A regional ASF image data processing method and system based on multi-modal data fusion

Abstract

The present application relates to the technical field of ASF image data processing, in particular to a regional ASF image data processing method and system based on multi-modal data fusion, which realizes efficient and accurate ASF image analysis through a three-step collaborative process: step one, fuse multi-modal data to generate regional ASF images and evaluate quality, ensure that subsequent analysis is based on qualified images, and reduce the interference of inferior data; step two, extract features and dynamically monitor and optimize the extraction process to improve feature accuracy and reliability, and avoid missing or misreporting key information; step three, fuse features and dynamic data to construct a time series matrix, quantify feature importance to output prediction results, and enhance the ability to capture ASF dynamic changes. Through multi-modal fusion and dynamic optimization, the method improves the comprehensiveness of image quality evaluation, the accuracy of feature extraction, and the reliability of prediction results, providing strong support for accurate analysis and prediction of regional ASF.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ASF image data processing, in particular to a regional ASF image data processing method and system based on multi-modal data fusion. BACKGROUND

[0002] The existing regional ASF image data processing technology based on multi-modal data fusion is to integrate multi-source clinical data such as patient demographic information, clinical history records, laboratory test indicators, and imaging examination results, to build a multi-dimensional analysis system from macro symptom characteristics to tissue structure. In terms of technical implementation, a fusion strategy at the feature level, decision level or mixed level is adopted, combined with deep learning models (such as the double attention mechanism of DRIFA-Net and the multi-task learning of PanDerm) to realize cross-modal information complementation, and rigid / elastic registration technology is used to solve the spatial misalignment problem between modalities. To adapt to the clinical scene, weakly supervised learning is introduced to reduce the dependence on labeling, domain adaptation technology is used to overcome equipment differences, and Monte Carlo Dropout and attention visualization are used to improve model interpretability

[0003] For example, the Chinese invention patent with publication number CN116645347B discloses a diabetic foot prediction method, device and equipment based on multi-view feature fusion, which includes the following steps: obtaining sample information of the diabetic foot to be predicted, including foot infrared thermal map and foot color image; constructing a deep learning model: based on a deep learning model generated using transfer learning, applying a receptive field optimization algorithm to construct a temperature distribution feature enhancement module, applying a receptive field optimization algorithm to construct a skin texture feature enhancement module, constructing a spatial attention enhancement module and a multi-view feature fusion module to obtain a deep learning model to be trained.

[0004] For example, the Chinese invention patent with publication number CN118396933A discloses a skin disease detection method based on YOLO-IRLSK lightweight model, which includes: obtaining a skin disease image data set to be detected; designing a YOLO-IRLSK model based on a YOLO model, which is composed of an MC module and a double-layer internal fusion block IRLSK. The double-layer internal fusion block can more completely extract the internal features of the picture, and the model optimization technology makes the backbone network more lightweight. Input the data in the training data set into the detection model to obtain a region prediction map, and perform iterative training. Input the skin disease picture to be recognized, and detect and recognize the skin disease through the YOLO-IRLSK network model obtained by training.

[0005] But in the process of implementing the embodiments of the application, the above-mentioned technologies at least have the following technical problems: the uneven surface and abnormal optical properties caused by skin edema and exudation of ASF patients, the difficulty of fixed parameters of superimposed imaging equipment to adapt to the diversity of lesions, and the interference of environmental micro-movement and light source fluctuation; low feature and background differentiation, difficulty in multi-modal data segmentation and screening, significant format and scale heterogeneity of image, spectrum, and clinical data, prominent data noise and missing value problems; dynamic changes are difficult to predict, and due to the time-dependent nature of ASF progression (such as the chain reaction of ischemia-necrosis) and the interference of sudden factors, the modeling ability of traditional static models is exceeded, which further limits the prediction accuracy. SUMMARY

[0006] In order to solve the technical problems in the prior art that the poor and unstable image quality of ASF leads to distortion of subsequent analysis, low feature extraction accuracy causes multi-modal data fragmentation and information redundancy / missing, traditional models cannot capture time dependence, and ASF dynamic changes are difficult to predict, the embodiments of the present application provide a regional ASF image data processing method and system based on multi-modal data fusion. The technical solution is as follows:

[0007] On the one hand, a regional ASF image data processing method based on multi-modal data fusion is provided, which comprises the following steps: step one, acquiring multi-modal data of a target region, and performing multi-modal data fusion to generate a regional ASF image, acquiring and analyzing quality parameters of the regional ASF image, and determining whether to mark the regional ASF image as a qualified image; step two, extracting a target feature set from the qualified image, monitoring extraction process parameters of the target feature set, and determining whether to perform data optimization on the extraction process of the target feature set; step three, performing data fusion on each feature in the target feature set and the acquired dynamic data, and constructing a time series matrix, thereby evaluating the importance score of each feature in the target feature set, and outputting a prediction result of the regional ASF image based on the importance score of each feature.

[0008] In another aspect, a regional ASF image data processing system based on multi-modal data fusion is provided, which comprises: a multi-modal fusion and quality screening module, configured to acquire multi-modal data of a target region, perform multi-modal data fusion, generate a regional ASF image, acquire and analyze quality parameters of the regional ASF image, and determine whether to mark the regional ASF image as a qualified image; a feature extraction and optimization control module, configured to extract a target feature set from the qualified image, monitor extraction process parameters of the target feature set, and determine whether to perform data optimization on the extraction process of the target feature set; and a dynamic fusion and prediction evaluation module, configured to perform data fusion on each feature in the target feature set and acquired dynamic data, construct a time series matrix, evaluate importance scores of each feature in the target feature set, and output a prediction result of the regional ASF image based on the importance scores of the features.

[0009] The technical scheme provided by the embodiments of the present application has at least the following beneficial effects:

[0010] (1) The present application provides a regional ASF image data processing method and system based on multi-modal data fusion, which breaks the information limitation of single data mode through a closed-loop process of "multi-modal fusion-feature optimization-dynamic evaluation", realizes full-chain quality control and dynamic optimization from raw data to prediction results, integrates multi-source information, accurately selects high-quality images, optimizes feature extraction accuracy, evaluates feature importance combined with dynamic change rules, and finally improves the objectivity, accuracy and timeliness of regional ASF image analysis, providing reliable quantitative basis for clinical monitoring and intervention of ASF lesions.

[0011] (2) The regional ASF image is generated by fusing the multi-modal data of the target region, and the image eligibility is determined based on the quality parameters, which can effectively avoid the one-sidedness of single modal data (such as neglecting subcutaneous structure information by relying only on visual images); in actual operation, imaging parameters can be dynamically adjusted for unqualified images to ensure that the images input for subsequent analysis meet the standards in terms of edge sharpness, stereoscopic shape restoration, and microscopic texture integrity, thereby reducing feature extraction deviations caused by image quality defects from the source and laying a high-quality data foundation for subsequent analysis.

[0012] (3) When extracting the target feature set from the qualified image, the extraction process parameters are monitored to timely detect abnormalities in feature extraction (such as artifact interference); for extraction deviations, data optimization can be performed by adjusting the imaging system parameters and introducing adaptive optical correction technology to ensure that the extracted target feature set truly reflects the core attributes of ASF lesions (such as necrotic focus range), avoiding deviations of subsequent evaluation results caused by feature distortion. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort based on these drawings.

[0014] Figure 1 is a flow chart of a regional ASF image data processing method based on multi-modal data fusion provided by the embodiments of the present application;

[0015] Figure 2 is a schematic diagram of a regional ASF image data processing system based on multi-modal data fusion provided by the embodiments of the present application;

[0016] Figure 3 is a flow chart of a method for judging whether a regional ASF image is a qualified image provided by the present application;

[0017] Figure 4 is a flow chart of a secondary unqualified image optimization process provided by the present application;

[0018] Figure 5 is a flow chart of a process for judging whether to perform data optimization on the extraction process of a target feature set provided by the present application;

[0019] Figure 6 is a flow chart of a data optimization process provided by the present application;

[0020] Figure 7 is a flow chart of a model construction and result output process provided by the present application;

[0021] Figure 8 is an unprocessed normal skin ultrasound image provided by the embodiments of the present application;

[0022] Figure 9 is a processed normal skin ultrasound image provided by the embodiments of the present application;

[0023] Figure 10 is an unprocessed ASF ultrasound image provided by the embodiments of the present application;

[0024] Figure 11 is a processed ASF ultrasound image provided by the embodiments of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the present application will be described below with reference to the drawings.

[0026] In the embodiments of the present application, the words such as "exemplary", "for example", etc. are used to represent an example, an illustration, or a description. Any embodiment or design solution described as "exemplary" in the present application should not be interpreted as more preferred or having more advantages than other embodiments or design solutions. Rather, the word "exemplary" is intended to present the concept in a specific manner. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two options.

[0027] In the embodiments of the present application, "image" and "picture" can be used interchangeably at times. It should be pointed out that the meanings expressed are consistent when the distinction is not emphasized. "Of", "corresponding" and "corresponding" can be used interchangeably at times. It should be pointed out that the meanings expressed are consistent when the distinction is not emphasized.

[0028] In the embodiments of the present application, subscripts such as W1 may be written in non-subscript form such as W1 at times. The meanings expressed are consistent when the distinction is not emphasized.

[0029] In order to make the technical problems, technical solutions and advantages of the present application clearer, specific embodiments will be described in detail below with reference to the accompanying drawings.

[0030] The embodiments of the present application provide a regional ASF image data processing method based on multi-modal data fusion, as shown in Figure 1A regional ASF image data processing method flowchart based on multi-modal data fusion is shown. The processing flow of the method can include the following steps: step one, acquiring multi-modal data of a target region, performing multi-modal data fusion, generating a regional ASF image, acquiring and analyzing quality parameters of the regional ASF image, and determining whether to mark the regional ASF image as a qualified image; the multi-modal data refers to data dynamically collected every 8 hours within 7 days after admission, specifically including: ultrasound-related data: peak systolic blood flow velocity (PSV), end diastolic blood flow velocity (EDV), mean blood flow velocity (Mean Velocity), time-averaged maximum velocity (TAMX), resistance index (RI), pulsatility index (PI), blood flow (Q), area index (CSA), and vascular diameter change rate of local tissue microarteries (palmar superficial arch artery, digital artery, and tibial artery); near-infrared spectral parameters: local regional blood oxygen saturation (rSO2), dynamic change (ArSO2), and rSO2 curve area (AUC); step two, extracting a target feature set from the qualified image, monitoring the extraction process parameters of the target feature set, and determining whether to optimize the extraction process of the target feature set; step three, performing data fusion on each feature in the target feature set and the acquired dynamic data, and constructing a time series matrix, thereby evaluating the importance score of each feature in the target feature set, and outputting a prediction result of the regional ASF image based on the importance score of each feature.

[0031] It should be explained that the English full name of ASF is Acute Skin Failure, and the Chinese name is acute skin failure. As a common skin failure symptom and skin complication in ICU, it is closely related to factors such as long-term bedridden of critically ill patients, local skin compression, blood circulation disorder, poor nutritional status, and often shows symptoms such as skin and subcutaneous tissue damage and necrosis, which has a greater impact on the prognosis and quality of life of patients; the above-mentioned regional ASF image should cover the morphological features (edge, contour) of the target region.

[0032] Specifically, whether to mark the regional ASF image as a qualified image is determined by analyzing the quality parameters of the regional ASF image, obtaining a regional ASF image quality index, and comparing it with a preset regional ASF image quality threshold in a prediction database; the regional ASF image quality threshold represents the minimum value of the regional ASF image quality index within a specified range.

[0033] If the regional ASF image quality index is greater than or equal to the regional ASF image quality threshold, the regional ASF image is marked as a qualified image.

[0034] If the regional ASF image quality index is less than the regional ASF image quality threshold value, it is judged that the regional ASF image is marked as an unqualified image, the unqualified image is denoised, the regional ASF image quality deviation value is obtained based on the regional ASF image quality index and the regional ASF image quality threshold value, and the spherical aberration coefficient reduction coefficient is matched based on the regional ASF image quality deviation value, so as to reduce the spherical aberration coefficient of the objective lens in the imaging system, significantly improve the image edge definition and detail resolution, reduce the blur, halo and other artifacts caused by spherical aberration, and reduce noise interference, so that the edge contrast, surface fitting error, fractal dimension and other quality parameters can more truly reflect the image characteristics, and the accuracy of the regional ASF quality index evaluation is improved. In addition, the stability and consistency of long-term imaging of the system can be enhanced, and a reliable basis is provided for subsequent image analysis and comparison.

[0035] The above denoising processing refers to denoising processing by using non-local mean filtering, which effectively suppresses noise (such as speckle noise), while better preserving the edge, texture and other detail characteristics of the image, avoiding excessive smoothing that leads to loss of key pathological information, and providing a more reliable image basis for subsequent quality deviation calculation and spherical aberration coefficient adjustment. The regional ASF image quality deviation value refers to the result of subtracting the regional ASF image quality index from the regional ASF image quality threshold value. The spherical aberration coefficient reduction coefficient is matched based on the regional ASF image quality deviation value, and the specific matching process is as follows: the spherical aberration coefficient reduction coefficients corresponding to each regional ASF image quality deviation value interval are stored in the prediction database, the obtained regional ASF image quality deviation value is input into the prediction database, the prediction database can match the corresponding regional ASF image quality deviation value interval, and then the spherical aberration coefficient reduction coefficient corresponding to the interval is the required reduction coefficient. The obtained spherical aberration coefficient reduction coefficient is multiplied by the original spherical aberration coefficient, and the result is the spherical aberration coefficient that needs to be adjusted. The spherical aberration coefficient reduction coefficient is less than 1, indicating the proportion of the spherical aberration coefficient of the objective lens in the imaging system that needs to be reduced. The spherical aberration coefficient is a quantitative parameter for describing the severity of spherical aberration in an optical system, which is used to measure the focusing deviation of different aperture rays caused by the fact that the surface of a lens or a mirror is a spherical surface (rather than an ideal aspherical surface). The larger the spherical aberration coefficient, the more significant the focusing deviation of different aperture rays, and the worse the imaging definition.

[0036] The regional ASF image quality index of the unqualified image after denoising processing is obtained, which is marked as a regional ASF image quality review index, and it is judged whether the unqualified image is marked as a secondary unqualified image. The quality parameters of the regional ASF image include the edge contrast factor of the regional ASF image, the surface fitting error factor of the regional ASF image, and the fractal dimension factor of the regional ASF image.

[0037] The edge contrast factor represents a deviation relationship between the edge contrast of the regional ASF image and a corresponding reference value; the surface fitting error factor represents a ratio between the surface fitting error value of the regional ASF image and a corresponding defined value; and the fractal dimension factor represents a ratio between the fractal dimension value of the regional ASF image and a corresponding defined value.

[0038] The action intensity coefficients of the edge contrast factor, the surface fitting error factor and the fractal dimension factor are set in the prediction database to quantify the influence contribution values of the factors to the regional ASF image quality index, and finally the regional ASF image quality index is obtained by weighting and comprehensively combining the influence contribution values, wherein the regional ASF image quality index represents a comprehensive advantage and disadvantage degree of the regional ASF image in the key quality dimension, and the specific evaluation method is as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] In the formula, QALK is the regional ASF image quality index, TFX is the edge contrast factor of the regional ASF image, RGN is the edge contrast of the regional ASF image, UHV is a preset reference edge contrast in the prediction database, CDE is the surface fitting error factor of the regional ASF image, XFT is the surface fitting error value of the regional ASF image, VSE is a defined surface fitting error value preset in the prediction database, MJY is the fractal dimension factor of the regional ASF image, NKP is the fractal dimension value of the regional ASF image, BHU is a preset reference fractal dimension value in the prediction database, gt is an action intensity coefficient corresponding to the edge contrast factor preset in the prediction database, gc is an action intensity coefficient corresponding to the surface fitting error factor preset in the prediction database, and gm is an action intensity coefficient corresponding to the fractal dimension factor preset in the prediction database.

[0044] The edge contrast represents the degree of differentiation of the edge, the edge is located by edge detection, the pixel value difference (such as the mean difference) of the two sides is extracted and calculated, the surface fitting error value reflects the approximation effect of the fitting model on the data, the discrete data points are fitted by selecting a surface model, the deviation (such as the root mean square error) between the actual value and the fitted value is calculated after solving the parameters by the least square method, and the fractal dimension represents the irregularity and complexity of the self-similarity of the complex structure (such as texture), the target area is covered by boxes of different sizes, the number of non-empty boxes is counted, and the absolute value of the slope is obtained by logarithmic linear regression fitting.

[0045] The reference edge contrast represents a reference value of the edge contrast; the defined surface fitting error value represents a maximum value of the surface fitting error value within a specified range; and the reference fractal dimension value represents a reference value of the fractal dimension.

[0046] A high fractal dimension generally corresponds to a complex structure (such as a multi-contour). If the structure contains a large number of strong edges (high contrast), the fractal dimension is high and the edge contrast is high. If it is a weak edge (such as a blurred texture), the fractal dimension is high but the edge contrast is low. A high edge contrast means that there is a significant gray level change locally, which is easy to deviate from the overall fitting surface (such as a smooth surface model), resulting in an increase in fitting error. A high fractal dimension corresponds to an irregular and complex structure, which is difficult to approximate by a simple surface (such as a quadratic surface), resulting in a larger fitting error.

[0047] The edge contrast factor corresponds to an action intensity coefficient, which represents the corresponding change in the regional ASF image quality index when the factor changes by a unit. The surface fitting error factor corresponds to an action intensity coefficient, which represents the corresponding change in the regional ASF image quality index when the factor changes by a unit. The fractal dimension factor corresponds to an action intensity coefficient, which represents the corresponding change in the regional ASF image quality index when the factor changes by a unit. The prediction database stores the mapping relationship between the edge contrast factor and the corresponding action intensity coefficient, the mapping relationship between the surface fitting error factor and the corresponding action intensity coefficient, and the mapping relationship between the fractal dimension factor and the corresponding action intensity coefficient. For example, the edge contrast factor, the surface fitting error factor, and the fractal dimension factor are input into the prediction database, and the prediction database generates the corresponding action intensity coefficient of the edge contrast factor, the action intensity coefficient of the surface fitting error factor, and the action intensity coefficient of the fractal dimension factor based on a preset mapping rule. The numerical range of each type of action intensity coefficient is strictly controlled between 0 and 1.

[0048] The smaller the edge contrast factor, the closer the edge is to the ideal state, and the better the edge quality of the regional ASF image. The smaller the surface fitting error factor, the smaller the actual error, and the closer the image structure (such as the spatial form) is to the expected regular surface (good fitting effect), and the better the edge quality of the regional ASF image. A high fractal dimension factor means that the image structure is too complex (such as noise interference), deviating from the normal form, and the order of the structure of the image is worse, and the edge quality of the regional ASF image is worse.

[0049] In a specific embodiment, the regional ASF image is generated by integrating multi-modal data of the target region, and whether the image is qualified is determined according to the quality parameter, which can fully absorb the advantages of different modal data and break through the limitations of a single data dimension; in actual operation, for unqualified images, the imaging strategy can be dynamically adjusted in combination with the specific deviation of the quality parameter, which can not only improve the edge sharpness, stereoscopic shape authenticity and microscopic texture details of the image, but also reduce the interference of artifacts through cross-validation of multi-modal data; this way ensures that the image input into subsequent analysis has both “structural integrity” and “information authenticity” from the data source, providing a reliable data foundation for accurately extracting ASF key features (such as subtle texture changes of early inflammation), and avoiding feature misjudgment or missed detection caused by image quality problems.

[0050] Specifically, whether to mark the unqualified image as a secondary unqualified image is determined by comparing the regional ASF image quality review index with the regional ASF image quality threshold; if the regional ASF image quality review index is greater than or equal to the regional ASF image quality threshold, the unqualified image is not marked as a secondary unqualified image, and is marked as a qualified image; if the regional ASF image quality review index is less than the regional ASF image quality threshold, the unqualified image is marked as a secondary unqualified image, and the secondary unqualified image is optimized.

[0051] As Figure 3 The method for determining whether the regional ASF image is a qualified image is shown in the flowchart, which starts with obtaining multi-modal data of the target region, fusing to generate a regional ASF image, collecting a regional ASF image quality index, and comparing it with a regional ASF image quality threshold: if the regional ASF image quality index is greater than or equal to the regional ASF image quality threshold, it is directly marked as a qualified image, and a feature extraction accuracy index in the qualified image is obtained; otherwise, it is marked as an unqualified image and is subjected to noise reduction processing, and a spherical aberration coefficient reduction coefficient is matched based on the regional ASF image quality deviation value to reduce the spherical aberration coefficient of the imaging system objective lens; then, a regional ASF image quality review index is obtained, if the regional ASF image review index is greater than or equal to the regional ASF image quality threshold, it is not marked as a secondary unqualified image, is marked as a qualified image, and a feature extraction accuracy index in the qualified image is obtained; otherwise, it is marked as a secondary unqualified image.

[0052] Further, the secondary substandard image is optimized, and the specific optimization process is as follows: based on the regional ASF image quality review index and the regional ASF image quality threshold, a quality review deviation value is obtained, a coma coefficient reduction coefficient is matched based on the quality review deviation value, so as to reduce the coma coefficient of the optical system, and eliminate the comet-like trailing blur of the off-axis image point--this phenomenon will cause the object details at the edge of the field of view (such as the edge features of the ASF lesion area) to appear asymmetric blur, interfere with the morphology judgment, improve the image contrast and resolution of the off-axis region, and make the fine structure at the edge of the field of view (such as the boundary of the irregular sinus) clear and distinguishable, while expanding the effective imaging field of view of the system (maintaining imaging consistency in a larger range), and reducing the difficulty of subsequent feature extraction (such as edge positioning); based on the quality review deviation value, a reduction amount of optical coherence length is matched, so as to reduce the optical coherence length of the regional ASF image, enhance the resolution of the image to the shallow structure, and after the optical coherence length is shortened, the imaging system can only capture the coherent signals within a specific depth range, which can effectively filter out the stray light interference of the deep tissue, reduce the image blur caused by the superposition of multi-layer structures, make the edge of the target area (such as the shallow necrotic focus of the ASF lesion) more sharp and the details more clear; at the same time, the shorter coherence length can improve the axial resolution, accurately distinguish the fine structure differences between adjacent tissues, avoid feature misjudgment caused by depth aliasing, further optimize the quality of the secondary substandard image, and thus improve the subsequent feature extraction accuracy index.

[0053] The quality review deviation value refers to the result of subtracting the regional ASF image quality review index from the regional ASF image quality threshold; the coma coefficient reduction coefficient is matched based on the quality review deviation value, and the specific matching process is as follows: the coma coefficient reduction coefficients corresponding to each quality review deviation value interval are stored in a prediction database, the obtained quality review deviation value is input into the prediction database, the prediction database can match the corresponding quality review deviation value interval, and thus the coma coefficient reduction coefficient corresponding to the interval is the required reduction coefficient; the coma coefficient reduction coefficient is less than 1, indicating that the proportion of the coma coefficient of the optical system needs to be reduced.

[0054] The reduction amount of the optical coherence length is matched based on the quality review deviation value, so as to reduce the optical coherence length of the regional ASF image, and the specific matching process is as follows: the reduction amount of the optical coherence length corresponding to each quality review deviation value interval is stored in a prediction database, the obtained quality review deviation value is input into the prediction database, the prediction database can match the corresponding reduction amount of the optical coherence length, and the original optical coherence length is reduced by the obtained reduction amount of the optical coherence length, and the result is the required adjusted optical coherence length.

[0055] The system obtains the regional ASF image quality index of the image that fails twice, marks it as the regional ASF image quality reassessment index, and determines whether to mark the image that fails twice as invalid. The specific determination process is as follows: the regional ASF image quality reassessment index is compared with the regional ASF image quality threshold; if the regional ASF image quality reassessment index is greater than or equal to the regional ASF image quality threshold, the image that fails twice is not marked as invalid and is recorded as a qualified image; if the regional ASF image quality reassessment index is less than the regional ASF image quality threshold, the image that fails twice is marked as invalid and an early warning is issued for the regional ASF image.

[0056] The aforementioned warning for regional ASF images refers to the appearance of a red pop-up window on the image analysis system interface, highlighting the thumbnail of the invalid image and the regional ASF image quality reassessment index, reminding operators to pay attention in real time.

[0057] like Figure 4 The schematic diagram of the secondary non-compliant image optimization process of this invention shows that the coma coefficient is reduced based on the quality review deviation value, and the optical coherence length is reduced by matching the reduction amount. Then, the regional ASF image quality reassessment index is obtained. If the regional ASF image quality reassessment index is greater than or equal to the regional ASF image quality threshold, it is not marked as an invalid image and is recorded as a qualified image. At the same time, the feature extraction accuracy index in the qualified image is obtained. Otherwise, it is marked as an invalid image and a warning is issued for the regional ASF image.

[0058] Specifically, such as Figure 5 The flowchart of this invention for determining whether to optimize the data extraction process of the target feature set is shown. The target feature set is extracted from a qualified image, the extraction process parameters are analyzed, a feature extraction accuracy index is obtained, and this index is compared with a preset feature extraction accuracy threshold in the prediction database. The feature extraction accuracy threshold refers to the minimum value of the feature extraction accuracy index. If the feature extraction accuracy index is greater than or equal to the feature extraction accuracy threshold, it is determined that no data optimization is performed on the target feature set extraction process. Simultaneously, the extracted target feature set is stored and associated with the imaging parameters of the regional ASF image to form a "feature-parameter" mapping database, used to construct a time-series feature data matrix. If the feature extraction accuracy index is less than the feature extraction accuracy threshold, it is determined that data optimization is performed on the target feature set extraction process.

[0059] The parameters for extracting the target feature set include the edge localization error factor of the target feature set, the micro-texture fractal dimension matching factor of the target feature set, and the signal-noise separation factor of the target feature set; and the final value of the comprehensive evaluation of the regional ASF image quality is obtained.

[0060] The edge positioning error factor represents a ratio of an edge positioning error value of the target feature set to a corresponding limit value. The micro-texture fractal dimension matching degree factor represents a ratio of a micro-texture fractal dimension matching degree of the target feature set to a corresponding limit value. The signal-noise separation degree factor represents a ratio of a signal-noise separation degree of the target feature set to a corresponding limit value. The regional ASF image quality comprehensive evaluation final value represents a final regional ASF image quality index in the judgment process.

[0061] The edge positioning error factor, the micro-texture fractal dimension matching degree factor, the signal-noise separation degree factor, and the regional ASF image quality comprehensive evaluation final value are set with action intensity coefficients in the prediction database, and the influence contribution values of the factors to the feature extraction accuracy index are quantified. Finally, the feature extraction accuracy index is obtained by weighting and comprehensively evaluating the influence contribution values. The feature extraction accuracy index represents the accuracy of extracting the target feature from the regional ASF image. The specific evaluation method is as follows:

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] In the formula, XLMH represents the feature extraction accuracy index of the target feature set, QALK_T represents the regional ASF image quality comprehensive evaluation final value, FVX represents the edge positioning error factor of the target feature set, HLM represents the edge positioning error value of the target feature set, DSC represents a preset limit edge positioning error value in the prediction database, WDH represents the micro-texture fractal dimension matching degree factor of the target feature set, TGD represents the micro-texture fractal dimension matching degree of the target feature set, RFG represents a preset limit micro-texture fractal dimension matching degree in the prediction database, TKN represents the signal-noise separation degree factor of the target feature set, DGK represents the signal-noise separation degree of the target feature set, ANM represents a preset limit signal-noise separation degree in the prediction database, fv represents an action intensity coefficient corresponding to the edge positioning error factor preset in the prediction database, fd represents an action intensity coefficient corresponding to the micro-texture fractal dimension matching degree factor preset in the prediction database, fk represents an action intensity coefficient corresponding to the signal-noise separation degree factor preset in the prediction database, and fa represents an action intensity coefficient corresponding to the regional ASF image quality comprehensive evaluation final value preset in the prediction database.

[0067] The edge positioning error value represents the deviation degree of the detected edge of the target region (e.g., the edge of an ASF lesion) from the actual physical edge. The detected edge of the target region in the image is extracted by an edge detection algorithm, and compared with the high-precision physical measurement true edge of the region to obtain the average Euclidean distance between the two. The micro-texture fractal dimension matching degree reflects the complexity of the micro-texture of the target region (e.g., the texture of the ASF lesion tissue) in the image and the degree of agreement with the reference standard (e.g., the texture of normal skin). The fractal dimension of the micro-texture of the target region is calculated by the box-counting method, and then compared with the pre-set reference fractal dimension range to obtain the deviation rate. The signal-noise separation degree measures the separation effect of the effective signal (e.g., the structure of the ASF lesion tissue) and the noise (e.g., sensor interference) in the image. The image is decomposed into effective signal and noise by wavelet transform, and the energy of each is calculated to obtain the ratio of the energy of the effective signal to the total energy.

[0068] The defined edge positioning error value represents the maximum value of the edge positioning error value within the specified range. The defined micro-texture fractal dimension matching degree represents the minimum value of the micro-texture fractal dimension matching degree within the specified range. The defined signal-noise separation degree represents the minimum value of the signal-noise separation degree within the specified range.

[0069] The smaller the edge positioning error (the more accurate the edge), the higher the accuracy of feature extraction (e.g., lesion boundary). The higher the matching degree of the texture fractal feature and the reference standard (the more real the texture feature), the more accurate the feature extraction of the microstructure of the lesion (e.g., the level of tissue damage). The better the signal-noise separation effect (the higher the proportion of effective signal), the less the noise interference, and the higher the accuracy of feature extraction (e.g., lesion density).

[0070] The action intensity coefficient corresponding to the edge positioning error factor represents the corresponding change amount of the feature extraction accuracy index when the factor changes by one unit. The action intensity coefficient corresponding to the micro-texture fractal dimension matching degree factor represents the corresponding change amount of the feature extraction accuracy index when the factor changes by one unit. The action intensity coefficient corresponding to the signal-noise separation degree factor represents the corresponding change amount of the feature extraction accuracy index when the factor changes by one unit. The action intensity coefficient corresponding to the final value of the regional ASF image quality comprehensive evaluation represents the corresponding change amount of the feature extraction accuracy index when the factor changes by one unit.

[0071] The prediction database stores the mapping relationship of the edge positioning error factor and its corresponding action intensity coefficient, the mapping relationship of the micro-texture fractal dimension matching degree factor and its corresponding action intensity coefficient, the mapping relationship of the signal-noise separation degree factor and its corresponding action intensity coefficient, and the mapping relationship of the regional ASF image quality comprehensive evaluation terminal value and its corresponding action intensity coefficient. For example, the edge positioning error factor, the micro-texture fractal dimension matching degree factor, the signal-noise separation degree factor, and the regional ASF image quality comprehensive evaluation terminal value are input into the prediction database, and the prediction database generates the corresponding action intensity coefficient corresponding to the edge positioning error factor, the corresponding action intensity coefficient corresponding to the micro-texture fractal dimension matching degree factor, the corresponding action intensity coefficient corresponding to the signal-noise separation degree factor, and the corresponding action intensity coefficient corresponding to the regional ASF image quality comprehensive evaluation terminal value based on the preset mapping rule, and the numerical range of each type of action intensity coefficient is strictly controlled between 0 and 1.

[0072] The smaller the edge positioning error factor, the lower the proportion of the actual error relative to the allowable error, the more reliable the edge positioning, and the higher the feature extraction accuracy. The larger the micro-texture fractal dimension matching degree factor value, the higher the degree of agreement between the texture feature and the standard, and the more accurate the feature extraction of the microstructure. The higher the signal-noise separation degree factor, the better the overall image quality, the more reliable the basic information provided for feature extraction, and the higher the accuracy.

[0073] Further, the data optimization of the extraction process of the target feature set is performed, and the specific optimization process is as follows:

[0074] Based on the feature extraction accuracy index and the feature extraction accuracy threshold, a feature extraction accuracy deviation value is obtained, and an objective lens numerical aperture increasing coefficient is matched based on the feature extraction accuracy deviation value, so as to increase the numerical aperture of the objective lens in the imaging system, improve the spatial resolution, more clearly capture the microstructure of the target feature set (such as the fine texture of the ASF area), enhance the condensing capacity of the objective lens, improve the image brightness and contrast, reduce the noise interference of weak signal features (such as low-contrast lesion texture), and at the same time, expand the effective imaging field of view, avoid the segmentation of the feature set due to the field of view limitation, provide more complete and fine original data for feature extraction, thereby increasing the feature extraction accuracy index; at the same time, an adaptive optical correction technology is introduced, which is an advanced technology for improving imaging quality by real-time detection and dynamic correction of wavefront distortion in the optical system; its core principle is to simulate the dynamic adjustment function of the human lens, compensate for the aberration of the optical system caused by various factors in real time through a closed-loop feedback system, and finally obtain a clear image close to the diffraction limit; which can significantly improve the imaging stability and clarity: first, it can accurately compensate for dynamic aberration, avoid blurring, shifting or distortion of target features (such as edge profiles) due to aberration, and ensure the consistency of feature extraction; second, for multi-layer tissue samples (such as skin and subcutaneous structures involved in ASF), it can correct the wavefront distortion caused by deep tissue scattering, making the imaging quality of deep features close to the surface, reducing feature misjudgment caused by aberration; in addition, this technology can work synergistically with high numerical aperture objective lenses to eliminate the problem of aberration sensitivity under high NA, fully release the resolution potential, and provide a more real and distortion-free image basis for feature extraction.

[0075] The above-mentioned feature extraction accuracy deviation value is obtained by subtracting the feature extraction accuracy index from the feature extraction accuracy threshold; the specific matching process of matching the objective lens numerical aperture increasing coefficient based on the feature extraction accuracy deviation value is as follows: the objective lens numerical aperture increasing coefficient corresponding to each feature extraction accuracy deviation value interval in the prediction database is stored, the obtained feature extraction accuracy deviation value is input into the database, the database can match the corresponding feature extraction accuracy deviation value interval, and then the objective lens numerical aperture increasing coefficient corresponding to the interval is the required increasing coefficient; the obtained objective lens numerical aperture increasing coefficient is multiplied by the original objective lens numerical aperture, and the result is the objective lens numerical aperture to be adjusted to; the above-mentioned objective lens numerical aperture increasing coefficient is greater than 1, indicating the numerical value of the objective lens numerical aperture increasing multiple.

[0076] The feature extraction accuracy index of the target feature set after data optimization is re-obtained, marked as the feature extraction accuracy re-evaluation index, and it is judged whether the target feature set is marked as a special set.

[0077] Specifically, whether the target feature set is marked as a special set is judged, and the specific judgment process is: comparing the feature extraction accuracy reevaluation index with the feature extraction accuracy threshold; if the feature extraction accuracy reevaluation index is greater than or equal to the feature extraction accuracy threshold, it is judged that the target feature set is not marked as a special set, the extracted target feature set is stored, and the imaging parameters of the regional ASF image are associated to form a "feature-parameter" mapping database for constructing a time series feature data matrix; if the feature extraction accuracy reevaluation index is less than the feature extraction accuracy threshold, the target feature set is marked as a special set, the regional ASF image and the optimization parameter are stored, and the artificial review instruction is triggered; the above-mentioned optimization parameter is specifically the numerical aperture size of the objective lens.

[0078] In a specific embodiment, when extracting the target feature set of the qualified image, by monitoring the edge positioning error, the fractal dimension matching degree and other parameters in real time, the feature missing detection or artifact interference and other abnormalities can be quickly identified; for the deviation, in addition to adjusting the numerical aperture of the objective lens, a multi-scale feature fusion algorithm can also be introduced to adaptively calibrate the feature extraction threshold; this way can not only enhance the image details (such as improving the edge sharpness of early lesions), but also can remove redundant information through feature complementation, ensure that the extracted necrosis focus three-dimensional morphology, blood vessel density gradient and other features truly reflect the nature of the lesion, provide reliable data support for subsequent analysis, and effectively reduce the risk of clinical misjudgment.

[0079] As Figure 6 , the flowchart of the data optimization of the present application is shown, the numerical aperture of the objective lens is increased based on matching the feature extraction accuracy deviation value with the numerical aperture increase coefficient of the objective lens, and the adaptive optical correction technology is introduced to reacquire the feature extraction accuracy reevaluation index; if the feature extraction accuracy reevaluation index is greater than or equal to the feature extraction accuracy threshold, the target feature set is not marked as a special set, and the parameters are stored and associated in the above-mentioned manner; otherwise, the target feature set is marked as a special set, the regional ASF image and the optimization parameter are stored, and the artificial review instruction is triggered.

[0080] Specifically, the dynamic data is fused with each feature in the target feature set, and the specific fusion process is: taking the spatial coordinates of the regional ASF image as the reference, the dynamic data is mapped to the same coordinate system through coordinate conversion for spatial alignment; the target feature set is added with a time stamp, and the dynamic data is time anchored according to the collection time and the time stamp of the image feature.

[0081] In a specific embodiment, the target features are fused with dynamic data and a time series matrix is constructed, which can track the dynamic evolution of ASF lesions in real time (such as the time sequence fluctuation of inflammatory infiltration range and the expansion rate of necrotic focus), breaking the limitations of static analysis; by quantifying the action intensity coefficient of each feature (such as the proportion of the influence of blood flow characteristics on prognosis at a certain stage), the key driving factors (such as the correlation between vascular permeability change and disease progression) can be locked, and the feature weight can be dynamically adapted; in practical application, combined with the time sequence rule to optimize the sensitivity of the model to short-term sudden change characteristics, the output prediction result can significantly improve the predictability of clinical decision-making.

[0082] Further, the time series matrix is constructed, and the specific construction process is: unifying the time granularity according to the collection frequency of dynamic data, the columns of the time matrix are all the fused features, the values of the same feature at different time points are standardized, and the values of different features in the matrix can be compared horizontally.

[0083] The above-mentioned dynamic data includes patient basic information (age, gender, weight, height, body mass index (BMI), past medical history, etc.), dynamic vital signs (mean arterial pressure (MAP), heart rate, blood pressure, etc.), peripheral perfusion index (PPI), laboratory indicators (serum lactic acid level, C-reactive protein (CRP), white blood cell count, procalcitonin (PCT), etc.), medical order information (vasoactive drug types and doses, fluid input and output, etc.), disease score (SOFA score, etc.), main observation outcome indicators (acute skin failure), clinical outcomes (death, hospitalization time), etc., based on multi-modal data, "feature-parameter" mapping database and dynamic data, a time series feature data matrix is constructed.

[0084] Outlier and error value monitoring and processing of the obtained data: identify and remove outliers by standard deviation outlier method, reduce the influence of data noise on the model; use interpolation method (less than 10% of missing values are filled by simple interpolation; missing values between 10% and 20% are filled by multiple interpolation; more than 20% of missing values are not filled) to fill the missing values in the time series, avoid the difference caused by missing data. Normalize the time series data to ensure that different parameters are analyzed on the same scale; arrange the cleaned data in chronological order to generate a complete time series matrix for each patient as input for feature selection and model training; convert the processed data into a time series matrix, each patient's monitoring data includes multiple key parameters; these data are recorded in chronological order to form a time series: X = {x1, x2,..., xT}, where xt represents all physiological characteristics at time step t, with dimension d, and T is the total length of the time series; in order to reduce the dimensional difference between different parameters, the input data needs to be normalized: Xt'= (xt-μt) / σt, where μt and σt are the mean and standard deviation of the feature at time step t, respectively. The processed time series data will be used as input for the LSTM model.

[0085] The time series matrix is input into the target model as input data, the target model includes a target number of trees, and the maximum depth and minimum sample partition number of each tree are set. Each tree will randomly select a subset of samples and a subset of features when constructing, and generate a tree structure in turn through the decision tree algorithm.

[0086] The above target model refers to a random forest (Random Fores), which is a supervised learning algorithm based on decision tree ensemble. It reduces the overfitting risk of a single tree by generating a large number of decision trees; the above target number refers to the total number of decision trees planned to be generated when constructing a random forest; in feature selection, the random forest will evaluate the importance of each feature to measure the influence of the feature on the target variable (ASF); feature importance is usually calculated by the "Gini importance" index.

[0087] Use the constructed time series matrix as input data; prepare input data set X and output label y, where X contains multi-modal feature data of patients, and the data set contains various physiological, clinical, and ultrasound and near-infrared spectrum acquisition parameters: X = {x1, x2,..., xN}, Xi ∈ R d , yi ∈ {0, 1}; where N is the number of patients, d is the number of features, Xi is the i-th patient, and yi is the label of the i-th patient. y is the label, indicating whether each patient has acute skin failure (ASF, 0 means no, 1 means yes), which needs to be explained, R ddenotes d-dimensional real number space, indicating that the features are all real values, that is: each sample xi is a d-dimensional vector containing the feature data of the i-th patient, which may include multiple modalities, such as physiological signals, ultrasound parameters, near-infrared spectral indicators, clinical scores, etc.; d represents the number of features extracted from each patient at a time point or a time sequence segment.

[0088] Random forest determines the importance of each feature by evaluating its contribution to the prediction performance of the model. Gini importance is used to measure it; according to the feature importance ranking, the top k features with high feature importance are selected for the training of the subsequent model; the dimensionality is reduced by setting a threshold for the importance score (retaining the top 20% important features) to retain key multi-modal parameters; the key features selected by the random forest will be used for subsequent model construction. These features can effectively capture the potential physiological signals of ASF, reduce redundant information, and enhance the interpretability of the model; the selected feature set Xselected can be represented as: Xselected={x1', x2',..., xk'}, k << d; At the same time, cross-validation (K-fold cross-validation) is used to evaluate the improvement of model performance after feature selection, and the effectiveness of the selected features in predicting sepsis ASF is verified by comparing the model performance before and after feature selection.

[0089] The output results of the integrated random forest model are marked as a feature set, and the feature set is input into the dynamic prediction model, which can output the corresponding results.

[0090] The above dynamic prediction model refers to the dynamic prediction model based on LSTM. Long Short-Term Memory (LSTM) is a type of Recurrent Neural Network (RNN) that can capture long-term dependencies in time series and is suitable for processing continuous physiological data. The core components of LSTM include input gates, forget gates, and output gates, which control how information flows in the sequence. The input layer takes the filtered key parameters as input features, with each time point corresponding to a set of data. The LSTM layer: multiple LSTM units are configured, and this model can capture both short-term and long-term dependencies in time series. Finally, the hidden state output by the last time step of the LSTM layer is passed to the fully connected layer for classification tasks. The output layer: the hidden state of LSTM is input into the Multi-Layer Perceptron (MLP) for binary classification tasks. The output layer uses the Sigmoid activation function to map the prediction results to a value between 0 and 1.

[0091] In one specific embodiment, the present application provides a regional ASF image data processing method based on multi-modal data fusion, taking a "multi-modal fusion-feature optimization-dynamic evaluation" closed-loop process as the core, breaking through the information barrier of a single data mode; by integrating multi-source data, selecting high-quality images, improving feature extraction accuracy, and combining dynamic rule evaluation of feature importance, the objectivity, accuracy and timeliness of regional ASF image analysis are significantly improved, providing solid quantitative support for clinical monitoring, evaluation and intervention decision-making of ASF lesions.

[0092] As Figure 7 shown in the flowchart of model construction and result output of the present application, based on multi-modal data, a "feature-parameter" mapping database (containing feature and parameter association data of non-special set) and dynamic data, a time series feature data matrix is constructed; the time series feature data matrix is input into a random forest model, the random forest model output results are integrated and labeled as a feature set, and then the feature set is input into a dynamic prediction model, and finally the corresponding prediction results are output by the dynamic prediction model, completing the whole process from image quality evaluation to feature application.

[0093] Referring to Figure 8 and Figure 9 , the unprocessed normal skin ultrasound image and the processed normal skin ultrasound image provided by the embodiment of the present application are shown, the objective indicators of the unprocessed normal skin ultrasound image are: RMS contrast is 0.071, Tenengrad sharpness is 2.794, and information entropy is 6.621; the objective indicators of the normal skin ultrasound image are: RMS contrast is 0.110, Tenengrad sharpness is 6.307, and information entropy is 7.119; it can be seen that after the median despeckling and contrast enhancement processing of the normal skin ultrasound image, the superficial fascia-subcutaneous-muscle layer is clear, the texture continuity is better, and the sensitivity of the model to speckle noise can be reduced.

[0094] Referring to Figure 10 and Figure 11 , the unprocessed ASF ultrasound image and the processed ASF ultrasound image provided by the embodiment of the present application are shown, the objective indicators of the unprocessed ASF ultrasound image are: RMS contrast is 0.071, Tenengrad sharpness is 2.844, and information entropy is 6.632; the objective indicators of the processed ASF ultrasound image are: RMS contrast is 0.108, Tenengrad sharpness is 7.442, and information entropy is 7.145; it can be seen that after the enhancement processing of the ASF ultrasound image, the contrast of the liquid dark area and the edema hypoechoic band is improved, the boundary of the tissue discontinuity is more easily identified, and the stability of detection or segmentation can be significantly improved.

[0095] Referring to Figure 2As shown, the second aspect of the present application provides a regional ASF image data processing system based on multi-modal data fusion, comprising: a multi-modal fusion and quality screening module, a feature extraction and optimization control module, a dynamic fusion and prediction evaluation module and a prediction database.

[0096] The multi-modal fusion and quality screening module is connected to the feature extraction and optimization control module, and the feature extraction and optimization control module is connected to the dynamic fusion and prediction evaluation module. The multi-modal fusion and quality screening module, the feature extraction and optimization control module and the dynamic fusion and prediction evaluation module are all connected to the prediction database. The prediction database is used to store various parameters involved in the regional ASF image data processing system based on multi-modal data fusion.

[0097] The multi-modal fusion and quality screening module is used to obtain multi-modal data of a target region, perform multi-modal data fusion, generate a regional ASF image, obtain and analyze quality parameters of the regional ASF image, and determine whether to mark the regional ASF image as a qualified image. The feature extraction and optimization control module is used to extract a target feature set from the qualified image, monitor extraction process parameters of the target feature set, and determine whether to perform data optimization on the extraction process of the target feature set. The dynamic fusion and prediction evaluation module is used to perform data fusion on each feature in the target feature set and obtained dynamic data, construct a time series matrix, evaluate importance scores of each feature in the target feature set, and output a prediction result of the regional ASF image based on the importance scores of the features.

[0098] It should be understood that the term "and / or" herein only describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent three cases of A alone, A and B together, and B alone, where A and B can be singular or plural. In addition, the character " / " herein generally represents an "or" relationship between the front and rear associated objects, but can also represent an "and / or" relationship. The specific meaning can be understood according to the context before and after.

[0099] It should be understood that in various embodiments of the present application, the size of the serial number of each process described above does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0100] Those skilled in the art can realize the units and algorithm steps of each example described in connection with the embodiments disclosed herein can be realized in electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints. Those skilled in the art can realize the described functions by different methods, and the implementation should not be considered beyond the scope of the present application.

[0101] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of the changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A regional ASF image data processing method based on multi-modal data fusion, characterized in that, The method comprises: Step one, acquiring multi-modal data of a target area, performing multi-modal data fusion to generate an area ASF image, acquiring and analyzing quality parameters of the area ASF image to determine whether to mark the area ASF image as a qualified image; Step two, extracting a target feature set from the qualified image, monitoring extraction process parameters of the target feature set to determine whether to optimize the extraction process of the target feature set, analyzing the extraction process parameters of the target feature set to obtain a feature extraction accuracy index, and comparing the feature extraction accuracy index with a preset feature extraction accuracy threshold in a prediction database, if the feature extraction accuracy index is less than the feature extraction accuracy threshold, it is determined to optimize the extraction process of the target feature set, the extraction process parameters of the target feature set include an edge positioning error factor of the target feature set, a micro-texture fractal dimension matching degree factor of the target feature set, and a signal-noise separation degree factor of the target feature set; Step three, performing data fusion on each feature in the target feature set and the acquired dynamic data, and constructing a time series matrix to evaluate the importance score of each feature in the target feature set, and outputting a prediction result of the area ASF image based on the importance score of each feature.

2. The regional ASF image data processing method based on multi-modal data fusion according to claim 1, characterized in that, The determination of whether to mark the area ASF image as a qualified image is specifically as follows: Analyze the quality parameters of the area ASF image to obtain an area ASF image quality index, and compare the area ASF image quality index with a preset area ASF image quality threshold in a prediction database; If the area ASF image quality index is greater than or equal to the area ASF image quality threshold, the area ASF image is determined to be a qualified image; If the area ASF image quality index is less than the area ASF image quality threshold, the area ASF image is determined to be an unqualified image, the unqualified image is denoised, based on the area ASF image quality index and the area ASF image quality threshold, an area ASF image quality deviation value is obtained, a spherical aberration coefficient reduction coefficient is matched based on the area ASF image quality deviation value, thereby reducing the spherical aberration coefficient of the objective lens in the imaging system; Obtain the area ASF image quality index of the denoised unqualified image, mark it as an area ASF image quality review index, and determine whether to mark the unqualified image as a secondary unqualified image; The quality parameters of the area ASF image include an edge contrast factor of the area ASF image, a surface fitting error factor of the area ASF image, and a fractal dimension factor of the area ASF image; In the prediction database, the edge contrast factor, the surface fitting error factor, and the fractal dimension factor are set to have an action intensity coefficient, the influence contribution value of each factor on the area ASF image quality index is quantified, and finally the area ASF image quality index is obtained by weighting and comprehensively considering the influence contribution value, wherein the area ASF image quality index represents the comprehensive advantages and disadvantages of the area ASF image in the key quality dimension.

3. The regional ASF image data processing method based on multi-modal data fusion according to claim 2, characterized in that, The specific determination process of determining whether to mark the unqualified image as a secondary unqualified image is as follows: The regional ASF image quality review index is compared with the regional ASF image quality threshold value; If the regional ASF image quality review index is greater than or equal to the regional ASF image quality threshold value, it is determined that the unqualified image is not marked as a secondary unqualified image; If the regional ASF image quality review index is less than the regional ASF image quality threshold value, it is determined that the unqualified image is marked as a secondary unqualified image, and the secondary unqualified image is optimized.

4. The regional ASF image data processing method based on multi-modal data fusion according to claim 3, characterized in that, The optimization of the secondary unqualified image is specifically as follows: Based on the regional ASF image quality review index and the regional ASF image quality threshold value, a quality review deviation value is obtained, a coma coefficient reduction coefficient is matched based on the quality review deviation value, so as to reduce the coma coefficient of the optical system, and a reduction amount of the optical coherence length is matched based on the quality review deviation value, so as to reduce the optical coherence length of the regional ASF image; The regional ASF image quality index of the secondary unqualified image is obtained, which is marked as a regional ASF image quality reevaluation index, and it is determined whether the secondary unqualified image is marked as an invalid image; The specific determination process is that the regional ASF image quality reevaluation index is compared with the regional ASF image quality threshold value; If the regional ASF image quality reevaluation index is greater than or equal to the regional ASF image quality threshold value, it is determined that the secondary unqualified image is not marked as an invalid image; If the regional ASF image quality reevaluation index is less than the regional ASF image quality threshold value, it is determined that the secondary unqualified image is marked as an invalid image, and the regional ASF image is warned.

5. The regional ASF image data processing method based on multi-modal data fusion according to claim 1, characterized in that, The specific determination process of determining whether to perform data optimization on the extraction process of the target feature set is as follows: If the feature extraction accuracy index is greater than or equal to the feature extraction accuracy threshold value, it is determined that data optimization is not performed on the extraction process of the target feature set, the extracted target feature set is stored, and the imaging parameters of the regional ASF image are associated to form a "feature-parameter" mapping database; A regional ASF image quality comprehensive evaluation terminal value is obtained; In the prediction database, the action intensity coefficients of the edge positioning error factor, the micro-texture fractal dimension matching degree factor, the signal-noise separation degree factor and the regional ASF image quality comprehensive evaluation terminal value are set, the influence contribution values of each factor on the feature extraction accuracy index are quantified, and finally the feature extraction accuracy index is obtained by weightedly combining each influence contribution value, wherein the feature extraction accuracy index represents the accuracy of extracting the target feature from the regional ASF image.

6. The regional ASF image data processing method based on multi-modal data fusion according to claim 5, characterized in that, The specific optimization process of performing data optimization on the extraction process of the target feature set is as follows: Based on the feature extraction accuracy index and the feature extraction accuracy threshold value, a feature extraction accuracy deviation value is obtained, and an objective lens numerical aperture increasing coefficient is matched based on the feature extraction accuracy deviation value, so as to increase the numerical aperture of the objective lens in the imaging system, and an adaptive optical correction technology is introduced; The feature extraction accuracy index of the target feature set after data optimization is reobtained, which is marked as a feature extraction accuracy reevaluation index, and it is determined whether the target feature set is marked as a special set.

7. The regional ASF image data processing method based on multi-modal data fusion according to claim 6, characterized in that, The specific determination process of determining whether to mark the target feature set as a special set is as follows: The feature extraction accuracy reevaluation index is compared with the feature extraction accuracy threshold value; If the feature extraction accuracy reevaluation index is greater than or equal to the feature extraction accuracy threshold value, it is determined that the target feature set is not marked as a special set, the extracted target feature set is stored, and the imaging parameters of the regional ASF image are associated to form a "feature-parameter" mapping database; If the feature extraction accuracy reevaluation index is less than the feature extraction accuracy threshold value, it is determined that the target feature set is marked as a special set, the regional ASF image and the optimization parameters are stored, and an artificial review instruction is triggered.

8. The regional ASF image data processing method based on multi-modal data fusion according to claim 1, characterized in that, The specific fusion process is as follows: The spatial coordinates of the regional ASF image are taken as the reference, and the dynamic data is mapped to the same coordinate system through coordinate conversion for spatial alignment; A timestamp is added to the target feature set, and the dynamic data is time-anchored by associating the collection time with the timestamp of the image feature.

9. The regional ASF image data processing method based on multi-modal data fusion according to claim 1, characterized in that, The specific construction process of the time sequence matrix is as follows: The time granularity is unified according to the collection frequency of the dynamic data, the columns of the time matrix are all the features after fusion, the values of the same feature at different time points are standardized to ensure that the values of different features in the matrix can be compared horizontally; The time sequence matrix is input into the target model as input data, the target model includes target trees, the maximum depth and minimum sample partition number of each tree are set, each tree randomly selects a subset of samples and a subset of features during construction, and a tree structure is generated by a decision tree algorithm in turn; The output results of the random forest model are integrated and marked as a feature set, and the feature set is input into the dynamic prediction model, which can output the corresponding results.

10. A system for applying the regional ASF image data processing method based on multi-modal data fusion according to any one of claims 1 to 9, characterized in that: It comprises: A multi-modal fusion and quality screening module for obtaining multi-modal data of a target region, performing multi-modal data fusion, generating a regional ASF image, obtaining and analyzing quality parameters of the regional ASF image, and determining whether to mark the regional ASF image as a qualified image; A feature extraction and optimization control module for extracting a target feature set from a qualified image, monitoring extraction process parameters of the target feature set to determine whether to optimize the extraction process of the target feature set, analyzing the extraction process parameters of the target feature set, obtaining a feature extraction accuracy index, and comparing the feature extraction accuracy index with a feature extraction accuracy threshold value preset in a prediction database, if the feature extraction accuracy index is less than the feature extraction accuracy threshold value, it is determined that the extraction process of the target feature set is optimized, the extraction process parameters of the target feature set include an edge positioning error factor of the target feature set, a microtexture fractal dimension matching degree factor of the target feature set, and a signal-noise separation degree factor of the target feature set; A dynamic fusion and prediction evaluation module for data fusion of each feature in the target feature set with acquired dynamic data, and construction of a time sequence matrix to evaluate the importance scores of each feature in the target feature set, and output of a prediction result of the regional ASF image based on the importance scores of each feature.

Images