A method for constructing a retinopathy of prematurity auxiliary diagnosis system
By acquiring a small number of key pixels and automatically selecting intermediate pixels, a retinal lesion path segmentation atlas label is constructed, which solves the problem of insufficient lesion path segmentation data for retinopathy of prematurity (ROP) and achieves high-precision and robust lesion path segmentation, supporting the auxiliary diagnosis of ROP in prematurity.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF COMPUTING TECH CHINESE ACAD OF SCI
- Filing Date
- 2026-04-23
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies struggle to acquire high-quality, large-scale segmentation data of retinopathy of prematurity lesion pathways, limiting the application of intelligent segmentation systems in this field, resulting in low segmentation accuracy, weak generalization ability, and poor robustness.
By acquiring a small number of key pixels and automatically selecting intermediate pixels, a retinal lesion path segmentation map label is constructed. A convolutional neural network is used to train the model to generate the lesion path segmentation map, and binary cross-entropy loss is used to optimize the model parameters.
It significantly reduces the difficulty of obtaining labeled data, ensures the accuracy and consistency of labeling, and trains a preterm infant retinopathy of prematurity auxiliary diagnostic system with high segmentation accuracy, strong generalization ability and good robustness, providing interpretable lesion path segmentation results.
Smart Images
Figure CN122289291A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, specifically to the field of intelligent segmentation of medical images, and more specifically, to a method for constructing an auxiliary diagnostic system for retinopathy of prematurity. Background Technology
[0002] Retinopathy of Prematurity (ROP) is a retinal vascular abnormality that primarily affects premature infants and is a major cause of visual impairment and irreversible blindness in children. The lesions in this disease spread along a linear path, and if not detected and intervened in the early stages, they can lead to serious consequences.
[0003] In recent years, with the popularization of medical imaging equipment and the continuous expansion of data scale, deep learning-based medical image analysis methods have been widely used in tasks such as disease detection and lesion region segmentation, thereby assisting doctors in diagnosis. Among these, lesion segmentation can characterize the spatial distribution of lesions at the pixel level, providing more accurate reference information for disease grading, progression assessment, and efficacy analysis. However, for retinopathy of prematurity (ROP) in premature infants, due to the current lack of sufficient high-quality labeled data to train high-performance intelligent segmentation systems, the application of artificial intelligence in this disease scenario is currently mainly limited to classification tasks such as staging or grading. Specifically, the scarcity of high-quality labeled data stems from the extremely high difficulty in segmenting the lesion path in this disease, mainly reflected in:
[0004] Compared with the imaging manifestations of common organs or masses, the imaging manifestations of retinopathy of prematurity often have characteristics such as blurred outlines, elongated shape, and complex structure. Therefore, the lesion path segmentation of retinopathy of prematurity is very easy to be wrong.
[0005] In actual annotation, the lesion pathways of retinopathy of prematurity (ROP) need to be delineated pixel by pixel. This process is extremely time-consuming and requires highly experienced physicians. However, these physicians are often burdened with heavy clinical workloads, making it virtually impossible to coordinate sufficient time for them to complete the extensive, pixel-by-pixel annotation work. Furthermore, because the boundaries of ROP lesions are subject to subjective judgment, discrepancies can easily arise between annotators, making it difficult to ensure annotation consistency.
[0006] In summary, due to the unique imaging characteristics of retinopathy of prematurity (ROP) and practical limitations, it is currently difficult to obtain sufficiently large and consistently labeled training and supervision data. Therefore, it is impossible to train an intelligent segmentation system specifically for ROP. Forcing training with limited data will only result in a system with low segmentation accuracy, weak generalization ability, and poor robustness, making it difficult to output truly valuable retinal lesion path segmentation maps for auxiliary diagnosis. This situation limits the clinical application of intelligent segmentation systems in ROP scenarios.
[0007] It should be noted that the background information presented here is only for illustrating relevant information about the present invention to aid in understanding the technical solution of the present invention, and does not imply that the relevant information is necessarily prior art. The relevant information was submitted and disclosed together with the present invention, and should not be considered prior art unless there is evidence that the relevant information was disclosed before the filing date of the present invention. Summary of the Invention
[0008] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing an auxiliary diagnostic system for retinopathy of prematurity.
[0009] The objective of this invention is achieved through the following technical solution:
[0010] According to a first aspect of the present invention, a method for constructing an auxiliary diagnostic system for retinopathy of prematurity is provided, comprising:
[0011] S1. Obtain the initial segmentation model based on the convolutional neural network.
[0012] S2. Obtain the training set, which includes multiple fundus images of premature infants and retinal lesion path segmentation atlas labels corresponding to each fundus image of premature infants. Each fundus image of a premature infant contains one or more actual lesion paths. The retinal lesion path segmentation atlas labels are obtained as follows: S21. Obtain multiple key pixels corresponding to the fundus images of premature infants. Each key pixel is an endpoint or inflection point on an actual lesion path. S22. According to a preset distance threshold, filter out all key pixels belonging to the same actual lesion path, and mark the two endpoints belonging to the same actual lesion path. Sort all key pixels corresponding to each actual lesion path by selecting the key pixel closest to the previous selected key pixel as the next key pixel, starting from any endpoint, until a key pixel at the other endpoint is selected. S23. For each key pixel on an actual lesion path, according to the sorting in S22, select multiple pixels between two adjacent key pixels in a preset manner to form the corresponding lesion path, and use all the formed lesion paths as labels for retinal lesion path segmentation atlas.
[0013] S3. Using the fundus image of the premature infant as input and the predicted retinal lesion path segmentation map as output, the initial segmentation model is trained to convergence using the acquired training set to obtain an auxiliary diagnostic system for retinopathy of prematurity. During the training process, the model parameters are updated with the goal of minimizing the binary cross-entropy loss between the predicted retinal lesion path segmentation map and the retinal lesion path segmentation map label.
[0014] This solution achieves at least the following beneficial technical effects: 1. By combining the acquisition of a small number of key pixels with the automatic selection of intermediate pixels, the lesion path of retinopathy of prematurity (ROP) is characterized, solving the problem of relying solely on doctors to delineate the lesion path pixel by pixel. This significantly reduces the difficulty of obtaining labeled data and ensures the accuracy and consistency of the labeling. 2. The retinal lesion path segmentation map labels can be directly used to train an intelligent segmentation model to obtain an auxiliary diagnostic system for ROP, enabling the system to automatically segment the lesion path of ROP.
[0015] Optionally, in S23, the preset method is as follows: S231, construct a path pixel pool for two adjacent key pixels belonging to the same actual lesion path. The path pixel pool is used to accommodate multiple pixels, including the two key pixels in the initial time step and multiple pixels selected continuously in subsequent time steps. S232, obtain multiple candidate pixels corresponding to the current reference pixel according to a preset search method. The current reference pixel is any key pixel in the path pixel pool in the initial time step and the pixel selected in the previous time step in subsequent time steps. S233, calculate the features of each candidate pixel corresponding to the current reference pixel and the average features of all pixels in the path pixel pool in the current time step. Select the candidate pixel with the smallest difference from the average features of all pixels in the path pixel pool in the current time step as the pixel selected in the current time step and add it to the path pixel pool. S234, repeat S232-S233 in subsequent time steps to select multiple pixels until all pixels in the path pixel pool form a corresponding lesion path.
[0016] Optionally, in S232, the preset search method is: taking the current reference pixel as the center, search for 8 spatially adjacent pixels in the horizontal, vertical and 45° and 135° diagonal directions of the center as multiple candidate pixels corresponding to the current reference pixel.
[0017] This scheme can achieve at least the following beneficial technical effects: by using the eight-neighbor method to determine candidate pixels, it can ensure the connectivity of the lesion path while making the expansion of the lesion path more directionally flexible, thereby better adapting to the morphological characteristics of the slender, curved or irregular distribution of retinopathy of prematurity lesions.
[0018] Optionally, in S233, the average feature of all pixels in the pixel pool of the current time step path is calculated as follows:
[0019]
[0020] in, This represents the average feature of all pixels in the pixel pool at the current time step. The feature representing the i-th pixel in the path pixel pool at the current time step. This represents the total number of pixels in the pixel pool at the current time step.
[0021] Optionally, in S233, except for the initial time step, the average feature of all pixels in the pixel pool of the current time step path is calculated as follows:
[0022]
[0023] in, This represents the average feature of all pixels in the pixel pool at the current time step. This represents the average feature of all pixels in the path pixel pool at the previous time step. The features representing the pixels selected in the previous time step, This represents the number of pixels in the pixel pool at the current time step.
[0024] This scheme can achieve at least the following beneficial technical effects: it can dynamically update the average feature value without retracing all pixels, thereby improving the efficiency of lesion path generation.
[0025] Optionally, in S233, the candidate pixel with the smallest average feature difference from all pixels in the current time step path pixel pool refers to:
[0026]
[0027] in, This represents the candidate pixel with the smallest average feature difference from all pixels in the current time step path pixel pool. The features representing the candidate pixels corresponding to the current reference pixel. It represents the average feature of all pixels in the pixel pool of the current time step path.
[0028] This scheme can achieve at least the following beneficial technical effects: it can make the generated lesion path extend along the pixel region that is closest to the overall features of the current path, so that the generated lesion path is more in line with the actual lesion path.
[0029] Optionally, in S3, the predicted retinal lesion path segmentation map consists of multiple predicted pixels, each with its own prediction probability. The binary cross-entropy loss is expressed as:
[0030]
[0031] in, Represents binary cross-entropy loss. This represents the total number of pixels involved in the loss calculation. Representing the The predicted probability of each predicted pixel Representing the The predicted pixel point corresponds to a true value of 0 or 1 in the retinal lesion path segmentation map label.
[0032] According to a first aspect of the present invention, a method for auxiliary diagnosis of retinopathy of prematurity is provided, comprising: T1, acquiring a fundus image of a premature infant; T2, processing the fundus image acquired in T1 using an auxiliary diagnostic system for retinopathy of prematurity constructed as described in the first aspect of the present invention to obtain a corresponding retinal lesion path segmentation map, and using the obtained retinal lesion path segmentation map for auxiliary diagnosis of retinopathy of prematurity.
[0033] Compared with existing technologies, the advantages of this invention are as follows: By automatically generating retinal lesion path segmentation atlas labels based on a small number of marked key pixels, this invention makes it easy to obtain training supervision data of stable quality and sufficient scale. Furthermore, based on this training supervision data, a preterm infant retinopathy of prematurity auxiliary diagnostic system with high segmentation accuracy, strong generalization ability, and good robustness can be trained, thereby enabling clinicians to obtain retinal lesion path segmentation atlases with real auxiliary diagnostic value. Attached Figure Description
[0034] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:
[0035] Figure 1 This is a schematic diagram illustrating the steps of constructing an auxiliary diagnostic system for retinopathy of prematurity according to an embodiment of the present invention;
[0036] Figure 2 This is a schematic diagram illustrating the steps of constructing retinal lesion path segmentation map tags in a method for constructing an auxiliary diagnostic system for retinopathy of prematurity according to an embodiment of the present invention.
[0037] Figure 3 This is a schematic diagram illustrating the steps for generating lesion paths between adjacent key pixels in a retinal lesion path segmentation atlas label according to an embodiment of the present invention. Detailed Implementation
[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.
[0039] As mentioned in the background section, due to the unique imaging characteristics of retinopathy of prematurity (ROP) and practical constraints, it is currently difficult to obtain training supervision data with stable annotation quality and sufficient scale. Therefore, it is impossible to train an intelligent segmentation system specifically for ROP. Forcing training with limited data will only result in a system with low segmentation accuracy, weak generalization ability, and poor robustness, making it difficult to output truly valuable retinal lesion path segmentation maps for auxiliary diagnosis. This situation limits the clinical application of intelligent segmentation systems in ROP scenarios.
[0040] To address the aforementioned problems, this invention provides a method for constructing an auxiliary diagnostic system for retinopathy of prematurity (ROP). The overall approach of this method is to automatically generate a large number of pixel-level segmentation labels consistent with the pathological distribution of ROP using only a small number of key pixels, without relying on professional physicians to annotate the pathology of ROP lesions pixel by pixel. This overcomes the difficulty in obtaining training and supervision data with stable annotation quality and sufficient scale, thereby enabling the construction of an auxiliary diagnostic system for ROP with good segmentation performance.
[0041] In summary, refer to Figure 1 The construction method provided by the present invention includes steps S1, S2 and S3, wherein each step performs the following operations:
[0042] In step S1, an initial segmentation model based on a convolutional neural network is obtained.
[0043] In step S2, a training set is obtained, which includes multiple fundus images of premature infants and retinal lesion path segmentation atlas labels corresponding to each fundus image of premature infants; wherein, each fundus image of premature infants contains one or more actual lesion paths, referring to... Figure 2 The retinal lesion path segmentation map labels are obtained as follows: S21. Obtain multiple key pixels corresponding to the fundus image of the premature infant. Each key pixel is an endpoint or inflection point on an actual lesion path. S22. According to a preset distance threshold, filter out all key pixels belonging to the same actual lesion path, and mark the two endpoints belonging to the same actual lesion path. Sort all key pixels corresponding to each actual lesion path. The sorting method is as follows: starting from the key pixel at any endpoint, select the key pixel closest to the previous selected key pixel as the next key pixel, until the key pixel at the other endpoint is selected. S23. For the key pixels on each actual lesion path, according to the sorting in S22, select multiple pixels between two adjacent key pixels one by one in a preset manner to form the corresponding lesion path, and use all the formed lesion paths as retinal lesion path segmentation map labels.
[0044] In step S3, the initial segmentation model is trained to convergence using the preterm infant fundus image as input and the predicted retinal lesion path segmentation map as output, and the acquired training set is used to obtain the preterm infant retinopathy of prematurity auxiliary diagnostic system. During the training process, the model parameters are updated with the goal of minimizing the binary cross-entropy loss between the predicted retinal lesion path segmentation map and the retinal lesion path segmentation map label.
[0045] As summarized in this invention, the present invention obtains training and supervision data (i.e., retinal lesion path segmentation atlas labels) for retinopathy of prematurity (ROP) by combining the acquisition of a small number of key pixels with the automatic selection of intermediate pixels. This solves the problem of relying solely on doctors to draw the lesion path pixel by pixel, significantly reducing the difficulty of acquiring labeled data while ensuring the accuracy and consistency of the labeling. Furthermore, the retinal lesion path segmentation atlas labels of this invention can be directly used to train intelligent segmentation models to obtain an auxiliary diagnostic system for ROP, enabling the system to automatically segment the lesion path of ROP, thus establishing a complete technical path from data labeling to clinical auxiliary diagnosis.
[0046] To better understand the present invention, each step will be described in detail below with reference to specific embodiments.
[0047] According to one embodiment of the present invention, the initial segmentation model in S1 is a segmentation structure based on a convolutional neural network (such as a U-Net structure or an improved structure thereof). The reason is that compared with a classification model that can only output diagnostic results, a segmentation model based on a convolutional neural network can output spatial segmentation results of lesions, thereby making the results interpretable and enhancing their reference value as auxiliary diagnostic reference material.
[0048] According to one embodiment of the present invention, the fundus images of premature infants in S2 can be color images or grayscale images, without limitation. Since these fundus images of premature infants may come from different imaging devices or data sources, their resolution and size may vary. Therefore, the fundus images of premature infants can be pre-compressed according to the degree of difference to reduce the computational load when generating lesion paths subsequently. For example, the nearest neighbor interpolation algorithm can be used to compress the fundus images of premature infants. The basic principle of the nearest neighbor interpolation algorithm is: when downsampling the target image, for each pixel position in the target image, the pixel value with the closest spatial distance in the original image is selected as the pixel value of that position, thereby completing the image size scaling. The advantage of using the nearest neighbor interpolation algorithm for compression is that the retinal lesion path segmentation map labels generated by the present invention are in binary form, and their pixel values only include 0 or 1. If bilinear interpolation or other weighted interpolation methods are used, intermediate values between 0 and 1 may be generated during resampling, thereby introducing unnecessary grayscale changes and affecting the stability and consistency of the label structure.
[0049] According to one embodiment of the present invention, the key pixels in S21 can be obtained by clinical experts based on fundus images of premature infants, and the obtained key pixels are stored in the form of two-dimensional coordinates. Each key pixel corresponds to a spatial location in the fundus image of premature infants. By associating the key pixels with the fundus image of premature infants, an input data structure for subsequent processing is formed.
[0050] According to one embodiment of the present invention, in S22, to filter out all key pixels belonging to the same actual lesion path, a distance threshold between key pixels needs to be set as a filtering criterion. When the spatial distance between two key pixels exceeds the preset distance threshold, they are considered to belong to different lesion regions. This distance threshold can be set based on the experience of clinical experts in the actual annotation process and the overall size of the fundus image of premature infants. For example, if the size of the input fundus image of premature infants is 512×512 pixels, then according to the experience of clinical experts, the corresponding distance threshold can be set to 1 / 10 of the image size, that is, about 50 pixels. In some special cases, such as if clinical experts believe that there is morphological ambiguity or distance abnormality between certain lesion regions during the annotation process, the distance threshold can also be appropriately adjusted according to the specific situation, thereby achieving flexible handling of special cases. This embodiment can ensure that the selection of key pixels along the same actual lesion path is more reasonable and can avoid misconnection between different lesion paths in subsequent processing.
[0051] According to one embodiment of the present invention, referring to Figure 3 The preset method in S23 is as follows: S231. Construct a path pixel pool for two adjacent key pixels belonging to the same actual lesion path. The path pixel pool is used to accommodate multiple pixels, including the two key pixels in the initial time step and multiple pixels selected continuously in subsequent time steps. S232. Obtain multiple candidate pixels corresponding to the current reference pixel according to a preset search method. The current reference pixel is any key pixel in the path pixel pool in the initial time step and the pixel selected in the previous time step in subsequent time steps. S233. Calculate the features of each candidate pixel corresponding to the current reference pixel and the average features of all pixels in the path pixel pool in the current time step. Select the candidate pixel with the smallest difference from the average features of all pixels in the path pixel pool in the current time step as the pixel selected in the current time step and add it to the path pixel pool. S234. Repeat S232-S233 in subsequent time steps to select multiple pixels until all pixels in the path pixel pool form a corresponding lesion path. It should be noted that this invention primarily relies on pixel feature similarity for path expansion during lesion path generation, rather than imposing fixed geometric constraints on the shape of the lesion path. For actual lesion paths that are curved or irregular in shape, the generation process can more accurately match the actual lesion path by increasing the density of key pixel annotations.
[0052] According to one embodiment of the present invention, in S232, the preset search method is to determine candidate pixels using an 8-connected neighborhood approach. Specifically, with the current reference pixel as the center, eight spatially adjacent pixels are searched in the horizontal, vertical, and 45° and 135° diagonal directions to serve as multiple candidate pixels corresponding to the current reference pixel. By using the 8-connected neighborhood approach to determine candidate pixels, the connectivity of the lesion path can be ensured while allowing for better directional flexibility in the expansion of the lesion path, thus better adapting to the elongated, curved, or irregularly distributed morphological characteristics of retinopathy of prematurity lesions.
[0053] According to an embodiment of the present invention, in S233, the average feature of all pixels in the current time step path pixel pool is calculated in the following manner:
[0054]
[0055] in, This represents the average feature of all pixels in the pixel pool at the current time step. The feature representing the i-th pixel in the path pixel pool at the current time step. This represents the total number of pixels in the pixel pool at the current time step.
[0056] According to an embodiment of the present invention, in S233, in addition to the initial time step, the average feature of all pixels in the current time step path pixel pool can also be calculated in the following way:
[0057]
[0058] in, This represents the average feature of all pixels in the pixel pool at the current time step. This represents the average feature of all pixels in the path pixel pool at the previous time step. The features representing the pixels selected in the previous time step, This represents the number of pixels in the pixel pool at the current time step. Using the calculation method in this embodiment, subsequent time steps can dynamically update the average feature value without retracing all pixels, thereby improving the efficiency of lesion path generation.
[0059] According to an embodiment of the present invention, in S233, the candidate pixel with the smallest average feature difference from all pixels in the current time step path pixel pool refers to:
[0060]
[0061] in, This represents the candidate pixel with the smallest average feature difference from all pixels in the current time step path pixel pool. The features representing the candidate pixels corresponding to the current reference pixel. This represents the average feature of all pixels in the pixel pool of the current time step path. The method provided in this embodiment allows the lesion path to extend along the pixel region that most closely matches the average feature of all pixels in the pixel pool of the current time step path, thereby making the generated lesion path more closely resemble the actual lesion path.
[0062] The aforementioned embodiments solve the problem of obtaining large-scale, high-quality retinal lesion path segmentation map labels in the training set. Next, the initial segmentation model is trained until convergence according to S3 to construct an auxiliary diagnostic system for retinopathy of prematurity with practical clinical application value.
[0063] According to an embodiment of the present invention, in S3, the predicted retinal lesion path segmentation map is composed of multiple predicted pixels, each of which has its own prediction probability. The binary cross-entropy loss is expressed as:
[0064]
[0065] in, Represents binary cross-entropy loss. This represents the total number of pixels involved in the loss calculation. Representing the The predicted probability of each predicted pixel Representing the The predicted pixel point corresponds to a true value of 0 or 1 in the retinal lesion path segmentation map label.
[0066] The preterm infant retinopathy of prematurity auxiliary diagnostic system constructed based on the above embodiments of the present invention can output a retinal lesion path segmentation map based on the fundus images of preterm infants. This map directly corresponds to the location of the lesion in the fundus images of preterm infants. Therefore, compared with auxiliary diagnostic systems that can only output disease categories, the output results of the system constructed according to the scheme of the present invention have stronger interpretability.
[0067] According to one embodiment of the present invention, the present invention also provides an auxiliary diagnostic method for retinopathy of prematurity (ROP), comprising: T1, acquiring a fundus image of a preterm infant; T2, processing the fundus image acquired in T1 using an auxiliary diagnostic system for ROP constructed according to the method described in the present invention to obtain a corresponding retinal lesion path segmentation map, and using the obtained retinal lesion path segmentation map for auxiliary diagnosis of ROP. Thus, the present invention removes the clinical application limitations of intelligent segmentation systems in the ROP scenario, fully exploring the auxiliary diagnostic value of intelligent segmentation systems in this scenario.
[0068] In summary, this invention simplifies the acquisition of stable and sufficiently large training and supervision data by automatically generating retinal lesion path segmentation atlas labels based on a small number of marked key pixels. Furthermore, based on this training and supervision data, a preterm infant retinopathy of prematurity (ROP) auxiliary diagnostic system with high segmentation accuracy, strong generalization ability, and good robustness can be trained, thereby enabling clinicians to obtain truly valuable retinal lesion path segmentation atlases.
[0069] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.
[0070] This invention can be a computer device or a computer program product. A computer program product mainly refers to a software product that implements this solution through a computer program.
[0071] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. A method for constructing an auxiliary diagnostic system for retinopathy of prematurity (ROP) in premature infants, wherein the auxiliary diagnostic system for ROP in premature infants is used to output a retinal lesion path segmentation map based on fundus images of premature infants, characterized in that... The construction method includes: S1. Obtain the initial segmentation model based on a convolutional neural network; S2. Obtain the training set, which includes multiple fundus images of premature infants and retinal lesion path segmentation atlas labels corresponding to each fundus image of premature infants; wherein, each fundus image of premature infants contains one or more actual lesion paths, and the retinal lesion path segmentation atlas labels are obtained in the following way: S21. Obtain multiple key pixels corresponding to the fundus image of the premature infant. Each key pixel is an endpoint or inflection point on an actual lesion path. S22. According to the preset distance threshold, filter out all key pixels belonging to the same actual lesion path, mark the two endpoints on the same actual lesion path, and sort all key pixels corresponding to each actual lesion path. The sorting method is as follows: starting from the key pixel at any endpoint, select the key pixel closest to the previous selected key pixel as the next key pixel, until the key pixel at the other endpoint is selected. S23. For each key pixel on an actual lesion path, according to the sorting in S22, select multiple pixels between two adjacent key pixels in a preset manner to form the corresponding lesion path, and use all the formed lesion paths as retinal lesion path segmentation map labels. S3. Using the fundus image of the premature infant as input and the predicted retinal lesion path segmentation map as output, the initial segmentation model is trained to convergence using the acquired training set to obtain an auxiliary diagnostic system for retinopathy of prematurity. During the training process, the model parameters are updated with the goal of minimizing the binary cross-entropy loss between the predicted retinal lesion path segmentation map and the retinal lesion path segmentation map label.
2. The method for constructing an auxiliary diagnostic system for retinopathy of prematurity according to claim 1, characterized in that, In S23, the default method is: S231. Construct a path pixel pool for two adjacent key pixels belonging to the same actual lesion path. The path pixel pool is used to accommodate multiple pixels, including the two key pixels at the initial time step and multiple pixels selected continuously in subsequent time steps. S232. Obtain multiple candidate pixels corresponding to the current reference pixel according to a preset search method. The current reference pixel is any key pixel in the path pixel pool in the initial time step and the pixel selected in the previous time step in subsequent time steps. S233. Calculate the features of each candidate pixel corresponding to the current reference pixel and the average features of all pixels in the current time step path pixel pool. Select the candidate pixel with the smallest difference from the average features of all pixels in the current time step path pixel pool as the pixel selected for the current time step and add it to the path pixel pool. S234. In subsequent time steps, repeat S232-S233 to select multiple pixels until all pixels in the path pixel pool form a corresponding lesion path.
3. The method for constructing an auxiliary diagnostic system for retinopathy of prematurity according to claim 2, characterized in that, In S232, the preset search method is as follows: taking the current reference pixel as the center, search for 8 spatially adjacent pixels in the horizontal, vertical and 45° and 135° diagonal directions of the center as multiple candidate pixels corresponding to the current reference pixel.
4. The method for constructing an auxiliary diagnostic system for retinopathy of prematurity according to claim 2, characterized in that, In S233, the average feature of all pixels in the pixel pool of the current time step path is calculated as follows: in, This represents the average feature of all pixels in the pixel pool at the current time step. The feature representing the i-th pixel in the path pixel pool at the current time step. This represents the total number of pixels in the pixel pool at the current time step.
5. The method for constructing an auxiliary diagnostic system for retinopathy of prematurity according to claim 2, characterized in that, In S233, except for the initial time step, the average feature of all pixels in the pixel pool of the current time step path is calculated in the following way: in, This represents the average feature of all pixels in the pixel pool at the current time step. This represents the average feature of all pixels in the path pixel pool at the previous time step. The features representing the pixels selected in the previous time step, This represents the number of pixels in the pixel pool at the current time step.
6. The method for constructing an auxiliary diagnostic system for retinopathy of prematurity according to claim 2, characterized in that, In S233, the candidate pixel with the smallest average feature difference from all pixels in the current time step path pixel pool is: in, This represents the candidate pixel with the smallest average feature difference from all pixels in the current time step path pixel pool. The features representing the candidate pixels corresponding to the current reference pixel. It represents the average feature of all pixels in the pixel pool of the current time step path.
7. The method for constructing an auxiliary diagnostic system for retinopathy of prematurity according to claim 1, characterized in that, In S3, the predicted retinal lesion path segmentation map consists of multiple predicted pixels, each with its own prediction probability. The binary cross-entropy loss is expressed as: in, Represents binary cross-entropy loss. This represents the total number of pixels involved in the loss calculation. Representing the The predicted probability of each predicted pixel Representing the The predicted pixel point corresponds to a true value of 0 or 1 in the retinal lesion path segmentation map label.
8. A method for auxiliary diagnosis of retinopathy of prematurity, characterized in that, include: T1. Obtain fundus images of premature infants; T2. The fundus image obtained in T1 is processed using the preterm infant retinopathy auxiliary diagnostic system constructed by any one of the methods described in claims 1-7 to obtain the corresponding retinal lesion path segmentation map, and the obtained retinal lesion path segmentation map is used for auxiliary diagnosis of preterm infant retinopathy.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-8.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-8.