Esophageal cancer radiotherapy and chemotherapy patient oral mucositis image grading method and system

Abstract

An esophageal cancer radiotherapy and chemotherapy patient oral mucositis image grading method and system relates to the technical field of medical image processing and auxiliary diagnosis. A color image of the oral mucosa is obtained, and image microblocks are obtained through clustering segmentation. According to the local color dispersion and spatial distribution, a healthy color reference point is extracted, the color relative distance of each microblock and the reference point is calculated, and a relative red shift color difference matrix is generated through one-way truncation mapping. Core seed points are extracted for region growing expansion and calculation of the spatial color difference gradient, and when the gradient is less than a threshold value, boundary truncation is triggered, and a red macule lesion connected domain is fused. The mucositis grading result is output based on the area ratio of the lesion, and the feedback optimization parameter is used to update the constraint condition. It can solve the problem of red macule lesion extraction missed and false detection caused by ignoring the individual mucosa color difference characteristics and not breaking through the absolute global color threshold assumption.

Description

Technical Field

[0001] This application relates to the field of medical image processing and auxiliary diagnostic technology, specifically to an image grading method and system for oral mucositis in patients undergoing radiotherapy and chemotherapy for esophageal cancer. Background Technology

[0002] In the clinical setting of concurrent chemoradiotherapy for esophageal cancer patients, oral mucositis is one of the most common and serious adverse reactions. Image-based grading and evaluation of oral mucositis is a core basis for doctors to adjust treatment plans and intervene in mucosal protection. Current mainstream technologies generally use a globally fixed red threshold based on the RGB or HSV color space for traversal segmentation. The underlying assumption is that pathological erythema possesses a uniform and constant absolute color signature across the entire population.

[0003] However, in real clinical settings, the baseline color of a patient's oral mucosa is significantly affected by multiple factors, including individual differences, pigment deposition from chemotherapy drugs, and capillary blood supply, resulting in substantial baseline shifts. Existing absolute thresholding methods confuse the two physical quantities of individual baseline color and pathological erythema increment, failing to recognize the inherent contradiction between absolute thresholds and individual baseline shifts. They cannot fundamentally eliminate background color interference because they still fit absolute pixel features and cannot understand the relative relationship between the individual baseline mucosa and erythematous lesions.

[0004] Because it cannot perceive and adapt to individual baseline color differences, the system consistently leads to serious clinical misdiagnosis under normal operating conditions. Its sensitivity is extremely unstable during the critical window period of the progression from primary to secondary mucosal inflammation, delaying optimal intervention. Therefore, overcoming the assumption of ignoring individual differences in mucosal base color, breaking through the absolute global color threshold, and avoiding systematic missed or false detections caused by individual baseline color drift in erythematous lesion extraction are urgent technical problems that need to be solved in this field. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for grading images of oral mucositis in patients undergoing radiotherapy and chemotherapy for esophageal cancer, aiming to solve the technical problems of systematic missed detection and false detection of erythematous lesions caused by ignoring individual differences in mucosal base color and failing to break through the assumption of absolute global color threshold.

[0006] The objective of this application can be achieved through the following technical solution: Firstly, a method for image grading of oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy, comprising the following steps:

[0007] Acquire color image data containing the patient's oral mucosa, and perform clustering and segmentation processing on it based on a preset region mask to obtain multiple image micro-blocks;

[0008] The target micro-block is obtained based on the local color dispersion features and spatial coordinate distribution features of the pixels within each of the image micro-blocks, and the healthy background color reference point is extracted based on the target micro-blocks.

[0009] The relative distance in chromaticity space between each of the image micro-blocks and the healthy background color reference point is obtained, and a one-way truncation mapping process is performed on the relative distance in chromaticity space based on preset constraints to generate a relative redshift color difference matrix;

[0010] Based on the relative redshift color difference matrix, the image micro-blocks that meet the preset sorting conditions are extracted as the core seed point set, and the region growing and expanding processing is performed based on the core seed point set to obtain the spatial color difference gradient between the image micro-blocks at the adjacent expansion front.

[0011] When the spatial color difference gradient is less than a preset attenuation rate threshold, a boundary truncation operation is triggered, and the traversed image micro-blocks are fused to generate a connected domain for erythema lesions.

[0012] The percentage of erythema area is obtained based on the connected domain of the erythema lesion and the preset region mask, and the mucositis grading result matching it is output. Feedback optimization parameters characterizing the confidence of the mucositis grading result are obtained, and the preset constraints are updated based on them.

[0013] Secondly, the image grading system for oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy includes the following modules:

[0014] The data acquisition module is used to acquire color image data containing the patient's oral mucosa and perform clustering and segmentation processing on it based on a preset region mask to obtain multiple image micro-blocks;

[0015] The extraction module is used to obtain target micro-blocks based on the local color dispersion features and spatial coordinate distribution features of pixels within each image micro-block, and to extract healthy background color reference points based on the target micro-blocks;

[0016] The matrix generation module is used to obtain the relative distance in chromaticity space between each of the image micro-blocks and the healthy background color reference point, and to perform a one-way truncation mapping process on the relative distance in chromaticity space based on preset constraints to generate a relative redshift color difference matrix.

[0017] The gradient acquisition module is used to extract the image micro-blocks that meet the preset sorting conditions based on the relative redshift color difference matrix as the core seed point set, and to perform region growing and expansion processing based on the core seed point set to obtain the spatial color difference gradient between the image micro-blocks at the adjacent expansion front.

[0018] The fusion module is used to trigger a boundary truncation operation when the spatial color difference gradient is less than a preset attenuation rate threshold, and to fuse the traversed image micro-blocks to generate erythema lesion connected regions.

[0019] The feedback optimization module is used to obtain the erythema area ratio based on the erythema lesion connected domain and the preset region mask, output the mucositis grading result matching it, obtain the feedback optimization parameters characterizing the confidence of the mucositis grading result, and update the preset constraint conditions based on them.

[0020] Thirdly, a computer storage medium storing computer-executable instructions, which, when executed, implement the image grading method for oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy as described in the first aspect.

[0021] Compared with the prior art, the beneficial effects of this application are:

[0022] By jointly optimizing the local color dispersion and spatial distribution of micro-blocks in the image, an adaptive anchoring healthy background color reference point is established, and a relative redshift color difference matrix is ​​constructed. The measurement standard is unified as the relative deviation from the individual's healthiest mucosa on that day, thereby completely eliminating the dual background interference caused by chemotherapy drug pigment deposition and radiotherapy-induced redness. This can fundamentally eliminate misjudgment of erythema caused by individual background color drift.

[0023] By selecting core seed points based on the relative redshift color difference matrix and using the spatial color difference gradient as the truncation criterion for region growth boundaries, the sensitivity to extremely faint primary erythema can be exponentially improved. By introducing feedback optimization parameters based on hierarchical confidence and updating constraints in a closed loop, the system can continuously adapt to individual differences, achieving self-verification and calibration solely based on the healthiest region within a single frame, without requiring historical data or additional healthy images. Attached Figure Description

[0024] Figure 1 A schematic diagram illustrating the steps of the image grading method for oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy according to this application;

[0025] Figure 2 This is a schematic diagram of the modules of the oral mucosal inflammation image grading system for esophageal cancer patients undergoing radiotherapy and chemotherapy according to this application. Detailed Implementation

[0026] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only to illustrate selected embodiments of this application.

[0027] Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item has been defined in one figure, it does not need to be further defined and explained in subsequent figures. The terms "first", "second", etc. are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0028] In existing image grading systems for oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy, fixed absolute color thresholds and globally uniform criterion rules cannot adapt to the dynamic drift and differential baseline shifts in the background color of the oral mucosa in individual patients. When patients experience overall mucosal pigmentation due to systemic toxicity of chemotherapy drugs, mucosal whitening due to anemia, or widespread mild dilation and reddening of the mucosal basal capillaries due to radiation, the existing system cannot establish a dynamic correlation between color feature criteria and the individual's physiological state on that day. This leads to absolute color threshold segmentation directly misclassifying the affected entire mucosa as erythema or failing to detect true early erythema. This static threshold mechanism fundamentally reduces the accuracy of erythema segmentation, causing the erythema area proportion received by the subsequent mucositis grading module to contain a large systematic bias, ultimately affecting the timeliness and specificity of clinical intervention decisions.

[0029] Taking the clinical scenario of concurrent chemoradiotherapy for esophageal cancer as an example, some patients receiving fluorouracil chemotherapy showed significant pigmentation on the buccal mucosa after the third cycle of treatment, making the originally healthy mucosa significantly redder. Simultaneously, other patients receiving high-dose radiotherapy experienced widespread dilation of the basal capillaries of the buccal mucosa, causing the entire mucosa to exhibit a mildly reddened state. Under these conditions, traditional systems still use a preset absolute red threshold for comprehensive segmentation, misclassifying the entire reddened area caused by chemotherapy or radiotherapy as erythema, resulting in a severely inflated erythema area ratio. The mucositis grading module then falsely reports a level two warning or even higher, leading to doctors over-adjusting treatment plans. Conversely, for cachexia patients with a whitish background, the actual, very pale level one early erythema is completely missed because the absolute color value fails to cross the fixed threshold. The mucositis grading module then fails to report this as no warning, missing the optimal intervention opportunity for mucosal protectants.

[0030] If the above problems are not addressed, systemic misjudgment of erythematous lesions will lead to a persistent misalignment between the radiotherapy and chemotherapy treatment plan and the patient's actual mucosal condition, exacerbating the risk of further mucosal ulceration, infection, and even treatment interruption. The rigidity of the fixed threshold mechanism will hinder the system from capturing the critical state of transformation from grade I erythema to grade II ulceration, delaying the optimal time for rehabilitation intervention, and ultimately creating a negative feedback loop that affects the safety and compliance of the entire esophageal cancer radiotherapy and chemotherapy treatment cycle.

[0031] To address the aforementioned issues, this application first considers establishing a dynamic correlation mechanism between color feature criteria and individual physiological states. Traditional systems use a globally fixed threshold to cut erythema, resulting in key boundary information being masked by absolute color signatures, failing to reflect individual baseline drift. To resolve this, this application attempts to construct a relative color difference field in the perceived color space, with the individual's healthiest mucosal region of the day as the dynamic zero point. The individual baseline is anchored by finding the minimum covariance value, and a relative redshift color difference matrix is ​​generated using this baseline as a subtraction factor. Further analysis reveals that relying solely on absolute color difference values ​​is insufficient to address boundary diffusion and missed detections caused by baseline drift. A spatial color difference gradient needs to be introduced as a boundary truncation criterion, with topological boundary delineation based on the physical diffusion attenuation law of inflammatory capillary dilation. By designing a linkage mechanism between the extraction module, matrix generation module, and gradient acquisition module, the erythema segmentation boundary adaptively adjusts with the individual's actual state of the day, thereby solving the problems of assessment mismatch and systematic misjudgment.

[0032] Therefore, such as Figure 1 As shown, this application provides an image grading method for oral mucositis in patients undergoing radiotherapy and chemotherapy for esophageal cancer, including the following steps:

[0033] Acquire color image data containing the patient's oral mucosa, and perform clustering and segmentation processing on it based on a preset region mask to obtain multiple image micro-blocks;

[0034] The target micro-block is obtained based on the local color dispersion features and spatial coordinate distribution features of the pixels within each of the image micro-blocks, and the healthy background color reference point is extracted based on the target micro-blocks.

[0035] The relative distance in chromaticity space between each of the image micro-blocks and the healthy background color reference point is obtained, and a one-way truncation mapping process is performed on the relative distance in chromaticity space based on preset constraints to generate a relative redshift color difference matrix.

[0036] Based on the relative redshift color difference matrix, the image micro-blocks that meet the preset sorting conditions are extracted as the core seed point set, and the region growing and expanding processing is performed based on the core seed point set to obtain the spatial color difference gradient between the image micro-blocks at the adjacent expansion front.

[0037] When the spatial color difference gradient is less than a preset attenuation rate threshold, a boundary truncation operation is triggered, and the traversed image micro-blocks are fused to generate a connected domain for erythema lesions.

[0038] The percentage of erythema area is obtained based on the connected domain of the erythema lesion and the preset region mask, and the mucositis grading result matching it is output. Feedback optimization parameters characterizing the confidence of the mucositis grading result are obtained, and the preset constraints are updated based on them.

[0039] In another implementation, such as Figure 2 As shown, this application also provides an image grading system for oral mucositis in patients undergoing radiotherapy and chemotherapy for esophageal cancer, including the following modules:

[0040] The data acquisition module is used to acquire color image data containing the patient's oral mucosa and perform clustering and segmentation processing on it based on a preset region mask to obtain multiple image micro-blocks;

[0041] The extraction module is used to obtain target micro-blocks based on the local color dispersion features and spatial coordinate distribution features of pixels within each image micro-block, and to extract healthy background color reference points based on the target micro-blocks;

[0042] The matrix generation module is used to obtain the relative distance in chromaticity space between each of the image micro-blocks and the healthy background color reference point, and to perform a one-way truncation mapping process on the relative distance in chromaticity space based on preset constraints to generate a relative redshift color difference matrix.

[0043] The gradient acquisition module is used to extract the image micro-blocks that meet the preset sorting conditions based on the relative redshift color difference matrix as the core seed point set, and to perform region growing and expansion processing based on the core seed point set to obtain the spatial color difference gradient between the image micro-blocks at the adjacent expansion front.

[0044] The fusion module is used to trigger a boundary truncation operation when the spatial color difference gradient is less than a preset attenuation rate threshold, and to fuse the traversed image micro-blocks to generate erythema lesion connected regions.

[0045] The feedback optimization module is used to obtain the erythema area ratio based on the erythema lesion connected domain and the preset region mask, output the mucositis grading result matching it, obtain the feedback optimization parameters characterizing the confidence of the mucositis grading result, and update the preset constraint conditions based on them.

[0046] I. Acquisition of image micro-blocks;

[0047] The data acquisition module acquires current-frame two-dimensional color image data of the patient's buccal mucosa using a standardized oral monocular camera. Simultaneously, a pre-defined region mask, such as a connected region mask of the buccal mucosa, is output by a pre-anatomical region segmentation module. Specifically, target mucosal image data is extracted based on the color image data and the pre-defined region mask. This target mucosal image data is then converted to the CIELAB color space, resulting in a three-dimensional feature matrix containing a lightness feature channel, a first chromaticity feature channel, and a second chromaticity feature channel.

[0048] In this step, the brightness feature channel corresponds to the L channel in the CIELAB color space, representing the brightness information of a pixel from pure black to pure white; the first chromaticity feature channel corresponds to the a channel in the CIELAB color space, representing the color bias of a pixel from green to red; and the second chromaticity feature channel corresponds to the b channel in the CIELAB color space, representing the color bias of a pixel from blue to yellow.

[0049] The reason for converting color image data from the original RGB color space to the CIELAB color space is that, under the conditions of conventional radiotherapy and chemotherapy clinics, the ring lighting easily produces reflections on the moist mucous membrane surface. Directly using the RGB space is easily affected by fluctuations in light intensity. The CIELAB color space, on the other hand, mathematically decouples the brightness of light from the inherent properties of color. Furthermore, the CIELAB space has a uniform and equidistant characteristic for perceiving red, green, yellow, and blue hues in biological tissues, effectively filtering out slight reflections from the mucous membrane surface. The essence of pathological erythema is subcutaneous capillary dilation and congestion, and its physical characteristics are mainly mapped to the positive value changes of the first chromaticity feature channel. Therefore, subsequent processing in the CIELAB space can maximize the separability of the erythema signal from background noise.

[0050] The three-dimensional feature matrix is ​​clustered and segmented using a linear iterative clustering algorithm to obtain multiple image micro-blocks. The mean values ​​of the first and second chroma feature channels of pixels within each micro-block, as well as the centroid coordinates of each micro-block, are extracted. The linear iterative clustering algorithm used in this embodiment is the SLIC algorithm, which combines spatial distance and color similarity to perform joint clustering and segmentation of the three-dimensional feature matrix. This clusters a continuous pixel array into several superpixel micro-blocks with edges conforming to tissue texture. Each image micro-block corresponds to an actual physical size of approximately 2 mm × 2 mm.

[0051] The reason for using superpixel clustering segmentation instead of pixel-by-pixel traversal is that oral images often contain sensor noise in dark areas and reflections from tiny saliva bubbles during actual acquisition, making it easy to generate spurious features during single-pixel feature calculation. Superpixel segmentation forces the mean distribution of locally adjacent similar pixels, achieving local smoothing and noise reduction while preserving the true boundaries of mucosal lesions. It can directly eliminate interference from tiny sensor noise and tiny saliva bubbles using the principle of spatial topological averaging.

[0052] II. Extraction of the baseline points for a healthy complexion;

[0053] After obtaining multiple image micro-blocks, the extraction module obtains the target micro-block based on the local color dispersion features and spatial coordinate distribution features of the pixels within each image micro-block, and extracts the healthy background color reference point based on it. This step is one of the core innovative nodes of this invention. It breaks the technical bias of regarding pathological erythema as having an absolutely uniform color signature, and instead uses the healthiest region within a single frame image to self-certify and define the boundary of the pathological region.

[0054] Specifically, the first variance of the first chroma feature channel of the pixels within the image micro-block is first obtained. The second variance of the second chromaticity feature channel The physical meaning is as follows: the pigment and blood flow distribution of healthy mucosa is extremely uniform at the local microscopic level, and its color variance approximates the background white noise of the sensor. In contrast, pathological erythema areas experience drastic fluctuations in internal tissue color due to congestion and exudation, resulting in a significantly increased variance. Ulcer or pseudomembrane areas also exhibit high variance due to uneven tissue exudation. Therefore, the sum of the first and second variances can quantitatively reflect the local color dispersion characteristics of the image micro-block; the smaller the dispersion, the smoother the local color, and the closer it is to healthy mucosal tissue unaffected by inflammation.

[0055] Next, the Euclidean distance between the centroid coordinates of the image micro-block and the geometric center of the preset region mask is obtained. And obtain its distribution coefficient. The distribution coefficient is determined by a value of 1 and a value determined by a preset penalty coefficient and Euclidean distance. The product terms formed by the preset distance thresholds are added together to obtain the result. The value is the Euclidean distance from the centroid of the image micro-block to the geometric center of the preset region mask. Its physical meaning is that non-mucosal artifacts are prone to appear near teeth, mucosal vignetting, or image edge dead zones. If the minimum variance is used as the sole optimization criterion without constraints, the system is highly likely to misclassify dark pixels in edge dead zones as healthy baselines. After introducing a spatial distance penalty weight, the closer an image micro-block is to the geometric center of the preset region mask, the smaller its distribution coefficient, and the more likely it is to be selected as a target micro-block; conversely, the closer an image micro-block is to the edge, the larger its distribution coefficient, and the lower its probability of being selected as a target micro-block. The calculation relationship of this distribution coefficient is as follows: ;

[0056] in, The distribution coefficient is... The preset penalty coefficient is, for example, 0.5; The Euclidean distance between the centroid of the image micro-block and the geometric center of the preset region mask; The preset distance threshold is, for example, the maximum span radius of the preset region mask.

[0057] Subsequently, the joint evaluation coefficients corresponding to the image micro-blocks are obtained. The joint evaluation coefficient The joint evaluation coefficient is obtained by multiplying the sum of the first and second variances by the distribution coefficient. It fuses local color dispersion features and spatial coordinate distribution features into a single optimization index through multiplicative coupling. Its physical meaning is: when the chromaticity fluctuation of pixels within an image micro-block approaches the background noise (i.e., the local color dispersion feature is minimal), and the physical coordinates of the image micro-block approach the geometric center of the preset region mask (i.e., the spatial coordinate distribution feature approaches 1), the calculated joint evaluation coefficient reaches a global minimum. This minimum value points to a unique physical state, namely, a healthy mucosa that is not eroded by inflammation and is far from the edge of the image dark corner. The calculation relationship of this joint evaluation coefficient is as follows: ;

[0058] Finally, the image micro-block with the smallest joint evaluation coefficient is selected as the target micro-block, and the mean values ​​of the first and second chroma feature channels of the target micro-block are used as the healthy background color reference point. This healthy background color reference point is characterized by two components, corresponding to the mean values ​​of the first and second chroma feature channels of the target micro-block, respectively.

[0059] In the actual clinical scenario of concurrent chemoradiotherapy for esophageal cancer, the aforementioned covariance minimization optimization logic possesses natural compatibility with both radiotherapy and chemotherapy, two different pathological attack mechanisms. On the one hand, the systemic toxicity of chemotherapy drugs mainly manifests as pigmentation or anemia and whitening of the mucous membranes throughout the body, which appears as a general background color shift in the image, without changing the color uniformity within the healthiest area. On the other hand, the basal dilation and congestion of the local capillary bed caused by radiation penetration appears as a general reddish background in the image, but similarly does not change the color uniformity within the smoothest area. Therefore, regardless of whether the patient's background color is reddish, whitish, or reddish, the extraction module can find the relatively smoothest and healthiest micro-block with the smallest variance on that day within the image as a benchmark anchor point, thus ensuring that subsequent relative color difference measurements are always based on the patient's true physiological baseline on that day.

[0060] III. Generation of the relative redshift color difference matrix;

[0061] After extracting the healthy background color reference point, the matrix generation module obtains the relative chromaticity distance between each image micro-block and the healthy background color reference point. Based on preset constraints, it performs a one-way truncated mapping process on these relative chromaticity distances to generate a relative redshift chromaticity difference matrix. Specifically, it obtains a first difference between the mean of the first chromaticity feature channel of the image micro-block and the mean of the first chromaticity feature channel of the healthy background color reference point; it also obtains a second difference between the mean of the second chromaticity feature channel of the image micro-block and the mean of the second chromaticity feature channel of the healthy background color reference point; and it performs a square root operation on the sum of the squares of the first and second differences to obtain the corresponding relative chromaticity distance. Physically, this means that on a two-dimensional chromaticity plane, the healthy background color reference point is considered the origin, and the two-dimensional Euclidean distance between the chromaticity mean vector of each image micro-block and this origin is used as a relative chromaticity difference metric. This transforms the algorithm from identifying absolute red in the image to identifying the chromaticity deviation relative to the health of the mucous membrane.

[0062] Next, a one-way truncation mapping process is performed on the relative distance in the chroma space based on preset constraints. The specific judgment logic is as follows: if the average value of the first chroma feature channel of the image micro-block is greater than the sum of the average value of the first chroma feature channel of the healthy background color reference point and a preset bias parameter, then its relative distance in the chroma space is used as the relative redshift increment; if the average value of the first chroma feature channel of the image micro-block is less than or equal to the sum of the average value of the first chroma feature channel of the healthy background color reference point and the preset bias parameter, then its relative distance in the chroma space is zeroed out and used as the relative redshift increment. The preset bias parameter is the core parameter of the preset constraints; its initial value is usually zero, and it is dynamically updated subsequently through a feedback optimization module.

[0063] The reason for introducing unidirectional truncated mapping is based on the clear pathological and physical characteristics of mucositis erythema: the essence of erythema is capillary dilation and congestion, which in the CIELAB color space will necessarily manifest as a positive increase in the first chromaticity feature channel, meaning it is redder than the healthy baseline. Conversely, non-erythematous lesions such as tissue ischemia, pseudomembrane coverage, and anemia show an inverse change in the first chromaticity feature channel, meaning they are whiter or yellower than the healthy baseline. Through unidirectional truncated mapping, the matrix generation module forcibly applies directional filtering, retaining only the color difference components consistent with the pathological characteristics of erythema, filtering out inverse color difference interference caused by non-erythematous factors, thereby preventing pale mucosal or pseudomembrane areas from being incorrectly included in the erythema determination range.

[0064] A relative redshift chromatic difference matrix is ​​generated based on the relative redshift increment mapping of each of the image micro-blocks. At the engineering implementation level, the matrix generation module further performs the following processing: obtaining the two-dimensional spatial size parameters of the preset region mask, and constructing an initial two-dimensional matrix with the same dimensions based on the two-dimensional spatial size parameters; extracting the pixel coordinate mapping set corresponding to each of the image micro-blocks in the preset region mask; assigning the relative redshift increment of each of the image micro-blocks to the matrix nodes corresponding to the pixel coordinate mapping sets in the initial two-dimensional matrix; after all image micro-blocks have been traversed and assigned values, the obtained two-dimensional matrix containing the spatial topology and redshift increment values ​​is used as the relative redshift chromatic difference matrix.

[0065] Through the above steps, the matrix generation module outputs a relative redshift color difference matrix that corresponds to the topological structure of the original image and retains only the relative pathological redshift increment. This matrix completely eliminates the global bias caused by differential baseline drift, such as individual baseline color being reddish, whitish, or reddish, and forces the measurement standard of all image micro-blocks to be unified to the relative deviation from the individual's healthy mucosa on that day, providing clean and stable input data for subsequent boundary truncation and erythema segmentation.

[0066] IV. Extraction of the core seed point set and acquisition of spatial color difference gradient;

[0067] The gradient acquisition module extracts the image micro-blocks that meet the preset sorting conditions based on the relative redshift color difference matrix as the core seed point set, and performs region growth and expansion processing based on them to obtain the spatial color difference gradient between the image micro-blocks at the adjacent expansion front.

[0068] Specifically, firstly, the matrix values ​​corresponding to each image micro-block in the relative redshift color difference matrix are extracted. A predetermined number of image micro-blocks are then extracted sequentially in descending order and pushed into a region growth queue to obtain a core seed point set. In this embodiment, the predetermined number is, for example, 5% of the total number of image micro-blocks in the relative redshift color difference matrix. That is, the top 5% of image micro-blocks with the largest relative redshift increments are extracted as the core seed point set. These image micro-blocks represent the core lesion region with the most severe redshift in the current image and are the starting point for subsequent region growth and expansion.

[0069] The reason for using a relative redshift increment descending sorting and then taking the first preset percentage to extract the core seed point set, instead of using an absolute color difference threshold for screening, is that the numerical amplitude of the relative redshift color difference matrix itself has already adaptively drifted with the individual's background color baseline, and a fixed absolute threshold would still reintroduce individual differences. On the other hand, using the sorting and then taking the percentage method ensures that the seed point extraction rules remain consistent across different patients, which is in line with the core design logic of this invention of replacing global absolute differences with individual relative differences.

[0070] Next, using the image micro-blocks in the core seed point set as parent node micro-blocks, a traversal expansion is performed on their adjacent child node micro-blocks. "Adjacent" refers to image micro-blocks that share edges or angles in the two-dimensional spatial topology. During each expansion, the gradient acquisition module starts from the parent node micro-block, visits each of its adjacent untraversed child node micro-blocks one by one, and calculates the corresponding spatial chromatic gradient for each pair of parent and child node micro-blocks.

[0071] Specifically, the matrix values ​​corresponding to the parent node microblock and its child node microblocks are obtained. , The absolute difference between them is used to obtain the centroid coordinates of the parent node microblock and its child node microblock. , The spatial chromatic aberration gradient between adjacent image micro-blocks at the expansion front can be obtained by dividing the absolute difference by the Euclidean distance between them. Its physical meaning is as follows: This spatial chromatic aberration gradient represents the slope of the relative redshift increment as physical distance increases, that is, the fading rate from the lesion core to healthy tissue. When the expansion front is located inside the erythematous lesion, the relative redshift increment between adjacent image micro-blocks is basically the same, and the spatial chromatic aberration gradient is small; when the expansion front is exactly at the boundary between the erythematous lesion and healthy tissue, the relative redshift increment drops sharply along the spatial direction, and the spatial chromatic aberration gradient reaches its maximum value; when the expansion front crosses the lesion boundary and enters the healthy tissue background, the relative redshift increment tends to stabilize again, and the spatial chromatic aberration gradient falls back to the noise level. The calculation relationship of this spatial chromatic aberration gradient is as follows:

[0072] ;

[0073] in, The spatial chromatic aberration gradient between the image micro-blocks at adjacent expansion fronts. The value of the parent node micro-block in the relative redshift color difference matrix is ​​the corresponding matrix value. The value of the child node micro-block in the relative redshift color difference matrix is ​​the matrix value. The centroid coordinates of the parent node's microblock. The coordinates of the centroid of the child node microblock.

[0074] V. Boundary truncation and the generation of connected regions of erythematous lesions;

[0075] After the gradient acquisition module outputs the spatial color difference gradient, the fusion module triggers a boundary truncation operation when the spatial color difference gradient is less than a preset attenuation rate threshold, fusing the traversed image micro-blocks to generate a connected component for the erythematous lesion. Specifically, the spatial color difference gradient when expanding towards the adjacent expansion front of the child node micro-block is numerically compared with the preset attenuation rate threshold. The preset attenuation rate threshold is a noise floor threshold reflecting the natural diffusion properties of capillaries. For example, it is taken as the mean of the spatial color difference gradient in the healthy mucosal background area obtained through long-term statistics plus three standard deviations. Its physical meaning represents the maximum allowable gradient amplitude generated by the baseline fluctuation of healthy tissue.

[0076] When the spatial color difference gradient is less than the preset attenuation rate threshold, it indicates that the rapid fading process of inflammation and congestion has ended, and the color of the block has been smoothly integrated into the background baseline fluctuation. At this time, the fusion module triggers the boundary truncation operation to terminate the region growth expansion towards the spatial topology direction where the child node micro-block is located. When the spatial color difference gradient is greater than or equal to the preset attenuation rate threshold, it indicates that the fading process of inflammation is still ongoing, and the child node micro-block still belongs to the transitional gradient area of ​​the erythematous lesion. At this time, the fusion module marks the child node micro-block as the image micro-block that has been traversed, and updates the child node micro-block as a new expansion front to continue the region growth expansion processing.

[0077] The reason for using spatial color difference gradient as the boundary truncation criterion, rather than the absolute color difference threshold commonly used in traditional methods, is that, based on fluid dynamics and pathological exudation models, capillary dilation caused by inflammatory factors has the physical characteristics of high concentration at the center and nonlinear gradual decay around the edges. The edge of pathological tissue is, topologically, essentially the inflection point of the gradient, rather than an absolute color discontinuity. Boundary delineation methods based on absolute color difference thresholds often suffer from incomplete cuts or excessive diffusion under background color drift conditions. In contrast, the topological boundary determination method based on spatial color difference gradient can delineate the true boundary according to the physical diffusion law of the pathological tissue itself, fundamentally solving the boundary delineation distortion problem caused by background color drift.

[0078] After the boundary truncation operation is triggered in all spatial topological directions of the core seed point set during region growth and expansion, the fusion module extracts the set of spatial physical coordinates of all image micro-blocks marked as traversed. Based on the set of spatial physical coordinates, spatial morphological stitching and fusion processing is performed on all traversed image micro-blocks to obtain a topological region with a closed outer contour boundary, which is then used as the connected domain of the erythema lesion. In this embodiment, the spatial morphological stitching and fusion processing can be implemented by combining morphological opening and closing operations with a connected domain labeling algorithm. For example, morphological closing operations are first performed on all traversed image micro-blocks to fill small holes and gaps, and then a connected domain labeling algorithm is performed to extract the largest closed outer contour boundary as the connected domain of the erythema lesion.

[0079] VI. Optimization of the output and feedback of mucositis grading results;

[0080] The feedback optimization module obtains the erythema area ratio based on the connected domain of the erythema lesion and a preset region mask, and outputs a matching mucositis grading result. It then obtains feedback optimization parameters characterizing the confidence level of the mucositis grading result and updates the preset constraints based on these parameters. Specifically, the feedback optimization module first extracts the geometric pixel area contained in the connected domain of the erythema lesion and the total pixel area contained in the preset region mask. It then divides the geometric pixel area by the total pixel area to obtain the corresponding erythema area ratio. The percentage of erythema area is the core quantitative indicator output by this invention for subsequent clinical decision-making, and its calculation relationship can be expressed by the following formula:

[0081] Next, the mucositis grading result matching the erythema area percentage is output. The specific grading rule is: if the erythema area percentage... Less than the first preset threshold The output will be a graded result of mucositis without warning, if the percentage of erythema area is... Greater than or equal to the second preset threshold If the output represents the grade of mucositis indicating a level II warning, then the output will be the grade of the mucositis. Then, the output will be the mucositis grading result representing the first-level warning, where, The first and second preset thresholds are set with reference to the clinical intervals of the NCI-CTCAE mucositis grading standard, for example... The value is 5%. The value is 30%.

[0082] As another key innovation of this invention, the feedback optimization module further introduces a closed-loop update mechanism based on the confidence level of mucositis grading results and preset constraints, so that the solution can continuously maintain its adaptive ability to individual differences and temporal drift under long-term operating conditions. Specifically, the feedback optimization module obtains continuous... Variance evaluation value of pixel features at the boundary of the connected domain of the erythematous lesion within a preset acquisition period The variance evaluation value From continuous The sum of the squared deviations of the average sequence of the relative redshift increments of the corresponding image micro-blocks at the boundary within a preset acquisition period and their global temporal mean, divided by [the value of the summation]. The calculation yields the following physical meaning: When the erythema boundary output by the system is stable and reliable, the relative redshift increment of the image micro-blocks at the boundary remains relatively consistent across multiple consecutive acquisition cycles, resulting in a smaller variance evaluation value; conversely, when the erythema boundary output by the system oscillates frequently due to improper preset constraint settings, the variance evaluation value is larger. The calculation relationship for this variance evaluation value is as follows:

[0083] ;

[0084] in, The variance evaluation value is... This represents the total number of consecutive preset data collection cycles. Indicates the first The average value of the relative redshift increment of the corresponding image micro-block at the boundary within a preset acquisition period. Indicates continuity The average value of the relative redshift increment of the corresponding image micro-block at the boundary within a preset acquisition period, i.e., the above The global mean of the sequence.

[0085] Based on the variance evaluation value The feedback optimization module further obtains feedback optimization parameters. The feedback optimization parameters Variance evaluation value With preset smoothing coefficient The reciprocal of the summation term. Its physical meaning is: the smaller the variance assessment value, the larger the feedback optimization parameter, indicating higher boundary stability and higher confidence in the current mucositis grading result; conversely, the larger the variance assessment value, the smaller the feedback optimization parameter, indicating more severe boundary oscillations and lower confidence in the current mucositis grading result. Preset smoothing coefficient. The purpose of this approach is to prevent numerical overflow of the feedback optimization parameter when the variance evaluation value approaches zero, thus ensuring the numerical stability of the calculation. The calculation relationship of this feedback optimization parameter is as follows: ;

[0086] The feedback optimization module will use the feedback optimization parameters With preset information threshold Perform numerical comparison. When the feedback optimization parameters... Less than the preset threshold When this occurs, it indicates that the current preset constraints are no longer sufficient to provide stable boundary partitioning under the new operating conditions, necessitating an update to the preset constraints. At this point, the feedback optimization module obtains the updated preset bias parameters. The updated preset bias parameters Based on preset gain coefficient Multiply by the preset threshold With feedback optimization parameters The larger the difference between the feedback optimization parameter and the preset confidence threshold, the greater the additional bias needs to be, thus pushing the criterion of the one-way truncation mapping processing to adjust in a more stringent direction, filtering out pseudo-boundary oscillations caused by temporary background fluctuations; when the feedback optimization parameter is greater than or equal to the preset confidence threshold, the maximum value operation outputs zero, the preset bias parameter remains zero, and the preset constraint conditions remain unchanged. This update relationship is as follows: ;

[0087] in, For the updated preset bias parameters, The preset gain coefficient can be set to a positive constant between 0 and 1. To preset the threshold, The parameters are optimized for the feedback.

[0088] The feedback optimization module applies the updated preset bias parameters to the preset constraints in the next preset acquisition cycle. This ensures that when generating the relative redshift color difference matrix in the next preset acquisition cycle, the criterion for unidirectional truncation mapping is updated to the mean value of the first chromaticity feature channel of the image micro-block is greater than the sum of the mean value of the first chromaticity feature channel of the healthy background color reference point and the updated preset bias parameters. This process is repeated continuously, forming an iterative correction loop for the preset constraints, enabling the solution to maintain its adaptive capability to individual differences and temporal drift under long-term operating conditions.

[0089] Through the aforementioned mechanism of updating preset constraints based on feedback-optimized closed-loop parameters, this invention achieves dual adaptive capabilities, from self-verification within a single frame image to multi-frame temporal self-calibration. Without relying on any prior historical data or additional healthy control images, the solution utilizes only the healthiest region within a single frame image and boundary stability indices between consecutive frames to achieve unified adaptive responses to multiple interfering factors, including an individual's daily physiological baseline, chemotherapy-induced pigmentation drift, radiotherapy-induced basal redness, and fluctuations in salivary bubble microreflectivity.

[0090] As a preferred embodiment, the specific implementation process of this application is as follows: In the esophageal cancer concurrent chemoradiotherapy clinic, an industrial-grade oral monocular camera is used to acquire images of the patient's buccal mucosa under standardized ring illumination conditions. The resolution is set to 1920x1080, and a preset region mask of the buccal mucosa is output through a pre-anatomical region segmentation module. The data acquisition module performs a bitwise AND operation between the color image and the preset region mask, converts it to the CIELAB color space, and performs superpixel clustering segmentation based on the SLIC algorithm at a physical scale of approximately 2 mm × 2 mm to obtain multiple image micro-blocks.

[0091] The extraction module traverses each image micro-block, calculates the variance of the first chromaticity feature channel and the variance of the second chromaticity feature channel of each pixel, and calculates the Euclidean distance from the centroid coordinates to the geometric center of the preset region mask. According to the distribution coefficient formula and the joint evaluation coefficient formula, the joint evaluation coefficient corresponding to each image micro-block is obtained. The mean of the first chromaticity feature channel and the mean of the second chromaticity feature channel of the image micro-block with the smallest joint evaluation coefficient are taken as the healthy background color benchmark point of the patient on that day.

[0092] The matrix generation module uses the healthy background color reference point as the subtraction reference, traverses each image micro-block to calculate its relative chromaticity distance to the reference point, and performs unidirectional truncation mapping based on preset constraints defined by preset bias parameters, ultimately generating a relative redshift chromatic difference matrix. The gradient acquisition module extracts the top 5% of image micro-blocks in descending order of matrix values ​​as the core seed point set, performs breadth-first region growing expansion on adjacent image micro-blocks, and calculates the spatial chromatic difference gradient between parent and child micro-blocks in real time. The fusion module triggers boundary truncation operations based on a preset attenuation rate threshold, and performs spatial morphological stitching and fusion processing on all untrunculated and traversed image micro-blocks to form the erythema lesion connected domain.

[0093] The feedback optimization module calculates the ratio of the geometric pixel area occupied by the connected region of the erythema lesion to the total pixel area occupied by the preset region mask to obtain the erythema area proportion, and outputs the mucositis grading result based on the interval mapping between the first preset threshold and the second preset threshold. Simultaneously, it continuously monitors the relative redshift increment sequence of image micro-blocks at the boundary of the connected region of the erythema lesion within multiple consecutive preset acquisition cycles. Based on the calculation relationship between the variance evaluation value and the feedback optimization parameters, it determines whether the current preset constraints need to be updated. If necessary, it generates a bias adjustment term based on the preset gain coefficient and the preset confidence threshold to update the preset constraints for the next cycle.

[0094] In another embodiment, this application also provides a computer storage medium storing computer-executable instructions, which, when executed, implement the method for grading images of oral mucositis in patients undergoing radiotherapy and chemotherapy for esophageal cancer.

[0095] The computer storage medium includes, but is not limited to, various types of disk storage media, optical disk storage media, USB flash drive storage media, solid-state drive storage media, and cloud storage media, as well as any type of media capable of storing computer-executable instructions. When the computer-executable instructions are executed by the processor, the processor sequentially executes all the processing steps described above, including image micro-block acquisition, healthy background color reference point extraction, relative redshift color difference matrix generation, spatial color difference gradient acquisition, erythema lesion connected component generation, mucositis grading result output and feedback optimization, to complete the image grading function completely equivalent to the method embodiment of the present invention.

[0096] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A method for grading oral mucositis images in patients with esophageal cancer undergoing radiotherapy and chemotherapy, characterized in that... Includes the following steps: Acquire color image data containing the patient's oral mucosa, and perform clustering and segmentation processing on it based on a preset region mask to obtain multiple image micro-blocks; The target micro-block is obtained based on the local color dispersion features and spatial coordinate distribution features of the pixels within each image micro-block, and a healthy background color reference point is extracted based on it. The relative distance in chromaticity space between each of the image micro-blocks and the healthy background color reference point is obtained, and a one-way truncation mapping process is performed on the relative distance in chromaticity space based on preset constraints to generate a relative redshift color difference matrix; Based on the relative redshift color difference matrix, the image micro-blocks that meet the preset sorting conditions are extracted as the core seed point set, and a region growing and expansion process is performed on them to obtain the spatial color difference gradient between adjacent expansion fronts of the image micro-blocks. When the spatial color difference gradient is less than a preset attenuation rate threshold, a boundary truncation operation is triggered, and the traversed image micro-blocks are fused to generate a connected domain for erythema lesions. Based on the connected domain of the erythema lesion and the preset region mask, the proportion of erythema area is obtained, and the mucositis grading result matching it is output. Feedback optimization parameters characterizing the confidence of the mucositis grading result are obtained, and the preset constraints are updated based on them. The process of obtaining image micro-blocks includes: Based on the color image data and the preset region mask, target mucosal image data is extracted, the target mucosal image data is converted to the CIELAB color space, and a three-dimensional feature matrix containing a lightness feature channel, a first chroma feature channel, and a second chroma feature channel is obtained. The three-dimensional feature matrix is ​​subjected to clustering and segmentation processing based on a linear iterative clustering algorithm to obtain multiple image micro-blocks. The mean values ​​of the first chromaticity feature channels and the mean values ​​of the second chromaticity feature channels of the pixels inside each image micro-block are extracted, as well as the centroid coordinates of each image micro-block. The process of extracting the baseline points for healthy skin tone includes: Obtain the first variance of the first chroma feature channel of the pixels within the image micro-block. The second variance of the second chromaticity feature channel ; Obtain the Euclidean distance between the centroid coordinates of the image micro-block and the geometric center of the preset region mask. And obtain its distribution coefficient. ,in, To preset the penalty coefficient, The preset distance threshold; Obtain the joint evaluation coefficients corresponding to the image micro-blocks. The image micro-block with the smallest joint evaluation coefficient is taken as the target micro-block, and its first chromaticity feature channel mean and second chromaticity feature channel mean are taken as the healthy background color reference point. The process of obtaining the spatial color difference gradient includes: Extract the matrix values ​​corresponding to each image micro-block in the relative redshift color difference matrix, extract a preset number of image micro-blocks in descending order, and push them into the region growth queue to obtain the core seed point set; Using the image micro-block in the core seed point set as the parent node micro-block, perform traversal expansion to the adjacent child node micro-blocks; Obtain the matrix values ​​corresponding to the parent node microblock and its child node microblocks. , The absolute difference between them is used to obtain the centroid coordinates of the parent node microblock and its child node microblock. , The spatial chromatic aberration gradient between adjacent image micro-blocks at the Euclidean distance is obtained. ; ； in, The spatial chromatic aberration gradient between the image micro-blocks at adjacent expansion fronts. The value of the parent node micro-block in the relative redshift color difference matrix is ​​the corresponding matrix value. The value of the child node micro-block in the relative redshift color difference matrix is ​​the matrix value. The centroid coordinates of the parent node's microblock. The coordinates of the centroid of the child node microblock.

2. The method for grading images of oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy according to claim 1, characterized in that, The process of obtaining relative distances in the color space includes: The difference between the mean value of the first chroma feature channel of the image micro-block and the mean value of the first chroma feature channel of the healthy background color reference point is obtained, and the difference between the mean value of the second chroma feature channel of the image micro-block and the mean value of the second chroma feature channel of the healthy background color reference point is obtained. The square root operation is performed on the sum of the squares of the above two differences to obtain the corresponding relative distance in the chroma space.

3. The method for grading images of oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy according to claim 2, characterized in that, The process of generating the relative redshift color difference matrix includes: If the average value of the first chromaticity feature channel of the image micro-block is greater than the sum of the average value of the first chromaticity feature channel of the healthy background color reference point and the preset bias parameter, then the relative distance in chromaticity space is taken as the relative redshift increment. If the mean value of the first chromaticity feature channel of the image micro-block is less than or equal to the sum of the mean value of the first chromaticity feature channel of the healthy background color reference point and the preset bias parameter, then the relative distance in its chromaticity space is zeroed out and used as the relative redshift increment. Based on the relative redshift increment of each image micro-block, a relative redshift color difference matrix is ​​generated.

4. The method for grading images of oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy according to claim 1, characterized in that, The process of outputting the mucositis grading results includes: Extract the geometric pixel area contained in the connected region of the erythema lesion and the total pixel area contained in the preset region mask, and divide the geometric pixel area by the total pixel area to obtain the corresponding erythema area ratio. ; If the percentage of the erythema area Less than the first preset threshold The output will be a graded result of mucositis without warning, if the percentage of erythema area is... Greater than or equal to the second preset threshold If the output represents the grade of mucositis indicating a level II warning, then the output will be the grade of the mucositis. Then, the output will be the mucositis grading result representing the first-level warning, where, .

5. The method for grading images of oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy according to claim 3, characterized in that, The process of updating preset constraints includes: Get continuous Variance evaluation value of pixel features at the boundary of the connected domain of the erythematous lesion within a preset acquisition period ; in, Indicates the first The average value of the relative redshift increment of the corresponding image micro-block at the boundary within a preset acquisition period. Indicates continuity The average value of the relative redshift increment of the corresponding image micro-block at the boundary within a preset acquisition period; Based on the variance evaluation value Obtain feedback optimization parameters ,in, As a preset smoothing coefficient, when the feedback optimization parameter Less than the preset threshold At that time, the updated preset bias parameters are obtained. And apply it to the next preset collection cycle, where, This is the preset gain coefficient.

6. An image grading system for oral mucositis in patients with esophageal cancer undergoing radiotherapy and chemotherapy, characterized in that, Includes the following modules: The data acquisition module is used to acquire color image data containing the patient's oral mucosa and perform clustering and segmentation processing on it based on a preset region mask to obtain multiple image micro-blocks; The extraction module is used to obtain the target micro-block based on the local color dispersion features and spatial coordinate distribution features of the pixels inside each image micro-block, and to extract the healthy background color reference point based on it. The matrix generation module is used to obtain the relative distance in chromaticity space between each of the image micro-blocks and the healthy background color reference point, and to perform a one-way truncation mapping process on the relative distance in chromaticity space based on preset constraints to generate a relative redshift color difference matrix. The gradient acquisition module is used to extract the image micro-blocks that meet the preset sorting conditions based on the relative redshift color difference matrix as the core seed point set, and to perform region growth and expansion processing based on them to obtain the spatial color difference gradient between the image micro-blocks at the adjacent expansion front. The fusion module is used to trigger a boundary truncation operation when the spatial color difference gradient is less than a preset attenuation rate threshold, and to fuse the traversed image micro-blocks to generate erythema lesion connected regions. The feedback optimization module is used to obtain the erythema area ratio based on the erythema lesion connected domain and the preset region mask, and output the mucositis grading result that matches it, obtain the feedback optimization parameters that characterize the confidence of the mucositis grading result, and update the preset constraint conditions based on them. The process of obtaining image micro-blocks includes: Based on the color image data and the preset region mask, target mucosal image data is extracted, the target mucosal image data is converted to the CIELAB color space, and a three-dimensional feature matrix containing a lightness feature channel, a first chroma feature channel, and a second chroma feature channel is obtained. The three-dimensional feature matrix is ​​subjected to clustering and segmentation processing based on a linear iterative clustering algorithm to obtain multiple image micro-blocks. The mean values ​​of the first chromaticity feature channels and the mean values ​​of the second chromaticity feature channels of the pixels inside each image micro-block are extracted, as well as the centroid coordinates of each image micro-block. The process of extracting the baseline points for healthy skin tone includes: Obtain the first variance of the first chroma feature channel of the pixels within the image micro-block. The second variance of the second chroma feature channel ; Obtain the Euclidean distance between the centroid coordinates of the image micro-block and the geometric center of the preset region mask. And obtain its distribution coefficient. ,in, To preset the penalty coefficient, The preset distance threshold; Obtain the joint evaluation coefficients corresponding to the image micro-blocks. The image micro-block with the smallest joint evaluation coefficient is taken as the target micro-block, and its first chromaticity feature channel mean and second chromaticity feature channel mean are taken as the healthy background color reference point. The process of obtaining the spatial color difference gradient includes: Extract the matrix values ​​corresponding to each image micro-block in the relative redshift color difference matrix, extract a preset number of image micro-blocks in descending order, and push them into the region growth queue to obtain the core seed point set; Using the image micro-block in the core seed point set as the parent node micro-block, perform traversal expansion to the adjacent child node micro-blocks; Obtain the matrix values ​​corresponding to the parent node microblock and its child node microblocks. , The absolute difference between them is used to obtain the centroid coordinates of the parent node microblock and its child node microblock. , The spatial chromatic aberration gradient between adjacent image micro-blocks at the Euclidean distance is obtained. ; ； in, The spatial chromatic aberration gradient between the image micro-blocks at adjacent expansion fronts. The value of the parent node micro-block in the relative redshift color difference matrix is ​​the corresponding matrix value. The value of the child node micro-block in the relative redshift color difference matrix is ​​the matrix value. The centroid coordinates of the parent node's microblock. The coordinates of the centroid of the child node microblock.

7. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement the image grading method for oral mucositis in esophageal cancer patients undergoing radiotherapy and chemotherapy as described in any one of claims 1-5.

Images