Method for training a recognition model for recognizing chronic mountain sickness and related products

Abstract

This application discloses a method and related products for training a model for recognizing chronic altitude sickness (CASS). The method includes: acquiring fundus images and classifying and labeling the fundus images for CASS to form a training set; using a preprocessing network to extract a first specific feature related to CASS recognition from the training set to obtain a feature vector containing the first specific feature; using a backbone network to extract a second specific feature related to CASS recognition from the feature vector to recognize CASS and obtain a predicted recognition result; and calculating a loss function based on the CASS classification and the predicted recognition result to train the CASS recognition model. Using the solution of this application, the efficiency and accuracy of CASS recognition can be improved.

Description

Technical Field

[0001] This application generally relates to the field of artificial intelligence technology. More specifically, this application relates to a method, apparatus, and computer-readable storage medium for training a recognition model for identifying chronic altitude sickness. Furthermore, this application also relates to an apparatus and computer-readable storage medium for identifying chronic altitude sickness. Background Technology

[0002] Chronic mountain sickness (CMS) is a clinical syndrome caused by prolonged exposure to altitudes above 2500 meters, leading to a gradual loss of physiological adaptation to hypoxic environments. Common clinical symptoms of CMS include headache, dizziness, shortness of breath, palpitations, sleep disturbances, fatigue, and localized cyanosis. A characteristic feature of the disease is that symptoms disappear after leaving the high-altitude area but reappear upon returning. In addition, excessive hemocytosis (Hb ≥ 19 g / dL for women, Hb ≥ 21 g / dL for men), severe hypoxemia, or potentially significant pulmonary hypertension and heart failure are common signs in patients with chronic mountain sickness.

[0003] Currently, the diagnosis of CMS still mainly relies on doctors to take medical history, conduct laboratory tests, or use the Qinghai scoring and statistical system. This results in low diagnostic efficiency and inaccurate diagnostic results, leading to CMS patients in high-altitude areas not receiving timely diagnosis and treatment.

[0004] In view of this, there is an urgent need to provide a scheme for training a model for identifying chronic mountain sickness in order to improve the efficiency and accuracy of the identification results. Summary of the Invention

[0005] In order to at least address one or more of the technical problems mentioned above, this application proposes a scheme for training an identification model for recognizing chronic mountain sickness in several aspects.

[0006] In a first aspect, this application provides a method for training a recognition model for identifying chronic altitude sickness (CASS), wherein the recognition model includes a preprocessing network and a backbone network, and the method includes: acquiring fundus images and performing CASS classification annotation on the fundus images to form a training set; using the preprocessing network to extract a first specific feature related to CASS identification from the training set to obtain a feature vector containing the first specific feature; using the backbone network to extract a second specific feature related to CASS identification from the feature vector to identify CASS and obtain a predicted recognition result; and calculating a loss function based on the CASS classification annotation and the predicted recognition result to train the recognition model for CASS identification.

[0007] In one embodiment, the method further includes: classifying and labeling the fundus image for chronic mountain sickness based on the scoring statistics of chronic mountain sickness.

[0008] In another embodiment, the first specific feature includes at least the fundus edema and discoloration feature, and the second specific feature includes at least the fundus vascular hemorrhage and dilation feature.

[0009] In yet another embodiment, before using the backbone network to extract a second specific feature from the feature vector that is associated with identifying chronic mountain sickness, the method further includes: sharpening the feature vector to obtain a sharpened feature vector.

[0010] In yet another embodiment, sharpening the feature vector includes: performing an edge detection operation on the feature vector to obtain edge detection information; extracting enhancement factors related to chronic mountain sickness based on the edge detection information; and sharpening the feature vector based on the enhancement factors.

[0011] In yet another embodiment, the enhancement factor includes edge response range and local contrast.

[0012] In another embodiment, the backbone network includes multiple convolutional layers, each convolutional layer including cascaded convolutional kernels, and the extraction of a second specific feature from the feature vector related to the identification of chronic altitude sickness using the backbone network to identify chronic altitude sickness and obtain a predicted identification result includes: extracting a second specific feature from the feature vector related to the identification of chronic altitude sickness using the cascaded convolutional kernels of each convolutional layer in the backbone network to identify chronic altitude sickness and obtain a predicted identification result.

[0013] In yet another embodiment, the preprocessing network includes a convolutional neural network and the backbone network includes a ResNet50 network.

[0014] In another embodiment, the method of calculating a loss function based on the chronic altitude sickness classification label and the predicted recognition result to train the recognition model for chronic altitude sickness includes: calculating a first loss function and a second loss function based on the chronic altitude sickness classification label and the predicted recognition result; and performing a weighted summation operation on the first loss function and the second loss function to obtain a total loss function to train the recognition model for chronic altitude sickness.

[0015] In a second aspect, this application provides an apparatus for training a recognition model for identifying chronic altitude sickness, comprising: a processor; and a memory storing program instructions for training the recognition model for identifying chronic altitude sickness, wherein when the program instructions are executed by the processor, the apparatus implements the various embodiments of the first aspect described above.

[0016] In a third aspect, this application provides an apparatus for identifying chronic altitude sickness, comprising: a processor; and a memory storing program instructions for identifying chronic altitude sickness, wherein when the program instructions are executed by the processor, the apparatus causes to perform the following operations: acquiring a fundus image to be identified; inputting the fundus image to be identified into an identification model trained according to the plurality of embodiments in the first aspect for identification, so as to obtain an identification result for chronic altitude sickness.

[0017] In a fourth aspect, this application provides a computer-readable storage medium having stored thereon computer-readable instructions for training a recognition model for recognizing chronic mountain sickness and for recognizing chronic mountain sickness, wherein the computer-readable instructions, when executed by one or more processors, implement the operations performed by the various embodiments of the first aspect or by the apparatus described in the third aspect.

[0018] Using the above-described scheme for training a model to identify chronic altitude sickness (CAS), this embodiment first employs a preprocessing network to extract first specific features related to CAS identification based on a training set formed by CAS classification and annotation of fundus images. This extracts features reflecting early CAS symptoms, resulting in a feature vector. Next, the feature vector is input into the backbone network to extract second specific features related to CAS identification, extracting more obvious features reflecting CAS, thus obtaining a predicted identification result. A loss function is calculated based on the predicted identification result and CAS classification annotations (i.e., ground truth labels) to train the identification model. Based on this, this embodiment improves the accuracy of the identification model in identifying different stages of CAS by transferring features of early CAS symptoms to the backbone network before feature extraction, resulting in more accurate identification results. By using a fast, low-cost, and non-invasive method based on fundus images, combined with a deep learning-based identification model, CAS identification efficiency is significantly improved.

[0019] Furthermore, this embodiment of the application also sharpens the feature vectors to improve the distinction between the features and the surrounding ocular tissue, thereby enhancing the recognition accuracy of lesion features and improving the recognition model's recognition precision. Even further, this embodiment of the application uses cascaded convolutional kernels on the backbone network to extract more detailed features for more accurate identification of chronic altitude sickness. In addition, this embodiment of the application further improves the accuracy of the recognition model by calculating a dual loss function (i.e., a first loss function and a second loss function). Attached Figure Description

[0020] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts, wherein:

[0021] Figure 1 This is an exemplary flowchart illustrating a method for training an identification model for recognizing chronic mountain sickness according to an embodiment of this application;

[0022] Figure 2 This is an exemplary schematic diagram showing fundus images of several different states of chronic mountain sickness according to embodiments of this application;

[0023] Figure 3 This is an exemplary flowchart illustrating a data cleaning operation according to an embodiment of this application;

[0024] Figure 4 This is an exemplary schematic diagram illustrating the overall process of training a recognition model for identifying chronic mountain sickness according to an embodiment of this application;

[0025] Figure 5 This is an exemplary structural block diagram illustrating a device for identifying chronic mountain sickness according to an embodiment of this application; and

[0026] Figure 6 This is an exemplary structural block diagram illustrating an apparatus for training a recognition model for identifying chronic mountain sickness according to an embodiment of this application. Detailed Implementation

[0027] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the embodiments described in this specification are only some embodiments provided by this application for the purpose of facilitating a clear understanding of the solutions and complying with legal requirements, and are not all embodiments that can be implemented in this application. Based on the embodiments disclosed in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] As described in the background section above, current diagnostic methods for CMS still primarily rely on physicians' history taking, laboratory tests, or the Qinghai scoring and statistical system. The aforementioned Qinghai scoring and statistical system includes various CMS-related symptoms and their severity levels, with corresponding scores, as detailed below:

[0029] Dyspnea and / or palpitations: 0 No dyspnea / palpitations; 1 Mild dyspnea / palpitations; 2 Moderate dyspnea / palpitations; 3 Severe dyspnea / palpitations.

[0030] Sleep disorders: 0. Normal sleep; 1. Sleep is not as good as normal; 2. Staying awake for a long time and poor sleep; 3. Difficulty falling asleep.

[0031] Cyanosis: 0 No cyanosis; 1 Mild cyanosis; 2 Moderate cyanosis; 3 Severe cyanosis.

[0032] Venous dilation: 0 No venous dilation; 1 Mild venous dilation; 2 Moderate venous dilation; 3 Severe venous dilation.

[0033] Localized sensory abnormalities: 0 No localized sensory abnormalities; 1 Mild localized sensory abnormalities; 2 Moderate localized sensory abnormalities; 3 Severe localized sensory abnormalities.

[0034] Headache: 0 No headache; 1 Mild headache; 2 Moderate headache; 3 Severe headache.

[0035] Tinnitus: 0 No tinnitus; 1 Mild tinnitus; 2 Moderate tinnitus; 3 Severe tinnitus.

[0036] Hemoglobin concentration (Hb): For males, a score of 0 is given if the concentration is >18 g / dl or <21 g / dl, and 3 if the concentration is ≥21 g / dl. For females, a score of 0 is given if the concentration is >16 g / dl or <19 g / dl, and 3 if the concentration is ≥19 g / dl.

[0037] In practice, CMS is determined by summing symptom scores and Hb scores, and the total score is used to classify CMS as follows: no CMS: score 0-5; mild CMS: score 6-10; moderate CMS: score 11-14; severe CMS: score ≥15. However, existing diagnostic methods are inefficient and inaccurate, leading to delayed diagnosis and treatment for CMS patients in high-altitude areas.

[0038] Currently, deep learning-based fundus image detection has made significant progress in the field of medical image analysis, providing a wealth of physiological information about patients. Existing research has shown that fundus images can predict hemoglobin concentration. This study utilized deep learning technology to predict hemoglobin concentration and screen for anemia through fundus photography, providing an innovative method for rapid, non-invasive, and widespread anemia screening. As mentioned earlier, the main characteristic of CMS is excessive erythrocyte proliferation, and current research has confirmed that fundus images contain hemoglobin information; therefore, diagnosing CMS through fundus images is theoretically feasible.

[0039] Based on this, this application proposes a scheme for training a recognition model for chronic altitude sickness (CASS) and for recognizing CASS based on the trained model. By classifying and labeling fundus images for CASS, a preprocessing network extracts first specific features reflecting early CASS symptoms. These features are then transferred to the backbone network to further extract second specific features, obtaining predicted recognition results. This improves the accuracy of the recognition model in identifying CASS at different stages, resulting in more accurate recognition results. Furthermore, the recognition model is trained using a loss function calculated based on the predicted recognition results and CASS classification labels. By utilizing the trained recognition model to identify CASS, recognition efficiency and accuracy can be improved, enabling CASS patients to receive timely diagnosis and treatment.

[0040] The following is in conjunction with the appendix Figures 1-6 The specific implementation methods of this application will be described in detail below.

[0041] Figure 1 This is an exemplary flowchart illustrating a method 100 for training a recognition model for identifying chronic altitude sickness according to an embodiment of this application. In this embodiment, the recognition model for identifying chronic altitude sickness includes a preprocessing network and a backbone network, and the method 100 for training the recognition model is as follows: Figure 1As shown, in step 101, fundus images are acquired and labeled for chronic altitude sickness classification to form a training set. In one embodiment, the aforementioned fundus images can be acquired using, for example, a fundus camera. In some embodiments, based on the acquired fundus images, data cleaning can first be performed to select high-quality fundus images. Specifically, the aforementioned data cleaning aims to remove fundus images containing noisy points. In other embodiments, textual data such as sociological indicators and detection indicators of the person to be identified can also be acquired, which can be, for example, indicators within the aforementioned sea-based statistical system. The aforementioned sociological indicators include, for example, age, gender, medical history, etc., and the medical history can include, for example, cyanosis, sleep disorders, palpitations, etc. The aforementioned detection indicators include, for example, hemoglobin concentration, etc. In one implementation scenario, for the aforementioned sociological indicators and detection indicators, invalid data can also be removed by data cleaning, and valid data can be matched with fundus images. That is, fundus images of the same person to be identified are bound with valid sociological indicators and detection indicators, and after manual verification, a high-quality dataset is finally obtained.

[0042] In one embodiment, chronic altitude sickness (CMS) classification and annotation can be performed on the acquired fundus images to form a training set. In one implementation scenario, CMS classification and annotation can be performed on the fundus images based on the CMS scoring statistics. These scoring statistics can be obtained by statistically analyzing the aforementioned sociological and detection indicators using the Qinghai scoring statistics system. In this scenario, the fundus images are labeled as having no CMS, mild CMS, moderate CMS, or severe CMS by calculating the scoring statistics of the person to be identified. As an example, assuming the scoring statistics of the person to be identified is 4 points, which falls within the range of 0-5 points, the fundus images are labeled as having no CMS; assuming the scoring statistics of the person to be identified is 16 points, which falls within the range of 15 points, the fundus images are labeled as having severe CMS. Similarly, when the scoring statistics of the person to be identified is within the range of 6-10 points, the fundus images are labeled as having mild CMS; when the scoring statistics of the person to be identified is within the range of 11-14 points, the fundus images are labeled as having moderate CMS, thus forming a training set.

[0043] Next, in step 102, a preprocessing network is used to extract a first specific feature related to the identification of chronic mountain sickness (CMS) from the training set to obtain a feature vector containing the first specific feature. In one embodiment, the aforementioned preprocessing network may be, for example, a convolutional neural network (“CNN network”), and the aforementioned first specific feature may include at least the fundus edema discoloration feature. It is understood that fundus edema is an early symptom of CMS, which is relatively mild and not easily extracted during learning. Therefore, this embodiment of the application further marks the fundus images in the training set with edema discoloration, and then extracts the first specific feature through a preprocessing network (e.g., a CNN network) to obtain a feature vector containing the fundus edema discoloration feature. Through this fundus edema discoloration feature, accurate identification of early CMS can be achieved.

[0044] After obtaining the feature vector containing the first specific feature, in step 103, the backbone network is used to extract the second specific feature related to the identification of chronic altitude sickness (CMS) from the feature vector to identify CMS and obtain a predicted identification result. That is, the aforementioned feature vector is used as the input of the backbone network and combined with the backbone network for transfer learning. In one embodiment, the aforementioned backbone network can be, for example, a ResNet50 network, and the aforementioned second specific feature can include at least the fundus vascular hemorrhage and dilation feature. It should be understood that ResNet50 is a technology that uses residual learning to alleviate the problem of decreased learning efficiency and ineffective improvement of accuracy caused by the increasing number of layers in deep learning models, thereby better completing the identification task of this application embodiment. As mentioned above, one of the main features of CMS is erythrocyte proliferation, which is manifested in fundus images as obvious color differences and capillary dilation, even accompanied by microbleeds. Therefore, in order to extract more accurate fundus vascular hemorrhage and dilation features, this application embodiment proposes to sharpen the feature vector before using the backbone network to extract the second specific feature related to the identification of CMS to obtain a sharpened feature vector. Based on this, the edges of capillaries and hemorrhages in fundus images can be enhanced, and the noise level in fundus images can be reduced, so as to reduce the interference of irrelevant features and improve the clarity of capillaries and hemorrhages.

[0045] In one embodiment, edge detection information is first obtained by performing an edge detection operation on the feature vector. Then, enhancement factors related to chronic altitude sickness are extracted based on the edge detection information, and the feature vector is sharpened based on these enhancement factors. In one implementation scenario, the aforementioned enhancement factors may include, but are not limited to, edge response range and local contrast. Specifically, edge detection information can be obtained by performing an edge detection operation on the feature vector using, for example, the Laplacian edge detection operator. The Laplacian edge detection operator is a second-derivative-based edge detection algorithm that detects edges by calculating the second derivative of the pixel values ​​in the feature vector. Specifically, edge detection can be performed using the following formula:

[0046]

[0047] Where L(z,y) represents the obtained edge detection information, and f(x,y) represents the value of the current feature vector.

[0048] Based on the obtained edge detection information, the edge response range can be adjusted. A larger edge response range can highlight edges to address areas with prominent edges, while a smaller edge response range avoids over-enhancement and is particularly advantageous for areas with less detail or blurred edges. In this embodiment, it shows good results for fundus images accompanied by edema. Furthermore, based on the obtained edge detection information, local contrast can also be acquired. Lower local contrast requires stronger sharpening to improve the clarity of the fundus image, while higher local contrast requires less sharpening.

[0049] Based on the sharpened feature vector, a backbone network (e.g., ResNet50 network) can be used to extract a second specific feature (e.g., fundus vascular hemorrhage and dilation feature) related to the identification of chronic altitude sickness (CMS) from the feature vector to identify CMS and obtain a predicted identification result. In one embodiment, the backbone network may include multiple convolutional layers, and each convolutional layer includes cascaded convolutional kernels. In one implementation scenario, the second specific feature related to the identification of CMS is extracted from the feature vector by using the cascaded convolutional kernels of each convolutional layer in the backbone network to identify CMS and obtain a predicted identification result. As mentioned above, one of the main features of CMS is erythrocyte proliferation, which manifests in fundus images as significant color differences and capillary dilation, even accompanied by microbleeds. Therefore, in order to extract the fundus vascular hemorrhage and dilation feature, this embodiment uses cascaded convolutional kernels, employing continuous convolution, so that small convolutional kernels can achieve the extraction effect of large convolutional kernels. Furthermore, continuous convolution through multiple convolutional layers can make the extracted features more detailed, thereby obtaining a more accurate identification result.

[0050] Further, in step 104, a loss function is calculated based on the chronic altitude sickness (CAS) classification labeling and prediction recognition results to train the CAS recognition model. In one embodiment, a first loss function and a second loss function can be calculated based on the CAS classification labeling and prediction recognition results, and then a weighted summation operation is performed on the first and second loss functions to obtain a total loss function for training the CAS recognition model. In some embodiments, the aforementioned first and second loss functions can be, for example, a sigmoid cross-entropy loss function and a binary cross-entropy loss function. For example, in one implementation scenario, the total loss function is denoted as L. total The total loss function can then be calculated using the following formula:

[0051] L total =w1L sig +w2L bc (1)

[0052] Among them, L sig L represents the sigmoid cross-entropy loss function. bc Let q1 and q2 represent the binary cross-entropy loss function, and q1 and q2 represent the weighting coefficients. That is, the embodiments of this application employ a dual loss function to further improve the accuracy of the recognition model.

[0053] As described above, this embodiment of the application forms a training set by classifying and annotating fundus images for chronic mountain sickness (CMS). First, a preprocessing network is used to extract first specific features reflecting early CMS symptoms to obtain feature vectors. These feature vectors are then used as input to the backbone network, and transfer learning is used to obtain a second specific feature reflecting more obvious CMS symptoms to obtain a predicted recognition result. Based on this, the recognition model can be improved to accurately identify different stages of CMS, thus making the recognition results obtained based on the model more accurate. Furthermore, this embodiment of the application sharpens the feature vectors input to the backbone network to enhance the clarity of features such as edema, capillaries, and hemorrhages, improving the recognition accuracy of lesion features and thus increasing the recognition model's accuracy. Even further, this embodiment of the application sets cascaded convolutional kernels in the backbone network to extract more detailed features for more accurate identification of CMS. In addition, this embodiment of the application calculates a loss function to further improve the accuracy of the recognition model.

[0054] Figure 2 This is an exemplary schematic diagram illustrating fundus images of several different states of chronic mountain sickness according to embodiments of this application. Figure 2Figures (a), (b), (c), and (d) show, in order, a normal fundus image, a fundus image with edema and congestion, a fundus image with retinal hemorrhage due to circulatory disturbance, and a fundus image with optic disc discoloration and vasodilation. As mentioned earlier, the aforementioned fundus images of the subject to be identified can be acquired, for example, using a fundus camera. Based on the acquired fundus images, CMS classification (including no CMS, mild CMS, moderate CMS, or severe CMS) can be performed using the scoring statistics obtained from the sociological and detection indicators of the subject to be identified through the aforementioned Qinghai scoring and statistical system, to form a training set. For more details on CMS classification, please refer to [reference needed]. Figure 1 The descriptions in the document are not repeated here.

[0055] In one embodiment, to extract fundus edema discoloration features reflecting early symptoms of CMS, this application embodiment further marks the fundus images in the training set with edema discoloration, for example, by... Figure 2 In the fundus images shown in (b) and (d), edema discoloration is marked so that subsequent first-specific feature extraction can be performed through a preprocessing network (e.g., a CNN network) to obtain a feature vector containing fundus edema discoloration features. Additionally, for Figure 2 Figures (b), (c), and (d) show congestion. These areas can be sharpened to enhance the clarity of capillaries and bleeding points, improving the recognition of lesion features and thus increasing the accuracy of the recognition model. The sharpened feature vectors can then be extracted using multi-layered cascaded convolutional kernels in the backbone network to obtain more detailed features and ultimately more accurate recognition results.

[0056] In one implementation scenario, by collecting fundus images of the subject to be identified, along with their sociological and detection metrics, a high-quality dataset can be obtained through data cleaning, matching, and manual verification. For example... Figure 3 As shown.

[0057] Figure 3 This is an exemplary flowchart illustrating data cleaning according to embodiments of this application. Figure 3As shown, in step 301, fundus images, sociological indicators, and detection indicators of the subject to be identified are collected. Sociological indicators include, for example, age, gender, and medical history; detection indicators include, for example, hemoglobin concentration. Specific indicators can be found in the aforementioned Qinghai scoring and statistical system. Next, in step 302, invalid data is cleaned from the sociological and detection indicators to obtain valid data. In step 303, the valid data is marked, for example, by recording the valid data for each subject to be identified. Further, in step 304, the sociological and detection indicators are matched with the fundus images. That is, the fundus images of the same subject to be identified are bound to valid sociological and detection indicators. After data matching, in step 305, manual verification is performed, for example, deleting fundus images containing noisy points, to obtain a high-quality dataset in step 306, thus completing data cleaning.

[0058] Figure 4 This is an exemplary schematic diagram illustrating the overall process of training a recognition model for identifying chronic mountain sickness according to an embodiment of this application. Figure 4 As shown, the recognition model 401 includes a preprocessing network 402 and a backbone network 403. When training the recognition model 401, fundus images 403 are acquired, cleaned, and CMS classification and annotation are performed to form a training set. Based on this training set, a first specific feature can be extracted via the preprocessing network 402 to obtain a feature vector containing the first specific feature. In one embodiment, the aforementioned preprocessing network can be, for example, a CNN network, and the aforementioned first specific feature includes at least fundus edema discoloration features. In the implementation scenario, fundus edema discoloration features are extracted by labeling the edema discoloration in the fundus images in the training set to achieve accurate early-stage CMS recognition. Next, the aforementioned feature vector is input into the backbone network 403 to extract a second specific feature to identify chronic altitude sickness and obtain a predicted recognition result. In one embodiment, the aforementioned backbone network can be, for example, a ResNet50 network, and the aforementioned second specific feature includes at least fundus vascular hemorrhage and dilation features.

[0059] In one implementation scenario, before using the backbone network to extract the second specific feature from the feature vector, the feature vector can be sharpened to obtain a sharpened feature vector. Specifically, edge detection is performed on the feature vector to obtain edge detection information, and enhancement factors such as edge response range and local contrast are extracted based on the edge detection information to achieve sharpening. For more details on the aforementioned sharpening process, please refer to [link to relevant documentation]. Figure 1 The descriptions in the document are not repeated here.

[0060] In one embodiment, the backbone network 403 may include multiple convolutional layers, and each convolutional layer includes cascaded convolutional kernels. For example, the figure exemplifies four convolutional layers 403-1, each of which includes convolutional kernels with sizes of 1*1, 3*3, and 1*1, forming cascaded convolutional kernels. The numbers following each convolutional kernel (e.g., 64, 128, 512, 256, 1024, and 2048) represent the size of the extracted features. By using the aforementioned cascaded convolutional kernels, the extraction effect of large convolutional kernels can be achieved by continuously convolving smaller convolutional kernels. Furthermore, continuous convolution through multiple convolutional layers can make the extracted features more detailed, resulting in more accurate prediction and recognition results. Further, by calculating the loss function 404 based on the obtained CMS classification and prediction results, the recognition model can be trained to obtain a trained recognition model 405. Preferably, in this embodiment, the loss function 404 is calculated by, for example, a sigmoid cross-entropy loss function and a binary cross-entropy loss function, and based on the above formula (1). Based on the trained recognition model 405, the recognition result of chronic altitude sickness can be directly obtained. Based on the aforementioned trained recognition model, labeled samples that were not involved in the training are used as test samples to verify the model performance. Experiments show that the AUC (area under the ROC curve) of the recognition model can reach 0.84, proving that the recognition model trained by this embodiment has practical feasibility. In the implementation scenario, chronic altitude sickness can be detected by inputting a fundus image.

[0061] Figure 5 This is an exemplary structural block diagram illustrating a device 500 for identifying chronic altitude sickness according to an embodiment of this application. Figure 5 As shown, the device 500 of this application may include a processor 501 and a memory 502, wherein the processor 501 and the memory 502 communicate via a bus. The memory 502 stores program instructions for identifying chronic altitude sickness (CMS). When the program instructions are executed by the processor 501, the device 500 performs the following operations: acquiring a fundus image to be identified and inputting the fundus image to be identified into a trained identification model for identification, thereby obtaining a result indicating CMS. For example, obtaining a result indicating no CMS, mild CMS, moderate CMS, or severe CMS. Based on this, by establishing a CMS identification model based on machine learning fundus image analysis, medical personnel can be helped to identify CMS more efficiently and accurately, thus contributing to the early screening and control of CMS.

[0062] Figure 6This is an exemplary structural block diagram illustrating a device 600 for training a recognition model for chronic mountain sickness according to an embodiment of this application. It will be understood that the device implementing the solution of this application may be a single device (e.g., a computing device) or a multifunctional device including various peripheral devices.

[0063] like Figure 6 As shown, the device of this application may include a central processing unit (“CPU”) 611, which may be a general-purpose CPU, a special-purpose CPU, or other information processing and program execution unit. Furthermore, the device 600 may also include a mass storage 612 and a read-only memory (“ROM”) 613, wherein the mass storage 612 may be configured to store various types of data, including various fundus images, algorithm data, intermediate results, and various programs required to run the device 600. The ROM 613 may be configured to store power-on self-test (POST) data for the device 600, initialization of various functional modules in the system, drivers for the system's basic input / output, and data and instructions required to boot the operating system.

[0064] Optionally, device 600 may also include other hardware platforms or components, such as the tensor processing unit (“TPU”) 614, graphics processing unit (“GPU”) 615, field-programmable gate array (“FPGA”) 616, and machine learning unit (“MLU”) 617 shown. It is understood that although various hardware platforms or components are shown in device 600, they are merely exemplary and not limiting, and those skilled in the art can add or remove appropriate hardware as needed. For example, device 600 may include only a CPU, associated storage devices, and interface devices to implement the method of this application for training a recognition model for chronic mountain sickness.

[0065] In some embodiments, to facilitate data transmission and interaction with external networks, the device 600 of this application further includes a communication interface 618, through which it can connect to a local area network / wireless local area network (“LAN / WLAN”) 605, and further through the LAN / WLAN to connect to a local server 606 or to the Internet (“Internet”) 607. Alternatively or additionally, the device 600 of this application can also directly connect to the Internet or cellular network via the communication interface 618 based on wireless communication technology, such as 3G (“3G”), 4G (“4G”), or 5G (“5G”) wireless communication technology. In some application scenarios, the device 600 of this application can also access the server 608 and database 609 of an external network as needed to obtain various known algorithms, data, and modules, and can remotely store various data, such as various data or instructions for presenting, for example, fundus images.

[0066] Peripherals of device 600 may include a display device 602, an input device 603, and a data transmission interface 604. In one embodiment, the display device 602 may include, for example, one or more speakers and / or one or more visual displays, configured to provide voice prompts and / or display images and videos to the recognition model for training and identifying chronic mountain sickness according to this application. The input device 603 may include, for example, a keyboard, mouse, microphone, posture capture camera, and other input buttons or controls, configured to receive audio data input and / or user commands. The data transmission interface 604 may include, for example, a serial interface, parallel interface, or Universal Serial Bus interface (“USB”), Small Computer System Interface (“SCSI”), Serial ATA, FireWire (“FireWire”), PCI Express, and High Definition Multimedia Interface (“HDMI”), configured for data transmission and interaction with other devices or systems. According to the scheme of this application, the data transmission interface 604 can receive fundus images acquired from a fundus camera and transmit fundus images or various other types of data or results to device 600.

[0067] The CPU 611, mass storage 612, ROM 613, TPU 614, GPU 615, FPGA 616, MLU 617, and communication interface 618 of the device 600 of this application can be interconnected via bus 619, and can interact with peripheral devices through this bus. In one embodiment, the CPU 611 can control other hardware components in the device 600 and its peripheral devices through bus 619.

[0068] The above combination Figure 6 This paper describes a device that can be used to train a recognition model for chronic mountain sickness, which is applicable to the present application. It should be understood that the device structure or architecture described herein is merely exemplary, and the implementation methods and entities of this application are not limited thereto, but can be modified without departing from the spirit of this application.

[0069] Based on the foregoing description in conjunction with the accompanying drawings, those skilled in the art will understand that the embodiments of this application can also be implemented by software programs. Therefore, this application also provides a computer-readable storage medium storing computer-readable instructions for training a recognition model for identifying chronic mountain sickness and for identifying chronic mountain sickness. When executed by one or more processors, these computer-readable instructions can be used to implement the embodiments of this application in conjunction with the accompanying drawings. Figure 1 and Figure 5 The method described is for training an identification model for recognizing chronic mountain sickness, and the operations performed by the device for recognizing chronic mountain sickness are also described.

[0070] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0071] It should be understood that when the terms "first," "second," "third," and "fourth," etc., are used in the claims, specification, and drawings of this application, they are used only to distinguish different objects and not to describe a specific order. The terms "comprising" and "including" as used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

[0072] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0073] Although the embodiments of this application are described above, the content is merely an example adopted for the purpose of facilitating understanding of this application and is not intended to limit the scope and application scenarios of this application. Any person skilled in the art described in this application may make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in this application, but the scope of patent protection of this application shall still be determined by the scope defined in the appended claims.

Claims

1. A method for training a recognition model for identifying chronic mountain sickness, wherein the recognition model includes a preprocessing network and a backbone network, and the method includes: Fundus images were acquired and classified and labeled for chronic altitude sickness to form a training set; The preprocessing network is used to extract a first specific feature related to the identification of early chronic mountain sickness from the training set to obtain a feature vector containing the first specific feature, wherein the first specific feature includes at least fundus edema discoloration features. The backbone network is used to extract a second specific feature from the feature vector that is related to the identification of chronic mountain sickness, so as to identify chronic mountain sickness and obtain a prediction identification result. The second specific feature includes at least the fundus vascular hemorrhage and dilation feature. as well as Based on the chronic mountain sickness classification and labeling and the predicted identification results, a loss function is calculated to train the identification model for chronic mountain sickness. Before using the backbone network to extract the second specific feature related to the identification of chronic mountain sickness from the feature vector, the method further includes: The feature vector is sharpened to obtain a sharpened feature vector; The backbone network comprises multiple convolutional layers, each convolutional layer comprising cascaded convolutional kernels, and uses the backbone network to extract a second specific feature from the feature vector related to the identification of chronic altitude sickness, in order to identify chronic altitude sickness and obtain a predicted identification result, including: The cascaded convolutional kernels of each convolutional layer in the backbone network are used to extract a second specific feature from the feature vector that is related to the identification of chronic altitude sickness, so as to identify chronic altitude sickness and obtain a predicted identification result.

2. The method according to claim 1, further comprising: The fundus images were classified and labeled for chronic altitude sickness based on the scoring statistics of chronic altitude sickness.

3. The method according to claim 1, wherein sharpening the feature vector comprises: Perform edge detection operation on the feature vector to obtain edge detection information; Based on the edge detection information, enhancement factors related to chronic mountain sickness are extracted and identified; as well as The feature vector is sharpened based on the enhancement factor.

4. The method of claim 3, wherein the enhancement factor includes edge response range and local contrast.

5. The method according to claim 1, wherein the preprocessing network comprises a convolutional neural network and the backbone network comprises a ResNet50 network.

6. The method according to claim 1, wherein calculating a loss function based on the chronic mountain sickness classification label and the predicted identification result to train the identification model for chronic mountain sickness includes: Calculate the first loss function and the second loss function based on the chronic mountain sickness classification label and the prediction and identification results; as well as A weighted summation operation is performed on the first loss function and the second loss function to obtain a total loss function, which is then used to train the identification model for chronic altitude sickness.

7. An apparatus for training a recognition model for chronic mountain sickness, comprising: processor; as well as A memory storing program instructions for training a recognition model for identifying chronic mountain sickness, wherein when the program instructions are executed by the processor, the device implements the method according to any one of claims 1-6.

8. A device for identifying chronic mountain sickness, comprising: processor; as well as A memory storing program instructions for recognizing chronic altitude sickness, which, when executed by the processor, cause the device to perform the following operations: Acquire fundus images of the eye to be identified; The fundus image to be identified is input into the identification model trained by the method according to any one of claims 1-6 for identification, so as to obtain the identification result of chronic altitude sickness.

9. A computer-readable storage medium storing thereon computer-readable instructions for training a recognition model for recognizing chronic mountain sickness and for recognizing chronic mountain sickness, wherein the computer-readable instructions, when executed by one or more processors, implement the method as described in any one of claims 1-6 or implement the operation performed by the apparatus as described in claim 8.

Images