Detailed Description
As shown in fig. 1, a corneal topography discrimination method based on deep learning includes: obtaining a corneal topography to be distinguished and carrying out image preprocessing to obtain corneal topography characteristic data; and inputting the corneal topography feature data into a corneal topography distinguishing model to obtain a corneal morphology result output by the corneal topography distinguishing model.
The invention discloses a corneal topography distinguishing method based on deep learning, which is characterized in that a corneal topography obtained in the prior art is preprocessed to obtain corneal topography feature data which can be processed by a corneal topography distinguishing model, the corneal topography feature data are input into the corneal topography distinguishing model, and a corneal morphology result is obtained through the corneal topography distinguishing model. The corneal topography distinguishing model analyzes the corneal topography to determine the form result, a doctor can directly determine the corneal form according to the result output by the corneal topography distinguishing model without judging according to the corneal topography, the work of the doctor is reduced, and the corneal topography distinguishing model is a convolutional neural network model obtained by training a large number of corneal topography with determined form results, so that the prediction accuracy is high, and the condition of wrong judgment possibly caused by insufficient clinical experience of the doctor in practice is avoided. The cornea topographic map distinguishing method based on deep learning provided by the invention utilizes the trained reliable convolutional neural network model to distinguish the shape of the cornea topographic map, and solves the problem that the cornea shape distinguishing technology for deep learning processing analysis of the cornea topographic map is lacked in the prior art. In this embodiment, the corneal morphology results include five results of normal cornea, suspected corneal morphology abnormality, early keratoconus, keratoconus and myopic cornea refractive surgery, and the risk of postoperative complications of patients with suspected corneal morphology abnormality is high, while the corneal refractive surgery is not recommended for the early keratoconus and the keratoconus. The corneal morphology result can be directly obtained according to the corneal topography to be distinguished, the corneal morphology of the patient is determined, and the refractive surgery decision is guided.
Further, the acquiring a corneal topography to be distinguished and performing image preprocessing to obtain corneal topography feature data includes: acquiring corneal topography data to be distinguished; and deleting invalid data in the corneal topographic data to obtain effective corneal topographic data, and combining the effective corneal topographic data to obtain corneal topographic feature data.
In a preferred embodiment of the invention, the obtained corneal topography is cut, only valid data is reserved, invalid data is deleted, the obtained corneal topography feature data is prevented from containing a large amount of useless data to influence the identification and judgment of a convolutional neural network, and meanwhile, the data processing amount is reduced, so that the model operation efficiency can be improved.
Modern corneal topography (corneal topography) detection is based on the projection of corneal images, and consists of Placido discs, and the principle is that a concentric circle of light rings is projected on the cornea, then a video camera captures images of the rings reflected by the tear film layer of the cornea, and finally the data are analyzed through a computer software system. The smaller the distance between the concentric rings, the higher the refractive power of the cornea there, and the computer can plot this data as a color thermal map. The Placido disc-based detection system, however, collects only the data for the anterior surface of the cornea, but not the posterior surface of the cornea, which is a more sensitive feature of ectasic disease. The Scheimpflug camera solves this problem by being a rotating camera that produces a three-dimensional image containing all the information from the anterior surface of the cornea to the posterior surface of the lens. Corneal tomography (kerneal tomography) is used to depict images generated based on Scheimpflug imaging instruments. But the meaning of corneal topography we often say clinically includes corneal tomograms. The system using the Scheimpflug camera technology can provide the topography and thickness of the whole cornea, namely, a cornea front surface height map, a cornea rear surface height map, a cornea thickness map and the like can be generated through computer software processing, and a cornea curvature map can also be derived from the data of the height maps. There are 4 major instruments currently used with the Scheimpflug camera, TMS-5 (Tomey, japan), pentacam HR (OCULUS, germany), sirius (CSO, italy), and Galilei (Ziemer, switzerland). The more common corneal topographers in our country are Pentacam and Sirius. Pentacam is a combination of a rotating Scheimpflug camera and a static camera, and a standard report page of the special optical lens can be reconstructed according to height data obtained by pictures of the special optical lens, wherein the standard report page comprises a cornea front surface axial curvature map, a cornea front surface height map, a cornea rear surface height map and a cornea thickness map. The topography system of Sirius is a combination of Placido disc and a single-rotation Scheimpflug system, and the standard report page includes the tangential curvature map of the anterior surface of the cornea, the elevation map of the posterior surface of the cornea, and the thickness map of the cornea, but these data have algorithms of their own systems, which cannot be compared with the corresponding maps in Pentacam. The literature reports that several corneal topographers have high repeatability and that the Pentacam results are more reliable in measuring corneal astigmatism. Since the Pentacam refractive tetragram better reflects the true morphology of the cornea, the Pentacam refractive tetragram was selected as the corneal topography data in this example.
The original picture of the Pentacam refractive quadruple map comprises two parts, wherein the left side is patient data and anterior segment parameters, the right side is a quadruple refractive composition map (an axial curvature map of the front surface of the cornea, a height map of the back surface of the cornea and a thickness map of the cornea, and the positions are sequentially from top left to top right to bottom right and from bottom left), and the interface presents four maps most direct for a clinician to screen patients with refractive surgery. The steep areas in the anterior corneal surface axial curvature map are shown in warm (red and orange) tones and the flat areas in cool (green and blue) tones. The morphology may be represented as symmetrical bowties, asymmetrical bowties, circles, ovals, irregularities, and the like. The map of corneal anterior-posterior surface heights is a detail describing the corneal height by matching the detected corneal surface to a reference surface. Dots above the reference surface are considered elevated, indicated by positive values, and dots below the reference surface are considered depressed, indicated by negative values, again indicating corneal morphology in a cool-warm tone. Since the corneal height is not affected by the axial direction, and corneal curvature, corneal abnormalities can be detected more accurately. The reference surface of the corneal elevation map is set to a default diameter of 8.0mm, which can be adjusted as needed. The pachymetry map is a graphical representation of the distribution of corneal thickness throughout. The measurements are also displayed on a color-coded topographical map, with red and orange representing thinner areas and blue and green representing thicker areas of the cornea.
In the prior art, a refractive surgeon needs to make a judgment on the corneal morphology according to comprehensive analysis of the shape, color characteristics and the like of four images. A system for assisting a doctor to judge the cornea morphology, such as BAD-D expansion analysis and a keratoconus ABCD grading method, is also arranged in the Pentacam anterior segment analysis system; the Sirius three-dimensional keratotopography system classifies corneal morphology into Normal (Normal), keratoconus (Suspect Keratoconus), keratoconus (Keratoconus compatible), abnormal or Post-treatment (Abnormal or treated), and Myopic Post-OP (Myopic Post-OP) according to SVM (support vector machine) algorithm. Domestic refractive surgery doctors can make analysis by combining the judgment of each system and make preoperative preliminary diagnosis. However, the criteria of the two systems are mainly based on the database of caucasian, and the judgment of the Chinese myopia population needs to be based on the experience of the operating doctor and the image reading process needing a certain time.
Therefore, in the present embodiment, a pentagram refractive quadruplet is selected as corneal topographic data in the present embodiment, and only the quadruple refractive composition map (the corneal anterior surface axial curvature map, the corneal anterior surface height map, the corneal posterior surface height map, and the corneal thickness map, which are the positions of upper left, upper right, lower right, and lower left, respectively) most important for the corneal morphology discrimination is retained in the original pentagram refractive quadruplet.
As shown in fig. 2, in this embodiment, the preprocessing step for the Pentacam refractive tetragram specifically includes cutting off a left useless region (i.e., deleting invalid data) of the refractive tetragram (i.e., original corneal topography data), and splicing a right valid region (i.e., merging valid data) to obtain a preprocessed image (i.e., corneal topography feature data); the pixel values [ 0, 255 ] are scaled to [ 0,1 ] for data normalization. The raw corneal topography data includes four image portions: anterior surface axial curvature map, anterior surface elevation map, posterior surface elevation map, corneal thickness map. And the final judgment result is a comprehensive analysis by combining big data characteristics and expert experience knowledge according to the shape, heat and BAD-D expansion analysis results of the four image parts. Since the effective region is extracted from the original corneal topography, the trained model can accurately and precisely determine the corneal morphology of the corneal topography.
Further, the obtaining step of the corneal topography distinguishing model comprises: obtaining a cornea identification sample data set, wherein the cornea identification sample data set comprises a training set, a verification set and a test set; inputting the training set and the verification set into a convolutional neural network model, and calculating the error between the output value of the model and the target value corresponding to the cornea identification sample data set and the model prediction accuracy; updating the model weight according to the residual error, inputting the training set and the verification set into the updated convolutional neural network model, and calculating the error between the output value of the model and the corresponding target value in the cornea identification sample data set and the model prediction accuracy; and taking the convolution neural network model with the updating times reaching the preset value as a corneal topography distinguishing model.
In a preferred embodiment of the present invention, the corneal topography discrimination model is a convolutional neural network model trained based on a large amount of corneal discrimination sample data. Because the convolutional neural network model has the characteristic of accurate classification, the convolutional neural network model is trained by a large number of corneal topography maps with determined corneal topography results, and the accuracy of the corneal topography map discrimination model can be ensured.
In this embodiment, the training process of the convolutional neural network model is as follows:
s1, initializing a weight by a network;
s2, inputting training set data, and carrying out forward propagation on a convolution layer, a pooling layer and a full-link layer to obtain an output value;
s3, solving the error between the output value and the target value of the network and the accuracy rate of the model monitoring index;
s4, repeating S2 and S3 for the verification set;
s5, updating the weight of the network according to the residual error;
and S6, repeating the steps S2, S3 and S4 until the iteration number reaches a preset value, and ending the training.
In this embodiment, the convolutional neural network model is a Resnet model, and its loss function formula is:
where weight { y }
j = t } represents the weight of the t-th class,
represents a penalty term that prevents overfitting, wherein>
And x represents an input picture, y represents a corresponding category, and m, n, kw and k represent batches.
Classical convolutional neural network CNN models include LeNet, alexNet, VGGNet, xception, inclusion, resNet, ZFNet, denseNet, and the like. Generally, the deeper the network, the better the effect, but in the neural network with larger depth, the deep learning causes gradient disappearance and gradient explosion due to the depth of the network. The traditional solutions are initialization (nonrelizationitiation) and regularization (batch nonrelization) of data, but the solution solves the problem of gradient, but causes degradation of network performance, that is, the depth is deepened, and the accuracy is lowered. One milestone event in the history of CNN is the occurrence of the ResNet model. The ResNet model can well solve degradation and gradient problems, so that a deeper CNN model can be trained, and higher accuracy can be realized. Therefore, the Resnet model with better performance is selected here.
Further, the acquiring the corneal discrimination sample data set includes: acquiring corneal topography data to be distinguished and corresponding classification results; deleting invalid data in the corneal topographic map data to obtain effective corneal topographic map data, combining the effective corneal topographic map data, and extracting to obtain corneal topographic feature data; and obtaining sample data in the cornea identification sample data set according to the corneal topographic feature data and the corresponding classification result.
In a preferred embodiment of the invention, the obtained corneal topography is cut, only valid data is reserved, invalid data is deleted, the obtained corneal topography feature data is prevented from containing a large amount of useless data to influence the identification and judgment of a convolutional neural network, and meanwhile, the data processing amount is reduced, so that the model operation efficiency can be improved.
In this embodiment, the preprocessing step for the Pentacam refractive tetragon map is specifically to crop the left useless area of the refractive tetragon map (i.e., the original corneal topographic map data) (i.e., delete the invalid data), and splice the right valid area (i.e., merge the valid data) to obtain a preprocessed image (i.e., corneal topographic feature data); dividing the processed image into a training set, a verification set and a test set; the pixel values [ 0, 255 ] are scaled to [ 0,1 ] for data normalization. The raw corneal topography data includes four image portions: anterior surface axial curvature map, anterior surface elevation map, posterior surface elevation map, corneal thickness map. And the final judgment result is the comprehensive analysis by combining big data characteristics and expert experience knowledge according to the forms and heat of the four image parts. Since the effective region is extracted from the original corneal topography, the trained model can accurately and precisely determine the corneal morphology of the corneal topography.
Further, the step of obtaining the corneal topography distinguishing model further comprises: inputting the test set into a convolutional neural network model with the model updating times reaching a preset value, calculating the model prediction accuracy, and when the model prediction accuracy is not less than a preset threshold value, determining the updated convolutional neural network model as a corneal topography distinguishing model; and when the model prediction accuracy is smaller than the preset threshold, expanding the number of training set samples, and training the convolutional neural network model again until the model prediction accuracy is not smaller than the preset threshold.
In a preferred embodiment of the invention, the test set is used for verifying the accuracy of the convolutional neural network model for judging the corneal morphology, and the convolutional neural network model with the accuracy is used as a corneal topography judgment model. And for the convolutional neural network model which does not meet the accuracy requirement, repeating the training steps in the embodiment on the convolutional neural network model by adopting more training samples so as to ensure that the obtained corneal topography distinguishing model has higher distinguishing accuracy.
As shown in fig. 3, a corneal topography discrimination method based on deep learning includes:
the data acquisition unit to be distinguished is used for acquiring a corneal topography map to be distinguished and carrying out image preprocessing to obtain corneal topography feature data;
and the cornea form distinguishing unit is used for inputting the cornea topography feature data into the cornea topography map distinguishing model and obtaining a cornea form result output by the cornea topography map distinguishing model.
The invention discloses a corneal topography distinguishing system based on deep learning, which is characterized in that a corneal topography obtained in the prior art is preprocessed to obtain corneal topography feature data which can be processed by a corneal topography distinguishing model, the corneal topography feature data are input into the corneal topography distinguishing model, and a corneal morphology result is obtained through the corneal topography distinguishing model. The corneal topography distinguishing model analyzes the corneal topography to determine the form result, a doctor can directly determine the corneal form according to the result output by the corneal topography distinguishing model without judging according to the corneal topography, the work of the doctor is reduced, and the corneal topography distinguishing model is a convolutional neural network model obtained by training a large number of corneal topography with determined form results, so that the prediction accuracy is high, and the condition of wrong judgment possibly caused by insufficient clinical experience of the doctor in practice is avoided. The cornea topographic map distinguishing system based on deep learning provided by the invention utilizes the trained reliable convolutional neural network model to distinguish the shape of the cornea topographic map, and solves the problem that the cornea shape distinguishing technology for deep learning processing analysis of the cornea topographic map is lacked in the prior art. In this embodiment, the corneal morphology results include five results of normal cornea, suspected corneal morphology abnormality, early keratoconus, and myopic cornea refractive surgery, and the corneal morphology results can be directly obtained according to the corneal topography to be distinguished to determine the corneal morphology of the patient.
As shown in fig. 4, the data to be discriminated acquisition unit includes:
the data acquisition module of the corneal topography to be distinguished is used for acquiring the corneal topography data to be distinguished;
and the corneal topography feature data to be distinguished acquisition module is used for deleting invalid data in the corneal topography data to obtain effective corneal topography data, and combining the effective corneal topography data to obtain corneal topography feature data.
In a preferred embodiment of the invention, the obtained corneal topography is cut, only valid data is reserved, invalid data is deleted, the obtained corneal topography feature data is prevented from containing a large amount of useless data to influence the identification and judgment of a convolutional neural network, and meanwhile, the data processing amount is reduced, so that the model operation efficiency can be improved. In this embodiment, a pentagram refractive quadruplet is selected as corneal topographic data in this embodiment, and only the quadruple refractive composition map (a corneal anterior surface axial curvature map, a corneal anterior surface height map, a corneal posterior surface height map, and a corneal thickness map, which are the most important for corneal morphology discrimination) is retained for the original pentagram refractive quadruplet. Cutting off a useless area on the left side of a refraction tetragram (namely original corneal topography map data) (namely deleting invalid data), splicing effective areas on the right side (namely combining effective data) to obtain a preprocessed image (namely corneal topography feature data); the pixel values [ 0, 255 ] are scaled to [ 0,1 ] for data normalization. Since the effective area is extracted from the original corneal topography, the model obtained by training can be more accurately positioned and the corneal morphology of the corneal topography can be judged.
As shown in fig. 4, the corneal morphology discriminating unit includes:
the system comprises a sample data set acquisition module, a cornea identification sample data set acquisition module and a cornea identification sample data set acquisition module, wherein the cornea identification sample data set comprises a training set, a verification set and a test set;
the model training module is used for inputting the training set and the verification set into a convolutional neural network model, and calculating the error between the output value of the model and the target value corresponding to the cornea discrimination sample data set and the model prediction accuracy;
the model updating module is used for updating the model weight according to the residual error, inputting the training set and the verification set into the updated convolutional neural network model, and calculating the error between the output value of the model and the target value corresponding to the cornea identification sample data set and the model prediction accuracy;
and the model generation module is used for obtaining a corneal topography distinguishing model according to the updated convolutional neural network model when the number of times of updating the model weight reaches a preset value.
In a preferred embodiment of the present invention, the corneal topography pattern recognition model is a convolutional neural network model trained according to a large amount of corneal recognition sample data. Because the convolutional neural network model has the characteristic of accurate classification, the convolutional neural network model is obtained by training a large number of corneal topography maps with determined corneal topography results, and the accuracy of the corneal topography map discrimination model can be ensured. Classical convolutional neural network CNN models include LeNet, alexNet, VGGNet, xception, inclusion, resNet, ZFNet, denseNet, and the like. Generally, the deeper the network, the better the effect, but in the neural network with larger depth, the deep learning causes gradient disappearance and gradient explosion due to the depth of the network. The traditional solutions are initialization (nonrelizationitiation) and regularization (batch nonrelization) of data, but the solution solves the problem of gradient, but causes degradation of network performance, that is, the depth is deepened, and the accuracy is lowered. One milestone event in the history of CNN is the occurrence of the ResNet model. The ResNet model can well solve degradation and gradient problems, so that a deeper CNN model can be trained, and higher accuracy can be realized. Therefore, the embodiment selects the Resnet model with better performance.
As shown in fig. 5, the sample data set obtaining module includes:
the sample corneal topography acquisition module is used for acquiring corneal topography data to be distinguished and corresponding classification results;
the sample characteristic data acquisition module is used for deleting invalid data in the corneal topographic map data to obtain effective corneal topographic map data, combining the effective corneal topographic map data and extracting to obtain corneal topographic characteristic data;
and the sample data generation module is used for obtaining the sample data in the cornea identification sample data set according to the corneal topographic feature data and the corresponding category result.
In a preferred embodiment of the invention, the obtained corneal topography is cut, only valid data is reserved, invalid data is deleted, the obtained corneal topography feature data is prevented from containing a large amount of useless data to influence the identification and judgment of a convolutional neural network, and meanwhile, the data processing amount is reduced, so that the model operation efficiency can be improved.
As shown in fig. 5, the corneal morphology discriminating unit further includes:
the model testing module is used for inputting the test set into the convolutional neural network model with the model updating times reaching a preset value, calculating the model prediction accuracy, and when the model prediction accuracy is not less than a preset threshold value, the updated convolutional neural network model is the corneal topography distinguishing model; and when the model prediction accuracy is smaller than the preset threshold, expanding the number of the training set samples, and training the convolutional neural network model again until the model prediction accuracy is not smaller than the preset threshold.
In a preferred embodiment of the invention, the test set is used for verifying the accuracy of the convolutional neural network model for judging the type of the cornea, and the convolutional neural network model with the accuracy is used as a corneal topography judgment model. And for the convolutional neural network model which does not meet the accuracy requirement, repeating the training steps in the embodiment on the convolutional neural network model by adopting more training samples so as to ensure that the obtained corneal topography distinguishing model has higher distinguishing accuracy.
While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention.