Intestinal tract cleaning grade judgment and recognition system for enteroscopy patient

Abstract

The invention relates to the field of computer-aided diagnosis, discloses a system for judging and recognizing the intestinal tract cleanliness level of an enteroscope patient, and aims to solve the problems that existing intestinal tract cleanliness evaluation is high in subjectivity, low in evaluation efficiency and lack of quantitative standards. The method comprises the following steps: acquiring and preprocessing enteroscopy image sequence data, and extracting multi-scale features; and the intestinal tract cleanliness is identified through a deep learning model, and a quantitative scoring report is generated. The system comprises an image data acquisition module, a preprocessing module, a multi-scale feature extraction module, a cleanliness identification and classification module and a report generation module. According to the technical scheme, evaluation objectivity, accuracy and efficiency can be improved, automation, standardization and high-precision quantification are achieved, and dependence on doctor experience is reduced.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided diagnosis, specifically relating to a system for judging and recognizing the intestinal cleanliness level of colonoscopy patients. Background Technology

[0002] With the development of medical technology, endoscopic diagnostic techniques, especially colonoscopy, have become an important means of screening and diagnosing colorectal diseases. By directly observing the intestinal mucosa, endoscopic techniques can effectively detect lesions and perform biopsies or treatments. However, ensuring the quality and diagnostic accuracy of colonoscopy depends on the patient's intestines achieving adequate cleanliness. The degree of intestinal cleanliness directly affects the field of view during examination, the lesion detection rate, and the subsequent treatment outcome.

[0003] Objective and accurate assessment of bowel cleanliness is a crucial step in the entire colonoscopy procedure. Traditional bowel cleanliness assessment relies primarily on visual judgment by the endoscopist during the examination, using standardized scales such as the Boston Bowel Preparation Scale (BBPS) for manual scoring. This assessment method aims to comprehensively evaluate bowel residue, stool volume, and fluid clarity to ensure the bowel is in optimal condition before the examination.

[0004] In current technologies, endoscopists rely on personal experience and visual observation to assess bowel cleanliness, which is subject to significant subjectivity and individual differences, making it difficult to guarantee consistency in assessment results among different physicians. This manual assessment method is not only inefficient and time-consuming, but also susceptible to factors such as physician fatigue and experience level, potentially leading to misjudgments of bowel cleanliness as too high or too low. Insufficient cleanliness assessment results in poor field of vision, increasing the risk of missed lesions and severely affecting diagnostic accuracy; while over-assessment may prolong the examination time and even require patients to repeat bowel preparation, imposing unnecessary burdens and discomfort. Furthermore, this assessment process, lacking standardization and automated assistance, is ill-suited to the needs of large-scale screening and precision medicine. There is an urgent technical challenge to solve the problem of how to objectively, accurately, and efficiently assess and identify the patient's bowel cleanliness level before or during colonoscopy. Summary of the Invention

[0005] This invention provides a system for judging and recognizing the intestinal cleanliness level of colonoscopy patients, aiming to solve the specific technical problems of existing intestinal cleanliness assessments, which rely heavily on physicians' subjective experience, leading to inconsistent assessment results, low efficiency, risk of missed diagnoses, and a lack of quantitative standards. This invention achieves automated, standardized, and high-precision quantitative assessment of intestinal cleanliness by constructing an intelligent system based on deep learning and fusion recognition of multimodal colonoscopy image features.

[0006] According to one aspect of the present invention, a method for judging and identifying the intestinal cleanliness level of patients undergoing colonoscopy is provided, comprising the following steps: acquiring colonoscopy image sequence data, the colonoscopy image sequence data comprising at least one video frame; performing preprocessing operations on the colonoscopy image sequence data to generate standardized intestinal image data; performing multi-scale feature extraction on the standardized intestinal image data to obtain intestinal cleanliness-related feature vectors; based on the intestinal cleanliness-related feature vectors, identifying and classifying intestinal cleanliness using a pre-trained deep learning model to obtain a cleanliness identification result; and generating a quantified intestinal cleanliness level score report based on the cleanliness identification result.

[0007] As a specific embodiment of the present invention, the acquisition of colonoscopy image sequence data includes: acquiring a high-definition color video stream of the inside of the intestine in real time through an endoscope probe, wherein the video stream is transmitted at a frame rate of at least 25 frames per second, each video frame has a resolution of at least 1,920 pixels by 1,080 pixels, and a color depth of 24-bit true color; the endoscope probe is connected to an image processing unit through an optical fiber, and the image processing unit converts the analog signal into a digital signal and transmits it to the main processing unit through a high-speed data bus.

[0008] As a specific embodiment of the present invention, the preprocessing operation of the colonoscopy image sequence data includes the following steps: noise filtering of the colonoscopy image sequence data, wherein the noise filtering uses a nonlocal mean filtering algorithm or a Gaussian filtering algorithm to reduce random noise generated during image acquisition and smooth the image; contrast enhancement of the colonoscopy image sequence data, wherein the contrast enhancement uses a limited contrast adaptive histogram equalization algorithm or a gamma correction algorithm to enhance the visual difference between intestinal mucosa and residue; image normalization of the colonoscopy image sequence data, wherein the image normalization unifies the brightness, contrast, and chromaticity of the image to a preset standard range, eliminating image differences caused by different colonoscopy equipment and lighting conditions; and region of interest segmentation of the colonoscopy image sequence data, wherein the region of interest segmentation uses a semantic segmentation model based on a deep convolutional network, such as U-Net or Mask R-CNN, to accurately identify and extract the intestinal wall region, excluding instrument shadows, blurred edges, and non-intestinal tissue regions. The input of the segmentation model is the normalized image, and the output is a binary mask image marking the intestinal wall region.

[0009] As a specific embodiment of the present invention, the multi-scale feature extraction of the standardized intestinal image data includes the following steps: Multi-level feature extraction of the standardized intestinal image data is performed using a deep convolutional neural network backbone structure. The backbone structure adopts ResNet or EfficientNet and consists of multiple convolutional layers, pooling layers, and activation function layers, progressively extracting low-level texture features, mid-level edge shape features, and high-level semantic features of the image; The multi-level features are fused using a feature pyramid network structure. The feature pyramid network structure fuses the different levels of features extracted by the backbone network from top to bottom and bottom to top, generating high-resolution low-level features and high-level features with high semantic information, thereby forming a multi-scale feature map set. This multi-scale feature map set can simultaneously capture both fine details of residue and the distribution of large pieces of residue in the image; Weights are assigned to the multi-scale feature map set using a spatial attention mechanism or a channel attention mechanism. The spatial attention mechanism focuses on specific regions in the image where residue may exist, while the channel attention mechanism enhances feature channels related to residue type, color, and texture, thereby generating feature vectors containing high-dimensional semantic information related to intestinal cleanliness.

[0010] As a specific embodiment of the present invention, the step of identifying and classifying intestinal cleanliness based on the intestinal cleanliness-related feature vector using a pre-trained deep learning model includes the following steps: performing temporal information modeling on the feature vector, wherein the temporal information modeling employs a long short-term memory network, a gated recurrent unit, or a temporal convolutional network to process feature vectors in a continuous video frame sequence, capturing dynamic information about intestinal cleanliness changes over time, such as the movement, removal, or addition of residue; and inputting the time-series information-modeled feature vector into a fully connected classification layer, wherein the fully connected classification layer contains at least three output nodes, each output node corresponding to the right node of the Powell Intestinal Preparedness Scale. The output node outputs the cleanliness score probability distribution for each of the intestinal, transverse, and left colon regions. The fully connected classification layer is pre-trained using a dataset containing a large number of colonoscopy image sequences annotated by experienced physicians. The pre-training is optimized using the cross-entropy loss function to minimize the difference between the predicted and actual scores. Based on the cleanliness score probability distribution for each region, a cleanliness score is determined for each intestinal region, ranging from zero to three points. Zero points indicate a large amount of solid residue, one point indicates a small amount of non-adsorbed residue, two points indicate a small amount of adsorbed or turbid liquid, and three points indicate that the intestinal mucosa is clearly visible with no residue or very little clear liquid.

[0011] As a specific embodiment of the present invention, generating a quantitative intestinal cleanliness level score report based on the cleanliness identification results includes the following steps: aggregating the cleanliness score values ​​of each intestinal region to calculate the overall intestinal cleanliness total score, wherein the overall intestinal cleanliness total score is the arithmetic mean of the score values ​​of the right colon, transverse colon, and left colon regions; mapping the overall intestinal cleanliness total score to a preset intestinal cleanliness level standard, wherein the level standard includes four levels: excellent, good, average, and poor, and the mapping relationship is as follows: a total score greater than or equal to eight points is excellent, a total score greater than or equal to six points and less than eight points is good, a total score greater than or equal to four points and less than six points is average, and a total score less than four points is poor; generating a visualization report containing the overall intestinal cleanliness total score, the cleanliness scores of each region, the intestinal cleanliness level, and a confidence assessment, wherein the visualization report highlights the identified residue areas on the original colonoscopy image and labels their type and size, providing visual auxiliary information.

[0012] According to another aspect of the present invention, a system for judging and recognizing the intestinal cleanliness level of colonoscopy patients is provided, comprising: an image data acquisition module for acquiring colonoscopy image sequence data, the colonoscopy image sequence data including at least one video frame; an image preprocessing module for preprocessing the colonoscopy image sequence data to generate standardized intestinal image data; a multi-scale feature extraction module for performing multi-scale feature extraction on the standardized intestinal image data to obtain intestinal cleanliness-related feature vectors; a cleanliness recognition and classification module for recognizing and classifying intestinal cleanliness based on the intestinal cleanliness-related feature vectors using a pre-trained deep learning model to obtain a cleanliness recognition result; and a cleanliness level report generation module for generating a quantified intestinal cleanliness level score report based on the cleanliness recognition result.

[0013] In one specific embodiment of the present invention, the image data acquisition module includes: a high-definition endoscope probe, an image processing unit, and a high-speed data transmission interface; the high-definition endoscope probe is used to acquire high-definition color video streams inside the intestine in real time, and the frame rate, resolution, and color depth of the video streams meet medical imaging standards; the image processing unit is used to convert the analog signals acquired by the endoscope probe into digital signals and perform preliminary image correction and encoding; the high-speed data transmission interface is used to transmit the digital image data to the subsequent processing module, and the interface adopts a serial digital interface or Ethernet interface standard.

[0014] In one specific embodiment of the present invention, the image preprocessing module includes: a noise filtering unit, a contrast enhancement unit, an image normalization unit, and a region of interest (ROI) segmentation unit; the noise filtering unit employs a non-local mean filter processor or a Gaussian filter processor to reduce image noise; the contrast enhancement unit employs an adaptive histogram equalization processor or a gamma correction processor to improve image contrast; the image normalization unit employs a brightness, contrast, and chromaticity normalization processor to standardize image parameters; and the ROI segmentation unit employs a neural network processor based on a U-Net or Mask R-CNN architecture to accurately segment the intestinal wall region.

[0015] In one specific embodiment of the present invention, the multi-scale feature extraction module includes: a deep convolutional backbone network, a feature pyramid network, and an attention mechanism unit; the deep convolutional backbone network adopts a ResNet or EfficientNet architecture to extract image features from different levels; the feature pyramid network is used to fuse the multi-level features extracted by the backbone network to generate a multi-scale feature map set; the attention mechanism unit adopts a spatial attention module or a channel attention module to perform weight allocation on the multi-scale feature map set, focusing on key regions and feature channels related to cleanliness assessment.

[0016] In one specific embodiment of the present invention, the cleanliness identification and classification module includes: a temporal feature modeling unit and a fully connected classification unit; the temporal feature modeling unit employs a long short-term memory network processor, a gated recurrent unit processor, or a temporal convolutional network processor to process feature vectors of continuous video frame sequences and capture dynamic changes in intestinal cleanliness; the fully connected classification unit employs a multilayer perceptron structure containing multiple output nodes, each output node corresponding to the cleanliness score probability of an intestinal region, and the fully connected classification unit acquires intestinal cleanliness identification and classification capabilities through pre-training.

[0017] In one specific embodiment of the present invention, the cleanliness level report generation module includes: a scoring aggregation unit, a level mapping unit, and a report visualization unit; the scoring aggregation unit is used to calculate the total score of intestinal cleanliness, which is the arithmetic mean of the scores of each region; the level mapping unit is used to map the total score to a preset intestinal cleanliness level standard; the report visualization unit is used to generate a graphical user interface report containing the total score, the scores of each region, the intestinal cleanliness level, and the confidence assessment, and to overlay the identified residue areas on the original colonoscopy image to provide intuitive visual feedback.

[0018] Compared with existing technologies, the advantages and positive effects of this invention are as follows: This invention provides an automated intestinal cleanliness level assessment and recognition system based on deep learning, overcoming the problems of strong subjectivity and poor consistency of assessment results in traditional doctor's visual observation, significantly improving the objectivity and accuracy of assessment. This invention effectively filters noise and enhances image contrast through high-precision image preprocessing technology, providing high-quality input data for subsequent analysis. This invention employs multi-scale feature extraction and attention mechanisms, enabling comprehensive capture of residue features of different sizes, shapes, and colors in intestinal images, avoiding the risk of missed diagnoses and misdiagnoses. This invention introduces temporal information modeling, enabling analysis of dynamic changes in intestinal cleanliness in continuous video frames, improving the ability to recognize dynamic scenes. This invention provides quantified cleanliness scores and standardized level reports, providing doctors with accurate diagnostic basis, and visually presenting the results through reports, greatly improving work efficiency. This invention reduces reliance on doctor experience, helping new doctors quickly master intestinal cleanliness assessment skills, promoting the standardization and intelligent development of colonoscopy, and realizing the automation, standardization, and high-precision quantification of intestinal cleanliness assessment. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical architecture of the intestinal cleanliness level judgment and recognition system for colonoscopy patients proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the intestinal cleanliness level judgment and identification method in this invention; Figure 3 This is a logical flow diagram of the image preprocessing module in this invention; Figure 4 This is a schematic diagram of the core principle framework of the multi-scale feature extraction module in this invention; Figure 5 This is a schematic diagram of the core principle framework of the cleanliness identification and classification module in this invention; Figure 6 This is a logical flowchart of the cleaning level report generation module in this invention. Detailed Implementation

[0020] Example 1 In the following description, specific terminology and structures are used to illustrate embodiments of the invention for ease of understanding. However, these descriptions are not intended to limit the invention, and those skilled in the art will understand that various modifications and substitutions can be made without departing from the spirit and scope of the invention. The accompanying drawings are for illustrative purposes only, are not drawn to scale, and should not be construed as limiting the invention.

[0021] With the development of medical diagnostic technology, colonoscopy has become an important means of detecting and preventing colorectal diseases. However, the effectiveness of colonoscopy is closely related to bowel cleanliness. Inadequate bowel preparation can limit the doctor's field of vision, potentially leading to missed lesions, increased misdiagnosis rates, and even the need for repeated examinations, causing unnecessary suffering and medical costs for patients. Traditional bowel cleanliness assessment mainly relies on the subjective experience of endoscopists. This approach suffers from low standardization, large individual differences, low judgment efficiency, and accuracy affected by doctor fatigue. Especially when faced with a large number of colonoscopy images, manual interpretation is prone to missed or misdiagnosis, seriously affecting diagnostic quality and patient experience. Furthermore, some existing auxiliary assessment methods are often based on single features or simple models, making it difficult to comprehensively capture the complex cleanliness state within the intestine. Their ability to identify minute residues or areas with blurred intestinal mucosa is limited, failing to provide refined cleanliness level judgments and thus failing to provide accurate guidance to doctors.

[0022] This application proposes a system and method for judging and recognizing the intestinal cleanliness level of colonoscopy patients, aiming to achieve an objective and accurate assessment of intestinal cleanliness through automation and standardization. This application acquires intestinal image data during colonoscopy using a high-resolution image acquisition device and performs a series of refined preprocessing operations on these raw images, including image enhancement, denoising, and standardization, to eliminate noise interference and improve image quality. Subsequently, this application utilizes advanced multi-scale feature extraction technology to capture visual information at different levels from the preprocessed images, such as color, texture, and shape, and constructs multi-dimensional feature representations. Based on this, this application introduces a deep learning model to classify and recognize the extracted features, dividing intestinal cleanliness into multiple fine levels, such as excellent, good, acceptable, and poor. Finally, the system generates a detailed cleanliness level judgment and recognition report based on the recognition results, providing endoscopists with objective and quantitative assessment basis, assisting them in making accurate diagnostic decisions, and providing a reference for subsequent treatment or follow-up examinations. The technical solution of this application, by integrating image processing, deep learning, and medical report generation, effectively overcomes the subjectivity and limitations of traditional manual judgment, improving the quality and efficiency of colonoscopy.

[0023] This application proposes a method for determining and identifying the intestinal cleanliness level of patients undergoing colonoscopy. Figure 2 This is a schematic diagram illustrating the core principle framework of the intestinal cleanliness level determination and identification method in this invention. (See diagram below.) Figure 1 and Figure 2As shown, the method for judging and identifying the intestinal cleanliness level of colonoscopy patients in this application includes the following steps: S100, acquiring colonoscopy image data; S200, performing image preprocessing on the colonoscopy image data; S300, extracting multi-scale features from the preprocessed image; S400, performing cleanliness identification and classification based on the extracted multi-scale features; S500, generating a cleanliness level judgment and identification report.

[0024] In the above-mentioned method for judging and identifying the intestinal cleanliness level of colonoscopy patients, step S100 involves acquiring colonoscopy image data. As one embodiment of the present invention, acquiring colonoscopy image data specifically includes: acquiring real-time video streams or static image sequences of the patient's intestines using a high-resolution colonoscopy device, and transmitting these digital image data to an image processing unit. This colonoscopy device integrates a miniature optical lens, a high-intensity cold light source, and an image sensor, enabling it to capture detailed information about the intestinal mucosa. During the acquisition process, the colonoscopy device continuously acquires visual information about the inside of the intestines at a preset frame rate, such as 30 frames per second or higher. Each frame carries timestamp information, device identification, and spatial positioning information. These raw image data are typically stored in a common digital image format, such as lossless PNG or highly compressed JPEG, with a resolution of at least 1920 pixels by 1080 pixels to ensure sufficient detail for subsequent analysis. To ensure the quality and integrity of the image data, the image acquisition unit integrates a data verification module, such as a cyclic redundancy check (CRC) mechanism, to detect and correct potential errors during data transmission. Furthermore, to ensure data security and patient privacy, all acquired image data is encrypted during transmission and storage, employing algorithms such as Advanced Encryption Standard (AES). In actual operation, the colonoscopy device connects to the main processing unit via a dedicated data cable, such as a high-speed Universal Serial Bus 3.0 interface or an Ethernet interface, enabling real-time, high-speed data transmission. During image acquisition, the system continuously monitors the colonoscopy device's operating status, including parameters such as light source brightness, probe position, and focal length settings, to ensure the stability and consistency of image quality. When poor image quality is detected, such as excessive blurring, underexposure, or overexposure, the system will issue a prompt and adjust the acquisition parameters, or mark the image data for that time period for subsequent special processing.

[0025] In the above-mentioned method for judging and identifying the intestinal cleanliness level of colonoscopy patients, step S200 involves preprocessing the colonoscopy image data. Figure 3This is a logical flowchart of the image preprocessing module in this invention. As one embodiment of this invention, the image preprocessing of colonoscopy image data specifically includes: S210, performing noise suppression processing on the acquired colonoscopy image data; S220, enhancing the brightness and contrast of the noise-suppressed image data; and S230, performing color correction and standardization on the brightness and contrast-enhanced image data. It should be understood that raw colonoscopy image data is often affected by various factors, including noise, uneven brightness, and color distortion. These problems reduce image quality, obscure the true details of the intestinal mucosa, and thus seriously interfere with the accuracy of subsequent feature extraction and cleanliness identification. Traditional image preprocessing methods may not be able to comprehensively and effectively solve these complex problems, or may excessively smooth image details while denoising, leading to the loss of valuable diagnostic information. Therefore, in the technical solution of this application, refined image preprocessing is performed on colonoscopy image data to improve image quality and feature expression capabilities. Through phased and targeted preprocessing operations, the aim is to fundamentally improve the usability and analytical value of the images, laying a solid foundation for subsequent high-precision identification.

[0026] Specifically, step S210 involves noise suppression processing of the acquired colonoscopy image data. In one embodiment of the invention, this noise suppression processing includes applying median filtering, nonlocal mean denoising algorithms, or wavelet thresholding to the original colonoscopy image to effectively remove shot noise, Gaussian noise, or salt-and-pepper noise caused by endoscope light source reflection, equipment vibration, or electromagnetic interference, while preserving image edges and details as much as possible. In practice, the system first performs noise type analysis on the input original colonoscopy image, for example, by calculating the image's grayscale histogram and Fourier transform spectrum to determine the main noise components. If a large number of randomly distributed pixels with abnormal brightness, such as salt-and-pepper noise, are detected in the image, the system will preferentially use a median filtering algorithm. Median filtering effectively eliminates abnormal bright and dark spots without blurring image edges by replacing the grayscale value of a pixel with the median of the grayscale values ​​of all pixels in its neighborhood. Its mathematical expression can be simplified to: Px,y′=Median{Pi,j|(i,j)∈N(x,y)} where Px,y represents the pixel value at coordinates x,y in the original image, Px,y′ represents the processed pixel value, and N(x,y) represents the neighborhood window centered at x,y. For Gaussian noise and shot noise, the system can use a nonlocal mean denoising algorithm. This algorithm achieves a more thorough denoising effect by finding image blocks similar to the current block across the entire image range and performing a weighted average of the pixels in these similar blocks, while better preserving the texture details of the image. In addition, for more complex noise patterns or scenarios requiring the preservation of fine structure, a wavelet thresholding denoising method can be used. This method decomposes the image into wavelet domains of different scales, performs thresholding on the wavelet coefficients to filter out noise, and then performs inverse wavelet transform to reconstruct the image. Denoising parameters, such as filter window size and threshold strength, will be dynamically adjusted based on the specific characteristics of the colonoscopy image and clinical requirements. For example, a smaller filter window will be selected for areas with rich gastrointestinal mucosa texture to avoid over-smoothing, while the window can be appropriately enlarged for large flat areas to enhance the denoising effect. The noise-suppressed image data will be stored in a temporary buffer along with a processing log, recording the algorithms and parameters used for quality traceability.

[0027] Specifically, step S220 involves enhancing the brightness and contrast of the image data after noise suppression processing. As one embodiment of the invention, enhancing the brightness and contrast of the image data after noise suppression processing specifically includes: employing adaptive histogram equalization, such as a contrast-limited adaptive histogram equalization algorithm or gamma correction technology, to adjust the overall brightness and local contrast of the image, making the visual information of the intestinal mucosa, vascular texture, and small lesions more clearly discernible, especially in areas of uneven lighting or shadow. During colonoscopy, due to the complex structure of the intestinal lumen and the limited illumination angle of the colonoscope light source, local brightness in the image often becomes too dark or too bright, making it difficult to observe details inside the intestine. Adaptive histogram equalization effectively solves the problem of uneven local lighting by dividing the image into multiple small regions and performing histogram equalization independently within each region, avoiding over-enhancement or noise amplification that may result from global histogram equalization. Specifically, contrast-limited adaptive histogram equalization avoids over-enhancement of noise areas and maintains the naturalness of the image by limiting the contrast amplification factor. Its core transformation function can be expressed as: T(v)=(L−1)∑i=0vNchc(i), where T(v) represents the gray level of the output image, v represents the gray level of the input image, L represents the total number of gray levels, hc(i) represents the histogram of a certain local region, and Nc represents the total number of pixels in that region. Furthermore, gamma correction adjusts the brightness response of the image through nonlinear transformation, effectively correcting the dark or bright image phenomena caused by light source attenuation or sensor response characteristics, thereby improving the visual effect of the image. The parameters of these enhancement operations, such as the contrast limiting factor and gamma value, are dynamically adjusted according to the color characteristics of the intestinal mucosa, the reflectivity of intestinal secretions, and the overall brightness distribution of the image. For example, for areas with a large amount of residual fluid in the intestinal lumen, the contrast will be increased to distinguish the boundary between the fluid and the mucosa; for deep areas with insufficient light, the gamma value will be increased to enhance dark details.

[0028] Specifically, step S230 involves color correction and standardization of the image data after brightness and contrast enhancement. As one embodiment of the invention, this color correction and standardization of the image data after brightness and contrast enhancement specifically includes: unifying the color representation of colonoscopy images by using a white balance algorithm or color space transformation, such as converting from RGB to CIELAB color space, eliminating color casts caused by different colonoscopy devices, lighting conditions, or individual patient differences, and converting the image to a standardized color space representation, such as CIELAB color space, so that subsequent feature extraction is not affected by color changes. In practical applications, the color of colonoscopy images is affected by various factors, such as differences in the spectral characteristics of light sources from different colonoscopy devices, the absorption and scattering of light by the intestinal environment, and color interference from intestinal residues or blood, resulting in color casts such as reddish, yellowish, or bluish tints. White balance algorithms adjust the overall tone of the image by identifying white or gray areas in the image and correcting them to standard white, making the colors closer to reality. Another method is to use color space transformation, such as converting the RGB image to the CIELAB color space. The CIELAB color space possesses perceptual uniformity, with its L channel representing luminance and the a and b channels representing red-green and yellow-blue color components, respectively. This separation allows luminance and color information to be processed independently, and the CIELAB space is more consistent with human visual perception, which is beneficial for subsequent color feature analysis. Color correction and standardization parameters, such as white balance gain coefficients and color matrix transformation parameters, will be based on a preset color reference table or learned from a large amount of standardized colonoscopy image data. For example, the system can analyze a batch of images labeled as "normal" intestinal mucosa, extract their average color distribution as a reference, and then correct the colors of the image to be processed to this reference. Standardized image data has a unified color benchmark, ensuring the comparability of features between different images and improving the generalization ability of subsequent recognition models.

[0029] In the above-mentioned method for judging and recognizing the intestinal cleanliness level of colonoscopy patients, step S300 involves extracting multi-scale features from the preprocessed image. Figure 4This is a schematic diagram illustrating the core principle framework of the multi-scale feature extraction module in this invention. As one embodiment of this invention, the extraction of multi-scale features from the preprocessed image specifically includes: S310, inputting the preprocessed colonoscopy image into a multi-scale convolutional neural network model; S320, extracting deep semantic features and shallow texture features of the image through the multi-scale convolutional neural network model; S330, fusing the deep semantic features and shallow texture features to obtain a multi-scale fused feature vector. It should be understood that the judgment of intestinal cleanliness depends on multi-level information in the image, including subtle mucosal textures, the shape and distribution of residues, and the overall clarity of the intestinal lumen. Traditional single-scale feature extraction methods, such as local binary mode or scale-invariant feature transform, often struggle to simultaneously capture this rich visual information from local to global and from low to high levels, resulting in insufficient feature expression capabilities and an inability to effectively distinguish subtle differences between different cleanliness levels, especially at the boundary between lesion areas and normal mucosa. Therefore, in the technical solution of this application, a multi-scale feature extraction method is adopted to comprehensively and robustly capture key information in colonoscopy images. Through deep learning architecture, features at different levels of abstraction in images can be automatically learned and extracted, thus providing more discriminative input for accurate cleanliness identification.

[0030] Specifically, step S310 involves inputting the preprocessed colonoscopy image into a multi-scale convolutional neural network model. As one embodiment of the invention, inputting the preprocessed colonoscopy image into the multi-scale convolutional neural network model specifically includes: feeding the colonoscopy image, which has undergone noise suppression, brightness and contrast enhancement, and color correction standardization, to a pre-trained multi-scale convolutional neural network model at a uniform size, for example, 256 pixels by 256 pixels. This multi-scale convolutional neural network model consists of multiple convolutional layers, activation layers, pooling layers, and skip connections, aiming to learn features at different spatial scales from the image. The input image is first feature-mapped through a first-layer convolutional kernel, for example, a 3x3 convolutional kernel, generating a preliminary feature map. Subsequently, the feature map undergoes a series of downsampling operations, such as max pooling and multiple convolutional operations, to gradually extract higher-level, more abstract semantic information. Simultaneously, the model includes parallel processing paths or dilated convolutional layers to capture features under different receptive fields, thereby achieving effective acquisition of multi-scale features. For example, one approach might focus on extracting subtle texture details, while another approach might focus on macroscopic intestinal structure and residue distribution.

[0031] Specifically, step S320 involves extracting deep semantic features and shallow texture features of the image using a multi-scale convolutional neural network model. As one embodiment of the invention, this extraction of deep semantic features and shallow texture features using a multi-scale convolutional neural network model specifically includes: extracting shallow feature maps from the first few layers of the multi-scale convolutional neural network model, such as the first to third convolutional layers. These feature maps mainly reflect local, low-level visual information such as edges, corners, color gradients, and textures of the image. Simultaneously, extracting deep feature vectors from the deeper layers of the model, such as the seventh to tenth convolutional layers or fully connected layers. These feature vectors encode high-level semantic concepts and global structural information such as intestinal residues, mucus, edema, and mucosal lesions. Shallow features are quantified, for example, through edge detection operator responses or local texture descriptors. These are highly sensitive to local changes in the image and can capture subtle folds in the intestinal mucosa or texture features on the surface of residues. Deep features, through the hierarchical abstraction of convolutional neural networks, automatically learn and encode more complex patterns, such as the overall shape of fecal residue, the contrast between its color and the surrounding mucosa, and the distribution of fluid within the intestinal lumen. These features are crucial for distinguishing different levels of cleanliness. For example, when a large amount of fecal residue is detected, deep features will exhibit specific patterns. The extraction of shallow texture features can be represented as: Ftexture = CNNshallow(Image). The extraction of deep semantic features can be represented as: Fsemantic = CNNdeep(Image), where CNNshallow and CNNdeep represent the parts of the multi-scale convolutional neural network model responsible for extracting shallow and deep features, respectively. The extracted shallow features are typically high-dimensional feature maps, while deep features can be low-dimensional vectors.

[0032] Specifically, step S330 involves fusing deep semantic features and shallow texture features to obtain a multi-scale fused feature vector. As one embodiment of the invention, this feature fusion process includes concatenating, weighted summing, or adaptively fusing deep semantic features and shallow texture features extracted from different levels with a comprehensive multi-scale fused feature vector. This fusion strategy aims to fully utilize the complementarity of features at different scales to overcome the limitations of single features. For example, shallow texture features can accurately describe the microstructure of the intestinal mucosa and the boundary information of residues, while deep semantic features can grasp the overall state of intestinal cleanliness from a macroscopic perspective. By concatenating features, the two feature vectors can be directly connected into a longer vector, thus preserving their original information. By weighted summing, different weights can be assigned to features based on their importance or reliability. For example, when the image quality is poor, the weight of shallow texture features can be reduced, while the weight of deep semantic features can be increased. Attention mechanisms enable more intelligent feature fusion, allowing the model to dynamically allocate the importance of different features during the fusion process. For example, when there are small, indistinguishable remnants in the intestines, the attention mechanism focuses more on superficial texture features to capture these minute details; while when there is a large amount of fluid in the intestinal lumen, the attention mechanism may focus more on deep semantic features to determine the impact of the fluid on cleanliness. The resulting multi-scale fused feature vector has stronger representational power and robustness, and can more comprehensively and accurately reflect intestinal cleanliness. For example, a typical fused vector may contain information reflecting multiple dimensions such as intestinal texture uniformity, the proportion of remnant area, and the degree of mucus coverage. The fused feature vector will be used as input to the subsequent cleanliness identification and classification module.

[0033] In the above-mentioned method for judging and identifying the intestinal cleanliness level of colonoscopy patients, step S400 involves classifying and identifying cleanliness based on extracted multi-scale features. Figure 5This is a schematic diagram of the core principle framework of the cleanliness identification and classification module in this invention. As one embodiment of this invention, the cleanliness identification and classification based on extracted multi-scale features specifically includes: S410, inputting the multi-scale fused feature vector into a deep classification model; S420, predicting the intestinal cleanliness level using the deep classification model; S430, evaluating the confidence level and performing post-processing on the predicted cleanliness level. It should be understood that the accurate determination of the intestinal cleanliness level directly affects the reliability of subsequent diagnoses. Traditional classification methods based on rules or simple machine learning models often struggle to achieve ideal accuracy and robustness when faced with complex and varied intestinal images, especially when distinguishing subtle differences in cleanliness or processing blurry images. Human experience-based judgment suffers from strong subjectivity and low efficiency. Therefore, in the technical solution of this application, a deep learning model is used to classify the multi-scale fused feature vector to achieve automated and high-precision cleanliness level identification. The deep classification model can learn the nonlinear mapping relationship between intestinal cleanliness level and visual features from complex features, thereby providing objective and reliable classification results.

[0034] Specifically, step S410 involves inputting the multi-scale fused feature vector into a deep classification model. As one embodiment of the invention, inputting the multi-scale fused feature vector into the deep classification model specifically includes: feeding the multi-scale fused feature vector generated in step S330 as input data into a pre-trained deep classification model, such as a fully connected neural network, support vector machine, or gradient boosting decision tree. This deep classification model is pre-trained on a large dataset of colonoscopy images annotated by professional doctors, learning feature patterns for different cleanliness levels, such as excellent, good, acceptable, and poor. Before input, the fused feature vector undergoes normalization processing, such as zero-mean unit variance normalization, to ensure consistent numerical ranges across different feature dimensions and avoid excessive influence of certain features on model training.

[0035] Specifically, step S420 involves predicting the intestinal cleanliness level using a deep classification model. As one embodiment of the invention, this prediction using a deep classification model includes: after receiving a multi-scale fused feature vector, the deep classification model performs nonlinear transformation and learning on the features through multiple hidden layers (e.g., fully connected layers) and activation functions (e.g., ReLU or Sigmoid), ultimately outputting a vector representing the probability distribution of each cleanliness level. Each element in the probability distribution vector corresponds to a predicted probability of an intestinal cleanliness level, such as excellent, good, acceptable, or poor. The system selects the level with the highest probability as the initial prediction result. For example, if the model outputs a probability distribution vector of, for example, 95% excellent, 3% good, 2% acceptable, and 0% poor, the initial prediction result is "excellent". This classification model is typically trained using, for example, a cross-entropy loss function and updated with, for example, an Adam optimizer to minimize the difference between the predicted result and the true label.

[0036] Specifically, step S430 involves performing a confidence assessment and post-processing on the predicted cleanliness level. As one embodiment of the invention, this confidence assessment and post-processing specifically includes: calculating the confidence level of the prediction result, for example, by analyzing the difference or entropy value between the highest and second-highest predicted probabilities to evaluate the reliability of the classification result; if the confidence level is lower than a preset threshold, such as 70%, the image is marked as having low confidence, and a manual review process may be triggered; simultaneously, post-processing is performed on the predicted cleanliness level, such as smoothing based on the contextual information of the image sequence. For example, in a continuous colonoscopy video stream, if the first ten frames and the last ten frames are predicted as "good," while a certain frame in the middle is predicted as "acceptable" with a low confidence level, the system may correct it to "good" based on the contextual information to ensure the continuity and stability of the evaluation results. The confidence level can be calculated using the reciprocal of information entropy: Confidence = 1 / (−∑i=1Npilog2(pi)), where pi represents the probability of predicting the i-th cleanliness level, and N is the total number of cleanliness levels. The post-processing mechanism also includes averaging or weighted averaging with other image frames to reduce the impact of single-frame noise or transient anomalies on the overall judgment. Finally, the final intestinal cleanliness level is obtained after confidence assessment and post-processing.

[0037] In the above-mentioned method for judging and identifying the intestinal cleanliness level of colonoscopy patients, step S500 generates a cleanliness level judgment and identification report. Figure 6This is a logical flowchart of the bowel cleanliness level report generation module in this invention. As one embodiment of this invention, generating the bowel cleanliness level assessment report specifically includes: S510, integrating the bowel cleanliness level assessment result with relevant auxiliary information; S520, formatting the integrated information into a structured report; S530, outputting and storing the structured report. It should be understood that the final bowel cleanliness level assessment result needs to be presented to medical staff in a clear and standardized form so that they can quickly and accurately understand the patient's bowel cleanliness status and make treatment decisions accordingly. Traditional report generation may rely on manual filling, which is inefficient and prone to errors, or the report content may lack unified standards, leading to inaccurate information transmission. Therefore, in the technical solution of this application, an automated report generation module is designed to integrate the bowel cleanliness level assessment result with other relevant information and output a structured report. This module aims to provide a comprehensive, traceable, and easy-to-understand assessment report, improving the efficiency and quality of clinical work.

[0038] Specifically, step S510 integrates the intestinal cleanliness level assessment results with relevant auxiliary information. As one embodiment of the invention, this integration includes: summarizing the final intestinal cleanliness level obtained in step S430 (e.g., excellent, good, acceptable, poor) and its corresponding confidence level, as well as basic patient information such as name, age, gender, medical record number, examination date, examination site (e.g., ascending colon, transverse colon, descending colon, sigmoid colon, rectum), the model of the colonoscope used, the intestinal preparation protocol (e.g., polyethylene glycol electrolyte powder), and key feature descriptions identified during image preprocessing and feature extraction, such as residue type, mucosal clarity score, and intestinal fluid accumulation. This auxiliary information can come from hospital information system interfaces, electronic medical record systems, or data automatically recorded by the colonoscope. For example, the system may extract basic demographic data of the patient from the patient database, obtain the name and dosage of intestinal preparation medications from the examination records, and, combined with image analysis results, include detailed descriptions such as "A small amount of yellow granular fecal matter exists in the rectum; cleanliness is assessed as acceptable, with a confidence level of 88%" in the report.

[0039] Specifically, step S520 involves formatting the integrated information into a structured report. As one embodiment of the invention, formatting the integrated information into a structured report specifically includes: organizing the integrated cleanliness level assessment results and auxiliary information into a standardized structure according to a predefined medical report template, such as following the Boston Bowel Preparation Scale or the Edinburgh Classification, including a report title, patient information area, examination details area, cleanliness assessment results area (e.g., including overall cleanliness level and cleanliness levels of each intestinal segment), key findings area (e.g., the distribution and type of identified residues), and physician recommendations area, etc. This template defines the names, data types, and display order of each field in the report, ensuring consistency and readability of the report content. For example, the system will display the overall cleanliness level in a prominent position in the report and visually demonstrate the cleanliness status of each intestinal segment using charts or color coding, such as green for "excellent," yellow for "fair," and red for "poor."

[0040] Specifically, step S530 involves outputting and storing a structured report. As one embodiment of the invention, outputting and storing the structured report specifically includes: outputting the formatted structured report in an editable and printable electronic document format, such as PDF or HL7 CDA, to a clinical workstation, hospital information system, or electronic medical record system, and simultaneously storing it in a secure medical database for subsequent querying, archiving, or statistical analysis. The output report can be transmitted to a designated receiving end via a network interface such as the DICOM protocol or RESTful API. During storage, the system generates a unique identifier, such as an MD5 hash value, for each report and records metadata such as the report's creation time and the user who operated it to ensure the report's integrity and traceability. The report data stored in the database is version controlled to ensure that any modifications are recorded. Simultaneously, the system supports report printing to meet the needs of traditional paper reports; printed reports include anti-counterfeiting watermarks and electronic signatures to further guarantee their legal validity.

[0041] This application also proposes a system for judging and recognizing the intestinal cleanliness level of colonoscopy patients. Figure 1 This is a schematic diagram of the overall technical architecture of the intestinal cleanliness level assessment and recognition system for colonoscopy patients proposed in this invention. Figure 1 As shown, the intestinal cleanliness level judgment and recognition system for colonoscopy patients of this application includes: an image acquisition unit, an image preprocessing module, a multi-scale feature extraction module, a cleanliness recognition and classification module, a report generation and display module, and a database module.

[0042] The image acquisition unit is used to acquire colonoscopy image data. Specifically, this unit integrates a high-resolution colonoscopy probe and an image sensor, enabling real-time acquisition of video streams or still images of the patient's intestinal tract. The image acquisition unit interacts with the image preprocessing module via a high-speed data transmission interface, such as Universal Serial Bus 3.0 or Gigabit Ethernet, ensuring the real-time nature and integrity of the raw image data. The unit also includes a light source control module for adjusting the illumination brightness of the colonoscopy tip to adapt to the lighting requirements of different intestinal environments.

[0043] The image preprocessing module, connected to the image acquisition unit, performs noise suppression, brightness and contrast enhancement, and color correction and standardization on colonoscopy image data. This module contains multiple processing sub-units, such as a median filtering unit, a non-local mean denoising unit, an adaptive histogram equalization unit, a gamma correction unit, and a white balance processing unit. Each sub-unit employs a high-performance image processing algorithm to eliminate various noises and distortions in the original image, improve image quality, and provide high-quality input for subsequent feature extraction. This module receives data from the image acquisition unit via a software interface and transmits the processed image data to the multi-scale feature extraction module via shared memory or a file system.

[0044] The multi-scale feature extraction module, connected to the image preprocessing module, inputs the preprocessed colonoscopy images into a multi-scale convolutional neural network model. The model extracts deep semantic features and shallow texture features from the images and fuses these features to obtain a multi-scale fused feature vector. The core of this module is a deep learning inference engine loaded with a pre-trained multi-scale convolutional neural network model. The model architecture includes multiple convolutional layers, pooling layers, and feature fusion layers, enabling the extraction of image features from different receptive fields and levels of abstraction. This module receives image data from the image preprocessing module and outputs a fused feature vector containing multi-scale information to the cleanliness identification and classification module. Internally, this module includes parallel computing units such as graphics processing units (GPUs) or application-specific integrated circuits (ASICs) to accelerate the feature extraction process.

[0045] The cleanliness identification and classification module, connected to the multi-scale feature extraction module, is used to input the multi-scale fused feature vector into a deep classification model. The deep classification model predicts the intestinal cleanliness level and performs confidence assessment and post-processing on the predicted cleanliness level. This module is the decision-making core of the system and contains a trained deep classifier, such as a multilayer perceptron or support vector machine. It receives the multi-scale fused feature vector as input and outputs the predicted intestinal cleanliness level and its corresponding confidence score. This module also includes a post-processing unit for smoothing and correcting the prediction results based on contextual information, ensuring the accuracy and stability of the final judgment.

[0046] The report generation and display module, connected to the cleanliness identification and classification module, integrates the intestinal cleanliness level assessment results with relevant auxiliary information, formats the integrated information into a structured report, and outputs and stores the structured report. This module is responsible for presenting the system analysis results in a user-friendly format. It includes a report template management unit for managing and selecting report templates that conform to clinical standards, a data integration unit for collecting data from various modules, and a document generation unit for populating the templates with data to generate electronic reports. This module displays the report to medical staff through a graphical user interface and provides printing and export functions.

[0047] The database module, connected to the image acquisition unit, image preprocessing module, multi-scale feature extraction module, cleanliness identification and classification module, and report generation and display module, stores raw colonoscopy image data, preprocessed image data, extracted feature vectors, cleanliness identification models, generated cleanliness level assessment reports, and patient basic information and examination records. This module employs a high-performance, high-security relational or non-relational database management system to ensure persistent data storage, efficient retrieval, and data integrity. The database module also handles data backup, recovery, and access control to comply with medical data security regulations.

[0048] In summary, this application provides an automated, objective, and standardized scheme for judging and identifying the intestinal cleanliness level of colonoscopy patients through the above-mentioned system and method. This significantly improves the quality and efficiency of colonoscopy, reduces errors caused by the subjectivity and fatigue of manual assessment, and provides reliable auxiliary support for the early diagnosis and treatment of colorectal diseases.

Claims

1. A method for determining and identifying the bowel cleanliness level of patients undergoing colonoscopy, characterized in that, include: Acquire colonoscopy image sequence data, which includes a high-definition color video stream of the inside of the intestine, acquired in real time through an endoscope probe, with a preset frame rate, resolution, and color depth; The colonoscopy image sequence data is subjected to noise suppression, contrast enhancement, image normalization, and region of interest segmentation to generate standardized intestinal image data. Multi-scale feature extraction is performed on the standardized intestinal image data to obtain a multi-scale fusion feature vector related to intestinal cleanliness. Based on the multi-scale fusion feature vector related to intestinal cleanliness, a pre-trained deep learning model is used to identify and classify intestinal cleanliness to obtain a cleanliness identification result containing the probability distribution of cleanliness scores for each intestinal region. According to the cleanliness identification result, the cleanliness scores of each intestinal region are aggregated and mapped to an intestinal cleanliness level standard to generate a quantitative report containing the overall intestinal cleanliness score, the scores of each region, and the intestinal cleanliness level.

2. The method for determining and identifying the intestinal cleanliness level of colonoscopy patients according to claim 1, characterized in that, The acquisition of colonoscopy image sequence data includes: real-time acquisition of a high-definition color video stream of the inside of the intestine through the endoscope probe, wherein the video stream is transmitted at a frame rate of at least 25 frames per second, each video frame has a resolution of at least 1,920 pixels by 1,080 pixels, and a color depth of 24-bit true color; the endoscope probe is connected to an image processing unit through an optical fiber, and the image processing unit converts analog signals into digital signals and transmits them to the main processing unit through a high-speed data bus.

3. The method for determining and identifying the intestinal cleanliness level of colonoscopy patients according to claim 1, characterized in that, The noise suppression, contrast enhancement, image normalization, and region of interest (ROI) segmentation processes for the colonoscopy image sequence data include: noise filtering of the colonoscopy image sequence data using a nonlocal mean filtering algorithm or a Gaussian filtering algorithm; contrast enhancement of the colonoscopy image sequence data using a contrast-limited adaptive histogram equalization algorithm or a gamma correction algorithm; image normalization of the colonoscopy image sequence data, unifying the brightness, contrast, and chroma of the image to a preset standard range; and region of interest (ROI) segmentation of the colonoscopy image sequence data using a semantic segmentation model based on a deep convolutional network, specifically a U-Net model or a Mask R-CNN model.

4. The method for determining and identifying the intestinal cleanliness level of colonoscopy patients according to claim 1, characterized in that, The multi-scale feature extraction of the standardized intestinal image data includes: extracting multi-level features from the standardized intestinal image data using a deep convolutional neural network backbone structure, wherein the backbone structure adopts ResNet or EfficientNet; fusing the multi-level features using a feature pyramid network structure, wherein the feature pyramid network structure fuses the different levels of features extracted by the backbone network from top to bottom and bottom to top to generate a multi-scale feature map set; and assigning weights to the multi-scale feature map set using a spatial attention mechanism or a channel attention mechanism to generate a feature vector containing high-dimensional semantic information related to intestinal cleanliness as the intestinal cleanliness-related multi-scale fused feature vector.

5. The method for determining and identifying the intestinal cleanliness level of colonoscopy patients according to claim 1, characterized in that, The step of identifying and classifying intestinal cleanliness based on the multi-scale fusion feature vector related to intestinal cleanliness using a pre-trained deep learning model includes: modeling the multi-scale fusion feature vector using temporal information, wherein the temporal information modeling employs a long short-term memory network, a gated recurrent unit, or a temporal convolutional network to capture the dynamic information of intestinal cleanliness changes over time; inputting the feature vector after the temporal information modeling into a fully connected classification layer, wherein the fully connected classification layer contains at least three output nodes, each output node corresponding to the right colon, transverse colon, and left colon regions of the Powell Intestinal Preparation Scale, and the output nodes output the probability distribution of cleanliness scores for each region; and determining the cleanliness score value for each intestinal region based on the probability distribution of cleanliness scores for each region.

6. The method for determining and identifying the intestinal cleanliness level of colonoscopy patients according to claim 5, characterized in that, The fully connected classification layer is pre-trained using a dataset containing a large number of colonoscopy image sequences annotated by experienced physicians. The pre-training is optimized using a cross-entropy loss function to minimize the difference between the predicted score and the actual score.

7. The method for determining and identifying the intestinal cleanliness level of colonoscopy patients according to claim 5, characterized in that, The cleanliness score for each intestinal region ranges from zero to three points, where zero points indicate a large amount of solid residue, one point indicates a small amount of non-adsorbent residue, two points indicate a small amount of adsorbent or turbid liquid, and three points indicate that the intestinal mucosa is clearly visible with no residue or very little clear liquid.

8. The method for determining and identifying the intestinal cleanliness level of colonoscopy patients according to claim 1, characterized in that, The step of aggregating the cleanliness scores of each intestinal region based on the cleanliness identification results and mapping them to an intestinal cleanliness level standard to generate a quantitative report containing an overall intestinal cleanliness score, scores of each region, and intestinal cleanliness level includes: aggregating the cleanliness scores of each intestinal region to calculate the overall intestinal cleanliness score, which is the arithmetic mean of the scores of the right colon, transverse colon, and left colon regions; mapping the overall intestinal cleanliness score to a preset intestinal cleanliness level standard, which includes four levels: excellent, good, average, and poor; and generating a visualization report containing the overall intestinal cleanliness score, cleanliness scores of each region, the intestinal cleanliness level, and a confidence assessment. The visualization report highlights the identified residue areas on the original colonoscopy image and labels their type and size.

9. A system for determining and identifying the intestinal cleanliness level of colonoscopy patients, characterized in that, include: The image data acquisition module is used to acquire colonoscopy image sequence data, which includes a high-definition color video stream of the inside of the intestine acquired in real time through the endoscope probe, with a preset frame rate, resolution and color depth; the image preprocessing module is used to perform noise suppression, contrast enhancement, image normalization and region of interest segmentation on the colonoscopy image sequence data to generate standardized intestinal image data. A multi-scale feature extraction module is used to extract multi-scale features from the standardized intestinal image data to obtain a multi-scale fusion feature vector related to intestinal cleanliness. A cleanliness identification and classification module is used to identify and classify intestinal cleanliness based on the multi-scale fusion feature vector related to intestinal cleanliness using a pre-trained deep learning model to obtain a cleanliness identification result containing the probability distribution of cleanliness scores for each intestinal region. A cleanliness level report generation module is used to aggregate the cleanliness scores of each intestinal region and map them to the intestinal cleanliness level standard according to the cleanliness identification result to generate a quantitative report containing the overall intestinal cleanliness score, the scores of each region, and the intestinal cleanliness level. The intestinal cleanliness level assessment and recognition system for colonoscopy patients according to claim 9 is characterized in that, The image data acquisition module includes: a high-definition endoscope probe for real-time acquisition of high-definition color video streams inside the intestine, wherein the frame rate, resolution, and color depth of the video stream meet medical imaging standards; an image processing unit for converting the analog signals acquired by the high-definition endoscope probe into digital signals and performing preliminary image correction and encoding; and a high-speed data transmission interface for transmitting the digital image data to subsequent processing modules, wherein the interface adopts a serial digital interface or an Ethernet interface standard.

Images