LD CytoMatrix 20 slide scanning image quality automatic evaluation method
By combining information entropy and deep learning models, the quality of pathological slide images is automatically evaluated, solving the problems of time-consuming, labor-intensive, and subjective influences in traditional manual evaluation. This achieves efficient and accurate image quality assessment, which is applicable to cervical cytology and other cytopathology studies.
Patent Information
- Application Number
- CN202410952912.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-07-16
AI Technical Summary
Traditional pathological slide image quality assessment relies on manual visual examination, which is time-consuming, labor-intensive, and easily affected by subjective factors. Existing deep learning models have long development cycles and their output results do not meet clinical needs.
By combining information entropy calculation with a deep learning model, digital images are acquired through full-slide scanning. Image quality is determined using information entropy, RGB threshold, and a three-class classification model based on focusing. Features are extracted using convolutional neural networks and residual networks, and a multi-task learning model is constructed for comprehensive evaluation.
It enables efficient and accurate automated assessment of the quality of pathological slide images, reduces human error, and improves the speed and accuracy of pathological diagnosis. It is applicable to cervical cytology and other cytopathological research.
Smart Images

Figure CN118941513A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of image quality assessment, and in particular to an automated assessment method for LD CytoMatrix 20 slide scanning image quality. Background Art
[0002] In the field of pathology, especially in liquid-based cytology (LBC), a branch of cytopathology, the preparation and analysis of cell samples are crucial to disease diagnosis. LBC technology collects cell samples in a liquid fixative, and then prepares them into slide specimens for microscopic examination through steps such as slide preparation and staining. One of the key links in this process is to ensure that the image quality of the prepared slide specimen is high enough so that pathologists can accurately identify and analyze cell morphology and structure.
[0003] Traditional pathology slide image quality assessment relies on manual visual inspection, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors, resulting in limited consistency and accuracy of the assessment results.
[0004] In recent years, with the development of digital pathology, the popularization of slide scanning technology, and the advancement of artificial intelligence (AI) and machine learning algorithms, automated image quality assessment has become possible, greatly improving the efficiency and reliability of pathology image analysis. However, deep learning models are often called "black boxes" because their decision-making process is not easy to understand.
[0005] The design and training of deep learning models usually involve multiple complex algorithmic steps, including preprocessing, feature extraction, model selection, and optimization. The efficiency of these steps directly affects the processing speed of the overall model and the quality of the output results. Unfortunately, existing deep learning frameworks and tool chains often fail to provide sufficiently intuitive and efficient solutions, resulting in a long model development cycle and the structure of the output results may not meet the needs of clinical practice. Summary of the invention
[0006] The main purpose of the present invention is to provide an automated evaluation method for the quality of LD CytoMatrix 20 slide scanning images, so as to solve the problem that the traditional pathology slide image quality evaluation relies on manual visual inspection, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors.
[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: an automated evaluation method for the quality of LD CytoMatrix 20 slide scanning images,
[0008] S1. First, obtain digital images of cervical cells by scanning the whole slide;
[0009] S2, input the 1024-size image into the information entropy function to calculate the amount of information of each image;
[0010] If the information volume is greater than 2.0, the image information volume is considered qualified;
[0011] Otherwise, it is regarded as a blank image or an image with less information;
[0012] The information entropy of all images will be recorded and sorted;
[0013] S3, based on the top five images ranked by information entropy in step S2, calculate the sum of the red, green and blue pixel values of these images, and compare them with the RGB threshold of the standard image to determine the staining quality of the image, and record the determination result;
[0014] S4, inputting all images with information entropy greater than 2.0 into a focus three-classification model, which can distinguish whether the image is well-focused, out of focus, or focused on the impurity layer;
[0015] S5. Finally, the image quality of the entire sample is determined by combining the staining and focusing conditions of steps S3 and S4. If both staining and focusing are good, the sample is determined to be qualified and given a corresponding image quality score.
[0016] In the preferred embodiment, the method for determining the staining quality of the image in step S3 is:
[0017] A1. Use the pre-trained convolutional neural network CNN to extract the high-dimensional feature vector F = f(I) of each image, where I represents the input image and f is the feature extraction function of CNN;
[0018] A2. Considering the diversity of color space, the RGB space is converted into HSV and Lab space, and the features F are extracted respectively. HSV and F Lab Then, these features are combined using the weighted fusion strategy to obtain the comprehensive feature F combined :
[0019] F combined =α·F+β·F HSV +γ·F Lab (1);
[0020] Among them, α, β, γ are weight parameters, which are determined by cross-validation optimization to satisfy α+β+γ=1;
[0021] A3. Determine the optimal staining threshold T using dynamic programming method color , based on the statistical characteristics of the current batch of image features, use the following formula:
[0022] T color =μ+k·σ (2);
[0023] Among them, μ is the mean value of the integrated feature vector, σ is its standard deviation, and k is a coefficient adjusted according to historical data to balance sensitivity and specificity;
[0024] A4. Determine the staining quality by comparing the comprehensive features with the threshold T color , to classify the staining quality. Define staining S color The score is expressed as:
[0025]
[0026] Among them, λ is the slope parameter of the Sigmoid function, which adjusts the steepness of the classification curve. color ≥0.7 was considered as good staining quality, otherwise it was poor.
[0027] In the preferred solution, the focusing model classification method used in step S4 is:
[0028] B1. Use the residual network ResNet to extract the focus feature F of the image focus ;
[0029] B2. Combined with the staining quality feature F combined and the focus feature F focus , building a multi-task learning model to simultaneously optimize the prediction of staining and focusing states;
[0030] B3. Use the softmax function to calculate the probability distribution P of the focusing state focus , P focus It is expressed as:
[0031] P focus =softmax(W·[F combined ,F focus ] T +b)(4);
[0032] Among them, W is the weight matrix, b is the bias term, [F combined ,F focus ] represents the concatenated vector of features, and the softmax function ensures that the sum of probabilities is 1;
[0033] B4. Select the category with the highest probability as the final focusing state. If the probability of the focusing category exceeds 0.8, it is judged as good focusing; if the probability of focusing on the impurity layer is the highest and exceeds 0.6, it is judged as poor focusing.
[0034] In the preferred solution, the method for giving the corresponding image quality score in step S5 is:
[0035] C1. The coloring quality is quantified by the color consistency index Qc. The Qc index is based on the difference between the sum of the RGB values of the image and the sum of the RGB values of the standard image. Qc is expressed as:
[0036]
[0037] Among them, Rs, Gs, Bs is the RGB average value of the standard image, and R, G, B is the RGB average value of the current image, σ c is the standard deviation of staining consistency, which is used to control the sensitivity of the formula;
[0038] C2, focus state through focus index Q f To evaluate, Q f The index is calculated using the Laplacian edge strength of the image and the entropy of the image, Q f It is expressed as:
[0039]
[0040] in, is the Laplace operator result of image I at point (x, y), and H(I) is the information entropy of image I;
[0041] C3, comprehensive dyeing quality Q c and focus state Q f , giving the image quality score Q t , the quality score Q is calculated using the weighted average formula t , Q t It is expressed as:
[0042] Q t =w c Q c +w f Q f (7);
[0043] Among them, w c and w f are the weight factors of staining quality and focusing state, respectively, satisfying w c +w f =1;
[0044] The weighting factors can be adjusted based on the experimental results to optimize the overall score;
[0045] C4. Set the qualified threshold T q , if the comprehensive score Q t Greater than or equal to T q , the image quality is judged to be qualified, otherwise it is unqualified.
[0046] The present invention provides an automated evaluation method for the quality of LD CytoMatrix 20 slide scan images. This algorithm design combines deep learning with traditional statistical methods, increases the complexity and adaptability of the model, and aims to more accurately evaluate the quality of cervical cell slide images. Through feature extraction, fusion, and adaptive threshold calculation, the staining quality of the image can be considered more comprehensively; deep learning is used to classify the focus state, combined with multi-task learning, to improve the accuracy of focus judgment. These steps work together to provide a more powerful algorithm framework for the quality evaluation of cervical cell slide images.
[0047] The LD CytoMatrix 20 slide scanning image quality automated assessment method can efficiently and accurately check the quality of cervical cytology slide images, reduce human errors, and improve the speed and accuracy of pathological diagnosis. This method is not only applicable to cervical cytology, but also widely used in other types of cytopathology and histopathology research. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0049] Figure 1 It is the algorithm flow chart of the present invention. DETAILED DESCRIPTION
[0050] Example 1
[0051] like Figure 1 As shown, an automated evaluation method for the image quality of LD CytoMatrix 20 slide scans is
[0052] S1. First, obtain digital images of cervical cells by scanning the whole slide;
[0053] S2, input the 1024-size image into the information entropy function to calculate the amount of information of each image;
[0054] If the information volume is greater than 2.0, the image information volume is considered qualified;
[0055] Otherwise, it is regarded as a blank image or an image with less information;
[0056] The information entropy of all images will be recorded and sorted.
[0057] S3, based on the top five images ranked by information entropy in step S2, calculate the sum of the red, green and blue pixel values of these images, and compare them with the RGB threshold of the standard image to determine the staining quality of the image, and record the determination result;
[0058] S4, inputting all images with information entropy greater than 2.0 into a focus three-classification model, which can distinguish whether the image is well-focused, out of focus, or focused on the impurity layer;
[0059] S5. Finally, the image quality of the entire sample is determined by combining the staining and focusing conditions of steps S3 and S4. If both staining and focusing are good, the sample is determined to be qualified and given a corresponding image quality score.
[0060] Example 2
[0061] Further illustrate with reference to Example 1, Figure 1 As shown, the method for determining the staining quality of the image in step S3 is:
[0062] A1. Use a pre-trained convolutional neural network (CNN) to extract a high-dimensional feature vector F = f(I) for each image, where I represents the input image and f is the feature extraction function of CNN. CNN can automatically learn and capture the hierarchical features in the image, which is crucial for subsequent staining quality analysis.
[0063] A2. Considering the diversity of color space, the RGB space is converted into HSV and Lab space, and the features F are extracted respectively. HSV and F Lab Then, these features are combined using the weighted fusion strategy to obtain the comprehensive feature F combined :
[0064] F combined =α·F+β·F HSV +γ·F Lab (1);
[0065] Among them, α, β, and γ are weight parameters, which are determined by cross-validation optimization to satisfy α+β+γ=1; these parameters will be optimized through cross-validation to achieve the best fusion effect.
[0066] A3. Determine the optimal staining threshold T using dynamic programming method color , based on the statistical characteristics of the current batch of image features, use the following formula:
[0067] T color =μ+k·σ (2);
[0068] Among them, μ is the mean value of the integrated feature vector, σ is its standard deviation, and k is a coefficient adjusted according to historical data to balance sensitivity and specificity;
[0069] A4. Determine the staining quality by comparing the comprehensive features with the threshold T color , to classify the staining quality. Define staining S colorThe score is expressed as:
[0070]
[0071] Among them, λ is the slope parameter of the Sigmoid function, which adjusts the steepness of the classification curve. color ≥0.7 was considered as good staining quality, otherwise it was poor.
[0072] The algorithm uses deep learning technology to extract and fuse features, combined with adaptive threshold decisions, to more accurately evaluate the staining quality of the image, thereby improving the overall evaluation accuracy of the sample image quality.
[0073] Example 3
[0074] Further illustrate with reference to Example 1, Figure 1 As shown, the focusing model classification method used in step S4 is:
[0075] B1. Use the residual network ResNet to extract the focus feature F of the image focus ;
[0076] B2. Combined with the staining quality feature F combined and the focus feature F focus , building a multi-task learning model to simultaneously optimize the prediction of staining and focusing states;
[0077] B3. Use the softmax function to calculate the probability distribution P of the focusing state focus , P focus It is expressed as:
[0078] P focus =softmax(W·[F combined ,F focus ] T +b)(4);
[0079] Among them, W is the weight matrix, b is the bias term, [F combined ,F focus ] represents the concatenated vector of features, and the softmax function ensures that the sum of probabilities is 1;
[0080] B4. Select the category with the highest probability as the final focusing state. If the probability of the focusing category exceeds 0.8, it is judged as good focusing; if the probability of focusing on the impurity layer is the highest and exceeds 0.6, it is judged as poor focusing.
[0081] The algorithm combines the advantages of deep learning with statistical methods to improve the complexity and flexibility of the model, thereby more accurately evaluating the focus quality of cervical cell slide images. Through deep feature extraction and multi-task learning models, the algorithm not only considers the staining quality of the image, but also optimizes the prediction of the focus state in a targeted manner, ensuring the comprehensiveness and accuracy of the quality assessment of cervical cell slide images.
[0082] Example 4
[0083] Further illustrate with reference to Example 1, Figure 1 As shown, the method for giving the corresponding image quality score in step S5 is:
[0084] C1. The coloring quality is quantified by the color consistency index Qc. The Qc index is based on the difference between the sum of the RGB values of the image and the sum of the RGB values of the standard image. Qc is expressed as:
[0085]
[0086] Among them, Rs, Gs, Bs is the RGB average value of the standard image, and R, G, B is the RGB average value of the current image, σ c is the standard deviation of staining consistency, which is used to control the sensitivity of the formula;
[0087] C2, focus state through focus index Q f To evaluate, Q f The index is calculated using the Laplacian edge strength of the image and the entropy of the image, Q f It is expressed as:
[0088]
[0089] in, is the Laplace operator result of image I at point (x, y), and H(I) is the information entropy of image I;
[0090] C3, comprehensive dyeing quality Q c and focus state Q f , giving the image quality score Q t , the quality score Q is calculated using the weighted average formula t , Q t It is expressed as:
[0091] Q t =w c Q c +w f Q f (7);
[0092] Among them, w c and w fare the weight factors of staining quality and focusing state, respectively, satisfying w c +w f =1;
[0093] The weighting factors can be adjusted based on the experimental results to optimize the overall score;
[0094] C4. Set the qualified threshold T q , if the comprehensive score Q t Greater than or equal to T q , the image quality is judged to be qualified, otherwise it is unqualified.
[0095] Perform steps C1 and C2 for each image and calculate Q c and Q f .
[0096] Calculate the comprehensive quality score Q using the formula in step C3 t .
[0097] Using the decision rule of step C4, according to Q t and T q Decide whether the image is eligible.
[0098] The function of formula Qc is to quantify the dyeing quality by calculating the similarity between the image and the standard image in the color space to ensure that the image is dyed evenly and close to the expected standard.
[0099] Formula Q f Function: By measuring the edge clarity and information entropy of the image, the focusing state of the image is quantified to ensure that the image is clear and contains sufficient information.
[0100] Formula Q t Function: By weighted fusion of staining quality and focus status, a comprehensive quality score is obtained, providing a quantitative measure of the overall quality of the image.
[0101] Through the above algorithms, the quality of cervical cell slide scanning images can be systematically evaluated, providing strong support for pathological diagnosis.
[0102] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.
Claims
1. An automated evaluation method for LD CytoMatrix 20 slide scanning image quality, characterized by: S1. First, obtain digital images of cervical cells by scanning the whole slide; S2, input the 1024-size image into the information entropy function to calculate the amount of information of each image; If the information volume is greater than 2.0, the image information volume is considered qualified; Otherwise, it is regarded as a blank image or an image with less information; The information entropy of all images will be recorded and sorted; S3, based on the top five images ranked by information entropy in step S2, calculate the sum of the red, green and blue pixel values of these images, and compare them with the RGB threshold of the standard image to determine the staining quality of the image, and record the determination result; S4, inputting all images with information entropy greater than 2.0 into a focus three-classification model, which can distinguish whether the image is well-focused, out of focus, or focused on the impurity layer; S5. Finally, combining the staining and focusing conditions of steps S3 and S4, the image quality of the entire sample is determined; If both staining and focus are good, the sample is judged to be qualified and given a corresponding image quality score.
2. The method for automatically evaluating the quality of LD CytoMatrix 20 slide scanning images according to claim 1, characterized in that: The method of determining the staining quality of the image in step S3 is: A1. Use the pre-trained convolutional neural network (CNN) to extract the high-dimensional feature vector F = f(I) of each image, where I represents the input image and f is the feature extraction function of CNN. A2. Considering the diversity of color space, the RGB space is converted into HSV and Lab space, and the features F are extracted respectively. HSV and F Lab Then, these features are combined using the weighted fusion strategy to obtain the comprehensive feature F combined : F combined =α·F+β·F HSV +γ·F Lab (1); Among them, α, β, γ are weight parameters, which are determined by cross-validation optimization to satisfy α+β+γ=1; A3. Determine the optimal staining threshold T using dynamic programming method color , based on the statistical characteristics of the current batch of image features, use the following formula: T color =μ+k·σ (2); Among them, μ is the mean value of the integrated feature vector, σ is its standard deviation, and k is a coefficient adjusted according to historical data to balance sensitivity and specificity; A4. Determine the staining quality by comparing the comprehensive features with the threshold T color , to classify the staining quality. Define staining S color The score is expressed as: Among them, λ is the slope parameter of the Sigmoid function, which adjusts the steepness of the classification curve. color ≥0.7 was considered as good staining quality, otherwise it was poor.
3. The method for automatically evaluating the quality of LD CytoMatrix 20 slide scanning images according to claim 1, characterized in that: The focusing model classification method used in step S4 is: B1. Use the residual network ResNet to extract the focus feature F of the image focus ; B2. Combined with the staining quality feature F combined and the focus feature F focus , building a multi-task learning model to simultaneously optimize the prediction of staining and focusing states; B3. Use the softmax function to calculate the probability distribution P of the focusing state focus , P focus It is expressed as: P focus =softmax(W·[F combined ,F focus ] T +b)(4); Among them, W is the weight matrix, b is the bias term, [F combined ,F focus ] represents the concatenated vector of features, and the softmax function ensures that the sum of probabilities is 1; B4. Select the category with the highest probability as the final focusing state. If the probability of the focusing category exceeds 0.8, it is judged as good focusing; if the probability of focusing on the impurity layer is the highest and exceeds 0.6, it is judged as poor focusing.
4. The method for automatically evaluating the quality of LD CytoMatrix 20 slide scanning images according to claim 1, characterized in that: The method for giving the corresponding image quality score in step S5 is: C1. The coloring quality is quantified by the color consistency index Qc. The Qc index is based on the difference between the sum of the RGB values of the image and the sum of the RGB values of the standard image. Qc is expressed as: Among them, Rs, Gs, Bs is the RGB average value of the standard image, and R, G, B is the RGB average value of the current image, σ c is the standard deviation of staining consistency, which is used to control the sensitivity of the formula; C2, focus state through focus index Q f To evaluate, Q f The index is calculated using the Laplacian edge strength of the image and the entropy of the image, Q f It is expressed as: in, is the Laplace operator result of image I at point (x, y), and H(I) is the information entropy of image I; C3, comprehensive dyeing quality Q c and focus state Q f , giving the image quality score Q t , the quality score Q is calculated using the weighted average formula t , Q t It is expressed as: Q t =w c ·Q c +w f ·Q f (7); Among them, w c and w f are the weight factors of staining quality and focusing state, respectively, satisfying w c +w f =1; The weighting factors can be adjusted based on the experimental results to optimize the overall score; C4. Set the qualified threshold T q , if the comprehensive score Q t Greater than or equal to T q , the image quality is judged to be qualified, otherwise it is unqualified.
Citation Information
Patent Citations
Traditional Chinese painting instance segmentation method based on digital culture entropy regression
CN117036687A
Analysis method based on ASCUS cervical sample pathological result
CN117809301A
Training method and apparatus for image processing model, and image processing method and apparatus
WO2021213188A1
Tumor tissue pathological classification system and method based on adaptive proportional learning
WO2023078041A1
