Quality evaluation method and system of digestive tract early cancer image based on artificial intelligence

Abstract

The invention provides a digestive tract early cancer image quality evaluation method and system based on artificial intelligence, and relates to the technical field of endoscopic image quality evaluation.The method comprises the following steps that a to-be-evaluated endoscopic image is obtained, and the endoscopic image is used for screening digestive tract early cancer; respectively carrying out distinguishable performance evaluation, color fidelity performance evaluation and image quantity and coverage evaluation on the acquired endoscopic image to obtain a distinguishable performance evaluation result, a color fidelity performance evaluation result and an image quantity and coverage evaluation result; and dynamically fusing the distinguishable performance evaluation result, the color fidelity performance evaluation result and the image quantity and coverage evaluation result based on an adaptive weight model, and outputting a comprehensive quality score. According to the method disclosed by the invention, the aspects of distinguishability, color fidelity, image quantity, coverage and the like are fused, so that more comprehensive, more accurate and more efficient quality evaluation of the endoscopic image is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of endoscopic image quality evaluation, and particularly to a method and system for evaluating the quality of early-stage digestive tract cancer images based on artificial intelligence. Background Art

[0002] Early-stage digestive tract cancer refers to early malignant tumors occurring in the digestive tract (such as the esophagus, stomach, duodenum, etc.). Its characteristic is that the lesion only involves the mucosal layer or the submucosal layer, without invading the deep tissues or metastasizing distantly. Early-stage digestive tract cancer has the clinical characteristic of hidden symptoms. Generally, endoscopic examination is the core means to detect early-stage digestive tract cancer. Based on this, the image quality of endoscopic examination plays a crucial role in the diagnosis and treatment of early-stage digestive tract cancer.

[0003] In the existing related technologies, generally, the quality of early-stage digestive tract cancer endoscopic images is evaluated manually. However, there are such problems: the current manual evaluation generally only performs artificial visual evaluation on the clarity or color restoration of endoscopic images. Not only are the evaluation indicators single, and both the accuracy and efficiency are not high, but also the evaluation effect highly depends on the experience level of the evaluator. In addition, it is difficult to form a standardized and generally applicable evaluation standard. Summary of the Invention

[0004] To solve the technical problems in the related technologies, the present invention provides a method and system for evaluating the quality of early-stage digestive tract cancer images based on artificial intelligence.

[0005] To achieve the above object, the technical solutions adopted by the present invention include:

[0006] According to the first aspect of the present invention, there is provided a method for evaluating the quality of early-stage digestive tract cancer images based on artificial intelligence, including the following steps:

[0007] Step S1: Obtain the endoscopic image to be evaluated, and the endoscopic image is used for early-stage digestive tract cancer screening;

[0008] Step S2: Respectively perform discriminability performance evaluation, color fidelity performance evaluation, and image quantity and coverage evaluation on the obtained endoscopic image to obtain a discriminability performance evaluation result, a color fidelity performance evaluation result, and an image quantity and coverage evaluation result;

[0009] Step S3: Dynamically fuse the discriminability performance evaluation result, the color fidelity performance evaluation result, and the image quantity and coverage evaluation result based on an adaptive weight model, and output a comprehensive quality score.

[0010] Optionally, in the step S2, the process of discriminability performance evaluation includes:

[0011] Step S2-1-1: Calculate the gradient corresponding to each pixel in each endoscopic image respectively:

[0012]

[0013] In the formula, G(x,y) is the gradient corresponding to the pixel pair (x,y), and G x is the horizontal edge kernel, and G y is the vertical edge kernel;

[0014] Step S2-1-2: Statistically calculate the proportion S G of the pixels whose gradient magnitude exceeds the set threshold T G :

[0015]

[0016] In the formula, δ(·) is the indicator function, whose value is 1 when the condition is satisfied, otherwise 0; N1 is the total number of pixels in the image;

[0017] Step S2-1-3: Segment the image into superpixel blocks, and calculate the contrast of the gray-level co-occurrence matrix GLCM for each superpixel block respectively:

[0018]

[0019] In the formula, C GLCM is the contrast of the gray-level co-occurrence matrix GLCM, and p(x,y) is the probability of the pixel pair (x,y) in the gray-level co-occurrence matrix;

[0020] Step S2-1-4: Statistically calculate the proportion S C of the superpixel blocks whose contrast is greater than the set threshold T C :

[0021]

[0022] In the formula, γ(·) is the indicator function, whose value is 1 when the condition is satisfied, otherwise 0; N2 is the total number of superpixel blocks in the image.

[0023] Optionally, before the step S2-1-1, there is also a step S2-1-0, first grayscale the acquired endoscopic image, and then segment the endoscopic image based on U-Net to focus on the mucosal surface and the vascular network, and / or,

[0024] Before the step S2-2-1, there is also a step S2-2-0, segment the endoscopic image based on U-Net to focus on the mucosal surface and the vascular network.

[0025] Optionally, in the step S2, the process of evaluating the color fidelity performance includes:

[0026] Step S2-2-1: Calculate the color deviation degree S between the endoscopic image and the standard color patch E :

[0027]

[0028] In the formula, ΔL is the brightness difference between the endoscopic image and the standard color patch, ΔC is the chromaticity difference between the endoscopic image and the standard color patch, ΔH is the hue difference between the endoscopic image and the standard color patch, K L , K C and K H are respectively the ambient light compensation coefficients, S L , S C and S H are respectively the weight factors;

[0029] Step S2-2-2: Compare the CIELAB color distribution of the endoscopic image with the target color gamut and calculate the coverage rate S cover :

[0030]

[0031] In the formula, S I is the intersection volume of the endoscopic color gamut and the target color gamut, S total is the volume of the target color gamut.

[0032] Optionally, in the step S2, the process of image quantity and coverage evaluation includes:

[0033] Step S2-3-1: Obtain the quantity i of the endoscopic images to be evaluated, and determine the quantity score value β according to the quantity of the endoscopic images:

[0034]

[0035] In the formula, β1, β2 and β3 are respectively different set score values, i1, i2 and i3 are respectively different quantity thresholds, and i1 < i2 < i3;

[0036] Step S2-3-2: Identify the quantity of the endoscopic images to be evaluated that include the recommended endoscopic parts; the recommended endoscopic parts include: the upper esophagus, the middle esophagus, the lower esophagus, the cardia, the fundus of the stomach, the body of the stomach, the gastric angle, the pylorus, the gastric antrum, the duodenal bulb and the descending part of the duodenum; and calculate the proportion S of the quantity of the non-appearing recommended endoscopic parts to the total quantity of the recommended endoscopic parts P :

[0037]

[0038] In the formula, N3 is the quantity of the non-appearing recommended endoscopic parts.

[0039] Optionally, the step S3 includes:

[0040] Step S3-1: Based on the adaptive weight model, calculate the comprehensive quality score Q according to the following formula z :

[0041] Q z = θ1·S G + θ2·S C + θ3·S E + θ4·S cover + θ5·β + θ6·S P

[0042] In the formula, θ1, θ2, θ3, θ4, θ5 and θ6 are adaptive weights, and their sum is 1.

[0043] Optionally, step S3 further includes:

[0044] Step S3-2: Determine the quality level L of the acquired endoscopic image according to the following formula:

[0045]

[0046] In the formula, L1, L2, L3 and L4 are different quality levels, and Q1, Q2 and Q3 are different quality score thresholds;

[0047] Step S3-3: Associate the corresponding decision reference according to the quality level; among them,

[0048] When the quality level of the acquired endoscopic image is L1, the corresponding decision reference is: the quality of the endoscopic image is poor, it cannot be used as a clinical reference, and re-examination must be carried out;

[0049] When the quality level of the acquired endoscopic image is L2, the corresponding decision reference is: the quality of the endoscopic image is medium, it can be used as a clinical reference, and re-examination is recommended;

[0050] When the quality level of the acquired endoscopic image is L3, the corresponding decision reference is: the quality of the endoscopic image is good, it can be used as a clinical reference, and re-examination is recommended;

[0051] When the quality level of the acquired endoscopic image is L4, the corresponding decision reference is: the quality of the endoscopic image is excellent and it can be used as a clinical reference;

[0052] Step S3-4: Output the quality level and the corresponding decision reference.

[0053] Optionally, the adaptive weights are respectively: θ1 = 0.263, θ2 = 0.263, θ3 = 0.232, θ4 = 0.104, θ5 = 0.069, θ6 = 0.069.

[0054] Optionally, step S1 specifically includes:

[0055] Step S1-1: Obtain all endoscopic images during a single endoscopic examination;

[0056] Step S1-2: Arrange all the endoscopic images obtained in Step S1-1 in the order of shooting time.

[0057] According to the second aspect of the present invention, there is also provided a quality evaluation system for early gastrointestinal cancer images based on artificial intelligence, which is used to execute the quality evaluation method for early gastrointestinal cancer images based on artificial intelligence described in any one of the technical solutions in the first aspect of the present invention. The quality evaluation system for early gastrointestinal cancer images based on artificial intelligence includes:

[0058] An image acquisition module, which is used to acquire endoscopic images to be evaluated;

[0059] An image evaluation module, which is used to respectively evaluate the discriminability performance, color fidelity performance, and image quantity and coverage of the acquired endoscopic images; and is used to dynamically fuse the evaluation results of discriminability performance, color fidelity performance, and image quantity and coverage based on an adaptive weight model;

[0060] An output module, which is used to output a comprehensive quality score.

[0061] Beneficial effects:

[0062] 1. Through the above technical solutions, first, the method of the present invention uses quantifiable indicators to replace manual visual judgment, which can not only effectively eliminate the influence of manual experience and eliminate subjective errors, but also effectively improve the accuracy, objectivity and efficiency of quality evaluation.

[0063] Second, the method of the present invention comprehensively evaluates endoscopic images from multiple indicators of endoscopic images (discriminability performance, color fidelity performance evaluation, image quantity and coverage evaluation), can be more targeted for early gastrointestinal cancer screening, and can effectively improve the detection rate of early gastrointestinal cancer.

[0064] 2. Other beneficial effects or advantages of the present invention will be described in detail in the specific implementation manners. Description of the Drawings

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative efforts.

[0066] Among them:

[0067] Figure 1 It is a schematic diagram of the step flow of a method for evaluating the quality of images of early-stage gastrointestinal cancer based on artificial intelligence provided by an exemplary embodiment of the present invention. Detailed implementation manners

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0069] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0070] In addition, the terms "including" and "having" and any variations thereof mentioned in the description of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes other steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products, or devices. Among them, it should also be noted that in the embodiments of the present invention, words such as "exemplary" or "for example" are used to represent examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0071] For the convenience of those skilled in the relevant art to have a clearer and more accurate understanding of the technical solutions of the present invention, the following will first be described in more detail.

[0072] First, due to the clinical characteristic that early-stage gastrointestinal cancer has hidden symptoms, most patients generally have no obvious symptoms in the early stage, and a small number of patients may have mild discomfort, such as indigestion, abdominal distension, changes in bowel habits, etc., which are easily mistaken for ordinary gastrointestinal problems.

[0073] Secondly, the characteristic of early-stage gastrointestinal cancer is that its lesion only involves the mucosal layer or the submucosal layer (generally, the gastrointestinal tract wall can be divided into the mucosal layer, submucosal layer, muscular layer, and serosa layer from the inside to the outside), and has not invaded the deep tissues or metastasized distantly. Therefore, endoscopic examination that can directly observe the mucosal condition has become the core means for diagnosing early-stage gastrointestinal cancer.

[0074] In the existing related technologies, the quality of endoscopic images of early gastrointestinal cancer is generally evaluated by artificial vision. The main aspects of the evaluation are the clarity of the endoscopic images (mainly because the lesions of early gastrointestinal cancer are generally small, so there are high requirements for clarity) and the degree of color restoration (generally, the lesions of early gastrointestinal cancer have certain color characteristics, for example, redness).

[0075] However, the artificial vision evaluation is inevitably restricted by the experience level of the evaluator. Its evaluation results highly depend on the experience level of the evaluator, and it is also difficult to form a standardized and generally applicable evaluation standard. At the same time, the evaluation indicators of artificial vision evaluation are relatively single (generally only clarity and color restoration degree), and both the accuracy and efficiency are low.

[0076] In view of this, the present invention provides a brand-new solution, that is, the quality evaluation method and system for early gastrointestinal cancer images based on artificial intelligence of the present invention. Among them, the core idea of the quality evaluation method for early gastrointestinal cancer images based on artificial intelligence of the present invention is: increasing evaluation indicators, integrating aspects such as distinguishability, color fidelity, number of images and coverage, and dynamically fusing the evaluation results of the above aspects by using an adaptive weight model for comprehensive evaluation, so as to achieve a more comprehensive, accurate and efficient quality evaluation of endoscopic images.

[0077] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0078] As Figure 1 shown, according to the first aspect of the present invention, there is provided a quality evaluation method for early gastrointestinal cancer images based on artificial intelligence, including the following steps:

[0079] Step S1: Obtain the endoscopic image to be evaluated, and the endoscopic image is used for early gastrointestinal cancer screening;

[0080] Step S2: Respectively perform distinguishability performance evaluation, color fidelity performance evaluation, number of images and coverage evaluation on the obtained endoscopic images to obtain distinguishability performance evaluation results, color fidelity performance evaluation results, and number of images and coverage evaluation results;

[0081] Step S3: Dynamically fuse the distinguishability performance evaluation results, color fidelity performance evaluation results, and number of images and coverage evaluation results based on the adaptive weight model, and output a comprehensive quality score.

[0082] Through the above technical solution, first, the method of the present invention uses quantifiable indicators to replace artificial visual judgment, which can not only effectively eliminate the influence of artificial experience and eliminate subjective errors, but also effectively improve the accuracy, objectivity and efficiency of quality evaluation.

[0083] Second, the method of the present invention comprehensively evaluates endoscopic images from multiple indicators of endoscopic images (discriminability performance, color fidelity performance evaluation, image quantity and coverage evaluation), can be more targeted and applicable to the screening of early gastrointestinal cancer, and can effectively improve the detection rate of early gastrointestinal cancer.

[0084] In an embodiment of the present invention, in step S2, the process of the discriminability performance evaluation of the present invention may include: Step S2-1-1: Calculate the gradient corresponding to each pixel in each endoscopic image respectively:

[0085]

[0086] wherein, G(x, y) is the gradient corresponding to the pixel pair (x, y), G x is the horizontal edge kernel, G y is the vertical edge kernel.

[0087] Step S2-1-2: Statistically calculate the proportion S G of the pixels whose gradient amplitude exceeds the set threshold T G :

[0088]

[0089] wherein, δ(·) is the indicator function, whose value is 1 when the condition is satisfied, otherwise 0; N1 is the total number of pixels in the image;

[0090] Step S2-1-3: Segment the image into superpixel blocks, and calculate the contrast of the gray-level co-occurrence matrix GLCM for each superpixel block respectively:

[0091]

[0092] wherein, C GLCM is the contrast of the gray-level co-occurrence matrix GLCM, and p(x, y) is the probability of the pixel pair (x, y) in the gray-level co-occurrence matrix;

[0093] Step S2-1-4: Statistically calculate the proportion S C of the superpixel blocks whose contrast is greater than the set threshold T C :

[0094]

[0095] wherein, γ(·) is the indicator function, whose value is 1 when the condition is satisfied, otherwise 0; N2 is the total number of superpixel blocks in the image.

[0096] In this embodiment, first, the process of discriminability performance evaluation is to evaluate the discriminability of endoscopic images. Through gradient analysis and gray-level co-occurrence matrix contrast, the detail clarity and detail retention ability of endoscopic images are quantified, so as to achieve a more objective and comprehensive evaluation of the clarity of endoscopic images.

[0097] Second, this embodiment adopts a dynamic threshold design (T G and T C , which can be selected and adjusted according to different parts. For example, for the esophagus and gastric antrum, different T G and T C ) can be used, which can effectively avoid the overfitting problem that may be caused by a fixed threshold.

[0098] Specifically, for gradient analysis, the present invention first calculates the pixel gradient magnitude of the endoscopic image, and then calculates the proportion of pixels whose gradient magnitude exceeds the set threshold T G to reflect the detail clarity of the image. It can be understood that the area with a high gradient magnitude corresponds to the sharp edges of the mucosal surface and blood vessels, and can effectively characterize the morphological features of early gastrointestinal cancer lesions.

[0099] For the gray-level co-occurrence matrix contrast, the endoscopic image is first segmented into superpixel blocks, the contrast of the gray-level co-occurrence matrix is calculated, and the proportion of superpixel blocks whose contrast exceeds the threshold T C is screened. To quantify the complexity and difference of the mucosal surface texture, such as the roughness of the mucosa or abnormal blood vessel distribution in early canceration. It can be understood that the GLCM contrast is a standardized index for texture analysis, which can distinguish the microscopic structural differences between normal tissues and lesion areas.

[0100] In this embodiment, the present invention combines the macroscopic aspect (gradient analysis, globally evaluating the overall clarity of endoscopic images) and the microscopic aspect (superpixel texture analysis, focusing on local details to facilitate the identification of microscopic mucosal lesions) to more comprehensively and accurately evaluate the quality of endoscopic images.

[0101] In an embodiment of the present invention, in step S2, the process of color fidelity performance evaluation includes: step S2-2-1: calculating the color deviation degree S E :

[0102]

[0103] In the formula, ΔL is the brightness difference between the endoscopic image and the standard color patch, ΔC is the chromaticity difference between the endoscopic image and the standard color patch, ΔH is the hue difference between the endoscopic image and the standard color patch, K L , K C and K HThey are the ambient light compensation coefficients, S L , S C and S H are the weight factors respectively;

[0104] Step S2-2-2: Compare the CIELAB color distribution of the endoscopic image with the target color gamut, and calculate the coverage rate S cover :

[0105]

[0106] In the formula, S I is the intersection volume of the endoscopic color gamut and the target color gamut, and S total is the volume of the target color gamut.

[0107] In this embodiment, for the calculation of the color deviation degree, the present invention calculates the brightness difference (ΔL), chromaticity difference (ΔC) and hue difference (ΔH) between the endoscopic image and the standard color block based on the CIELAB color space, and dynamically adjusts the color deviation degree score through the ambient light compensation coefficients (K L , K C and K H ) and the weight factors (S L , S C and S H ). It synthesizes the sensitivity responses of brightness, chromaticity and hue, and can effectively adapt to the evaluation of the color deviation degree in the clinical scenario.

[0108] For the calculation of the color gamut coverage rate, the present invention compares the CIELAB color distribution of the endoscopic image with the target color gamut (for example, the typical color gamut of early-stage gastrointestinal cancer lesions), and calculates the proportion of the intersection volume. In this way, it can be quantified whether the endoscopic image covers all the color ranges required for diagnosis, and avoid missed detection caused by color gamut missing.

[0109] In this way, through the above technical solutions, first, the present invention can effectively improve the color restoration accuracy. Specifically, the present invention adjusts the color difference calculation under different illumination conditions through the ambient light compensation coefficient. For example, the brightness weight can be reduced in a low-brightness environment to avoid color distortion caused by overexposure or underexposure. At the same time, for the typical color gamut of early-stage gastrointestinal cancer (such as mucosal erythema, vascular dark areas, etc.), the target color gamut range can be set to ensure the accurate restoration of the lesion color.

[0110] Second, the method of the present invention has better generality. Specifically, the present invention unifies the color output of different endoscopic devices through the CIELAB space, and solves the color deviation problem caused by manufacturer differences.

[0111] In addition, the method of the present invention can be through S E and S coverThe quantifiable numerical output replaces the manual inspection and evaluation in the existing related technologies, and is more objective. At the same time, through the calculation of the gamut coverage rate, the key color areas that are not covered can be quickly identified (for example, the absence of a specific red tone), so as to facilitate the operator to operate the endoscope for reshooting.

[0112] In an embodiment of the present invention, before step S2-1-1, there may also be step S2-1-0, first grayscale the acquired endoscope image, and then segment the endoscope image based on U-Net to focus on the mucosal surface and the vascular network, and / or, before step S2-2-1, there is also step S2-2-0, segment the endoscope image based on U-Net to focus on the mucosal surface and the vascular network.

[0113] In this embodiment, step S2-1-0 converts the endoscope image into a grayscale image to simplify the image complexity and reduce the interference of color channels on gradient calculation and texture analysis. At the same time, grayscaling can also highlight the edges and texture information of the mucosal surface and the vascular network by weighted fusion of the RGB channels. In steps S2-1-0 and S2-2-0, the U-Net network is also used to segment the grayscaled image, separate the mucosal surface and the vascular network regions, and exclude non-target regions (such as bubbles and mucus).

[0114] First, the accuracy of discriminability performance evaluation can be effectively improved. Specifically, after focusing on the mucosal area by U-Net segmentation, the gradient calculation (step S2-1-1) and the gray-level co-occurrence matrix contrast analysis (step S2-1-3) are only for the target area, avoiding the noise introduced by non-target areas (such as reflection or mucus-covered areas), and effectively improving the anti-interference ability. At the same time, the grayscaling process can reduce the influence of color information on edge detection, so that the gradient amplitude statistics can more truly reflect the clarity of the mucosal structure.

[0115] Second, the reliability of color fidelity evaluation can be effectively improved. Specifically, after U-Net segmentation, the color fidelity evaluation can be only for the mucosal surface and the vascular network regions, avoiding the influence of color deviations in other regions (such as bubbles or non-lesion tissues) on the overall score. At the same time, the color evaluation parameters of the segmented target area can be dynamically adjusted according to different parts (such as the esophagus or the gastric antrum), so as to further improve the scene adaptability and reliability of the evaluation.

[0116] In an embodiment of the present invention, in step S2, the process of image quantity and coverage evaluation includes: step S2-3-1: Obtain the quantity i of the endoscope images to be evaluated, and determine the quantity score value β according to the quantity of the endoscope images:

[0117]

[0118] Wherein, β1, β2, and β3 are respectively different set scoring values, i1, i2, and i3 are respectively different quantity thresholds, and i1 < i2 < i3;

[0119] Thus, by setting increasing thresholds (i1, i2, and i3), the number of endoscopic images i is divided into different intervals, and respectively corresponds to different set scoring values (β1, β2, and β3). For example, when i ≤ i1, the quantity scoring value β is the lowest (β1) to reflect the risk of lesion omission caused by insufficient images; when i2 < i ≤ i3, the quantity scoring value β is the highest (β3) to ensure that there are sufficient endoscopic images to cover all possible lesions.

[0120] Step S2-3-2: Identify the number of endoscopic images to be evaluated that include the recommended endoscopic sites; the recommended endoscopic sites include: the upper esophagus, the middle esophagus, the lower esophagus, the cardia, the fundus of the stomach, the body of the stomach, the gastric angle, the pylorus, the gastric antrum, the duodenal bulb, and the descending part of the duodenum; and calculate the proportion S of the number of non-appearing recommended endoscopic sites to the total number of recommended endoscopic sites P :

[0121]

[0122] Wherein, N3 is the number of non-appearing recommended endoscopic sites.

[0123] Thus, the method of the present invention identifies 11 recommended endoscopic sites from the esophagus to the duodenum (such as the gastric angle, the pylorus, etc.), covering the high-incidence areas of early gastrointestinal cancer. At the same time, by calculating the proportion of the uncovered sites, the coverage omission rate is quantified, directly reflecting the defects in operation standardization.

[0124] In this embodiment, the method of the present invention can effectively reflect the comprehensiveness of the coverage of endoscopic images. Specifically, the method of the present invention first forcibly requires a minimum number of images (i1) through segmented scoring, which is beneficial to avoiding missed detection due to too few images. Then, by calculating the proportion S of the number of non-appearing recommended endoscopic sites to the total number of recommended endoscopic sites P , the coverage rate of endoscopic images for the high-incidence areas of early gastrointestinal cancer (11 recommended endoscopic sites) is quantified. Compared with the existing related technologies that rely on the experience of evaluators to judge whether the number of endoscopic images is sufficient, it can effectively eliminate the subjective judgment differences of different physicians and reduce the missed detection rate.

[0125] In an embodiment of the present invention, step S3 of the present invention may include:

[0126] Step S3-1: Based on the adaptive weight model, calculate the comprehensive quality score Q according to the following formula z :

[0127] Q z = θ1·SG + θ2·S C + θ3·S E + θ4·S cover + θ5·β + θ6·S P

[0128] Wherein, θ1, θ2, θ3, θ4, θ5 and θ6 are adaptive weights, and their sum is 1.

[0129] Thus, in this embodiment, first, the present invention integrates six types of parameters (S G , S C , S E , S cover , β and S P ) through the adaptive weight model, which can effectively avoid the one-sidedness of evaluation indicators and improve the accuracy and objectivity of evaluation results.

[0130] Second, the adaptive weights (θ1, θ2, θ3, θ4, θ5 and θ6) of the present invention can be dynamically adjusted according to the individual differences of patients (such as esophageal stenosis, etc.) and actual situations (when S P is relatively large, that is, the proportion of the endoscopic image that does not cover the recommended endoscopic site is relatively large) to ensure the referenceability of the comprehensive quality score.

[0131] In an embodiment of the present invention, step S3 further includes: step S3-2: determining the quality grade L of the acquired endoscopic image according to the following formula:

[0132]

[0133] Wherein, L1, L2, L3 and L4 are different quality grades, and Q1, Q2 and Q3 are different quality score thresholds;

[0134] Step S3-3: associating the corresponding decision reference according to the quality grade; wherein,

[0135] When the quality grade of the acquired endoscopic image is L1, the corresponding decision reference is: the quality of the endoscopic image is poor, it cannot be used as a clinical reference, and reexamination must be carried out; when the quality grade of the acquired endoscopic image is L2, the corresponding decision reference is: the quality of the endoscopic image is medium, it can be used as a clinical reference, and reexamination is recommended; when the quality grade of the acquired endoscopic image is L3, the corresponding decision reference is: the quality of the endoscopic image is good, it can be used as a clinical reference, and reexamination is recommended; when the quality grade of the acquired endoscopic image is L4, the corresponding decision reference is: the quality of the endoscopic image is excellent, it can be used as a clinical reference;

[0136] Step S3-4: outputting the quality grade and the corresponding decision reference.

[0137] In this embodiment, the present invention divides the calculated comprehensive quality score into Q z corresponding to four different quality levels (L1, L2, L3, and L4), and corresponding to different clinical decision references. Among them, when Q z ≤Q1, the quality level is L1, which will directly trigger a mandatory re-examination recommendation to avoid misleading diagnosis by low-quality endoscopic images. When Q1 < Q z ≤Q2 and Q2 < Q z ≤Q3, the quality levels are L2 and L3, indicating a recommendation for re-examination, but also indicating that the endoscopic image can be used as a clinical reference to balance clinical resources and the risk of missed detection. When Q3 < Q z , the quality level is L4, and the corresponding endoscopic image can be directly used as a clinical reference to exempt from re-examination, which can effectively improve the screening efficiency.

[0138] In addition, the output decision reference can be docked with the information management system of existing medical institutions to automatically generate a corresponding decision reference report for the evaluator to quickly know.

[0139] In an embodiment of the present invention, the adaptive weights can be set respectively as: θ1 = 0.263, θ2 = 0.263, θ3 = 0.232, θ4 = 0.104, θ5 = 0.069, θ6 = 0.069.

[0140] In the existing manual evaluation method, to improve the manual evaluation efficiency, a sampling analysis method is generally used to improve the evaluation efficiency. For example, the evaluator will selectively evaluate endoscopic images containing specific parts (such as the gastric antrum or pylorus, etc.). However, this may lead to missed detection of some parts. For example, the endoscopic image of the gastric angle generally only appears in some consecutive frames and is easily missed.

[0141] In view of this, in an embodiment of the present invention, step S1 may specifically include: step S1-1: obtaining all endoscopic images during a single endoscopic examination;

[0142] step S1-2: arranging all the endoscopic images obtained in step S1-1 in the order of shooting time.

[0143] In this embodiment, first, the method of the present invention obtains all endoscopic images (including invalid frames, blurred frames, etc.) during a single endoscopic examination through step S1-1, which can effectively avoid missing endoscopic images and can provide a complete data basis for subsequent quality evaluation. At the same time, the method of the present invention requires obtaining all endoscopic images, which can effectively eliminate the distortion of the evaluation results caused by sampling deviation compared with the sampling analysis method used in the existing related manual evaluation.

[0144] Second, the method of the present invention arranges the images in chronological order, which is conducive to establishing the chronological mapping relationship during the digestive tract examination process, facilitating the computer to identify the continuous changes in the morphology of the lesions, such as the spread of bleeding points or the gradual change of mucosal color. At the same time, it also helps to analyze the endoscopic operation path through time series analysis (for example, whether it is covered according to the standard process, or there are local omissions). In addition, it also helps to assist in judging whether the recommended endoscopic parts that are not covered (refer to step S2-3-2 in the above-mentioned embodiment) are caused by endoscopic operation omissions or the limitations of the equipment itself.

[0145] According to the second aspect of the present invention, there is also provided a quality evaluation system for early-stage cancer images of the digestive tract based on artificial intelligence, which is used to execute the quality evaluation method for early-stage cancer images of the digestive tract based on artificial intelligence in any one of the technical solutions in the first aspect of the present invention. The quality evaluation system for early-stage cancer images of the digestive tract based on artificial intelligence includes an image acquisition module, an image evaluation module, and an output module. Among them, the image acquisition module is used to acquire the endoscopic images to be evaluated. The image evaluation module is used to respectively evaluate the discriminability performance, color fidelity performance, image quantity and coverage of the acquired endoscopic images; and is used to dynamically fuse the evaluation results of the discriminability performance, color fidelity performance, and image quantity and coverage based on an adaptive weight model. The output module is used to output the comprehensive quality score.

[0146] Through the above technical solutions, first, the system of the present invention uses quantifiable indicators to replace manual visual judgment, which can not only effectively eliminate the influence of manual experience and eliminate subjective errors, but also effectively improve the accuracy, objectivity, and efficiency of quality evaluation.

[0147] Second, the system of the present invention comprehensively evaluates the endoscopic images from multiple indicators of the endoscopic images (discriminability performance, color fidelity performance evaluation, image quantity and coverage evaluation), can be more targeted and applicable to the early screening of early-stage cancer of the digestive tract, and can effectively improve the detection rate of early-stage cancer of the digestive tract.

[0148] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A quality evaluation method for early-stage gastrointestinal cancer images based on artificial intelligence, characterized in that It includes the following steps: Step S1: Obtain the endoscopic images to be evaluated, where the endoscopic images are used for early gastrointestinal cancer screening; Step S2: Respectively perform discriminability performance evaluation, color fidelity performance evaluation, and image quantity and coverage evaluation on the obtained endoscopic images to obtain a discriminability performance evaluation result, a color fidelity performance evaluation result, and an image quantity and coverage evaluation result; Step S3: Dynamically fuse the discriminability performance evaluation result, the color fidelity performance evaluation result, and the image quantity and coverage evaluation result based on an adaptive weight model, and output a comprehensive quality score.

2. The quality evaluation method for early gastrointestinal cancer images based on artificial intelligence according to claim 1, wherein In the step S2, the process of discriminability performance evaluation includes: Step S2-1-1: Calculate the gradient corresponding to each pixel in each endoscopic image respectively: where G(x, y) is the gradient corresponding to the pixel pair (x, y), G x is the horizontal edge kernel, and G y is the vertical edge kernel; Step S2-1-2: Count the proportion S of pixels whose gradient magnitude exceeds the set threshold T G G :​ In the formula, δ(·) is an indicator function, whose value is 1 when the condition is satisfied, otherwise it is 0; N1 is the total number of pixels in the image; Step S2-1-3: Divide the image into superpixel blocks, and calculate the contrast of the gray-level co-occurrence matrix GLCM for each superpixel block respectively: Where C GLCM is the contrast of the gray-level co-occurrence matrix GLCM, and p(x, y) is the probability of the pixel pair (x, y) in the gray-level co-occurrence matrix; Step S2-1-4: Count the proportion S of superpixel blocks with a contrast greater than the set threshold T C C :​ In the formula, γ(·) is an indicator function, whose value is 1 when the condition is satisfied, otherwise it is 0; N2 is the total number of superpixel blocks in the image.

3. The quality evaluation method for early gastrointestinal cancer images based on artificial intelligence according to claim 2, wherein In the step S2, the process of color fidelity performance evaluation includes: Step S2-2-1: Calculate the color deviation degree S between the endoscopic image and the standard color patch E : Where ΔL is the brightness difference between the endoscopic image and the standard color patch, ΔC is the chromaticity difference between the endoscopic image and the standard color patch, ΔH is the hue difference between the endoscopic image and the standard color patch, K L 、K C and K H are the ambient light compensation coefficients respectively, and S L 、S C and S H are the weight factors respectively; Step S2-2-2: Compare the CIELAB color distribution of the endoscopic image with the target color gamut and calculate the coverage rate S cover : Where S I is the volume of the intersection of the endoscopic color gamut and the target color gamut, and S total is the volume of the target color gamut.

4. The quality evaluation method for early gastrointestinal cancer images based on artificial intelligence according to claim 3, characterized in that Before the step S2-1-1, there is also a step S2-1-0, first grayscale the obtained endoscopic images, and then segment the endoscopic images based on U-Net to focus on the mucosal surface and the vascular network, and / or Before the step S2-2-1, there is also a step S2-2-0, segment the endoscopic images based on U-Net to focus on the mucosal surface and the vascular network.

5. The quality evaluation method for early-stage digestive tract cancer images based on artificial intelligence according to claim 3, wherein In the step S2, the process of image quantity and coverage evaluation includes: Step S2-3-1: Obtain the quantity i of the endoscopic images to be evaluated, and determine the quantity score value β according to the quantity of the endoscopic images: In the formula, β1, β2, and β3 are different set score values respectively, i1, i2, and i3 are different quantity thresholds respectively, and i1 < i2 < i3; Step S2-3-2: Identify the number of endoscopic images to be evaluated that include the recommended endoscopic sites; the recommended endoscopic sites include: upper esophagus, middle esophagus, lower esophagus, cardia, gastric fundus, gastric body, gastric angle, pylorus, gastric antrum, duodenal bulb, and descending duodenum; and calculate the proportion S of the number of absent recommended endoscopic sites to the total number of recommended endoscopic sites P : In the formula, N3 is the quantity of the recommended endoscopic parts that do not appear.

6. The quality evaluation method for early-stage gastrointestinal cancer images based on artificial intelligence according to claim 5, wherein, The step S3 includes: Step S3-1: Based on the adaptive weight model, calculate the comprehensive quality score Q according to the following formula z :[[]]END]] Q z = θ1·S G + θ2·S C + θ3·S E + θ4·S cover + θ5·β + θ6·S P In the formula, θ1, θ2, θ3, θ4, θ5, and θ6 are adaptive weights, and their sum value is 1.

7. The quality evaluation method for early-stage digestive tract cancer images based on artificial intelligence according to claim 6, wherein The step S3 also includes: Step S3-2: Determine the quality grade L of the obtained endoscopic images according to the following formula: In the formula, L1, L2, L3, and L4 are different quality grades, and Q1, Q2, and Q3 are different quality score thresholds; Step S3-3: Associate the corresponding decision reference according to the quality grade; among them, When the quality grade of the obtained endoscopic images is L1, the corresponding decision reference is: The quality of the endoscopic images is poor, they cannot be used as clinical references, and reexamination is necessary; When the quality grade of the obtained endoscopic images is L2, the corresponding decision reference is: The quality of the endoscopic images is medium, they can be used as clinical references, and reexamination is recommended; When the quality grade of the obtained endoscopic images is L3, the corresponding decision reference is: The quality of the endoscopic images is good, they can be used as clinical references, and reexamination is recommended; When the quality level of the acquired endoscopic image is L4, the corresponding decision reference is: the quality of the endoscopic image is excellent and can be used as a clinical reference; Step S3-4: Output the quality level and the corresponding decision reference.

8. The quality evaluation method for early-stage digestive tract cancer images based on artificial intelligence according to claim 6, characterized in that, The adaptive weights are respectively: θ1 = 0.263, θ2 = 0.263, θ3 = 0.232, θ4 = 0.104, θ5 = 0.069, θ6 = 0.

069.

9. The quality evaluation method for early gastrointestinal cancer images based on artificial intelligence according to claim 1, characterized in that, The specific steps of step S1 include: Step S1-1: Acquire all endoscopic images during a single endoscopic examination; Step S1-2: Arrange all the endoscopic images acquired in step S1-1 in chronological order of shooting.

10. An artificial intelligence-based quality evaluation system for early-stage gastrointestinal cancer images, characterized in that, For implementing the quality evaluation method of the early-stage cancer image of the digestive tract based on artificial intelligence according to any one of claims 1-9, the quality evaluation system of the early-stage cancer image of the digestive tract based on artificial intelligence includes: An image acquisition module for acquiring endoscopic images to be evaluated; An image evaluation module for respectively performing discriminability performance evaluation, color fidelity performance evaluation, and image quantity and coverage evaluation on the acquired endoscopic images; and for dynamically fusing the discriminability performance evaluation result, color fidelity performance evaluation result, and image quantity and coverage evaluation result based on an adaptive weight model; An output module for outputting a comprehensive quality score.

Images