A tumor medical pattern recognition method and system based on image detection

Abstract

The application provides a tumor medical pattern recognition method and system based on image detection, relates to medical image processing and artificial intelligence assisted diagnosis, generates a local quality evaluation map to recognize a quality impaired area by blocking a digital pathology section image, evaluating color distribution, light and dark contrast and microscopic structure geometric features of each block, generates a local diagnosis judgment map based on a tumor recognition model analysis, and directly reflects the model's diagnosis confidence for each area; the low confidence area and the quality impaired area in the two maps are analyzed in association, and if the overlapping degree meets the standard, a quality association explanation is generated, and it is clear that the uncertainty is caused by the image quality; finally, an explainability report containing an original map, two evaluation maps and an explanation is generated, which provides clear basis for doctors, avoids traditional compensation defects, improves recognition reliability and transparency, assists accurate clinical decision making, and has practical value.

Description

Technical Field

[0001] This application relates to the fields of medical image processing and artificial intelligence-assisted diagnosis, and more specifically, to a method and system for tumor medical pattern recognition based on image detection. Background Technology

[0002] In the field of medical image analysis, tumor medical pattern recognition systems based on image recognition and detection have become important tools for assisting pathological diagnosis. These systems aim to improve the efficiency and accuracy of tumor diagnosis by automating the analysis of digital pathological slide images. However, in actual clinical applications, due to the long-term high-load operation of the objective lens of a digital slide scanner, the performance of its internal precision optical coating may experience localized and non-uniform degradation. This degradation can lead to irregular color deviations and reduced contrast in local areas of the scanned image, thus affecting the original image quality. Traditional image preprocessing methods often struggle to effectively compensate for this local optical distortion and may even introduce new artifacts.

[0003] For example, in routine clinical practice, the objective lenses of digital slide scanners bear a heavy workload, continuously acquiring a large number of pathological slide images. As the core component of optical imaging, the performance of the objective lens directly determines the initial image quality. However, under the cumulative effect of long-term, high-frequency scanning tasks, the precision optical coatings inside the objective lens will gradually experience localized and non-uniform performance degradation due to continuous optical wear and environmental factors. This degradation does not occur suddenly but slowly. For example, the coating layer may develop microscopic scratches or localized thinning, resulting in a slight reduction in the transmittance of specific wavelengths of light and potentially producing a slight stray light effect. This wear is not uniformly distributed across the entire lens surface but may be concentrated in certain areas, such as the lens edge or specific quadrants, causing its optical properties to exhibit irregular spatial variations.

[0004] Due to the attenuation of the objective lens coating and stray light effects, pathological images acquired by the scanner may exhibit slight color deviations and reduced contrast in local areas, particularly at the edges of the image or within certain specific fields of view. This deviation is not globally uniform but rather presents an irregular and unpredictable "mottled" distribution. For example, one corner of the image may appear slightly yellowish, while another area may have slightly lower contrast, directly affecting the original quality of the image. This localized distortion means that the mapping relationship between pixel values ​​in different areas of the image and the actual tissue structure has undergone a non-linear, spatially varying change.

[0005] When deep learning models perform pattern recognition in tumor medicine, the weights of their stored knowledge are learned on a large number of standardized, uniformly colored training images without significant optical distortion. When the input image contains irregular local color shifts that persist even after preprocessing, the model's extraction of specific texture or morphological features of the tumor becomes unstable. For example, the model might incorrectly identify slight color anomalies caused by optical distortion in normal tissue areas as signs of early lesions, resulting in false positives; or it might ignore crucial information within the actual tumor area that is less obvious due to color deviation, leading to false negatives. This difference in the distribution between the input data and the model's training data significantly reduces the model's accuracy and robustness when dealing with images from real-world applications.

[0006] In real-world clinical settings, when images with potentially distorted information are continuously input into tumor medical pattern recognition systems, the model's confidence levels in identifying marginal, early-stage, or atypical tumor lesions fluctuate or remain low. This unstable performance is often mistaken by clinicians as a limitation of the model itself, rather than a data input quality issue, potentially affecting diagnostic reliability and increasing the risk of missed diagnoses. Therefore, improving the transparency and interpretability of model diagnostic results under such complex and uncertain image quality conditions, and helping physicians accurately understand the basis of the model's judgments, has become a pressing technical challenge. Summary of the Invention

[0007] This application provides a tumor medical pattern recognition method and system based on image detection, which aims to solve the problem that irregular color deviation and reduced contrast occur in local areas of digital pathological slide images, thereby affecting the original quality of the image and causing a decrease in the confidence of tumor recognition models.

[0008] In a first aspect, this application discloses a tumor medical pattern recognition method based on image detection, comprising:

[0009] The digital pathological slide image is segmented into multiple image blocks, and the color distribution, brightness contrast, and geometric features of the microstructure in each image block are evaluated to obtain a local quality assessment map; wherein the color or value of each image block region in the local quality assessment map represents the degree of image quality impairment corresponding to that image block region.

[0010] Based on the analysis of the digital pathological slide images by the tumor recognition model, a local diagnostic judgment map is generated; wherein the color or value of each image block region in the local diagnostic judgment map represents the confidence of the tumor recognition model in making a tumor diagnosis at that location, and the judgment confidence represents the probability value or confidence score of the image block being identified as a specific tumor type.

[0011] The correlation analysis is performed between the low-confidence region in the local diagnostic judgment image and the quality-damaged region in the local quality assessment image. If the overlap between the low-confidence region and the quality-damaged region meets a preset condition, a quality correlation interpretation is generated. The quality correlation interpretation is used to explain that the uncertainty in the judgment of the low-confidence region stems from image quality issues.

[0012] Based on the quality correlation interpretation, an interpretability report is generated, which includes the original pathological image, local diagnostic judgment map, local quality assessment map, and a visual presentation of the quality correlation interpretation.

[0013] Secondly, this application discloses a tumor medical pattern recognition system based on image detection, the system comprising:

[0014] The quality information acquisition module is used to segment digital pathological slide images into multiple image blocks, and to evaluate the color distribution, brightness contrast, and geometric features of the microstructure in each image block to obtain a local quality assessment map.

[0015] The confidence acquisition module is used to analyze the digital pathological slide image based on the tumor recognition model and generate a local diagnostic judgment map.

[0016] The correlation interpretation generation module is used to correlate and analyze the low-confidence region in the local diagnostic judgment image and the quality-damaged region in the local quality assessment image. If the overlap between the low-confidence region and the quality-damaged region meets the preset conditions, a quality correlation interpretation is generated. The quality correlation interpretation is used to explain that the uncertainty in the judgment of the low-confidence region is due to image quality problems.

[0017] The association explanation presentation module is used to generate an interpretability report based on the quality association explanation. The interpretability report includes the original pathological image, local diagnostic judgment map, local quality assessment map, and a visual presentation of the quality association explanation.

[0018] Beneficial effects

[0019] This application discloses a tumor medical pattern recognition method and system based on image detection, effectively solving the problem in existing technologies where image quality issues caused by local performance degradation of digital slide scanner objective lenses, thus affecting the judgment confidence and diagnostic accuracy of tumor recognition models, are hindered. By segmenting digital pathological slide images into multiple image blocks and evaluating the color distribution, brightness contrast, and microstructural geometric features of each block, a local quality assessment map is generated. This accurately identifies quality-damaged areas in the image, such as local color deviations and reduced contrast caused by objective lens performance degradation. Simultaneously, based on the tumor recognition model, the digital pathological slide image is analyzed to generate a local diagnostic judgment map, which intuitively reflects the diagnostic confidence of the tumor recognition model for each region. Through correlation analysis, low-confidence areas in the local diagnostic judgment map and quality-damaged areas in the local quality assessment map are correlated. If the overlap between the two meets preset conditions, a quality correlation explanation is generated, clearly indicating that the uncertainty in the judgment of low-confidence areas stems from image quality issues. Finally, an interpretable report is generated, containing the original pathological image, the local diagnostic judgment map, the local quality assessment map, and the quality correlation explanation. This application provides physicians with clear and interpretable diagnostic evidence, avoiding the problems of traditional methods failing to effectively compensate for local optical distortions or even introducing artifacts. This method not only improves the reliability and transparency of pattern recognition in tumor medicine, but also, through visualization, enables physicians to intuitively understand the sources of uncertainty in model judgments, thereby making more accurate clinical decisions. It has significant practical value and represents a technological advancement. Attached Figure Description

[0020] To illustrate this application more clearly, the accompanying drawings used in the embodiments will be briefly described below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0021] Figure 1 The diagram above illustrates a flowchart of a tumor medical pattern recognition method based on image detection.

[0022] Figure 2 The diagram above illustrates a schematic of a tumor medical pattern recognition system based on image detection.

[0023] Figure reference numerals: 100, Tumor medical pattern recognition system based on image detection; 10, Quality information acquisition module; 20, Judgment confidence acquisition module; 30, Association explanation generation module; 40, Association description presentation module. Detailed Implementation

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0025] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] Traditional tumor medical pattern recognition systems, when analyzing digital pathological slide images, may experience localized and non-uniform performance degradation in the precision optical coatings of the digital slide scanner's objective lens due to prolonged high-load operation. This degradation leads to irregular color deviations and reduced contrast in localized areas of the scanned image, affecting the original image quality and reducing the diagnostic confidence of tumor recognition models in these areas, potentially resulting in misdiagnosis or missed diagnosis. Traditional image preprocessing methods often struggle to effectively compensate for this localized optical distortion and may even introduce new artifacts, thus failing to fundamentally address the impact of image quality issues on diagnostic uncertainty.

[0027] like Figure 1 As shown, an exemplary flowchart of an image detection-based tumor medical pattern recognition method is illustrated. By introducing a correlation analysis mechanism between image quality assessment and diagnostic confidence, the source of uncertainty in tumor recognition model judgments can be effectively identified and explained, thereby providing pathologists with more credible and interpretable diagnostic evidence and significantly enhancing the clinical application value of the tumor medical pattern recognition system. This application proposes an image detection-based tumor medical pattern recognition method, including:

[0028] S10, the digital pathological slide image is segmented into multiple image blocks, and the color distribution, brightness contrast, and geometric features of the microstructure in each image block are evaluated to obtain a local quality assessment map; the color or value of each image block region in the local quality assessment map represents the degree of image quality damage corresponding to that image block region.

[0029] S20, Based on the analysis of digital pathological slide images by the tumor recognition model, a local diagnostic judgment map is generated; wherein the color or value of each image block region in the local diagnostic judgment map represents the confidence of the tumor recognition model in making a tumor diagnosis at that location, and the judgment confidence represents the probability value or confidence score of the image block being identified as a specific tumor type.

[0030] S30, Correlation analysis is performed to determine the low-confidence region in the local diagnostic image and the quality-damaged region in the local quality assessment image. If the overlap between the low-confidence region and the quality-damaged region meets the preset conditions, a quality correlation interpretation is generated. The quality correlation interpretation is used to explain that the uncertainty in the judgment of the low-confidence region stems from image quality issues.

[0031] S40, based on quality correlation interpretation, generates an interpretability report, which includes the original pathological image, local diagnostic judgment map, local quality assessment map, and a visual presentation of quality correlation interpretation.

[0032] Digital pathology slide images are digitized images obtained by scanning pathology slides, typically featuring high resolution and multi-layered structure. An image patch refers to a smaller region within a digital pathology slide image, serving as the basic unit for local analysis. A local quality assessment map is a visual representation where the color or value of each image patch reflects the degree of image quality impairment at that location; for example, red may indicate severe impairment, while green indicates good quality. A local diagnostic judgment map is generated by a tumor recognition model after analyzing the digital pathology slide image. The color or value of each image patch represents the model's confidence in diagnosing a tumor in that region; for example, darker colors or higher values ​​indicate stronger confidence. Judgment confidence can be understood as the probability value or confidence score that the model considers the image patch to belong to a specific tumor type. Low-confidence regions refer to areas where the model has low confidence in its diagnostic judgment, potentially indicating uncertainty in the model's judgment in that region. Impaired quality regions refer to areas where the image quality assessment results show poor image quality. Quality correlation interpretation is an explanation used to explain why the uncertainty in the model's judgment is due to image quality issues when low-confidence regions overlap with impaired quality regions. Interpretable reports integrate and visualize original pathological images, local diagnostic judgment maps, local quality assessment maps, and quality-related interpretations, aiming to provide pathologists with comprehensive diagnostic information and interpretations.

[0033] Specifically, the method proposed in this application first involves segmenting a digital pathological slide image into multiple image blocks, and then assessing the color distribution, contrast, and geometric features of the microstructures within each image block to obtain a local quality assessment map. In one embodiment, image segmentation can employ a fixed-size grid, for example, dividing the entire digital pathological slide image into 1024x1024 pixel image blocks. For the quality assessment of each image block, the color of each block can be manually checked for deviations from the normal range, for example, by visually inspecting whether the image block exhibits a significant reddish or bluish tint. Simultaneously, the contrast of the image block can be manually assessed, for example, by observing whether the boundary between the cell nucleus and cytoplasm is clear. Furthermore, the geometric dimensions and shapes of specific microstructures within the image block can be manually measured and compared with known normal morphologies to determine if distortion or deformation exists. The local quality assessment map can be generated by overlaying the results of the manual assessment onto the original image in a color-coded manner, for example, marking areas of poor quality as red and areas of normal quality as green.

[0034] In another implementation, image segmentation can employ content-adaptive segmentation methods. For example, image processing algorithms can be used to identify the main tissue structures in the image, and segmentation can be performed using these structures as boundaries. For evaluating the color distribution and contrast of each image patch, histogram analysis and color space conversion (e.g., RGB to HSV) can be used to calculate parameters such as average brightness, contrast, and saturation of the image patch, and compare them with preset normal ranges. For example, the average R, G, and B values ​​of the image patch can be calculated and compared with the average R, G, and B values ​​of a standard pathological slide to quantify color deviation. Contrast can be obtained by calculating the standard deviation of the pixel values ​​of the image patch or using local contrast enhancement algorithms. For evaluating the geometric features of microstructures in the image patch, image segmentation algorithms (such as U-Net and Mask R-CNN) can be used to automatically identify microstructures such as cell nuclei and glands in the image patch and extract their area, perimeter, aspect ratio, and roundness. These features can be compared with preset normal geometric parameters to quantify the degree of shape distortion. The generation of a local quality assessment chart can be achieved by integrating these quantitative indicators into a comprehensive quality score through weighted summation or other methods, and then mapping the scores to different colors or values ​​to form a visualization chart.

[0035] Next, the method proposed in this application generates a local diagnostic judgment map based on the analysis of digital pathological slide images using a tumor recognition model. In one implementation, the tumor recognition model can be a pre-trained deep learning model, such as a model based on a convolutional neural network (CNN). This model receives the digital pathological slide image as input and analyzes each image patch in the image. The model's output can be the probability value of each image patch belonging to different tumor types. For example, the model might output that a certain image patch has a probability of 0.8 for "adenocarcinoma," 0.1 for "squamous cell carcinoma," and 0.1 for "normal tissue." The local diagnostic judgment map can be generated by directly displaying these probability values ​​numerically on the corresponding image patch area, or by mapping the probability values ​​to color codes; for example, a higher probability value corresponds to a darker color, indicating a stronger confidence level in the judgment.

[0036] In another implementation, the tumor identification model can be an ensemble learning model, combining the prediction results of multiple deep learning models with different architectures (such as ResNet and InceptionNet) to improve diagnostic robustness. The model's analysis of digital pathological slide images can employ a sliding window approach, independently judging each local region in the image. The model's output can be a confidence score for each local region being identified as a specific tumor type; this score can be output from the softmax layer within the model or obtained through calibrated probability values. The generation of the local diagnostic judgment map can be achieved by normalizing these confidence scores and mapping them to a preset color gradient, for example, from blue (low confidence) to red (high confidence), based on the normalized score range, thus visually presenting the model's judgment confidence in different regions.

[0037] Subsequently, the method proposed in this application correlates low-confidence regions in the local diagnostic judgment image with quality-damaged regions in the local quality assessment image. If the overlap between the low-confidence regions and the quality-damaged regions meets a preset condition, a quality correlation interpretation is generated. In one embodiment, the identification of low-confidence regions can be achieved by setting a fixed judgment confidence threshold, for example, marking regions with a judgment confidence below 0.6 as low-confidence regions. The identification of quality-damaged regions can also be achieved by setting a fixed quality damage threshold, for example, marking regions with a quality score below 0.5 as quality-damaged regions. The overlap can be calculated by simply calculating the intersection area of ​​the two regions and comparing it with the area of ​​the low-confidence region. For example, if the intersection area accounts for more than 50% of the area of ​​the low-confidence region, the preset condition is considered met. The generation of the quality correlation interpretation can be a simple text prompt, such as "The low diagnostic confidence in this region is due to image quality degradation."

[0038] In another implementation, low-confidence regions can be identified using an adaptive thresholding method. For example, the threshold can be dynamically adjusted based on the overall image confidence distribution, or clustering algorithms can be used to identify image patches with significantly lower confidence levels than their surrounding areas. Quality-impaired regions can be identified using a multi-level thresholding strategy, identifying different levels of quality-impaired regions based on varying degrees of impairment (mild, moderate, severe). Overlap calculations can employ more complex spatial analysis methods, such as calculating the Jaccard similarity coefficient or Dice coefficient between two regions and comparing it to a preset overlap ratio threshold. The generation of a quality association interpretation can be a more detailed report, explaining specific image quality issues (such as color deviation or blurring) and their specific impact on the model's confidence, and providing visual evidence, such as highlighting overlapping regions.

[0039] Finally, the method proposed in this application generates an interpretability report based on quality correlation interpretation. The interpretability report includes a visual representation of the original pathological image, local diagnostic judgment map, local quality assessment map, and quality correlation interpretation. In one embodiment, the interpretability report can be a PDF document containing thumbnails of the original pathological image, static images of the local diagnostic judgment map and local quality assessment map, and a textual description of the quality correlation interpretation. This content can be formatted according to a fixed layout for easy reference by pathologists.

[0040] In another implementation, the interpretability report can be an interactive digital report that allows pathologists to view detailed information about different areas by clicking or zooming. The original pathology image can support high-resolution zooming and panning, and local diagnostic judgment maps and local quality assessment maps can be overlaid on the original image with adjustable transparency, allowing pathologists to simultaneously observe image content and model judgment / quality information. Quality-related interpretations can be presented as pop-ups or sidebars, automatically displaying detailed explanatory information when the pathologist hovers the mouse over areas of low confidence and impaired quality. This information may indicate the degree of color deviation, the percentage decrease in contrast, and the distortion of microstructure in that area, and provide possible repair suggestions or prompts for further examination. This interactive report can greatly enhance the pathologist's understanding and trust in the diagnostic results.

[0041] The tumor medical pattern recognition method proposed in this application, based on image detection, segments digital pathological slide images into multiple image patches and performs local quality assessment to generate a local quality assessment map, enabling refined identification of quality issues within the image. Simultaneously, based on a tumor recognition model, the image is analyzed to generate a local diagnostic judgment map, intuitively demonstrating the model's confidence in different regions. Crucially, this application uses correlation analysis to link low-confidence regions in the local diagnostic judgment map with quality-impaired regions in the local quality assessment map. Once the overlap meets a preset condition, a quality correlation interpretation is generated, clearly indicating that the root cause of the model's uncertainty lies in image quality issues. Finally, by generating an interpretable report containing the original pathological image, the local diagnostic judgment map, the local quality assessment map, and the quality correlation interpretation, comprehensive, intuitive, and interpretable diagnostic evidence is provided to pathologists.

[0042] For example, suppose a digital pathology slide image has a localized area where the scanner objective lens performance is compromised, resulting in a yellowish tint and blurred contrast. Traditional tumor identification models, when analyzing this area, might fail to accurately identify tumor cells due to image quality issues, leading to a low confidence level. However, without an explanation for the low confidence, pathologists may struggle to determine whether the problem stems from insufficient model recognition capabilities or image quality issues, thus increasing diagnostic complexity and uncertainty.

[0043] The method described in this application effectively addresses the aforementioned problems. First, by performing a local quality assessment on the digital pathological slide image, the yellowish, blurry area is identified as a quality-impaired region and marked with a specific color or value in the local quality assessment map. Simultaneously, the tumor recognition model's analysis of this region generates a low confidence level, which is represented by a corresponding color or value in the local diagnostic judgment map. Subsequently, the method in this application correlates these two maps, discovering a high degree of overlap between the low-confidence region and the quality-impaired region. Based on this, the system generates a quality correlation explanation, clearly informing the pathologist: "The low confidence level for the tumor diagnosis in this region is mainly due to quality issues such as color deviation and contrast blurring in the image." Finally, in the interpretability report, the pathologist can simultaneously view the original image, the model's confidence map, the image quality assessment map, and the clear quality correlation explanation. Thus, the pathologist can clearly understand the specific reasons for the model's uncertainty in this region, allowing for more cautious assessment during diagnosis and potentially necessitating further image processing or rescanning to avoid misdiagnosis or missed diagnosis due to image quality issues. This mechanism significantly improves the transparency and reliability of tumor medical pattern recognition systems, providing pathologists with a more reliable auxiliary diagnostic tool.

[0044] The core innovation of this application lies in introducing a correlation analysis mechanism between image quality assessment and diagnostic confidence. Specifically, this application segments digital pathological slide images into multiple image blocks and performs quality assessments on the color distribution, contrast, and microstructural geometric features of each block, generating a local quality assessment map. This allows the system to finely identify locally impaired areas in the image, rather than simply making a global quality judgment. Simultaneously, a local diagnostic judgment map is generated through a tumor recognition model, intuitively presenting the model's confidence in different regions. Most importantly, this application generates a quality correlation interpretation by correlating low-confidence areas in the local diagnostic judgment map with quality-impaired areas in the local quality assessment map. If the overlap meets preset conditions, this interpretation provides a clear explanation for the pathologist. This mechanism allows the system to explicitly attribute model uncertainty to image quality issues, thus providing a clear explanation for the pathologist.

[0045] For example, in the closest existing technology, when a tumor recognition model has low diagnostic confidence in a certain area, the system may only indicate "insufficient diagnostic confidence" without further explanation. Pathologists, faced with such an indication, may need to spend considerable time manually checking image quality and making judgments based on their own experience. However, the method of this application can directly provide a quality-related explanation such as "the low diagnostic confidence in this area is due to image color deviation and contrast blurring." This explicit explanation greatly reduces the workload of pathologists and improves diagnostic efficiency. Furthermore, by generating an interpretable report containing the original pathological image, local diagnostic judgment map, local quality assessment map, and quality-related explanation, this application provides pathologists with a comprehensive, intuitive, and interpretable auxiliary diagnostic tool, significantly improving the system's transparency and credibility. Therefore, this application not only solves the problem of insufficient interpretability in traditional systems when faced with image quality issues but also effectively improves the clinical application value and diagnostic accuracy of tumor medical pattern recognition systems by providing explicit quality attribution.

[0046] Specifically, the step of generating a quality correlation interpretation can be further refined into the following specific operations, where the low-confidence region in the local diagnostic judgment map and the quality-damaged region in the local quality assessment map overlap with the quality-damaged region if the overlap meets a preset condition.

[0047] In some embodiments, the step of generating a quality correlation interpretation includes, if the overlap between the low-confidence region in the local diagnostic judgment map and the quality-impaired region in the local quality assessment map meets a preset condition, the above correlation analysis includes:

[0048] The low-confidence regions are identified by determining areas in the local diagnostic judgment map where the judgment confidence is lower than a preset judgment confidence threshold.

[0049] Identify regions in the local quality assessment map whose quality damage level exceeds a preset quality damage threshold to obtain the quality damage regions;

[0050] Calculate the intersection area of ​​the low confidence region and the quality-impaired region;

[0051] If the ratio of the intersection area to the area of ​​the low-confidence region exceeds a preset overlap ratio threshold, then the generated quality association interpretation is generated.

[0052] Specifically, during the correlation analysis, the first step is to identify regions in the local diagnostic judgment image where the judgment confidence is below a preset judgment confidence threshold, thus obtaining low-confidence regions. This preset judgment confidence threshold can be set according to the actual application scenario, the performance of the tumor recognition model, and the tolerance for diagnostic uncertainty. For example, it can be set to 0.7 or 0.8; regions below this value are considered to have insufficient model judgment confidence. Simultaneously, regions in the local quality assessment image where the degree of quality impairment exceeds a preset quality impairment threshold are also identified, thus obtaining quality-impaired regions. This preset quality impairment threshold is used to define which regions have reached a level of image quality issues requiring attention; for example, it can be set based on a comprehensive score of color deviation, contrast deviation, and shape distortion. After identifying the low-confidence regions and quality-impaired regions, the intersection area of ​​these two regions needs to be calculated. The size of the intersection area directly reflects the degree of spatial overlap between insufficient model judgment confidence and image quality impairment. Subsequently, the intersection area is compared with the area of ​​the low-confidence regions, specifically, the ratio between the two is calculated. If this ratio exceeds a preset overlap ratio threshold, the judgment uncertainty of the low-confidence region is considered to be significantly correlated with the image quality problem, thereby generating a quality correlation interpretation. The preset overlap ratio threshold can be determined based on experience or through experiments. For example, it can be set to 0.5, which means that a quality correlation is considered to exist only when at least half of the area in the low confidence region overlaps with the quality-impaired region.

[0053] Through the above technical solution, this application provides a more accurate and reliable method to identify the sources of uncertainty in tumor recognition model judgments. Specifically, by clearly defining low-confidence regions and quality-impaired regions and introducing overlap analysis, misjudgments can be effectively avoided. This ensures that a quality correlation explanation is generated only when image quality issues are indeed spatially correlated with a decrease in model judgment confidence. This not only improves the accuracy of quality correlation explanations but also makes subsequent interpretability reports more persuasive, helping doctors or users quickly understand the root causes of model judgment uncertainty, thereby enhancing the credibility and practicality of the entire tumor medical pattern recognition system.

[0054] In some of the embodiments described above in this application, in order to more specifically implement the steps of segmenting digital pathological slide images into multiple image blocks and evaluating the color distribution, brightness contrast, and geometric features of the microstructure in each image block to obtain a local quality assessment map, this application further proposes a detailed implementation plan.

[0055] The steps described above, which involve segmenting a digital pathological slide image into multiple image patches and evaluating the color distribution, contrast, and geometric features of the microstructure within each patch to obtain a local quality assessment map, include:

[0056] The digital pathological slide image is divided into multiple image blocks;

[0057] For each image block, its color distribution and brightness contrast are obtained and compared with preset standard reference values ​​to obtain color deviation and brightness contrast deviation;

[0058] For each image patch, the microstructure is identified, the geometric features of the microstructure are extracted, and the geometric features are compared with a preset geometric reference benchmark to obtain the degree of shape distortion. The preset geometric reference benchmark refers to the set of geometric parameters of the natural normal microstructure in a standard pathological slide under the optimal state of the objective lens, which is used to determine whether the actual microstructure geometric features are distorted due to image quality issues.

[0059] The local quality assessment map is generated based on the color deviation, the brightness contrast deviation, and the degree of shape distortion.

[0060] Specifically, digital pathological slide images are segmented into multiple image blocks. This can be achieved using a fixed-size grid, such as dividing the entire digital pathological slide image into 1024x1024 pixel or 256x256 pixel blocks. Alternatively, content-adaptive segmentation methods can be employed, such as using edge detection or region growing algorithms to group regions with similar features into a single image block. This approach aims to decompose large pathological images into manageable units for refined local analysis.

[0061] For each image patch, its color distribution and contrast are acquired and compared with preset standard reference values ​​to obtain color deviation and contrast deviation. Color distribution refers to the statistical characteristics of pixels within the image patch in RGB, HSV, or Lab color space, such as mean, variance, and histogram. Contrast deviation refers to the dynamic range, local contrast, or gradient information of pixel brightness values ​​within the image patch. The preset standard reference values ​​are obtained based on statistical analysis of a large number of high-quality, non-destructive standard pathological slide image patches, representing ideal color and contrast performance. By comparing the color distribution and contrast of the current image patch with these preset standard reference values, color deviation and contrast deviation can be quantified, for example, by calculating Euclidean distance, KL divergence, or percentage difference. These deviation values ​​reflect quality issues in the image patch regarding color and brightness.

[0062] Furthermore, for each image patch, the microstructures within them are identified, their geometric features are extracted, and these features are compared with a preset geometric reference to determine the degree of shape distortion. Microstructures can include morphological features of diagnostic significance in pathology, such as cell nuclei, cytoplasm, glandular structures, and blood vessels. Identification of these microstructures can be achieved through image segmentation, feature point detection, or deep learning models. Extracted geometric features can include the area, perimeter, aspect ratio, roundness, eccentricity, and texture features of the microstructure. The preset geometric reference is a set of geometric parameters of natural, normal microstructures in standard pathological sections scanned under optimal objective lens conditions, such as the average diameter and roundness range of normal cell nuclei. By comparing the geometric features of the identified microstructures in the actual image patch with this preset geometric reference, the degree of shape distortion can be quantified, for example, by calculating feature differences or deviations. The degree of shape distortion indicates whether the microstructure has been deformed due to problems during image acquisition or preparation (such as defocus, distortion, uneven slice thickness, etc.).

[0063] Finally, a local quality assessment map is generated based on the color deviation, the contrast deviation, and the shape distortion. These three quality indicators can be weighted and combined, or fused using a machine learning model, to generate a comprehensive quality score or grade. This quality score or grade can be mapped to the color or numerical value of each image patch region in the local quality assessment map, where the intensity or magnitude of the color or numerical value indicates the degree of image quality degradation corresponding to that patch region. For example, a red area indicates severe quality degradation, and a green area indicates good quality.

[0064] Through the above technical solution, this application provides a more refined and accurate method for local image quality assessment. Compared to relying solely on a single or coarse quality indicator, this solution comprehensively considers color deviation, contrast deviation, and the degree of distortion in microstructure shape, and introduces objective reference benchmarks for quantification, significantly improving the accuracy and reliability of local quality assessment maps. Therefore, when a tumor recognition model exhibits low confidence in a specific region, it can be more accurately attributed to image quality issues, thus providing doctors with more convincing and interpretable reports, avoiding the risk of misdiagnosis or missed diagnosis due to image quality problems, and enhancing the diagnostic assistance capability and credibility of the entire tumor medical pattern recognition system.

[0065] In some embodiments, the step of generating a local diagnostic judgment map based on the analysis of digital pathological slide images using a tumor recognition model may include the following:

[0066] Receive the original judgment map output by the tumor recognition model, wherein the original judgment map contains the probability value of each local region being the target object;

[0067] The probability values ​​are mapped to color codes to generate a local diagnostic judgment map, in which high probability values ​​correspond to high confidence regions and low probability values ​​correspond to low confidence regions.

[0068] The original judgment image can be understood as the raw, unprocessed result directly output by the tumor recognition model after analyzing the digital pathological slide image. This raw result is usually represented in numerical form, such as the probability value or confidence score of each pixel or image patch region being identified as a specific tumor type. These probability values ​​or confidence scores directly reflect the model's preliminary judgment on the diagnosis of that region.

[0069] Furthermore, mapping the probability values ​​to color coding means converting them into visually distinguishable colors based on the probability values ​​of each local region in the original judgment map. For example, a color gradient can be set, mapping high probability values ​​(e.g., close to 100%) to one color (such as green or red, representing high confidence), low probability values ​​(e.g., close to 0%) to another color (such as blue or yellow, representing low confidence), and intermediate probability values ​​to a gradient color between the two. Through this color coding, abstract probability values ​​can be intuitively transformed into visual information that is easily recognized and understood by the human eye, thereby generating a local diagnostic judgment map. In the color coding, high probability values ​​are set to correspond to high confidence regions, and low probability values ​​are set to correspond to low confidence regions, enabling users to quickly identify the degree of certainty in the model's judgment.

[0070] The aforementioned technical solution transforms the complex internal judgments of tumor recognition models—the raw probability values—into an intuitive and easily understood color-coded format. This visualization method allows doctors or users to quickly and accurately understand the model's diagnostic confidence level in different regions of digital pathology slide images, avoiding the difficulty of directly interpreting complex numerical values. Especially with the setting of high probability values ​​corresponding to high confidence regions and low probability values ​​corresponding to low confidence regions, the degree of certainty in the judgment is immediately apparent, greatly improving the readability and practicality of the diagnostic judgment chart. This helps users quickly identify the reliability of the model's judgment, thereby assisting in making more accurate medical decisions.

[0071] In some embodiments, the method for obtaining the above-mentioned standard reference values ​​and preset geometric reference benchmarks can be further specified as follows:

[0072] The standard reference value is obtained by scanning a series of standard pathological sections with the objective lens in optimal condition, extracting the color distribution feature value and brightness contrast feature value of the local image blocks of each section, and then statistically analyzing them.

[0073] The preset geometric reference benchmark refers to the set of standard geometric parameters determined after analyzing the geometric characteristics of naturally occurring normal microstructures in a series of standard pathological sections when the objective lens is in its optimal state.

[0074] The standard reference values ​​are important benchmarks for evaluating the quality of digital pathological slide images. Their acquisition process aims to establish an objective and stable reference system to reflect image characteristics under ideal imaging conditions. Specifically, under ideal optical conditions—without distortion, chromatic aberration, and accurate focal length—multiple rigorously prepared and representative standard pathological slides are scanned. Color distribution characteristics of local image patches extracted from these slides, such as the mean, standard deviation, and histogram distribution of the RGB channels, as well as contrast characteristics, such as local contrast and grayscale dynamic range, are collected and statistically analyzed. In this way, a statistically significant set of reference values ​​is obtained, which can be used for subsequent quantitative comparison of the color and contrast of actual pathological slide images.

[0075] Furthermore, the preset geometric reference benchmark is a key standard for evaluating whether the geometric features of microstructures are distorted due to image quality issues. Establishing this benchmark also requires the objective lens to be in optimal condition. By performing detailed geometric feature analysis on naturally occurring normal microstructures in a series of standard pathological sections, such as cell nuclei, cytoplasm, and glandular structures, the ideal set of geometric parameters for these structures can be determined. These geometric parameters may include, but are not limited to, shape factor, aspect ratio, perimeter, area, roundness, and statistical characteristics of internal texture. This set serves as the basis for judging whether the actual geometric features of microstructures deviate from the normal state, thereby enabling the identification of structural distortions caused by factors such as optical distortion and improper section preparation.

[0076] Through the above technical solution, this application provides a more accurate and reliable method for image quality assessment. Since the standard reference values ​​and preset geometric reference benchmarks are established under ideal conditions, they can serve as the gold standard for measuring actual image quality, effectively avoiding assessment biases caused by unclear or inaccurate reference standards. Therefore, image quality problems can be identified more accurately and correlated with low-confidence regions in tumor recognition models, thereby improving the accuracy of interpretability reports and their clinical guidance value. This clear benchmark definition helps improve the standardization level and diagnostic reliability of the entire tumor medical pattern recognition system.

[0077] In some embodiments described above, the degree of shape distortion is obtained by comparing the geometric features of the microstructure with a preset geometric reference. However, in its implementation, simply comparing geometric features may not effectively distinguish between different sources of deformation, such as those caused by tissue preparation and staining processes, or by defects in the optical system (e.g., objective lens). This indiscriminate deformation assessment may lead to inaccurate attribution of image quality problems, thereby affecting the accuracy of subsequent explanations of the reasons for decreased confidence in tumor recognition models and limiting the possibility of targeted improvements to image quality problems.

[0078] In response, this application further proposes the following steps for each image block: identifying naturally occurring microstructures, extracting the geometric features of the microstructures, and comparing the geometric features with a preset geometric reference to obtain the degree of shape distortion:

[0079] Obtain the spatial distribution characteristics and internal morphological characteristics of the microstructure;

[0080] The spatial distribution characteristics are compared with a preset tissue preparation deformation pattern to obtain a first comparison result; the internal morphological characteristics are compared with a preset optical deformation pattern to obtain a second comparison result; wherein, the preset tissue preparation deformation pattern refers to a reference model established by statistically analyzing the microstructural deformation patterns caused by the thickness or staining of tissue preparation sections in standard pathological sections; the preset optical deformation pattern refers to a reference model established by statistically analyzing the microstructural deformation patterns caused by objective lens distortion and stray light effects;

[0081] Based on the first comparison result and the second comparison result, the deformation of the microstructure is attributed to a deformation attribution result, which is used to indicate whether the deformation is mainly caused by tissue preparation or staining process, or mainly caused by objective lens deformation.

[0082] Based on the deformation attribution results, the deformation caused by tissue preparation or staining process is compensated or eliminated; and the compensationd or eliminated deformation is converted into the degree of shape distortion.

[0083] Specifically, when acquiring the spatial distribution and internal morphological features of microstructures, spatial distribution features can include macroscopic layout information such as the position, density, and distance between microstructures in the image, while internal morphological features can include microscopic details such as the shape, size, texture, and edge sharpness of the microstructure itself. These features can be extracted using image processing and pattern recognition algorithms, such as edge detection, region growing, and morphological operations, to identify and quantify these features.

[0084] The preset tissue preparation deformation mode and the preset optical deformation mode are two different reference models, used to identify and quantify microstructural deformations caused by different factors. The preset tissue preparation deformation mode is established by analyzing a large number of standard pathological sections and statistically analyzing the deformation patterns of microstructures (such as cell nuclei and glandular structures) caused by tissue preparation and staining process issues such as uneven section thickness and uneven staining, resulting in compression, stretching, blurring, or contrast changes. For example, a model can be established to describe the average ellipticity change of cell nuclei under different section thicknesses. The preset optical deformation mode is established by analyzing the microstructural deformation patterns caused by optical system problems such as objective lens distortion (such as pincushion distortion and barrel distortion), stray light effects, and defocus. For example, a model can be established to describe the degree of curvature of microstructural edges or the degree of blurring of internal textures under specific optical distortions.

[0085] In practical applications, comparing the spatial distribution characteristics of the microstructure with a preset tissue preparation deformation pattern can assess whether it conforms to the deformation pattern caused by tissue preparation, yielding the first comparison result. Simultaneously, comparing the internal morphological characteristics of the microstructure with a preset optical deformation pattern can assess whether it conforms to the deformation pattern caused by the optical system, yielding the second comparison result. These comparisons can be performed by calculating feature similarity, distance metrics, or classifier discrimination.

[0086] Therefore, based on the first and second comparison results, the causes of microstructural deformation can be attributed. For example, if the first comparison result shows a high degree of match with the deformation pattern of tissue preparation, while the second comparison result shows a lower degree of match, the deformation attribution result will indicate that the deformation is mainly caused by tissue preparation or staining processes. Conversely, if the second comparison result shows a high degree of match, it indicates that it is mainly caused by objective lens deformation. This attribution helps to more accurately understand the source of deformation.

[0087] Furthermore, based on the deformation attribution results, deformations caused by tissue preparation or staining processes are compensated for or excluded. This is because some deformations caused by tissue preparation may be unavoidable to some extent, or their impact on the confidence of tumor recognition models can be compensated for using specific image processing techniques, such as restoring their original shape through deformation correction algorithms or excluding their influence in subsequent analysis. Deformations caused by objective lens deformation, due to their different nature, may require different processing strategies or be directly marked as uncompensable quality issues. Finally, the compensated or excluded deformations are quantified into the degree of shape distortion. This quantified value can more accurately reflect the true quality problems in the image caused by non-compensable or uncompensated factors.

[0088] Through the above technical solution, this application provides a more refined and accurate method for assessing the degree of shape distortion. This method not only quantifies the degree of deformation in microstructures, but more importantly, it can attribute the causes of deformation, distinguishing between deformation caused by tissue preparation or staining processes and deformation caused by the optical system. This attribution capability allows the system to take targeted measures for deformation from different sources, such as compensating for or eliminating compensable deformations, thereby obtaining a more realistic and diagnostically significant degree of shape distortion. Consequently, the generated local quality assessment map can more accurately reflect the actual quality problems of the image, avoiding misjudgments of image quality due to non-critical or correctable deformations, thus improving the interpretability and reliability of tumor recognition models. Furthermore, this clear attribution of the causes of deformation also provides specific guidance for subsequent image quality improvement or diagnostic process optimization. For example, if a large amount of deformation is found to originate from tissue preparation, the preparation process can be optimized; if it originates from the optical system, the equipment can be inspected or calibrated.

[0089] The following is a specific example to illustrate this. Suppose that when analyzing a digital pathological slide image, the system identifies a significant deformation in a glandular structure within an image patch. First, the system acquires the spatial distribution characteristics of the glandular structure (e.g., the spacing between glands and their orientation) and its internal morphological characteristics (e.g., the shape of the glandular lumen and the ellipticity of the epithelial cell nuclei).

[0090] Next, the acquired spatial distribution features are compared with preset tissue preparation deformation patterns. For example, if the gland structure exhibits features highly similar to the gland compression or stretching pattern caused by uneven slice thickness, the first comparison result indicates that it may originate from a tissue preparation problem. Simultaneously, the internal morphological features are compared with preset optical deformation patterns. For example, if there is slight curvature at the gland edge, but its degree does not match the objective lens distortion pattern well, while the ellipticity change of the cell nucleus matches the tissue preparation pattern more closely, a second comparison result is obtained.

[0091] Based on these two comparison results, the system performs deformation attribution. If the match of the first comparison result is significantly higher than that of the second comparison result, the deformation attribution result will indicate that the deformation of the gland structure is mainly caused by tissue preparation or staining process.

[0092] Based on this, the system compensates for deformations caused by tissue preparation or staining processes according to the deformation attribution results. For example, an algorithm based on morphological or geometric transformations can be applied to correct compressed or stretched glandular structures, restoring them to a shape closer to normal. After compensation, the degree of shape distortion of the corrected glandular structure is quantified to obtain a more accurate shape distortion index. This index reflects the true degree of distortion that still exists in the glandular structure after excluding the influence of tissue preparation, which may be caused by other uncompensable factors. Finally, this compensated or excluded degree of shape distortion is used to generate a local quality assessment map, thereby providing more reliable image quality information for the tumor recognition model.

[0093] In some embodiments described above in this application, a scheme is proposed to compensate for or exclude deformation caused by tissue preparation or staining processes, and to quantify the compensated or excluded deformation as the degree of shape distortion. However, in practical applications, simply quantifying the degree of shape distortion may not fully reflect its specific impact on the confidence of tumor identification models, especially when faced with different pathological types and multiple deformation characteristics. This quantification method may lack sufficient precision and specificity, thus affecting the accurate interpretation of the source of uncertainty in the judgment.

[0094] In this regard, this application further proposes the following steps for converting the compensated or excluded deformation into the degree of shape distortion:

[0095] Identify the pathological type to which the microstructure belongs;

[0096] Obtain the first parameter set corresponding to the pathological type. The first parameter set defines the response weight of the pathological type to the deformation feature and the degree of influence of the deformation feature on the judgment confidence of the tumor recognition model.

[0097] Based on the first parameter set, multiple geometric features extracted from the microstructure are weighted and combined to obtain the shape distortion index;

[0098] The shape distortion index is divided into multiple levels;

[0099] Obtain the diagnostic impact factor corresponding to each level, whereby the diagnostic impact factor represents the expected contribution ratio of the level distortion to the decrease in the judgment confidence of the tumor identification model;

[0100] The shape distortion index and the diagnostic influence factors at each level are quantified into the degree of shape distortion.

[0101] Specifically, identifying the pathological type of a microstructure refers to automatically identifying the histological or cytological type of a specific microstructure (such as cell nuclei, glands, blood vessels, etc.) in a digital pathological slide image using image analysis techniques, such as deep learning-based image segmentation and classification models. This step aims to provide contextual information for subsequent deformation assessment, as different pathological types may have significantly different sensitivities to deformation features and their impact on diagnostic judgment.

[0102] The acquisition of the first parameter set corresponding to each pathological type can be understood as pre-establishing or dynamically adjusting a set of parameters for each identified pathological type. This first parameter set defines in detail the response weights of that pathological type to various deformation features (such as blurred edges, irregular shapes, and texture distortion), as well as the proportion of contribution of these deformation features to the decrease in confidence of the tumor recognition model at different degrees. Its purpose is to make the quantification of shape distortion more targeted and accurate, reflecting the actual impact of deformation on diagnostic confidence under specific pathological backgrounds.

[0103] In practical applications, the shape distortion index is obtained by weighted combination of multiple geometric features extracted from the microstructure based on the first parameter set. This means that after identifying the microstructure and acquiring its geometric features, the weights defined in the first parameter set are used to perform linear or nonlinear combination calculations on these geometric features to obtain a comprehensive numerical value, namely the shape distortion index. This index can more comprehensively reflect the degree of geometric deformation of the microstructure caused by various factors. For example, for the cell nucleus, its aspect ratio, roundness, and perimeter-to-area ratio, etc., can be weighted and summed according to the weights in the first parameter set to obtain a comprehensive shape distortion index.

[0104] Furthermore, dividing the shape distortion index into multiple levels means discretizing it into several predefined grades based on its numerical range, such as "slight distortion," "moderate distortion," and "severe distortion." This aims to simplify the understanding and processing of the degree of distortion and provide a foundation for subsequent diagnostic correlation of influencing factors.

[0105] Building upon this, obtaining the diagnostic impact factor for each level involves assigning a numerical value to each of the aforementioned shape distortion levels. This value represents the proportion of decreased confidence the tumor recognition model is expected to experience due to the distortion at that level. For example, slight distortion might correspond to a 5% decrease in confidence, moderate distortion to 15%, and severe distortion to 30%. These factors can be obtained through model training, expert experience, or statistical analysis, with the aim of transforming the abstract degree of shape distortion into a quantitative contribution to the decrease in diagnostic confidence.

[0106] Ultimately, quantifying the shape distortion index and diagnostic impact factors at each level into the degree of shape distortion means comprehensively considering the specific shape distortion index of the microstructure and the corresponding diagnostic impact factors at its level to form a final quantitative indicator that can directly reflect the impact of deformation on the model's confidence in its judgment. This indicator not only includes the degree of deformation itself but also incorporates its expected contribution to the uncertainty of diagnostic results, thus making the expression of "degree of shape distortion" more clinically and interpretably meaningful for the model.

[0107] Through the above technical solution, this application can achieve a more accurate and clinically significant quantification of the degree of microstructural deformation in digital pathological slide images. Specifically, by considering the pathological type of the microstructure and its response weight to deformation features, as well as the degree of influence of deformation on the judgment confidence of the tumor recognition model, the obtained shape distortion degree can more accurately reflect the contribution of image quality problems to the model's diagnostic uncertainty. This not only improves the accuracy and reliability of quality association interpretation, but also provides doctors with deeper insights, enabling them to understand the specific reasons for the model's low confidence judgment, thereby enhancing the interpretability and clinical practical value of the entire tumor medical pattern recognition system.

[0108] For example, suppose that when analyzing a digital pathological slide image, a glandular structure is identified in an image patch, and its pathological type is identified as "adenocarcinoma cell cluster". In this case, the system will obtain a first set of parameters corresponding to the "adenocarcinoma cell cluster" pathological type. This parameter set may define the weights of geometric features such as the roundness of the cell nucleus, the smoothness of the cell boundaries, and the regularity of the glandular cavities on the confidence of the tumor identification model for the adenocarcinoma cell cluster. For example, abnormal roundness of the cell nucleus may be assigned a higher weight because it is usually associated with malignancy, while abnormal regularity of the glandular cavities may be assigned a medium weight.

[0109] Next, the system extracts multiple geometric features from the glandular cell cluster, such as the aspect ratio, perimeter, area of ​​the cell nucleus, and local curvature of the cell boundary. These geometric features are weighted and combined according to the weights defined in the first parameter set to calculate a shape distortion index. For example, if the aspect ratio of the cell nucleus deviates significantly from the normal value and the cell boundary exhibits obvious serrations, the resulting shape distortion index after weighted combination will be higher.

[0110] Subsequently, the shape distortion index is divided into several preset levels, such as 0-0.2 for slight distortion, 0.2-0.5 for moderate distortion, and above 0.5 for severe distortion. Assuming the calculated shape distortion index is 0.6, it is classified into the "severe distortion" level. The system will then obtain the diagnostic impact factor corresponding to the "severe distortion" level; for example, this factor indicates that "severe distortion" is expected to cause a 30% decrease in the tumor recognition model's confidence in judging this region.

[0111] Ultimately, the shape distortion index (0.6) and its corresponding diagnostic impact factor (30%) were quantified as the degree of shape distortion of the image patch. This quantification not only indicates the severity of the deformation itself but also directly points to the extent to which this deformation is expected to negatively impact the diagnostic confidence of the tumor recognition model. In this way, when the tumor recognition model has low confidence in judging this region, the system can more accurately explain that this low confidence judgment is largely due to the severe shape distortion of the adenocarcinoma cell cluster, thus providing doctors with clear and actionable explanatory information.

[0112] Specifically, the steps for obtaining the first parameter set corresponding to the pathological type, which defines the response weight of the pathological type to the deformation feature and the degree of influence of the deformation feature on the judgment confidence of the tumor recognition model, include the following:

[0113] First, the confidence level of the tumor recognition model in judging different shape distortion features is assessed under its current operating state. These different shape distortion features can be understood as various microstructural deformations that may impair image quality and affect the model's judgment. For example, microstructural edge curvature caused by objective lens distortion may result in unnatural curvature of cell or tissue boundaries; blurring of internal textures caused by stray light or defocus may reduce the clarity of cell nuclei or cytoplasmic details; cell nucleus compression or stretching caused by uneven tissue slice thickness may alter the normal morphology of cells; and structural boundary contrast distortion caused by uneven staining may make certain tissue structures difficult to identify. This step aims to evaluate the model's sensitivity to these specific deformation features in real-time during practical applications.

[0114] Secondly, the confidence decline of the tumor recognition model in different shape distortion features is obtained on different datasets. This step aims to establish a generalized response pattern of the model to various shape distortion features by analyzing a large number of pre-collected and labeled datasets. These datasets can contain various pathological types and different degrees of image quality problems, thus providing a broader, statistically significant benchmark for model performance.

[0115] Finally, based on the decrease in confidence of the tumor recognition model regarding different shape distortion features under the current operating state and the decrease in confidence of the tumor recognition model regarding different shape distortion features on different datasets, the first parameter set is adjusted to obtain a calibrated first parameter set. The calibrated first parameter set defines the response weight of the pathological type to the deformation feature and the degree of influence of the deformation feature on the tumor recognition model's judgment confidence. Specifically, this adjustment process can employ methods such as weighted averaging, machine learning model calibration, or expert rule systems to combine real-time observed model performance with performance on historical datasets, thereby generating a more accurate parameter set adapted to the current operating environment.

[0116] Through the above technical solution, this application can provide a highly adaptive and accurate first parameter set, which accurately reflects the actual impact of different deformation features on the confidence of tumor recognition models under specific pathological types. This dynamic calibration mechanism overcomes the limitations of traditional fixed parameter sets and avoids evaluation bias caused by changes in the model operating environment or differences in data distribution. Therefore, the generated quality correlation interpretation will be more reliable and convincing, more effectively guiding physicians to understand the true source of uncertainty in model judgments, thereby improving the interpretability and clinical practical value of the entire tumor medical pattern recognition system.

[0117] In some embodiments, the steps described above for obtaining the degree of decrease in confidence in judging different distorted features of the tumor recognition model under the current operating state can be further refined.

[0118] Identify the pathological type to which the microstructure in the digital pathological slide image belongs;

[0119] For each identified pathological type, a set of test images with the characteristics of the pathological type and containing different shape distortion features are selected;

[0120] The test image is input into the tumor recognition model, and the confidence level of the tumor recognition model in judging the test image is recorded.

[0121] A set of reference images corresponding to the pathological type and without shape distortion is obtained, and the reference images are input into the tumor recognition model. The confidence of the tumor recognition model in judging the reference images is recorded.

[0122] By comparing the confidence level of the test image with that of the reference image, the decrease in confidence level of the tumor recognition model for different shape distortion features is calculated for each pathological type.

[0123] Specifically, identifying the pathological type of microstructures in digital pathology slide images involves using image analysis techniques or pre-trained classification models to identify and classify microstructures such as cells and tissues in digital pathology slide images to determine their specific pathological category, such as normal tissue, benign lesions, or specific types of malignant tumors. The purpose is to ensure that subsequent assessments of the decline in confidence are performed against a specific pathological context, as different pathological types may have varying sensitivities to image quality issues.

[0124] Furthermore, for each identified pathological type, a set of test images exhibiting the characteristics of that pathological type and containing different shape distortion features are selected. These test images are carefully selected or artificially synthesized images that not only display the typical morphological features of a specific pathological type but also intentionally incorporate various known shape distortion features, such as microstructural edge curvature caused by objective lens distortion, blurring of internal microstructural textures caused by stray light or defocus, cell nucleus compression or stretching caused by uneven tissue section thickness, and structural boundary contrast distortion caused by uneven staining. These test images are used to simulate various quality problems that may occur in actual pathological images.

[0125] Subsequently, the test image is input into the tumor recognition model, and the model's confidence level in judging the test image is recorded. After analyzing the test image, the tumor recognition model outputs a confidence value, which characterizes the model's confidence in the presence of a specific tumor type in the image. Recording these confidence values ​​can intuitively reflect the changes in the model's diagnostic certainty under different shape distortion features.

[0126] Simultaneously, a set of reference images corresponding to the stated pathological type and without shape distortion are acquired, and these reference images are input into the tumor recognition model. The model's confidence level in the reference images is recorded. The reference images are those corresponding to the same pathological type as the test images, but with ideal or near-ideal image quality, i.e., containing little or no shape distortion. These reference images serve as benchmarks to obtain the model's confidence level under optimal image quality conditions.

[0127] Finally, the confidence level of the test image is compared with that of the reference image to calculate the decrease in confidence level of the tumor recognition model for different shape distortion features under each pathological type. By comparing the confidence level of the test image with shape distortion with that of the reference image without shape distortion, the specific decrease in confidence level of the tumor recognition model caused by each particular shape distortion feature can be quantified.

[0128] The aforementioned technical solution allows for precise quantification of the specific impact of different shape distortion features on the judgment confidence of the tumor recognition model under its current operating state. This makes the attribution of model performance degradation clearer and helps identify the image quality issues that have the greatest impact on the model's judgment confidence. Consequently, it provides a refined, data-driven basis for subsequent adjustments to the first parameter set, thereby more accurately defining the response weights of pathological types to deformation features and the degree of influence of deformation features on the judgment confidence of the tumor recognition model, ultimately improving the robustness and interpretability of the entire tumor medical pattern recognition method.

[0129] like Figure 2 As shown in the illustration, this application also discloses an image detection-based tumor medical pattern recognition system 100, which includes:

[0130] The quality information acquisition module 10 is used to segment the digital pathological slide image into multiple image blocks, and to evaluate the color distribution, brightness contrast, and geometric features of the microstructure in each image block to obtain a local quality assessment map; wherein the color or value of each image block region in the local quality assessment map represents the degree of image quality damage corresponding to that image block region.

[0131] The confidence acquisition module 20 is used to analyze the digital pathological slide image based on the tumor recognition model and generate a local diagnostic judgment map; wherein the color or value of each image block region in the local diagnostic judgment map represents the confidence of the tumor recognition model in making a tumor diagnosis at that location, and the judgment confidence represents the probability value or confidence score of the image block being identified as a specific tumor type.

[0132] The association interpretation generation module 30 is used to analyze the association between the low-confidence region in the local diagnostic judgment image and the quality-damaged region in the local quality assessment image. If the overlap between the low-confidence region and the quality-damaged region meets the preset conditions, a quality association interpretation is generated. The quality association interpretation is used to explain that the uncertainty in the judgment of the low-confidence region is due to image quality problems.

[0133] The association description presentation module 40 is used to generate an interpretability report based on the quality association interpretation. The interpretability report includes the original pathological image, local diagnostic judgment map, local quality assessment map, and a visual presentation of the quality association interpretation.

[0134] The core innovation of this application lies in the introduction of a correlation analysis mechanism between the image quality assessment module and the diagnostic confidence acquisition module in its system architecture. Specifically, the system uses the quality information acquisition module to perform refined local quality assessment on digital pathological slide images, generating a local quality assessment map that enables the system to identify locally compromised areas in the image. Simultaneously, the confidence acquisition module generates a local diagnostic judgment map using a tumor recognition model, intuitively presenting the model's confidence level in different regions. Most importantly, the correlation interpretation generation module analyzes the correlation between low-confidence areas in the local diagnostic judgment map and quality-impaired areas in the local quality assessment map. If the overlap meets preset conditions, a quality correlation interpretation is generated. This mechanism allows the system to explicitly attribute model uncertainty to image quality issues, thus providing pathologists with a clear explanation.

[0135] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A tumor medical pattern recognition method based on image detection, characterized in that, include: The digital pathological slide image is segmented into multiple image blocks, and the color distribution, brightness contrast, and geometric features of the microstructure in each image block are evaluated to obtain a local quality assessment map; wherein the color or value of each image block region in the local quality assessment map represents the degree of image quality impairment corresponding to that image block region. Based on the analysis of the digital pathological slide images by the tumor recognition model, a local diagnostic judgment map is generated; wherein the color or value of each image block region in the local diagnostic judgment map represents the confidence of the tumor recognition model in making a tumor diagnosis at that location, and the judgment confidence represents the probability value or confidence score of the image block being identified as a specific tumor type. The correlation analysis is performed between the low-confidence region in the local diagnostic judgment image and the quality-damaged region in the local quality assessment image. If the overlap between the low-confidence region and the quality-damaged region meets a preset condition, a quality correlation interpretation is generated. The quality correlation interpretation is used to explain that the uncertainty in the judgment of the low-confidence region stems from image quality issues. Based on the quality correlation interpretation, an interpretability report is generated, which includes the original pathological image, local diagnostic judgment map, local quality assessment map, and a visual presentation of the quality correlation interpretation. The steps for generating a quality correlation interpretation include: Associating low-confidence regions in the local diagnostic judgment map with quality-impaired regions in the local quality assessment map; if the overlap between the low-confidence regions and the quality-impaired regions meets a preset condition. The low-confidence regions are identified by determining areas in the local diagnostic judgment map where the judgment confidence is lower than a preset judgment confidence threshold. Identify regions in the local quality assessment map whose quality damage level exceeds a preset quality damage threshold to obtain the quality damage regions; Calculate the intersection area of ​​the low confidence region and the quality-impaired region; If the ratio of the intersection area to the area of ​​the low confidence region exceeds a preset overlap ratio threshold, then the generated quality association interpretation is generated. The steps of segmenting a digital pathological slide image into multiple image blocks and evaluating the color distribution, contrast, and geometric features of the microstructure in each image block to obtain a local quality assessment map include: The digital pathological slide image is divided into multiple image blocks; For each image block, its color distribution and brightness contrast are obtained and compared with preset standard reference values ​​to obtain color deviation and brightness contrast deviation; For each image patch, the microstructure is identified, the geometric features of the microstructure are extracted, and the geometric features are compared with a preset geometric reference benchmark to obtain the degree of shape distortion. The preset geometric reference benchmark refers to the set of geometric parameters of the natural normal microstructure in a standard pathological slide under the optimal state of the objective lens, which is used to determine whether the actual microstructure geometric features are distorted due to image quality issues. The local quality assessment map is generated based on the color deviation, the brightness contrast deviation, and the degree of shape distortion.

2. The tumor medical pattern recognition method according to claim 1, characterized in that, The step of analyzing the digital pathological slide image based on the tumor recognition model to generate a local diagnostic judgment map includes: Receive the original judgment map output by the tumor recognition model, wherein the original judgment map contains the probability value of each local region being the target object; The probability values ​​are mapped to color codes to generate a local diagnostic judgment map, in which high probability values ​​correspond to high confidence regions and low probability values ​​correspond to low confidence regions.

3. The tumor medical pattern recognition method according to claim 1, characterized in that, The preset standard reference value is obtained by scanning a series of standard pathological sections with the objective lens in optimal condition, extracting the color distribution feature value and brightness contrast feature value of the local image block of each section, and then statistically analyzing them. The preset geometric reference benchmark refers to the set of standard geometric parameters determined after analyzing the geometric characteristics of naturally occurring normal microstructures in a series of standard pathological sections when the objective lens is in its optimal state.

4. The tumor medical pattern recognition method according to claim 1, characterized in that, The steps of identifying the microstructure within each image patch, extracting the geometric features of the microstructure, and comparing the geometric features with a preset geometric reference to obtain the degree of shape distortion include: Obtain the spatial distribution characteristics and internal morphological characteristics of the microstructure; The spatial distribution characteristics are compared with a preset tissue preparation deformation pattern to obtain a first comparison result; the internal morphological characteristics are compared with a preset optical deformation pattern to obtain a second comparison result; wherein, the preset tissue preparation deformation pattern refers to a reference model established by statistically analyzing the microstructural deformation patterns caused by the thickness or staining of tissue preparation sections in standard pathological sections; the preset optical deformation pattern refers to a reference model established by statistically analyzing the microstructural deformation patterns caused by objective lens distortion and stray light effects; Based on the first comparison result and the second comparison result, the deformation of the microstructure is attributed to a deformation attribution result, which is used to indicate whether the deformation is mainly caused by tissue preparation or staining process, or mainly caused by objective lens deformation. Based on the deformation attribution results, the deformation caused by tissue preparation or staining process is compensated or eliminated; and the compensationd or eliminated deformation is converted into the degree of shape distortion.

5. The tumor medical pattern recognition method according to claim 4, characterized in that, The step of converting the compensated or excluded deformation into the degree of shape distortion includes: Identify the pathological type to which the microstructure belongs; Obtain the first parameter set corresponding to the pathological type. The first parameter set defines the response weight of the pathological type to the deformation feature and the degree of influence of the deformation feature on the judgment confidence of the tumor recognition model. Based on the first parameter set, multiple geometric features extracted from the microstructure are weighted and combined to obtain the shape distortion index; The shape distortion index is divided into multiple levels; Obtain the diagnostic impact factor corresponding to each level, whereby the diagnostic impact factor represents the expected contribution ratio of hierarchical distortion to the decrease in confidence of the tumor identification model. The shape distortion index and the diagnostic influence factors at each level are quantified into the degree of shape distortion.

6. The tumor medical pattern recognition method according to claim 5, characterized in that, The step of obtaining the first parameter set corresponding to the pathological type, wherein the first parameter set defines the response weight of the pathological type to the deformation feature and the degree of influence of the deformation feature on the judgment confidence of the tumor recognition model, includes: The determination of the tumor recognition model's confidence level for different distorted features under the current operating state is obtained; wherein, the different distorted features include the bending of microstructure edges caused by objective lens distortion, the blurring of internal texture of microstructure caused by stray light or defocus, the compression or stretching of cell nuclei caused by uneven tissue slice thickness, and the distortion of structural boundary contrast caused by uneven staining. The confidence level of the tumor recognition model for different distorted features was obtained on different datasets. Based on the decrease in confidence of the tumor recognition model for different shape distortion features under the current operating state and the decrease in confidence of the tumor recognition model for different shape distortion features on different datasets, the first parameter set is adjusted to obtain a calibrated first parameter set. The calibrated first parameter set defines the response weight of the pathological type to the deformation feature and the degree of influence of the deformation feature on the judgment confidence of the tumor recognition model.

7. The tumor medical pattern recognition method according to claim 4, characterized in that, The step of obtaining the decrease in confidence level of the tumor recognition model for different shape distortion features in the current operating state includes: Identify the pathological type to which the microstructure in the digital pathological slide image belongs; For each identified pathological type, a set of test images with pathological type characteristics and containing different shape distortion features are selected; The test image is input into the tumor recognition model, and the confidence level of the tumor recognition model in judging the test image is recorded. A set of reference images corresponding to the pathological type and without shape distortion is obtained, and the reference images are input into the tumor recognition model. The confidence of the tumor recognition model in judging the reference images is recorded. By comparing the confidence level of the test image with that of the reference image, the decrease in confidence level of the tumor recognition model for different shape distortion features is calculated for each pathological type.

8. A tumor medical pattern recognition system based on image detection, characterized in that, The system includes: The quality information acquisition module is used to segment digital pathological slide images into multiple image blocks, and to evaluate the color distribution, brightness contrast, and geometric features of the microstructure in each image block to obtain a local quality assessment map. The confidence acquisition module is used to analyze the digital pathological slide image based on the tumor recognition model and generate a local diagnostic judgment map. The correlation interpretation generation module is used to correlate and analyze the low-confidence region in the local diagnostic judgment image and the quality-damaged region in the local quality assessment image. If the overlap between the low-confidence region and the quality-damaged region meets the preset conditions, a quality correlation interpretation is generated. The quality correlation interpretation is used to explain that the uncertainty in the judgment of the low-confidence region is due to image quality problems. The association explanation presentation module is used to generate an interpretability report based on the quality association explanation. The interpretability report includes the original pathological image, local diagnostic judgment map, local quality assessment map, and a visual presentation of the quality association explanation. The correlation analysis involves comparing low-confidence regions in the local diagnostic judgment map with quality-impaired regions in the local quality assessment map. If the overlap between the low-confidence regions and the quality-impaired regions meets a preset condition, a quality correlation interpretation is generated, including: The low-confidence regions are identified by determining areas in the local diagnostic judgment map where the judgment confidence is lower than a preset judgment confidence threshold. Identify regions in the local quality assessment map whose quality damage level exceeds a preset quality damage threshold to obtain the quality damage regions; Calculate the intersection area of ​​the low confidence region and the quality-impaired region; If the ratio of the intersection area to the area of ​​the low confidence region exceeds a preset overlap ratio threshold, then the generated quality association interpretation is generated. The process involves segmenting the digital pathological slide image into multiple image patches, and then evaluating the color distribution, contrast, and geometric features of the microstructure within each patch to obtain a local quality assessment map, including: The digital pathological slide image is divided into multiple image blocks; For each image block, its color distribution and brightness contrast are obtained and compared with preset standard reference values ​​to obtain color deviation and brightness contrast deviation; For each image patch, the microstructure is identified, the geometric features of the microstructure are extracted, and the geometric features are compared with a preset geometric reference benchmark to obtain the degree of shape distortion. The preset geometric reference benchmark refers to the set of geometric parameters of the natural normal microstructure in a standard pathological slide under the optimal state of the objective lens, which is used to determine whether the actual microstructure geometric features are distorted due to image quality issues. The local quality assessment map is generated based on the color deviation, the brightness contrast deviation, and the degree of shape distortion.

Images