type includes: A baseline feature region is selected from medical image features as the initial analysis object and included in the set to be analyzed. Check if there are any valid analysis objects in the set to be analyzed. If there are no valid analysis objects, terminate the feature analysis; if there are valid analysis objects, perform subsequent operations. Select the region with the best comprehensive feature parameters from the set to be analyzed as the current analysis region, and then determine and process the surrounding related feature regions of the current analysis region one by one; Move the current analysis area into the completed analysis set and check whether the target defect feature area is included in the set to be analyzed; if not included, continue feature analysis; if included, terminate feature retrieval. Starting from the target defect feature area, the system traces back to the initial analysis object according to the reference benchmark corresponding to each area to determine the quality defect type of the medical image to be evaluated and to determine the severity of the corresponding quality defect type.

4. The deep learning-based medical image quality assessment method according to claim 3, characterized in that, The step of selecting the region with the optimal comprehensive feature parameters from the set to be analyzed as the current analysis region, and then determining and processing the surrounding related feature regions of the current analysis region one by one, includes: If the surrounding related feature regions have already been included in the completed analysis set, no operation will be performed; If the surrounding related feature regions are not included in the set to be analyzed, then the surrounding related feature regions will be included in the set to be analyzed, and the comprehensive feature parameters and basic feature parameters of the surrounding related feature regions will be calculated using the current analysis region as a reference. If the surrounding related feature regions are already within the set to be analyzed, and the basic feature parameters of the surrounding related feature regions are better than the original values ​​after using the current analysis region as a reference, then the reference is updated, and the comprehensive feature parameters and basic feature parameters of the surrounding related feature regions are recalculated.

5. The deep learning-based medical image quality assessment method according to claim 3, characterized in that, Starting from the target defect feature region, the process involves tracing back to the initial analysis object according to the reference benchmark corresponding to each region to determine the type of quality defect in the medical image to be evaluated, and judging the severity of the corresponding quality defect type, including: Extract all feature parameters of the target defect feature region and simultaneously retrieve the reference benchmark information corresponding to each region on the tracing path; By tracing back layer by layer based on the reference benchmark information, the complete feature association link from the initial analysis object to the target defect feature area is restored; All feature parameters within the feature association link are compared and matched with a preset standard defect sample library, and the type of quality defect corresponding to the medical image to be evaluated is determined based on the matching results. Based on the preset defect severity grading index and combined with the deviation of characteristic parameters from the standard value, the severity of the quality defect type is determined.

6. The deep learning-based medical image quality assessment method according to claim 5, characterized in that, The step of comparing and matching all feature parameters within the feature association link with a preset standard defect sample library, and determining the type of quality defect corresponding to the medical image to be evaluated based on the matching results, includes: Extract all feature parameters within the feature association link, and perform data normalization and format unification processing on all feature parameters; Call the preset standard defect sample library and load the standard feature samples corresponding to various defects; After data normalization and format unification, all feature parameters are compared with standard feature samples one by one to obtain the feature fit of all feature parameters with standard feature samples. Feature classification is performed based on feature fit, and the type of quality defect corresponding to the medical image to be evaluated is determined.

7. The deep learning-based medical image quality assessment method according to claim 5, characterized in that, The determination of the severity of a quality defect type based on a preset defect severity grading index, combined with the deviation of characteristic parameters from standard values, includes: Retrieve pre-configured defect severity grading indicators and organize the standard range of characteristic parameters and grading rules corresponding to each severity level; Extract all feature parameters from the feature association link after data normalization and format unification. Using the standard feature parameters as a reference, calculate the offset of each feature parameter relative to the standard value to obtain the magnitude of the feature parameter's deviation from the standard value. The deviation of the characteristic parameters from the standard value is compared with the defect severity grading index one by one to determine the grade range to which the deviation belongs; By combining the grading rules with the grading range matching results, the grading is completed, and the severity of the quality defect type is determined.

8. The deep learning-based medical image quality assessment method according to claim 1, characterized in that, The quantitative analysis of the types of quality defects in the medical images to be evaluated and their corresponding severity identifies key factors affecting the imaging quality of the medical images to be evaluated, including: A medical image feature quantitative analysis system was constructed. All feature parameters after data normalization and format unification were initialized to an unanalyzed state. At the same time, the matching marks of quality defects corresponding to each feature parameter were uniformly set to an unmatched state. Select initial baseline feature parameters from all feature parameters, complete the feature parameter grouping based on the preset feature matching degree threshold, and calculate the ratio of the representation contribution to the matching strength of each feature parameter group after removing redundant interference. Feature parameters with a matching degree greater than the preset feature matching degree threshold are classified into the same feature group. The currently selected baseline feature parameters are marked as analyzed. Among all unanalyzed feature parameters, the feature parameters with the smallest ratio of characterization contribution to matching strength and the highest defect abnormality tendency are selected as the baseline feature parameters for the next stage. The quality defect types of the medical images to be evaluated and their corresponding severity are simultaneously associated. The process of selecting baseline feature parameters and dividing feature groups is repeated iteratively until all feature parameters have been traversed and analyzed. By combining the quality defect type, defect severity and defect abnormality tendency corresponding to each feature group, the key factors affecting the imaging quality of the medical image to be evaluated are comprehensively identified.

9. The deep learning-based medical image quality assessment method according to claim 8, characterized in that, The process of selecting initial baseline feature parameters from all feature parameters, classifying feature parameter groups based on a preset feature matching degree threshold, and calculating the ratio of representation contribution to matching strength for each feature parameter group after removing redundant interference includes: Parameters with benchmark characterization properties are selected from all feature parameters and used as initial benchmark feature parameters; Based on the preset feature matching threshold, all other feature parameters are matched and verified with the initial benchmark feature parameters. Based on the matching results, all feature parameters are grouped and formed into multiple independent feature parameter groups. Redundancy and interference removal are performed on each of the divided feature parameter groups one by one to remove invalid, duplicate and interference parameter data within the group, and retain the effective feature parameters that can reflect image quality defects. Based on each feature parameter group after removing redundant interference, the ratio of overall representation contribution to matching strength for each feature parameter group is calculated.

10. A deep learning-based medical image quality assessment system, characterized in that, include: The image processing module is used to preprocess the pre-acquired medical images to be evaluated and extract medical image features. The defect analysis module is used to perform global feature correlation analysis on medical image features, determine the type of quality defect in the medical image to be evaluated, and determine the severity of the corresponding quality defect type. The factor analysis module is used to quantitatively analyze the types of quality defects in the medical images to be evaluated and their corresponding severity, and to identify the key factors that affect the imaging quality of the medical images to be evaluated. The evaluation result output module is used to conduct a comprehensive assessment based on key factors and preset medical image quality evaluation indicators, and output the final quality evaluation result of the medical image to be evaluated.

Images