Intelligent diagnosis auxiliary identification system for endometrial lesions

Abstract

The invention provides an intelligent diagnosis and auxiliary recognition system for endometrial lesions, and relates to the technical field of artificial intelligence, and the system obtains eye parameters of a to-be-detected eye image through a CDFI image detection module, and also obtains the classification of eye diseases corresponding to the to-be-detected eye image through a CDFI image feature extraction module. In addition, the medical image analysis method and the medical image analysis system can process a large number of patient cases, and effectively relieve the working pressure of diagnosis personnel in a high-intensity working environment. Through automatic image analysis and parameter recording, the efficiency of a clinical working process is remarkably improved, the possibility of recording errors in the process of manually recording image parameters can also be reduced, meanwhile, the workload of doctors is also reduced, the learning cost of medical staff and other users is reduced, and through visual output, the efficiency is greatly improved. A doctor can better understand the illness state and the blood vessel state of a patient, and the method is very important for analysis and diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent diagnosis and auxiliary recognition system for endometrial lesions. Background Art

[0002] With the rapid development of computer science and image processing technology, the analysis and interpretation of color Doppler ultrasound images have become more precise and efficient. The introduction of artificial intelligence and machine learning methods has enabled doctors to more accurately identify and quantitatively evaluate blood flow velocity, direction, and abnormal blood flow patterns, thereby detecting cardiovascular diseases, liver diseases, etc. earlier. This technological advancement not only reduces the workload of doctors, but also improves the accuracy of clinical decision-making, providing strong support for patients to develop personalized treatment plans.

[0003] However, although color Doppler ultrasound images have high resolution and rich information, doctors and diagnosticians still face a series of challenges in actual clinical applications. In a high-intensity working environment, diagnosticians often need to deal with a large number of patient cases, which limits the efficiency of clinical workflow. In the diagnostic process, in addition to accurate and detailed image analysis, doctors also need to manually record parameter information in the image, which is often time-consuming and error-prone. When faced with a large number of patients, this working mode undoubtedly increases the workload of doctors and limits diagnostic efficiency and accuracy. Although color Doppler ultrasound images provide a powerful tool for clinical diagnosis with their rich details and large amount of information, manual processing and analysis of these image data, especially in high-intensity and emergency medical scenarios, is still a challenge in efficiency and accuracy. Therefore, finding a method that can automatically, quickly and accurately process and analyze Doppler ultrasound images has become the key to optimizing the current medical diagnosis process and improving diagnostic accuracy and efficiency.

[0004] Faced with this challenge, automated image analysis and information extraction technology has become the key to improving the current situation and enhancing diagnostic quality and efficiency. With the help of advanced computer vision and artificial intelligence technology, we are expected to achieve automated analysis of color Doppler ultrasound images, quickly and accurately extract key diagnostic information and parameters from them, freeing up doctors' time and energy, allowing them to focus more on clinical decision-making and patient care. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to propose an intelligent diagnosis and auxiliary identification system for endometrial lesions, including an identity recognition module, a function selection module, a CDFI image feature extraction module, a CDFI image detection module and a data storage module;

[0006] The identity recognition module is used to receive identity information input by a user, the identity information input by the user includes an account number and a password, verify the account number and the password, obtain a verification result, and output an error prompt message when the verification result indicates that the verification fails, and call the function selection module when the verification result indicates that the verification passes;

[0007] The function selection module is used to receive a function selection operation input by a user, and when the function selection operation represents the execution of CDFI image feature extraction, the CDFI image feature extraction module is called; when the function selection operation represents the execution of CDFI image detection, the CDFI image detection module is called;

[0008] The CDFI image detection module is used to receive an eye image to be detected, input the eye image to be detected into a pre-trained image detection model, obtain eye parameters in the eye image to be detected, the eye parameters include a blood flow parameter N, and a parameter characterizing whether the eye in the eye image to be detected is a left eye or a right eye, and visualize and output the eye parameters and the eye image to be detected;

[0009] The pre-trained image detection model includes a first target detection model and a text recognition model. The image detection model is obtained by model training based on multiple first training samples. The first training sample includes a first input sample and a first output sample. The first input sample includes a historical eye image and first annotation information for the historical eye image. The first annotation information includes an annotation of an area where a blood flow parameter N is located, and an annotation of an area where a parameter representing whether an eye in the historical eye image is a left eye or a right eye is located. The first output sample is a true value of the blood flow parameter N in the historical eye image, and a true value of the parameter representing whether an eye in the historical eye image is a left eye or a right eye.

[0010] The CDFI image feature extraction module is used to receive an eye image to be detected, input the eye image to be detected into a pre-trained image feature extraction model, obtain the classification of eye diseases corresponding to the eye image to be detected, and visualize and output the eye image to be detected and the classification of eye diseases corresponding to the eye image to be detected; wherein the image feature extraction model includes a second target detection model and a disease classification model;

[0011] The data storage module is used to store the eye parameters output by the CDFI image feature extraction module and the eye image to be detected, or the eye image to be detected output by the CDFI image detection module and the classification of eye diseases corresponding to the eye image to be detected.

[0012] Optionally, the eye image to be detected is input into a pre-trained image detection model to obtain eye parameters in the eye image to be detected, including:

[0013] The eye image to be detected is input into the first target detection model to obtain an image corresponding to the blood flow parameter N and an image corresponding to the parameters representing whether the eye in the eye image to be detected is a left eye or a right eye. The image corresponding to the blood flow parameter N and the image corresponding to the parameters representing whether the eye in the eye image to be detected is a left eye or a right eye are input into the text recognition model to obtain the eye parameters in the eye image to be detected.

[0014] Optionally, the YOLOv5 model and the PP-OCRv3 model are trained by using the first training sample to obtain a first object detection model and a text recognition model, wherein the first object detection model is a model trained by the YOLOv5 model, and the text recognition model is a model trained by the PP-OCRv3 model;

[0015] Based on this, when the YOLOv5 model and the PP-OCRv3 model are trained through the first training sample, it includes:

[0016] Input the first input sample into the YOLOv5 model to obtain an image corresponding to the blood flow parameter N and an image corresponding to the parameter representing whether the eye in the historical eye image is a left eye or a right eye;

[0017] Annotating an image corresponding to the blood flow parameter N and an image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye to obtain second annotation information, wherein the second annotation information includes annotations of text in the image corresponding to the blood flow parameter N and annotations of text in the image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye, inputting the image corresponding to the blood flow parameter N, the image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye, and the second annotation information into the PP-OCRv3 model to obtain a predicted value of the blood flow parameter N in the historical eye image and a predicted value of the parameter representing whether the eye in the historical eye image is the left eye or the right eye;

[0018] Based on the true value of the blood flow parameter N in the historical eye images and the predicted value of the blood flow parameter N in the historical eye images, as well as the true value of the parameter characterizing whether the eye in the historical eye images is the left eye or the right eye and the predicted value of the parameter characterizing whether the eye in the historical eye images is the left eye or the right eye, the parameters in the YOLOv5 model and the PP-OCRv3 model are updated through the loss function until the YOLOv5 model and the PP-OCRv3 model converge, thereby obtaining the trained first target detection model and text recognition model.

[0019] Optionally, in the CDFI image feature extraction module, the eye image to be detected is input into a pre-trained image feature extraction model to obtain the classification of eye diseases corresponding to the eye image to be detected, specifically including:

[0020] The eye image to be detected is input into the second target detection model to obtain an image of the orbital area in the eye image to be detected, and the image of the orbital area is input into the disease classification model to obtain the classification of the eye disease corresponding to the eye image to be detected.

[0021] Optionally, the second target detection model is obtained by training a YOLOv5 model based on multiple second training samples, the second training samples include a second input sample and a second output sample, the second input sample includes a historical eye image and annotations of the eye socket area in the historical eye image, and the second output sample is an image of the eye socket area.

[0022] Optionally, the disease classification model is obtained by training a VanillaNet model based on multiple third training samples, the third training samples include third input samples and third output samples, the third input samples include images of historical orbital areas, and the third output samples are classifications of eye diseases corresponding to the images of historical orbital areas.

[0023] Optionally, the first labeling information further includes a labeling of an area where the ophthalmic artery OA is located, a labeling of an area where the central retinal artery CRA is located, and a labeling of an area where the posterior ciliary artery PCA is located;

[0024] Based on this, the first input sample is input into the YOLOv5 model, and the image corresponding to the ophthalmic artery OA, the image corresponding to the central retinal artery CRA, and the image corresponding to the posterior ciliary artery PCA can also be obtained.

[0025] The beneficial effects of adopting the above technical solution are:

[0026] (1) The intelligent diagnosis and identification system for endometrial lesions provided by the present invention can handle a large number of patient cases, effectively alleviating the work pressure of diagnostic personnel in a high-intensity working environment. Through automated image analysis and parameter recording, the efficiency of clinical workflow is significantly improved.

[0027] (2) In the traditional manual recording of image parameters, errors are prone to occur and it is time-consuming. The automation function of the intelligent diagnosis and auxiliary identification system for intraocular membrane lesions provided by the present invention reduces the possibility of such errors and also reduces the workload of doctors. It also reduces the learning cost of medical staff and other users.

[0028] (4) The intelligent diagnosis and auxiliary identification system for intraocular membrane lesions provided by the present invention has high-precision CDFI medical image target detection and feature extraction functions, and can provide accurate medical image analysis results, which is crucial for doctors to diagnose and formulate treatment plans.

[0029] (5) The intelligent diagnosis and identification system for endometrial lesions provided by the present invention provides intuitive visualization results, enabling doctors to better understand the patient's condition and vascular status, which is very important for analysis and diagnosis. This visualization function helps doctors make more accurate judgments and treatment decisions.

[0030] (6) Users can easily save and manage test images, which facilitates later search and analysis and improves work efficiency.

[0031] (7) The intelligent diagnosis and identification system for endometrial lesions provided by the present invention is not only suitable for professional medical personnel, but also for ordinary users who are not familiar with medicine and want to understand their own eye color ultrasound images. This makes the system have a wider range of applications and audiences. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a schematic diagram of the structure of an intelligent diagnosis and auxiliary identification system for endometrial lesions in an embodiment of the present invention;

[0033] Figure 2 Schematic diagram of the output result of the CDFI image detection module in an embodiment of the present invention;

[0034] Figure 3 A schematic diagram of the process flow of the CDFI image detection module in an embodiment of the present invention;

[0035] Figure 4 Schematic diagram of the training indicator changes of the YOLOv5s model in the CDFI image detection module in an embodiment of the present invention, wherein Figure (a) is a graph of bounding box loss, Figure (b) is a graph of target loss, Figure (c) is a graph of classification loss, Figure (d) is a graph of accuracy, Figure (e) is a graph of recall rate, Figure (f) is a graph of validation set bounding box loss, Figure (g) is a graph of validation set target loss, Figure (h) is a graph of validation set classification loss, Figure (i) is a graph of average accuracy when the IOU threshold is 0.5, and Figure (j) is a graph of average accuracy within the range of IOU thresholds of 0.5 to 0.95;

[0036] Figure 5 Schematic diagram of the output result of the CDFI image feature extraction module in an embodiment of the present invention;

[0037] Figure 6 is a schematic diagram of a second training sample and a third training sample in an embodiment of the present invention;

[0038] Figure 7 Figure (a) is a schematic diagram of a normal eye, Figure (b) is a schematic diagram of vitreous opacity in the eye, Figure (c) is a schematic diagram of vitreous hemorrhage organization in the eye, Figure (d) is a schematic diagram of silicone oil filling in the eye, Figure (e) is a schematic diagram of choroidal detachment in the eye, Figure (f) is a schematic diagram of retinal detachment in the eye, and Figure (g) is a schematic diagram of stellate degeneration in the eye. DETAILED DESCRIPTION

[0039] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0040] In view of the problems existing in the prior art, the present invention provides an intelligent diagnosis and auxiliary identification system for endometrial lesions. Figure 1 , including an identity recognition module, a function selection module, a CDFI image feature extraction module, a CDFI image detection module and a data storage module;

[0041] The identity recognition module is used to receive identity information input by a user, the identity information input by the user includes an account number and a password, verify the account number and the password, obtain a verification result, and output an error prompt message when the verification result indicates that the verification fails, and call the function selection module when the verification result indicates that the verification passes;

[0042] In specific implementation, a login interface can be set for the identity recognition module. After entering the account and password and clicking the "Login" button, you can jump to the function selection interface. Click "Exit Program" and a prompt box will pop up, showing "Do you want to exit the program?" Select "YES" to close the program immediately. If the account or password is entered incorrectly, a prompt box will pop up, showing "Incorrect user name or password."

[0043] The function selection module is used to receive a function selection operation input by a user, and when the function selection operation represents the execution of CDFI image feature extraction, the CDFI image feature extraction module is called; when the function selection operation represents the execution of CDFI image detection, the CDFI image detection module is called;

[0044] In the specific implementation, an interface is set for the function selection module, including a CDFI medical image target detection button and a CDFI medical image feature extraction button. Different modal recognition methods can be selected according to different user needs. If you want to know the target of the detection, you can locate the detection position through the weight file, as well as the category and the corresponding confidence of the detection. If you want to understand the blood flow information based on the CDFI image, you can choose CDFI medical image feature extraction. Users can choose one of them according to their needs. If you choose one of the two systems, you will jump to the corresponding recognition module.

[0045] The CDFI image detection module is used to receive an eye image to be detected, input the eye image to be detected into a pre-trained image detection model, obtain eye parameters in the eye image to be detected, the eye parameters include a blood flow parameter N, and a parameter characterizing whether the eye in the eye image to be detected is a left eye or a right eye, and visualize the eye parameters and the eye image to be detected. In the specific output process, the output can also be combined with the eye image to be detected, such as Figure 2 ;

[0046] The pre-trained image detection model includes a first target detection model and a text recognition model. The image detection model is obtained by model training based on multiple first training samples. The first training sample includes a first input sample and a first output sample. The first input sample includes a historical eye image and first annotation information for the historical eye image. The first annotation information includes an annotation of an area where a blood flow parameter N is located, and an annotation of an area where a parameter representing whether an eye in the historical eye image is a left eye or a right eye is located. The first output sample is a true value of the blood flow parameter N in the historical eye image, and a true value of the parameter representing whether an eye in the historical eye image is a left eye or a right eye.

[0047] The first annotation information is obtained by semi-automatic annotation using a semi-automatic graphic annotation tool suitable for the OCR field.

[0048] The step of inputting the eye image to be detected into a pre-trained image detection model to obtain eye parameters in the eye image to be detected includes:

[0049] The eye image to be detected is input into the first target detection model to obtain an image corresponding to the blood flow parameter N and an image corresponding to the parameters representing whether the eye in the eye image to be detected is a left eye or a right eye. The image corresponding to the blood flow parameter N and the image corresponding to the parameters representing whether the eye in the eye image to be detected is a left eye or a right eye are input into the text recognition model to obtain the eye parameters in the eye image to be detected.

[0050] In specific implementation, combined with Figure 3, the first target detection model also outputs the image corresponding to the ophthalmic artery OA, the image corresponding to the central retinal artery CRA, and the image corresponding to the posterior ciliary artery PCA, that is, Figure 3 The image in the first row of the second column in the figure is the image corresponding to the blood flow parameter N, and the third row of the second column is the image corresponding to the parameter representing whether the eye in the eye image to be detected is the left eye or the right eye, where OS represents the left eye and OD represents the right eye; the third column is the eye parameters.

[0051] The YOLOv5 model and the PP-OCRv3 model are trained by the first training sample to obtain a first object detection model and a text recognition model, wherein the first object detection model is a model trained by the YOLOv5 model, and the text recognition model is a model trained by the PP-OCRv3 model;

[0052] Based on this, when training the YOLOv5 model and the PP-OCRv3 model through the first training sample, combined with Figure 2 ,include:

[0053] Input the first input sample into the YOLOv5 model to obtain an image corresponding to the blood flow parameter N and an image corresponding to the parameter representing whether the eye in the historical eye image is a left eye or a right eye;

[0054] Annotating an image corresponding to the blood flow parameter N and an image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye to obtain second annotation information, wherein the second annotation information includes annotations of text in the image corresponding to the blood flow parameter N and annotations of text in the image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye, inputting the image corresponding to the blood flow parameter N, the image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye, and the second annotation information into the PP-OCRv3 model to obtain a predicted value of the blood flow parameter N in the historical eye image and a predicted value of the parameter representing whether the eye in the historical eye image is the left eye or the right eye;

[0055] Based on the true value of the blood flow parameter N in the historical eye images and the predicted value of the blood flow parameter N in the historical eye images, as well as the true value of the parameter characterizing whether the eye in the historical eye images is the left eye or the right eye and the predicted value of the parameter characterizing whether the eye in the historical eye images is the left eye or the right eye, the parameters in the YOLOv5 model and the PP-OCRv3 model are updated through the loss function until the YOLOv5 model and the PP-OCRv3 model converge, thereby obtaining the trained first target detection model and text recognition model.

[0056] The first labeling information further includes a labeling of the area where the ophthalmic artery OA is located, a labeling of the area where the central retinal artery CRA is located, and a labeling of the area where the posterior ciliary artery PCA is located;

[0057] Based on this, the first input sample is input into the YOLOv5 model, and the image corresponding to the ophthalmic artery OA, the image corresponding to the central retinal artery CRA, and the image corresponding to the posterior ciliary artery PCA can also be obtained.

[0058] When training the YOLOv5 model, the YOLOv5 series networks were used for testing. Each network was trained for 200 batches. The Early Stopping strategy was used during the training process. Usually, the network performance converged in batches of about 100-150, and the training could be terminated early.

[0059] Figure 4The figure shows the change of YOLOv5s model training indicators, where the bounding box loss measures the difference between the bounding box predicted by the model and the true bounding box. The lower the loss, the closer the bounding box predicted by the model is to the bounding box of the actual target. As can be seen from Figure (a), the Box loss gradually decreases during the training process, indicating that the model is becoming more and more accurate in locating the target. The target loss is used to measure whether each predicted box contains the target. YOLO assigns a confidence to each predicted box (indicating whether there is a target in the box), and this loss is calculated based on the confidence. As shown in Figure (b), the target loss gradually decreases, indicating that the model's performance in judging whether a box contains a target is constantly improving. The classification loss measures the model's performance in correctly classifying the target category. The lower the classification loss, the more accurate the model is in predicting the target category. As can be seen from Figure (c), as the number of training rounds increases, the classification loss decreases significantly, indicating that the model's ability to distinguish different categories is gradually improving. Accuracy refers to the proportion of correct predictions made by the model among all predictions, that is, how many of the predicted positive samples are true positive samples. High accuracy means that the model has fewer false positives. As can be seen from Figure (d), the accuracy rises rapidly and stabilizes during the training process, indicating that the false alarm rate of the model decreases during training and the prediction is more accurate. The recall rate indicates the proportion of all actual positive samples that the model successfully detects. A high recall rate means fewer missed detections. As can be seen from Figure (e), the recall rate also rises rapidly and stabilizes in the early stage of training, indicating that the model performs well in identifying all targets and has fewer missed detections. The bounding box loss on the validation set reflects the positioning accuracy of the model on unseen data. This indicator is used to evaluate the generalization ability of the model. As shown in Figure (f), the Box loss of the validation set is also decreasing, indicating that the bounding box prediction accuracy of the model on the validation set is also improving. The target loss of the validation set indicates the accuracy of the model in judging whether there is an object in the box on the validation set. As can be seen from Figure (g), the target loss on the validation set also decreases, indicating that the model's judgment on the existence of the target on new data is becoming more and more accurate. The classification loss of the validation set measures the accuracy of the model's prediction of the target category on the validation set. As can be seen from Figure (h), the loss also gradually decreases, indicating that the model can effectively distinguish different categories on the validation set and has good generalization performance. mAP@0.5 represents the average precision when the IOU (intersection over union) threshold is 0.5, which is mainly used to measure the overall detection effect of the model. As can be seen from Figure (i), the mAP@0.5 curve rises rapidly in the early stage, indicating that the overall detection performance of the model improves rapidly during the training process and tends to stabilize in the later stage. mAP@0.5:0.95 (average precision within the range of IOU thresholds from 0.5 to 0.95) is a more stringent evaluation indicator, which represents the average precision within the range of IOU thresholds from 0.5 to 0.95. This indicator can reflect the comprehensive detection performance of the model at different IOU thresholds.As can be seen from Figure (j), mAP@0.5:0.95 rises rapidly in the early stage of training and then gradually stabilizes, indicating that the model can maintain high detection accuracy at different IOU thresholds.

[0060] In general, these 10 indicators cover the capabilities of the YOLOv5 model in target positioning, target detection, category differentiation, etc. The training graph shows that the performance of the model continues to improve while the loss decreases, and eventually stabilizes, showing good training and verification performance.

[0061] Based on this, the training results under the YOLOv5s model show that in the model training stage, through pre-training weights, after 77 rounds of training, the model has achieved high accuracy. In the training model, the proportion of correctly predicted positive samples to all predicted positive samples P reached 95.35%, the proportion of predicted positive samples in R positive samples was 97.43%, and the average precision mAP reached 97.43%. The GIOU loss function value is reduced, the target detection loss mean and the classification loss mean are relatively small and accurate, and its network performance is good.

[0062] After cutting the marked part N times, we found that the image in the specified area became blurred after the image was enlarged, which directly affected the accuracy of text recognition and made the effect very poor. Therefore, in order to solve this problem, the PP-OCRv3 text recognition pre-trained model was used for direct prediction and fine-tune training. The pre-trained model address, learning rate, and data set parameters to be used were set in the configuration file. The PP-OCRv3 Chinese and English ultra-lightweight recognition pre-trained model was fine-tuned for evaluation, and the acc reached 99%, which improved the text recognition performance in this challenging image enlargement situation.

[0063] For the CDFI image detection module, in the specific implementation, three functional partitions are set in the front-end interface, which are divided into the function selection area, the detection result area and the detection information area. The following is a detailed description:

[0064] The function selection area consists of weight selection, model initialization, and image detection. The detection result area consists of a display box. The detection information area consists of three text boxes: location, category, and confidence. There are also two buttons for continuing and ending detection.

[0065] Click the "Select Weight" button and select the trained detection weight in the folder. If no selection is made or the selection is wrong, the weight opening will fail. Select the correct weight and click Initialize Model to start loading the weight. After loading, the model loading is complete. Then click the image you want to detect, and the image will be displayed in the display box. The front-end interface can also include a detection authentication area, in which the detection results, i.e., eye parameters, will be retained in the display box, and the detection location information, category, and corresponding confidence will also be displayed in the corresponding text box. As can be seen from the detection test results, the accuracy is high and can be used with confidence.

[0066] The CDFI image feature extraction module is used to receive an eye image to be detected, input the eye image to be detected into a pre-trained image feature extraction model, obtain the classification of eye diseases corresponding to the eye image to be detected, and visualize and output the eye image to be detected and the classification of eye diseases corresponding to the eye image to be detected. In the specific output process, the eye image to be detected can also be combined for output, such as Figure 5 ; Wherein, the image feature extraction model includes a second target detection model and a disease classification model;

[0067] The step of inputting the eye image to be detected into a pre-trained image feature extraction model to obtain the classification of eye diseases corresponding to the eye image to be detected specifically includes:

[0068] The eye image to be detected is input into the second target detection model to obtain an image of the orbital area in the eye image to be detected, and the image of the orbital area is input into the disease classification model to obtain the classification of the eye disease corresponding to the eye image to be detected.

[0069] Among them, the second target detection model is obtained by training the YOLOv5 model based on multiple second training samples, the second training samples include second input samples and second output samples, the second input samples include historical eye images and annotations of the eye socket area in the historical eye images, and the second output samples are images of the eye socket area.

[0070] Among them, the disease classification model is obtained by training the VanillaNet model based on multiple third training samples, the third training samples include third input samples and third output samples, the third input samples include images of historical orbital areas, and the third output samples are the classifications of eye diseases corresponding to the images of historical orbital areas.

[0071] Among them, combined Figure 6 , the second training sample is the target detection dataset in the first row of the figure, and the third training sample is the image classification dataset in the second row;

[0072] Among them, there are seven types of eye diseases, combined with Figure 7 , (a)-(g) in the figure are normal, vitreous opacity, vitreous hemorrhage and organization, silicone oil filling, choroidal detachment, retinal detachment, and stellate degeneration. From the 7 pictures, it can be seen that pictures (a), (b), and (d) are relatively similar, and pictures (c), (e), and (f) are relatively similar, which brings difficulties to diagnosis and also makes it difficult to classify images.

[0073] Among them, for the selection of the model to be trained corresponding to the disease classification model, the present invention combines VanillaNet, ResNet101 and GoogleNet with YOLOv5 respectively. After each network is trained for 200 batches, in the target classification task, VanillaNet achieved the highest accuracy (98.4%) and the loss value was (0.019), but the overall performance was better than other networks. ResNet101 performed second in accuracy (90.8%), and the loss value was (0.011), indicating that it is relatively balanced in terms of precision and loss. GoogleNet has a relatively low accuracy (86.4%), and the loss value is close to ResNet101 (0.010). Overall, VanillaNet performs best in accuracy and is the optimal classification network. The combination of VanillaNet and YOLOv5 performs best in the target detection task. The lightweight structure of VanillaNet significantly reduces the computational cost, making it superior to ResNet101 and GoogleNet in terms of real-time and resource efficiency. It also performs well in detection accuracy. Therefore, the VanillaNet model is selected and trained to obtain a disease classification model.

[0074] Among them, the CDFI image feature extraction module, during its specific implementation, has a front-end interface for selecting test images, a text box for entering test results, and an area for saving test images. After clicking the Select Test button, the user can arbitrarily select ophthalmic images in the computer. The user can integrate the images to be tested into a folder as much as possible to facilitate user testing. After selecting the image interface, the test images are integrated into a folder.

[0075] Among them, for the results output by the CDFI image feature extraction module and / or the CDFI image detection module and the data storage module, the key information of the blood flow parameters is extracted just below the color Doppler ultrasound image, and it is judged whether the blood flow parameters are normal. According to the high or low parameters, it can be judged whether the blood flow parameters are normal or reduced. The normal range is usually determined according to relevant medical standards and previous studies. If the blood flow parameters are high or low, this may indicate abnormal vascular function and further analysis is required. Visualization results are crucial for analysis and diagnosis. Through visualization, the blood flow of the eye can be viewed intuitively for preliminary evaluation. Visualization provides important visual information, allowing doctors to better understand the patient's condition and vascular status, so as to develop a more precise treatment plan. During the entire analysis process, it is very important to accurately identify the visualization results of the specified area. This helps doctors to more accurately judge the abnormality of blood flow parameters, further develop treatment plans, and provide better medical services to patients. At the same time, ensure continuous monitoring and diagnosis to track the development of the disease and the effect of treatment.

[0076] The data storage module is used to store the eye parameters output by the CDFI image feature extraction module and the eye image to be detected, or the eye image to be detected output by the CDFI image detection module and the classification of eye diseases corresponding to the eye image to be detected.

[0077] When the data storage module is implemented, a front-end interface is provided, in which the user can observe whether the feature extraction of the CDFI image is successful. If successful, the user can click the "Save Test Image" button to save it. After clicking Save in the previous step, you can select the save path for the feature-extracted image. Enter the name of the image in the input box at "File Name". You can also choose to save it in a folder that already exists in the computer. Here, it is best for users to save the images in the same folder so that they can easily find the desired images later.

[0078] The system has undergone a rigorous testing and verification process, proving that its stability and reliability have reached excellent standards. In the rare case that a user encounters an error, the problem can be solved by simply restarting or checking the accuracy of the file path, which will not cause any loss of stored data or other user information. It is worth noting that it is not recommended to force close the interface during system operation to avoid unnecessary problems and errors.

[0079] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.

Claims

1. An intelligent diagnosis and identification system for endometrial lesions, characterized in that: It includes an identity recognition module, a function selection module, a CDFI image feature extraction module, a CDFI image detection module and a data storage module; The identity recognition module is used to receive identity information input by a user, the identity information input by the user includes an account number and a password, verify the account number and the password, obtain a verification result, and output an error prompt message when the verification result indicates that the verification fails, and call the function selection module when the verification result indicates that the verification passes; The function selection module is used to receive a function selection operation input by a user, and when the function selection operation represents the execution of CDFI image feature extraction, the CDFI image feature extraction module is called; when the function selection operation represents the execution of CDFI image detection, the CDFI image detection module is called; The CDFI image detection module is used to receive an eye image to be detected, input the eye image to be detected into a pre-trained image detection model, obtain eye parameters in the eye image to be detected, the eye parameters include a blood flow parameter N, and a parameter characterizing whether the eye in the eye image to be detected is a left eye or a right eye, and visualize and output the eye parameters and the eye image to be detected; The pre-trained image detection model includes a first target detection model and a text recognition model. The image detection model is obtained by model training based on multiple first training samples. The first training sample includes a first input sample and a first output sample. The first input sample includes a historical eye image and first annotation information for the historical eye image. The first annotation information includes an annotation of an area where a blood flow parameter N is located, and an annotation of an area where a parameter representing whether an eye in the historical eye image is a left eye or a right eye is located. The first output sample is a true value of the blood flow parameter N in the historical eye image, and a true value of the parameter representing whether an eye in the historical eye image is a left eye or a right eye. The CDFI image feature extraction module is used to receive an eye image to be detected, input the eye image to be detected into a pre-trained image feature extraction model, obtain the classification of eye diseases corresponding to the eye image to be detected, and visualize and output the eye image to be detected and the classification of eye diseases corresponding to the eye image to be detected; wherein the image feature extraction model includes a second target detection model and a disease classification model; The data storage module is used to store the eye parameters output by the CDFI image feature extraction module and the eye image to be detected, or the eye image to be detected output by the CDFI image detection module and the classification of eye diseases corresponding to the eye image to be detected.

2. The intelligent diagnosis and auxiliary identification system for endometrial lesions according to claim 1 is characterized in that: Inputting the eye image to be detected into a pre-trained image detection model to obtain eye parameters in the eye image to be detected, including: The eye image to be detected is input into the first target detection model to obtain an image corresponding to the blood flow parameter N and an image corresponding to the parameters representing whether the eye in the eye image to be detected is a left eye or a right eye. The image corresponding to the blood flow parameter N and the image corresponding to the parameters representing whether the eye in the eye image to be detected is a left eye or a right eye are input into the text recognition model to obtain the eye parameters in the eye image to be detected.

3. The intelligent diagnosis and auxiliary identification system for endometrial lesions according to claim 1 is characterized in that: The YOLOv5 model and the PP-OCRv3 model are trained by the first training sample to obtain a first object detection model and a text recognition model, wherein the first object detection model is a model trained by the YOLOv5 model, and the text recognition model is a model trained by the PP-OCRv3 model; Based on this, when the YOLOv5 model and the PP-OCRv3 model are trained through the first training sample, it includes: Input the first input sample into the YOLOv5 model to obtain an image corresponding to the blood flow parameter N and an image corresponding to the parameter representing whether the eye in the historical eye image is a left eye or a right eye; Annotating an image corresponding to the blood flow parameter N and an image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye to obtain second annotation information, wherein the second annotation information includes annotations of text in the image corresponding to the blood flow parameter N and annotations of text in the image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye, inputting the image corresponding to the blood flow parameter N, the image corresponding to the parameter representing whether the eye in the historical eye image is the left eye or the right eye, and the second annotation information into the PP-OCRv3 model to obtain a predicted value of the blood flow parameter N in the historical eye image and a predicted value of the parameter representing whether the eye in the historical eye image is the left eye or the right eye; Based on the true value of the blood flow parameter N in the historical eye images and the predicted value of the blood flow parameter N in the historical eye images, as well as the true value of the parameter characterizing whether the eye in the historical eye images is the left eye or the right eye and the predicted value of the parameter characterizing whether the eye in the historical eye images is the left eye or the right eye, the parameters in the YOLOv5 model and the PP-OCRv3 model are updated through the loss function until the YOLOv5 model and the PP-OCRv3 model converge, thereby obtaining the trained first target detection model and text recognition model.

4. The intelligent diagnosis and auxiliary identification system for endometrial lesions according to claim 1 is characterized in that: In the CDFI image feature extraction module, the eye image to be detected is input into a pre-trained image feature extraction model to obtain the classification of eye diseases corresponding to the eye image to be detected, specifically including: The eye image to be detected is input into the second target detection model to obtain an image of the orbital area in the eye image to be detected, and the image of the orbital area is input into the disease classification model to obtain the classification of the eye disease corresponding to the eye image to be detected.

5. The intelligent diagnosis and auxiliary identification system for endometrial lesions according to claim 1 is characterized in that: The second target detection model is obtained by training the YOLOv5 model based on multiple second training samples, the second training samples include a second input sample and a second output sample, the second input sample includes a historical eye image and an annotation of the eye socket area in the historical eye image, and the second output sample is an image of the eye socket area.

6. The intelligent diagnosis and auxiliary identification system for endometrial lesions according to claim 1 is characterized in that: The disease classification model is obtained by training the VanillaNet model based on multiple third training samples, the third training samples include third input samples and third output samples, the third input samples include images of historical orbital areas, and the third output samples are classifications of eye diseases corresponding to the images of historical orbital areas.

7. The intelligent diagnosis and auxiliary identification system for endometrial lesions according to claim 3 is characterized in that: The first labeling information also includes a labeling of the area where the ophthalmic artery OA is located, a labeling of the area where the central retinal artery CRA is located, and a labeling of the area where the posterior ciliary artery PCA is located; Based on this, the first input sample is input into the YOLOv5 model, and the image corresponding to the ophthalmic artery OA, the image corresponding to the central retinal artery CRA, and the image corresponding to the posterior ciliary artery PCA can also be obtained.

Images