WSI image quantitative evaluation system based on cross-scale feature cooperation and multi-dimensional quality factor fusion

Abstract

The invention discloses a WSI image quantitative evaluation system based on cross-scale feature collaboration and multi-dimensional quality factor fusion, belongs to the technical field of digital pathological image analysis, and aims at solving the problems that an existing evaluation method is single in dimension and insufficient in robustness and efficiency. The system obtains a WSI image through scanning, randomly samples a tissue area under a basic amplification factor, screens out an unclear area by adopting a definition detection method, and feeds back a re-scanning instruction. And for the clear area, synchronously extracting image slices with different objective lens magnifications under the same hierarchy of the pyramid structure, and constructing a multi-magnification slice data set. And carrying out cell identification and counting based on high-magnification slices, carrying out bubble area detection and staining defect identification based on low-magnification slices, and respectively generating cell number, slice integrity and staining quality sub-item scores. And finally, performing weighted fusion on each item score to obtain a slice comprehensive quality score, and realizing objective quantitative evaluation on the WSI image quality.

Description

Technical Field

[0001] This invention designs a WSI image quantization evaluation system based on cross-scale feature collaboration and multi-dimensional quality factor fusion. Background Technology

[0002] In recent years, with the rapid development of digital pathology technology, traditional microscopic pathological diagnosis is gradually moving towards digitalization and intelligence. However, in the actual acquisition process, WSI (whole slide image scanning) images are often affected by factors such as scanning equipment performance, lighting conditions, tissue preparation, and staining processes, easily resulting in quality defects such as blurring, uneven color, stains, bubbles, cracks, and wrinkles. These problems can lead to the loss of cell structural information or morphological distortion, thereby affecting the accuracy of subsequent automatic analysis and diagnosis of pathological images. Therefore, before entering the diagnostic stage, the quality of WSI images must be rigorously and reliably evaluated and screened.

[0003] Currently, quality control of pathological images mainly relies on manual inspection, where experienced pathologists make subjective judgments based on indicators such as sharpness, staining consistency, and tissue integrity. However, this method is time-consuming, costly, and highly subjective, and consistency among different pathologists is difficult to guarantee, making it unsuitable for the automated and standardized management of large-scale pathological images. To address this, researchers have proposed automated quality assessment methods based on image processing and deep learning. For example, traditional sharpness indicators such as Laplacian variance and Tenengrad gradient are used to quantify image blur, or convolutional neural networks (CNNs) are employed to automatically extract features, achieving blur detection and staining quality assessment.

[0004] Although the above methods have improved the level of automation to some extent, there are still obvious shortcomings: (1) The evaluation dimension is singular, often focusing only on clarity or staining quality, lacking comprehensive consideration of stains, cell density and tissue integrity; (2) The algorithm has limited robustness and is difficult to adapt to changes in different scanning equipment, staining conditions and tissue types; (3) There is a lack of interpretable and standardized quantitative scoring system, which cannot be directly correlated with clinical quality control standards; (4) There is a lack of quality closed-loop control mechanism, which cannot realize automatic feedback and rescanning of unqualified slides.

[0005] Therefore, there is an urgent need to construct a comprehensive WSI image quality assessment system. This system should be able to objectively, quantitatively, and automatically evaluate various indicators of pathological slides, such as clarity, stain consistency, stain distribution, and cell density, and establish an interpretable sub-item and total score scoring mechanism, thereby providing solid technical support for the standardized quality control and intelligent analysis of pathological images. Summary of the Invention

[0006] This invention aims to address the problems of existing digital pathology whole-slice image (WSI) quality assessment methods, such as limited assessment dimensions, lack of robustness, and low efficiency. This invention aims to provide a WSI image quality assessment system that integrates multiple indicators, combines rapid sampling with multi-model collaborative analysis, and achieves objective and multi-level evaluation of slide quality.

[0007] The aforementioned objectives are primarily achieved through the following technical solutions:

[0008] S1. Scan the pathological sample with a scanner to acquire WSI whole slide images with different quality levels, and select tissue regions of a preset size as input candidate regions for subsequent quality assessment using a random sampling strategy at the basic magnification of the whole slide images.

[0009] S2. Construct a clarity detection method that integrates depth features and edge information, and determine the clarity of the sampling area determined in step S1. If it is determined to be unclear, i.e. unqualified, the module system outputs a rescan feedback command for the WSI full slice image.

[0010] S3. Based on the clear WSI sampling area determined in step S2, simultaneously extract image slice units of preset size corresponding to different objective magnifications at a specific number of layers in the pyramid structure, and perform manual quality annotation on the obtained multi-magnification slice unit set to construct a slice dataset for subsequent quality assessment tasks.

[0011] S4. Perform tissue region preprocessing on the high-magnification slice dataset, remove non-target tissue regions with abnormal cell area through morphological opening operation, identify and count cells in the processed tissue regions, and remove the interfering slices with the most and fewest cells based on the counting results. For the finally retained effective slices, generate cell number sub-scores based on preset number thresholds, and generate corresponding cell number slice scores in combination with preset weight coefficients.

[0012] S5. In parallel, a pre-trained deep learning instance segmentation model is used to detect bubble regions in the slice dataset at low magnification. Based on the bubble detection results, the ratio of the bubble region area to the original tissue region area is calculated. A bubble stain sub-score is generated based on the ratio, and a corresponding slice integrity score is generated by combining the preset weight coefficients.

[0013] S6. In parallel, a pre-trained deep learning classification model is used to evaluate the Newbury staining quality of the slice dataset at low magnification. The model is used to identify staining defect patterns, including excessively dark cell nucleus staining and uneven background staining. The model outputs a staining defect degree index to determine whether it meets the preset qualified standard, and combines the preset weight coefficients to generate a staining quality sub-score.

[0014] S7. Perform a comprehensive quality evaluation of the slices, summarize the scores of each item obtained from S4 to S6 and perform weighted fusion to obtain the final comprehensive quality score, and give the final quality judgment result based on the comparison of the final comprehensive quality score with the preset threshold.

[0015] Invention Effects

[0016] This invention provides an automated, multi-dimensional quality assessment system for WSI images. By constructing a comprehensive evaluation framework covering multiple indicators such as sharpness, cell count, air bubbles and staining quality, it effectively solves the problems of single-dimensional WSI image quality assessment, low efficiency, high dependence on manual intervention, and insufficient robustness in existing technologies.

[0017] First, this invention employs a random sampling strategy combined with a comprehensive sharpness determination mechanism at 1x objective magnification to achieve efficient pre-screening of WSI, effectively avoiding the high latency problem caused by redundant calculations in traditional methods. Simultaneously, by utilizing multi-magnification slicing units, the overall tissue structure and local cellular details can be analyzed simultaneously within a unified framework, significantly enhancing the system's generalization ability.

[0018] Secondly, the cell count evaluation module combines morphological preprocessing with cell identification and counting methods to quickly eliminate areas with abnormal cell counts and generate reliable cell count scores; the bubble detection module uses a deep learning instance segmentation model to automatically identify bubble occlusion artifacts in the slides and achieves objective quantification through area ratio calculation; the staining quality assessment module is based on a deep learning classification model to quantify key staining defects, effectively reducing the subjective bias of manual interpretation.

[0019] Finally, this invention constructs a unified comprehensive quality scoring system by weighted fusion of three scores: cell number, bubble stain area, and staining quality. It also obtains highly reliable effective tissue areas through logical intersection, providing high-quality input data for subsequent AI diagnosis, image analysis, and pathological modeling. In summary, this invention significantly improves the speed, accuracy, and robustness of WSI quality assessment, and has broad application prospects and practical promotion value. Attached Figure Description

[0020] Figure 1 A general flowchart of a WSI image quality assessment system provided in this application embodiment;

[0021] Figure 2 A flowchart for determining the sharpness of a WSI sampling area is provided in this application embodiment;

[0022] Figure 3 This is a schematic diagram of the sharpness discrimination result of a WSI sampling area provided in an embodiment of this application;

[0023] Figure 4 A schematic diagram illustrating the process of constructing a multi-scale slice unit dataset provided in an embodiment of this application;

[0024] Figure 5 This is a schematic diagram of cell counting results in a sliced ​​tissue region provided in an embodiment of this application;

[0025] Figure 6 This is a schematic diagram illustrating the detection and identification results of a sliced ​​bubble region provided in an embodiment of this application.

[0026] Figure 7 A flowchart for identifying defects in the quality of ThinPrep staining of tissue sections is provided for embodiments of this application.

[0027] Figure 8 A flowchart illustrating the comprehensive quality scoring and final determination of slices provided in this application embodiment. Specific implementation methods

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Specific implementation method one:

[0031] This application provides a WSI image quality assessment system, the overall flowchart of which is shown below. Figure 1 As shown, it includes the following steps:

[0032] S1. Scan the pathological sample with a scanner to acquire WSI whole slide images with different quality levels, and select a 1024×1024 pixel tissue region at 1x magnification of the whole slide image as the input candidate region for subsequent quality assessment using a random sampling strategy.

[0033] S2. Construct a comprehensive slice sharpness determination module that combines deep learning methods and traditional image processing methods to determine the sharpness of the sampling area determined in step S1. If the image is determined to be unclear (unqualified), the module system outputs a rescan feedback instruction for the WSI full slice image.

[0034] S3. Based on the sampling area of ​​the clear WSI determined in step S2, simultaneously extract the 512×512 pixel image slice units at 1x and 10x objective lens magnification corresponding to the third layer of the pyramid structure, and perform manual quality annotation on the obtained multi-magnification slice unit set to construct a slice dataset for subsequent quality assessment tasks.

[0035] S4. Perform tissue region preprocessing on the slice dataset under 10x objective magnification. Remove non-target tissue regions with abnormal cell area by morphological opening operation. Identify and count cells in the processed tissue regions. Based on the counting results, remove the interfering slices with the most and fewest cells. For the finally retained effective slices, generate a cell number sub-score based on a preset number threshold and generate a corresponding cell number slice score based on a preset weight coefficient.

[0036] S5. In parallel, a pre-trained deep learning instance segmentation model is used to detect bubble regions in the slice dataset under 1x objective lens magnification. Based on the bubble detection results, the ratio of the bubble region area to the original tissue region area is calculated. Based on the ratio, a bubble stain sub-score is generated, and a corresponding slice integrity score is generated by combining the preset weight coefficients.

[0037] S6. In parallel, a pre-trained deep learning classification model is used to evaluate the Newbury staining quality of the slice dataset at 1x magnification. The model is used to identify staining defect patterns, including excessively dark cell nucleus staining and uneven background staining. The model outputs a staining defect degree index to determine whether it meets the preset qualified standard, and generates a staining quality sub-score by combining the preset weight coefficients.

[0038] S7. Perform a comprehensive quality evaluation of the slices, summarize the scores of each item obtained from S4 to S6 and perform weighted fusion to obtain the final comprehensive quality score, and give the final quality judgment result based on the comparison of the final comprehensive quality score with the preset threshold.

[0039] The embodiments of the present invention will be described in detail below:

[0040] S1. The specific implementation method of WSI image acquisition and candidate region selection is as follows:

[0041] The system first acquires WSI images for quality assessment. To ensure the generalization ability of the training and evaluation model, the acquired WSI images should be diverse, specifically including: scanner diversity, acquiring WSI images from various models and manufacturers of digital pathology scanners; tissue type diversity, including various pathological tissue types; and staining batch diversity, covering sections from different staining batches to cover variations in staining depth, uniformity, etc.

[0042] The WSI full-slice image is processed at 1x magnification. Based on a random sampling strategy, the system selects several 1024 × 1024 pixel tissue regions from the tissueed areas as input candidate regions for subsequent quality assessment.

[0043] This random sampling strategy aims to ensure that subsequent sharpness determination (step S2) and other quality assessments (steps S3 to S6) cover representative areas of the full-slice image.

[0044] S2. The specific working principle of the slice sharpness detection method is as follows:

[0045] The system first divides the 1024 × 1024 pixel candidate region into multiple sub-image blocks, and then processes each sub-image block in parallel.

[0046] First, a pre-trained ResNet-34 model is used to extract features from each sub-image patch, and the sharpness of each sub-image patch is initially classified based on the feature vectors to obtain a preliminary determination of each image unit based on depth features.

[0047] Simultaneously, a Laplacian variance algorithm is introduced to extract edge features from the aforementioned tissue regions. Based on the comparison between the calculated Laplacian variance value and a preset threshold, it is determined whether the sub-image patch is a blurred patch. The Laplacian variance calculation formula is as follows:

[0048] (1)

[0049] In the formula, Total number of pixels It is the first Laplacian value of each pixel, The mean of the Laplacian values ​​for all pixels;

[0050] The principle for determining the sharpness of the entire image is as follows: calculate the Laplace variance for each slice. and compare it with a preset threshold. When comparing, The sub-image block is determined to be a blurred block; the proportion of all blurred blocks to the total number of blocks is counted. :

[0051] (2)

[0052] In the formula, The number of blurred sub-image blocks, This represents the total number of sub-image blocks. When the scale... Greater than or equal to the preset ratio threshold At that time, the entire slice was determined to be a blurry image.

[0053] The sharpness detection results of the aforementioned depth features and edge information are comprehensively evaluated. Based on the fused result, a final sharpness assessment and feedback instruction are given. The sharpness assessment flowchart is as follows: Figure 2 As shown in the diagram, the sharpness discrimination is as follows: Figure 3 As shown.

[0054] S3. The specific implementation method for constructing the slice dataset is as follows:

[0055] This step aims to construct a multi-magnification, multi-label slice dataset based on the qualified and clear regions identified in step S2, for subsequent evaluation. The specific flowchart for constructing the dataset is shown below. Figure 3 As shown.

[0056] In the clear WSI sampling area determined in step S2, 512×512 pixel image slice units at 1x and 10x objective lens magnification corresponding to the third layer of the pyramid structure are extracted simultaneously to form a multi-magnification slice unit set.

[0057] The slice unit set was manually quality-labeled, primarily using image annotation tools such as Labelme. The annotations included labels for cell sufficiency, bubble region boundary masks, and defect classification labels for ThinPrep staining quality. Based on the annotation results, the dataset was divided into training, validation, and test sets in a ratio of 7:2:1.

[0058] The deep learning instance segmentation model in step S5 and the deep learning classification model in step S6 are trained and their parameters optimized using the training set. The training process aims to improve the model's fit by iteratively optimizing the model parameters and minimizing the loss function between the model's predicted output and the manually labeled data. The detection accuracy of each model and the overall quality assessment of the system are evaluated independently using the test set.

[0059] S4. The principle of the cell number estimation method is as follows:

[0060] This module aims to estimate and filter cell counts in slices of WSI images and generate cell count quality scores based on representative count results. Figure 4 This is a schematic diagram of the cell counting results.

[0061] Morphological opening operations were performed on the section images at 10x objective magnification to remove non-target tissue regions with abnormal area. Subsequently, automatic cell identification and counting were performed on the processed tissue region images to calculate the number of cells in each 512×512 section under the same WSI. , obtain the set Based on the statistical results of the cell count, 5% of the interfering slices with the highest and lowest cell counts were removed to obtain the retain set. To ensure the quality representativeness of the tissue sections ultimately used for scoring, the mean of the retained set is taken as the representative number of cells.

[0062] (3)

[0063] Will Substitute into the cell number sub-score formula for calculation :

[0064] (4)

[0065] in, The minimum number of qualified cells, To determine the ideal cell count, a threshold standard was established based on the TBS standard and empirical calibration method. It is the lowest score. This is the highest score, usually taken as 0 or 100;

[0066] If the number of cells is less than The score is 0; if the number of cells is between and Between these points, the score increases linearly proportionally; if the number of cells reaches... Or higher, with a score of 100;

[0067] Finally, the cell count will be scored separately. With preset weights Weighted scores were applied to obtain cell quantity and quality scores.

[0068] (5)

[0069] S5. The principle of the bubble detection method is as follows:

[0070] This module aims to detect bubble regions in low-magnification slices of WSI images and generate a bubble quality score based on the proportion of effective tissue area. A schematic diagram of the slice bubble detection and recognition results is shown below. Figure 5 As shown.

[0071] After extracting the tissue region from the 1x objective lens slice image, a pre-trained instance segmentation model is used to detect bubble regions. The areas of all detected bubble regions are then merged and denoted as... Let the area of ​​the original tissue region be denoted as The effective organization area The calculation formula is:

[0072] (6)

[0073] Calculate the ratio of the effective tissue region area to the original tissue region area to obtain the effective region proportion. :

[0074] (7)

[0075] Based on the proportion of effective areas Generate a score for the bubble stain item. :

[0076] (8)

[0077] When the effective tissue percentage is less than 70%, it indicates severe bubble occlusion, and the score is 0. When the effective tissue percentage is between 70% and 98%, the score is calculated according to a linear ratio. When the effective tissue percentage reaches or exceeds 98%, the bubble stain is considered to have minimal impact, and the score is 100.

[0078] Ultimately With preset weights Weighted scores are applied to obtain the slice integrity score. :

[0079] (9)

[0080] S6. The principle of staining evaluation is as follows:

[0081] This module aims to evaluate the staining quality of WSI slices using a deep learning classification model and generate a quantified staining quality score based on the defect probability output by the model. The staining identification flowchart is shown below. Figure 6 As shown.

[0082] Based on a pre-trained deep learning classification model, feature extraction is performed on slice images under 1x objective lens magnification. The model is used to identify multiple staining defects, including excessively dark cell nuclei and uneven background staining. The model outputs the probability that the slice has this type of defect. ;

[0083] Based on the output of the deep learning classification model, calculate the total defect probability score representing the degree of staining defects. :

[0084] (10)

[0085] in, The closer this value is to 1, the higher the degree of staining defects in the slice;

[0086] Based on the total defect probability score Generate staining sub-scores :

[0087] (11)

[0088] A linear penalty function is used for scoring to ensure that the score is strictly negatively correlated with the defect probability.

[0089] Ultimately With preset weights Weighted scores were applied to obtain the staining quality score of the slides. :

[0090] (12)

[0091] S7. The principle for evaluating the overall quality of the slices is as follows:

[0092] This module is the final step in the quality assessment system. It aims to integrate various quality scoring indicators, calculate a comprehensive score, and provide a final pass / fail determination based on preset thresholds. The comprehensive quality scoring flowchart is shown below. Figure 7 As shown.

[0093] This step summarizes the quality scores obtained in steps S4, S5, and S6, and performs a weighted average calculation on these scores based on preset weighting coefficients to obtain the final comprehensive quality score of the slice.

[0094] (13)

[0095] in, The final comprehensive quality score is compared with a preset quality pass threshold. A comparison is made to determine whether the slice is ultimately of acceptable or unacceptable quality. The acceptable quality threshold is... The preset value is 85 to ensure that the slides meet the minimum quality requirements for pathological diagnosis in three dimensions: cell count, air bubbles / stains, and staining quality. If the judgment result is acceptable, the WSI image can proceed to the subsequent pathological diagnosis process; if the judgment result is unacceptable, it should be rescanned or subjected to other processing.

Claims

1. A WSI image quantization evaluation system based on cross-scale feature collaboration and multi-dimensional quality factor fusion, characterized in that, include: S1. Scan the pathological sample with a scanner to acquire WSI whole slide images with different quality levels, and select tissue regions of a preset size as input candidate regions for subsequent quality assessment using a random sampling strategy at the basic magnification of the whole slide images. S2. Construct a clarity detection method that integrates depth features and edge information, and determine the clarity of the sampling area determined in step S1. If it is determined to be unclear, i.e. unqualified, the module system outputs a rescan feedback command for the WSI full slice image. S3. Based on the clear WSI sampling area determined in step S2, simultaneously extract image slice units of preset size corresponding to different objective magnifications at a specific number of layers in the pyramid structure, and perform manual quality annotation on the obtained multi-magnification slice unit set to construct a slice dataset for subsequent quality assessment tasks. S4. Perform tissue region preprocessing on the high-magnification slice dataset, remove non-target tissue regions with abnormal cell area through morphological opening operation, identify and count cells in the processed tissue regions, and remove the interfering slices with the most and fewest cells based on the counting results. For the finally retained effective slices, generate cell number sub-scores based on preset number thresholds, and generate corresponding cell number slice scores in combination with preset weight coefficients. S5. In parallel, a pre-trained deep learning instance segmentation model is used to detect bubble regions in the slice dataset at low magnification. Based on the bubble detection results, the ratio of the bubble region area to the original tissue region area is calculated. A bubble stain sub-score is generated based on the ratio, and a corresponding slice integrity score is generated by combining the preset weight coefficients. S6. In parallel, a pre-trained deep learning classification model is used to evaluate the Newbury staining quality of the slice dataset at low magnification. The model is used to identify staining defect patterns, including excessively dark cell nucleus staining and uneven background staining. The model outputs a staining defect degree index to determine whether it meets the preset qualified standard, and combines the preset weight coefficients to generate a staining quality sub-score. S7. Perform a comprehensive quality evaluation of the slices, summarize the scores of each item obtained from S4 to S6 and perform weighted fusion to obtain the final comprehensive quality score, and give the final quality judgment result based on the comparison of the final comprehensive quality score with the preset threshold.

2. The WSI image quality assessment system as described in claim 1, characterized in that, The method for constructing candidate regions for quality assessment in step S1 is as follows: S1.

1. Collect WSI images from multiple scanners, multiple pathological tissue types, and multiple staining batches to ensure the diversity of data sources; S1.2 At a low magnification of the WSI full slice image, several tissue regions of preset size are selected according to a random sampling strategy as candidate regions for subsequent quality assessment.

3. The WSI image quality assessment system as described in claim 1, characterized in that, The principle of the sharpness detection method that fuses deep features and edge information in step S2 is as follows: S2.1 For the tissue region in S1.2, use the pre-trained ResNet-34 model to extract features for each sub-image block, and perform preliminary sharpness classification for each sub-image block based on the feature vector to obtain preliminary determination of each image unit based on depth features. S2.

2. A parallel Laplacian variance algorithm is introduced to simultaneously extract edge features from the tissue region in S1.

2. Based on the comparison between the calculated Laplacian variance value and a preset threshold, it is determined whether the sub-image patch is a blurred patch. The Laplacian variance calculation formula is as follows: （1） In the formula, Total number of pixels It is the first Laplacian value of each pixel, The mean of the Laplacian values ​​for all pixels is used to calculate the Laplacian variance for each slice. and compare it with a preset threshold. When comparing, When a sub-image block is determined to be a blurred block, the proportion of all blurred blocks to the total number of blocks is counted. When the proportion is greater than or equal to a preset proportion threshold, the entire slice is determined to be a blurred image. S2.

3. The sharpness detection results of the above depth features and edge information are comprehensively judged, and the final sharpness evaluation and feedback instructions are given based on the fused comprehensive results.

4. The WSI image quality assessment system as described in claim 1, characterized in that, The specific method for constructing the slice dataset in step S3 is as follows: S3.1 In the clear WSI sampling area determined in step S2, image slice units of preset size under two different objective magnifications corresponding to the specified level of the pyramid structure are extracted simultaneously to form a multi-magnification slice unit set, and the slice units are quality labeled. S3.2 Based on the annotation results, the dataset is divided into a training set, a validation set, and a test set. The training set is used to perform supervised training and parameter optimization on the deep learning instance segmentation model and the deep learning classification model. The test set is used to evaluate the detection accuracy of each model and the overall performance of the system.

5. The WSI image quality assessment system as described in claim 1, characterized in that, The principle of the cell number estimation method in step S4 is as follows: S4.

1. Morphological processing is performed on the high-magnification slice images to remove non-target tissue regions with abnormal areas. Then, automatic cell identification and counting are performed on the processed tissue region images to calculate the number of cells in each preset-size slice under the same WSI. Then, interference slices located in the preset extreme value range are removed to obtain the retained set. To ensure the quality representativeness of the tissue sections used in the final scoring, the mean of the retained set was taken as the representative number of cells. (2) S4.2, will Substitute into the cell number sub-score formula for calculation : （3） in, The minimum number of qualified cells, To determine the ideal cell count, a threshold standard was established based on the TBS standard and empirical calibration method. It is the lowest score. This is the highest score, usually taken as 0 or 100; S4.3 Finally, score the cell count item separately. With preset weights Weighted scores were applied to obtain cell quantity and quality scores. : （4）。 6. The WSI image quality assessment system as described in claim 1, characterized in that, The principle of the bubble detection method in step S5 is as follows: S5.1 After extracting the tissue region from the 1x objective lens slice image, a pre-trained instance segmentation model is used to detect bubble regions. The areas of all detected bubble regions are merged and denoted as... Let the area of ​​the original tissue region be denoted as The effective organization area The calculation formula is: （5） Calculate the ratio of the effective tissue region area to the original tissue region area to obtain the effective region proportion. : （6） S5.2, Based on the proportion of effective areas Generate a score for the bubble stain item. : （7） S5.3 Finally, score the bubble stains separately. With preset weights Weighted scores are applied to obtain the slice integrity score. : （8）。 7. The WSI image quality assessment system as described in claim 1, characterized in that, The principle of staining evaluation in step S6 is as follows: S6.

1. A pre-trained deep learning classification model is used to extract features from slice images at 1x objective magnification. This model is used for multiple types of staining defects. The model outputs the probability that the slice has this type of defect. ; S6.

2. Based on the output of the deep learning classification model, calculate the total defect probability score representing the degree of staining defects. : （9） in, The closer this value is to 1, the higher the degree of staining defects in the slice; S6.3, Based on the total defect probability score Generate staining sub-scores : （10） S6.4, Finally With preset weights Weighted scores were applied to obtain the staining quality score of the slides. : （11）。

Images