Corneal ulcer lesion three-dimensional topological morphology feature evaluation device based on deep learning

Abstract

The application discloses a kind of corneal inflammation focus three-dimensional topological form feature evaluation device based on deep learning, comprising: obtaining the anterior segment optical coherence tomography image of corneal inflammation patient, and marking corneal inflammation focus, corneal contour line and two sides scleral spur, to construct image sample set;Based on deep convolutional neural network, construct the segmentation model for segmenting corneal inflammation focus, cornea and scleral spur, and train the segmentation model using image sample set;The trained segmentation model is used to segment the anterior segment optical coherence tomography image to be measured, and the segmented corneal inflammation focus, cornea and two scleral spur are obtained;Based on the boundary of cornea and two scleral spur, polar coordinate system and reference line are established, and corneal inflammation focus and cornea are calculated based on polar coordinate system and reference line, to determine the three-dimensional topological form feature of corneal inflammation focus.The evaluation is high in accuracy, high in efficiency, and provides technical support for the establishment of new clinical corneal inflammation diagnosis and treatment standard.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a device for evaluating the three-dimensional topological morphological features of keratitis lesions based on deep learning. Background Technology

[0002] The cornea is a transparent tissue that bears two-thirds of the eye's refractive power. Infectious keratitis is an inflammatory condition caused by pathogenic microorganisms invading the cornea. It is a general term for various infectious corneal inflammations, mainly including viral keratitis, bacterial keratitis, fungal keratitis, and Acanthamoeba keratitis. Its main precipitating factors include contact lens wear, eye trauma, and ocular surface diseases. Keratitis can cause corneal scarring or perforation, leading to visual impairment and even blindness. It is currently the fifth leading cause of blindness worldwide and the leading cause of corneal blindness.

[0003] Early diagnosis, treatment, and timely monitoring are crucial to preventing complications of keratitis, such as visual impairment and blindness. Currently, the gold standard for clinical diagnosis of keratitis is corneal scraping and pathogen culture. This method typically requires invasive procedures and a long culture time, which may lead to prolonged and worsening of the condition. Therefore, early empirical drug treatment and disease monitoring are more meaningful for patients. Currently, slit-lamp microscopy is routinely used clinically to assess the location of infiltration, the horizontal and vertical dimensions of ulcers and infiltrative lesions. However, slit-lamp examination is difficult to assess the three-dimensional characteristics of lesions, such as infiltration depth and related corneal edema, which significantly limits the real-time monitoring of changes in keratitis lesions and the observation of treatment efficacy.

[0004] Anterior segment optical coherence tomography (ASOCT), a novel imaging technique that emerged in recent years, can non-invasively display cross-sectional details of the anterior segment, providing clinicians with basic anterior segment anatomical features and related disease lesions. This imaging method can acquire values ​​such as infiltration thickness (IT) and corneal thickness (CT) in patients with keratitis, helping to assess the depth of inflammatory infiltration and providing a new approach for clinical monitoring and early treatment of keratitis. However, in clinical practice, due to time and manpower constraints, analysis is usually performed by the examiner subjectively selecting only one section of the panoramic scan. This provides limited information, lacking the ability to measure multiple objective parameters or perform longitudinal follow-up comparisons, which to some extent affects the diagnosis and treatment of keratitis. Furthermore, the relatively singular measurement of corneal lesions, relying primarily on the examiner's subjective selection and judgment, may lead to significant errors between different operators.

[0005] Deep learning-based automatic image analysis methods possess high objectivity, repeatability, and accuracy, and are increasingly widely used in clinical settings with broad application prospects. In ophthalmology, most techniques and algorithms for optical coherence tomography (OCT) images focus on analyzing fundus images, lacking a device for evaluating the three-dimensional topological morphology of lesions in the anterior segment of keratitis patients using OCT images, to assist clinicians in the early diagnosis, assessment, and treatment of keratitis. Summary of the Invention

[0006] In view of the above, the purpose of this invention is to provide a deep learning-based device for evaluating the three-dimensional topological features of keratitis lesions, which uses a deep convolutional neural network to accurately segment the cornea and keratitis lesions, and to automatically evaluate and analyze the three-dimensional topological features of keratitis lesions.

[0007] To achieve the above-mentioned objectives, an embodiment provides a deep learning-based three-dimensional topological morphological feature evaluation device for keratitis lesions, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps:

[0008] Obtain anterior segment optical coherence tomography images of keratitis patients and annotate keratitis lesions, corneal contours, and bilateral scleral ridges to construct an image sample set;

[0009] A segmentation model for segmenting keratitis lesions, cornea, and scleral processes was constructed based on a deep convolutional neural network, and the segmentation model was trained using an image sample set.

[0010] The trained segmentation model was used to segment the anterior segment optical coherence tomography image of the subject to obtain the segmented keratitis lesion, cornea and two scleral processes;

[0011] A polar coordinate system and reference lines are established based on the corneal boundary and the two scleral ridges. The keratitis lesion and cornea are calculated based on the polar coordinate system and reference lines to determine the three-dimensional topological morphological features of the keratitis lesion. The three-dimensional topological morphological features of the keratitis lesion include: the corneal thickness at the deepest point of the keratitis infiltration lesion, the infiltration depth of the lesion, the lesion width, the infiltration angle of the lesion, and the lesion eccentricity angle.

[0012] Preferably, the segmentation model includes upper and lower segmentation networks. The upper segmentation network includes a U-NET network, a multi-level feature enhancement module (HFEM), and a first pseudo-label repair module (PLRM1). The input of the multi-level feature enhancement module is connected to the output of the downsampling stage of the U-NET network, and its output is connected to the input of the upsampling stage of the U-NET network. The U-NET network and HFEM are combined to generate a first uncertainty map and two feature maps corresponding to the image sample pairs based on image sample pairs including labeled and unlabeled images, and through multi-level feature enhancement processing. The first uncertainty map and the feature maps corresponding to the unlabeled image samples are concatenated and input into PLRM1 for pseudo-label repair, generating a first segmentation result. The lower segmentation network includes at least one Transformer network consisting of an encoder and a decoder, and a second pseudo-label repair module (PLRM2). The input image sample pairs are processed by the Transformer network to generate a second uncertainty map and two other feature maps. The second uncertainty map and the feature maps corresponding to the unlabeled image samples in the other two feature maps are concatenated and input into PLRM2 for pseudo-label repair, generating a second segmentation result.

[0013] During training, the following four dicelosses are constructed: the diceloss between the feature map of the labeled image sample generated by the upper segmentation network and the ground truth label, which serves as the first segmentation result under supervision label, and the diceloss between the feature map of the unlabeled image sample in the lower segmentation network; the diceloss between the feature map of the labeled image sample generated by the lower segmentation network and the ground truth label, which serves as the second segmentation result under supervision label, and the diceloss between the feature map of the unlabeled image sample in the upper segmentation network. These four dicelosses together constitute the loss function of the segmentation model, which is used to optimize the parameters of the segmentation model.

[0014] Preferably, in each sampling convolutional layer of the U-NET network, the number of downsampling convolutional layers is equal to the number of upsampling convolutional layers. The features generated by the downsampling convolutional layers are input to a multi-level feature enhancement module. The enhanced features output after feature enhancement are then input to the corresponding upsampling convolutional layer. That is, the UNET hop connections of the U-NET network undergo feature enhancement through a multi-level feature enhancement module.

[0015] The multi-level feature enhancement module includes a semantic feature module, a spatial attention module, and a channel attention module. The semantic feature module is used to obtain the contextual multi-scale receptive field of the input features by dilated convolution with different dilation rates, thereby obtaining features at different scales. Then, the features at multiple scales are fused and input into the spatial attention module and the channel attention module respectively.

[0016] The feature maps output by large-scale downsampling convolutional layers focus more on spatial features, i.e., features are extracted by the SA module in HFEM. The feature maps output by small-scale downsampling convolutional layers focus more on channel features, i.e., features are extracted by the CA module in HFEM.

[0017] Preferably, establishing a polar coordinate system and reference line based on the corneal boundary and the two scleral ridges includes:

[0018] Let the two scleral protrusions be points A and C. Draw the perpendicular bisector of AC, which intersects the lower boundary of the cornea at point B. Draw the perpendicular bisector of AB, which intersects the perpendicular bisector of AC at point O. Establish a polar coordinate system with point O as the origin. Between OA and OC, draw rays at regular intervals as reference lines.

[0019] Preferably, the corneal thickness at the deepest point of the keratitis infiltrate is calculated in the following way:

[0020] The deepest point of the keratitis infiltrate lesion on CT scan shows the corneal thickness. low Let D be the distance between the two intersection points of the polar coordinate system reference line passing through the deepest point D of the keratitis lesion and the upper and lower boundaries of the cornea. The method for determining the deepest point D of the keratitis lesion is as follows: Let N be the intersection point of the reference line in the polar coordinate system and the lower boundary of the keratitis lesion. Then, the deepest point D of the keratitis lesion is point N when the ON value is the smallest.

[0021] Preferably, the lesion invasion depth is calculated in the following way:

[0022] lesion infiltration depth IT low Let be the distance between the polar coordinate reference line passing through the deepest point D of the keratitis lesion and the two intersection points of the upper boundary of the cornea and the lower boundary of the keratitis lesion.

[0023] Preferably, the width of the lesion is calculated in the following way:

[0024] The lesion width IW is the length of the curve segment between the intersection of ray OL passing through the left edge point L of the lesion and ray OR passing through the right edge point R of the lesion with the upper boundary of the cornea.

[0025] Preferably, the lesion infiltration angle is calculated in the following way:

[0026] The lesion infiltration angle IR is the angle between the ray OL passing through the left edge point L of the lesion and the ray OR passing through the right edge point R of the lesion.

[0027] Preferably, the lesion offset angle is calculated in the following way:

[0028] The right eccentricity angle EAR of the lesion is the angle between the ray OR passing through the right edge point R of the lesion and OB;

[0029] The left eccentricity angle EAL of the lesion is the angle between the ray OL passing through the left edge point L of the lesion and OB.

[0030] To achieve the above-mentioned objectives, the embodiments also provide a deep learning-based device for evaluating the three-dimensional topological morphology features of keratitis lesions, including a sample construction unit, a model training unit, a segmentation unit, and a feature evaluation unit.

[0031] The sample construction unit is used to acquire optical coherence tomography images of the anterior segment of patients with keratitis, and to annotate the keratitis lesions, corneal contours and bilateral scleral ridges to construct an image sample set;

[0032] The training unit is used to construct a segmentation model based on a convolutional neural network for segmenting keratitis lesions, the cornea, and the scleral protrusion, and to train the segmentation model using an image sample set;

[0033] The segmentation unit is used to segment the anterior segment optical coherence tomography image of the subject using a trained segmentation model to obtain the segmented keratitis lesion, cornea and two scleral processes.

[0034] The feature evaluation unit is used to establish a polar coordinate system and reference line based on the corneal boundary and the two scleral ridges. Based on the polar coordinate system and reference line, the keratitis lesion and cornea are calculated to determine the three-dimensional topological morphological features of the keratitis lesion. The three-dimensional topological morphological features of the keratitis lesion include: the corneal thickness at the deepest point of the keratitis infiltration lesion, the infiltration depth of the lesion, the lesion width, the infiltration angle of the lesion, and the lesion eccentricity angle.

[0035] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0036] This invention uses a deep convolutional neural network to segment lesions in ASOCT images of keratitis patients, obtaining lesion segmentation maps of keratitis lesions more efficiently. This provides a foundation for accurately obtaining more complete lesion information, and the automatic identification and segmentation make it possible to comprehensively utilize the patient's complete sequence (128 ASOCT images) of image information, providing technical support for providing more accurate and comprehensive disease information in clinical practice.

[0037] This invention establishes a polar coordinate system method for quantifying keratitis lesions. Compared with manual measurement methods, it has better repeatability and stability, and is more scientific and accurate. Furthermore, this method can provide numerical values ​​and information on some keratitis lesions that are currently unmeasurable in clinical practice, thus enriching the information available for clinical decision-making. This not only assists in clinical diagnosis and treatment decisions but also provides technical support for the establishment of new standards for the diagnosis and treatment of keratitis. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of the method for evaluating the three-dimensional topological features of keratitis lesions based on deep learning, provided in the embodiment.

[0040] Figure 2 This is a schematic diagram of the segmentation model provided in the embodiment;

[0041] Figure 3 This is a schematic diagram of the establishment of the polar coordinate system and reference lines provided in the embodiment;

[0042] Figure 4 This is a schematic diagram illustrating the morphological characteristics of the keratitis lesion provided in the embodiment;

[0043] Figure 5 This is a schematic diagram of the structure of the deep learning-based three-dimensional topological morphology feature evaluation device for keratitis lesions provided in the embodiment. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0045] The inventive concept of this invention addresses the problems of low efficiency and inaccuracy in existing manual assessments of keratitis. This invention provides a deep learning-based three-dimensional topological morphology assessment device for keratitis lesions. It utilizes a segmentation model built on a deep convolutional neural network to accurately segment the keratitis lesion, cornea, and scleral processes. Based on this, a polar coordinate system and reference lines are constructed, and the keratitis lesion is automatically identified and its topological morphological parameters are automatically measured and calculated. This is a key technology for achieving objective and effective assessment of keratitis lesions. The assessment is highly efficient and accurate, and has significant implications for the clinical monitoring and treatment of keratitis.

[0046] like Figure 1 As shown in the embodiment, the method for evaluating the three-dimensional topological morphology features of keratitis lesions based on deep learning includes the following steps:

[0047] S110: Acquire anterior segment optical coherence tomography images of keratitis patients and annotate keratitis lesions, corneal contours, and bilateral scleral ridges to construct an image sample set.

[0048] In this embodiment, anterior segment optical coherence tomography (ASOCT) images of keratitis patients with typical lesions were acquired. A total of 3337 ASOCT images of keratitis patients were collected from a specific hospital ophthalmology center. Patients with ocular rupture or a history of eye surgery, as well as images of poor quality such as blurry images, were excluded from this study. The images were captured using a swept-source Casia SS-1000AS-OCT (Tomey Corporation, Nagoya, Japan), yielding images with a resolution of 1657*1000. A dataset of anterior segment cross-sectional images of keratitis patients was constructed using these images.

[0049] Ophthalmologists analyzed the anterior segment cross-sectional image dataset of keratitis to identify keratitis lesions, corneal contours, and bilateral scleral ridges, thus constructing an image sample set.

[0050] S120 is a segmentation model based on a deep convolutional neural network for segmenting keratitis lesions, cornea, and scleral protrusions, and the segmentation model is trained using an image sample set.

[0051] like Figure 2 As shown, the constructed segmentation model includes two segmentation networks: an upper and a lower network. The upper network comprises a U-NET network, a multi-level feature enhancement module (HFEM), and a first pseudo-label repair module (PLRM1). The input of the multi-level feature enhancement module is connected to the output of the downsampling stage of the U-NET network, and its output is connected to the input of the upsampling stage of the U-NET network. The U-NET network and HFEM are combined to generate a first uncertainty map and two feature maps corresponding to the image sample pairs based on image sample pairs including labeled and unlabeled images and multi-level feature enhancement processing. The first uncertainty map and the feature maps corresponding to the unlabeled image samples are concatenated and then input into PLRM1 for pseudo-label repair, generating the first segmentation result. The lower network comprises at least one Transformer network consisting of an encoder and a decoder, and a second pseudo-label repair module (PLRM2). The input image sample pairs are processed by the Transformer network to generate a second uncertainty map and two other feature maps. The second uncertainty map and the feature maps corresponding to the unlabeled image samples in the other two feature maps are concatenated and then input into PLRM2 for pseudo-label repair, generating the second segmentation result.

[0052] The pattern sample pair includes two image samples, one image sample is labeled (L) and the other image sample is unlabeled (U). The two feature maps generated in this way correspond to the labeled image sample and the unlabeled image sample, respectively.

[0053] In this embodiment, in each sampling convolutional layer of the U-NET network, the number of downsampling convolutional layers is equal to the number of upsampling convolutional layers. The features generated by the downsampling convolutional layers are input to HFEM. After feature enhancement, the output enhanced features are then input to the corresponding upsampling convolutional layer. That is, the UNET hop connections of the U-NET network undergo feature enhancement through HFEM.

[0054] like Figure 2 As shown, HFEM includes a Semantic Feature Module (ASPP), a Spatial Attention Module (SA), and a Channel Attention Module (CA). ASPP uses dilated convolutions with different dilation rates to obtain multi-scale receptive fields of the input features, resulting in features at different scales. These features are then fused and input into SA and CA respectively. Specifically, the dilation rates of ASPP are 1, 2, and 3. A dilation rate of 1 results in a 3x3 receptive field, with one data point between every two receptive fields, and a dilation rate of 3 results in two data points between every two receptive fields. Although the convolutional kernels are the same in each layer, the different dilation rates result in larger receptive fields, allowing for better capture of features at various scales. The three features are then fused. For low-level features, the resolution is high but the number of channels is low; for deep features, the resolution is low but the number of channels is high. Low-level features, i.e., the output of the first two layers after passing through the Semantic Feature Module, are represented by spatial attention-enhanced features, while deep features, i.e., the output of the last two layers after passing through the Semantic Feature Module, are represented by channel attention-enhanced features.

[0055] In SA, the input feature map is processed by the maximum and average calculations to obtain a one-dimensional attention map of the compressed channel. Then, the maps are concatenated and processed by a convolutional layer to extract spatial attention weights. Finally, the sigmoid activation function is used to obtain the final output, which is then input into the corresponding upsampling convolutional layer.

[0056] In CA, the input feature map is processed by max pooling and average pooling, and then through a fully connected layer (the structure consists of one convolutional layer, one ReLU activation function, and one convolutional layer) to obtain attention maps with different channel weights. Finally, the maps are concatenated and processed by the sigmoid activation function to obtain the final output, which is then input into the corresponding upsampling convolutional layer.

[0057] It's also important to note that feature maps output by large-scale downsampling convolutional layers focus more on spatial features, meaning they extract features through the SA module in HFEM. Conversely, feature maps from small-scale downsampling convolutional layers focus more on channel features, meaning they extract features through the CA module in HFEM. Figure 2 The U-NET network shown consists of four downsampled convolutional layers and four upsampled convolutional layers. Downsampled convolutional layers 1 and 2, due to their large size, focus more on spatial features. That is, after the features are extracted by the SA module in HFEM, they are input into the corresponding upsampled convolutional layers 8 and 7. Downsampled convolutional layers 3 and 4, due to their small size, focus more on channel features. That is, after the features are extracted by the CA module in HFEM, they are input into the corresponding upsampled convolutional layers 6 and 5.

[0058] The first pseudo-label repair module (PLRM1) and the first pseudo-label repair module (PLRM2) adopt the same structure, which consists of multiple (convolutional block-max pooling)-convolutional block-classification convolutional blocks. Furthermore, for each (convolutional block-max pooling), its input and output are connected by residuals. This structure enables pseudo-label repair of the input image.

[0059] Before the above segmentation model can be applied, it needs to be optimized. During optimization, the following four dice losses are constructed: the dice loss between the feature map of the labeled image sample generated by the upper segmentation network and the ground truth label; the dice loss between the first segmentation result of the supervision label and the feature map of the unlabeled image sample in the lower segmentation network; the dice loss between the feature map of the labeled image sample generated by the lower segmentation network and the ground truth label; and the dice loss between the second segmentation result of the supervision label and the feature map of the unlabeled image sample in the upper segmentation network. These four dice losses together constitute the loss function of the segmentation model and are used to optimize the parameters of the segmentation model.

[0060] During training, the obtained image sample set is input into the segmentation model, and the parameters are adjusted according to the set learning rate to obtain the trained segmentation model, which is used for the segmentation of keratitis lesions, cornea, and scleral processes in ASOCT images. Specifically, the learning rate is set to 0.0001, the training iterations are 70 rounds, the learning rate is fixed, the parameters are iterated in each round, and the validation result index is recorded. When the index is better than the previous result, the model parameters are saved.

[0061] S130 uses a trained segmentation model to segment the anterior segment optical coherence tomography image of the subject to obtain segmented keratitis lesions, cornea, and two scleral processes.

[0062] In this embodiment, the ASOCT image to be tested is input into the trained segmentation model. The forward derivation of the upper segmentation network of the segmentation model calculates the classification probability of each pixel in the output image as a keratitis lesion, cornea and scleral ridge. Pixels with classification probabilities greater than a preset threshold are taken as the segmentation result and output. The segmentation result includes keratitis lesion, cornea and two scleral ridges.

[0063] S140 establishes a polar coordinate system and reference lines based on the corneal boundary and the two scleral ridges.

[0064] In the embodiments, such as Figure 3 As shown, the specific process of constructing a polar coordinate system includes: Let the two scleral protrusions be points A and C, draw the perpendicular bisector of AC intersecting the lower boundary of the cornea at point B, draw the perpendicular bisector of AB intersecting the perpendicular bisector of AC at point O, establish a polar coordinate system with point O as the origin, and draw rays between OA and OC at certain intervals as reference lines.

[0065] S150 calculates the keratitis lesion and cornea based on polar coordinates and reference lines to determine the three-dimensional topological morphology of the keratitis lesion.

[0066] In this embodiment, the three-dimensional topological morphological features of the keratitis lesion include: corneal thickness at the deepest point of the keratitis infiltration lesion, lesion infiltration depth, lesion width, lesion infiltration angle, and lesion eccentricity angle. For example... Figure 4 As shown, each morphological feature is determined in the following way:

[0067] Let N be the intersection of the reference line and the lower boundary of the keratitis lesion in the polar coordinate system. Then, the deepest point of the keratitis infiltrate is defined as point N when the ON value is the smallest, and is denoted as point D.

[0068] CT scan of corneal thickness at the deepest point of the keratitis infiltrate. low CT low The distance between the polar coordinate reference line passing through the deepest point D of the keratitis lesion and the two intersection points of the upper and lower corneal boundaries is used. If there is a depression defect at the upper corneal boundary, intelligent generation and completion are performed, and the CT is calculated after completion. low It can also calculate the distance (CT) between the two intersection points of the polar coordinate system reference line and the upper and lower boundaries of the cornea at n°. n It can also calculate the thickness ratio of CT. low / CT n .

[0069] IT targeting lesion infiltration depth low IT low This is the distance between the polar coordinate reference line passing through the deepest point D of the keratitis lesion and the two intersection points of the upper and lower boundaries of the cornea. If there is a depression at the upper boundary of the cornea, intelligent generation and completion are performed, and the IT is calculated after completion. low It can also calculate the distance IT between the two intersection points of the polar coordinate system reference line and the upper boundary of the cornea and the lower boundary of the keratitis lesion at n°. n It can also calculate the width ratio IT. low / IT n .

[0070] Regarding the lesion width IW, IW is the length of the curve segment between the intersection of ray OL passing through the left edge point L of the lesion and ray OR passing through the right edge point R of the lesion with the upper boundary of the cornea.

[0071] Regarding the lesion infiltration angle IR, IR is the angle between the ray OL passing through the left edge point L of the lesion and the ray OR passing through the right edge point R of the lesion. The right eccentricity angle EAR of the lesion is the angle between the ray OR passing through the right edge point R of the lesion and OB; the left eccentricity angle EAL of the lesion is the angle between the ray OL passing through the left edge point L of the lesion and OB.

[0072] Based on the same inventive concept, such as Figure 5 As shown, the embodiment also provides a deep learning-based device for evaluating the three-dimensional topological morphological features of keratitis lesions, including a sample construction unit 410, a model training unit 420, a segmentation unit 430, and a feature evaluation unit 440.

[0073] The sample construction unit 410 is used to acquire optical coherence tomography (OCT) images of the anterior segment of keratitis patients and annotate the keratitis lesions, corneal contours, and bilateral scleral ridges to construct an image sample set. The model training unit 420 is used to construct a segmentation model based on a convolutional neural network for segmenting the keratitis lesions, cornea, and scleral ridges, and to train the segmentation model using the image sample set. The segmentation unit 430 is used to segment the anterior segment OCT image to be tested using the trained segmentation model to obtain the segmented keratitis lesions, cornea, and two scleral ridges. The feature evaluation unit 440 is used to establish a polar coordinate system and reference lines based on the corneal boundary and the two scleral ridges, and to calculate the keratitis lesions and cornea based on the polar coordinate system and reference lines to determine the three-dimensional topological morphological features of the keratitis lesions.

[0074] It should be noted that the deep learning-based 3D topological morphology feature evaluation device for keratitis lesions provided in the above embodiments should be illustrated using the above-described functional unit divisions when performing 3D topological morphology feature evaluation of keratitis lesions. The functions can be assigned to different functional units as needed, i.e., the internal structure of the terminal or server can be divided into different functional units to complete all or part of the functions described above. Furthermore, the deep learning-based 3D topological morphology feature evaluation device for keratitis lesions provided in the above embodiments and the deep learning-based 3D topological morphology feature evaluation method for keratitis lesions belong to the same concept. For details of its implementation, please refer to the deep learning-based 3D topological morphology feature evaluation method for keratitis lesions, which will not be repeated here.

[0075] Based on the same inventive concept, the embodiment also provides a deep learning-based three-dimensional topological morphological feature evaluation device for corneal lesions, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps:

[0076] S110: Acquire anterior segment optical coherence tomography images of keratitis patients and annotate keratitis lesions, corneal contours, and bilateral scleral ridges to construct an image sample set;

[0077] S120 is a segmentation model for segmenting keratitis lesions, cornea and scleral protrusions based on a deep convolutional neural network, and the segmentation model is trained using an image sample set.

[0078] S130: The trained segmentation model is used to segment the anterior segment optical coherence tomography image of the subject to be tested, and the segmented keratitis lesion, cornea and two scleral processes are obtained.

[0079] S140 establishes a polar coordinate system and reference lines based on the corneal boundary and the two scleral ridges. The keratitis lesion and cornea are calculated based on the polar coordinate system and reference lines to determine the three-dimensional topological morphological characteristics of the keratitis lesion.

[0080] In practical applications, computer memory can be volatile memory at the near end, such as RAM, or non-volatile memory, such as ROM, FLASH, floppy disks, hard disks, etc., or even remote storage cloud. The computer processor can be a central processing unit (CPU), microprocessor (MPU), digital signal processor (DSP), or field-programmable gate array (FPGA), meaning that these processors can be used to implement the steps for evaluating the three-dimensional topological morphology of corneal lesions.

[0081] The methods and apparatus provided in the above embodiments achieve automatic segmentation of the cornea and keratitis lesions through deep learning. The established automatic quantitative analysis method for the topological morphology of keratitis lesions using polar coordinates has good scientific validity and reproducibility. It can further provide numerical values ​​and information on some keratitis lesions that are currently unmeasurable in clinical practice, providing technical support for the establishment of new standards for the diagnosis and treatment of keratitis in clinical practice, assisting in clinical diagnosis and treatment decisions, and can be applied to fields such as automated disease diagnosis, disease progression assessment, and clinical treatment decision-making.

[0082] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A device for evaluating the three-dimensional topological morphology of keratitis lesions based on deep learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it performs the following steps: Obtain anterior segment optical coherence tomography images of keratitis patients and annotate keratitis lesions, corneal contours, and bilateral scleral ridges to construct an image sample set; A segmentation model for segmenting keratitis lesions, cornea, and scleral processes was constructed based on a deep convolutional neural network, and the segmentation model was trained using an image sample set. The trained segmentation model was used to segment the anterior segment optical coherence tomography image of the subject to obtain the segmented keratitis lesion, cornea and two scleral processes; Based on the boundary of the cornea and the two scleral ridges, establish a polar coordinate system and reference lines, including: let the two scleral ridges be points A and C, draw the perpendicular bisector of AC intersecting the lower boundary of the cornea at point B, draw the perpendicular bisector of AB intersecting the perpendicular bisector of AC at point O, establish a polar coordinate system with point O as the origin, and draw rays between OA and OC at certain intervals as reference lines. The keratitis lesion and cornea are calculated based on polar coordinates and reference lines to determine the three-dimensional topological morphology of the keratitis lesion. The three-dimensional topological morphology of the keratitis lesion includes: corneal thickness at the deepest point of the keratitis infiltrate, lesion infiltration depth, lesion width, lesion infiltration angle, and lesion eccentricity angle. The corneal thickness at the deepest point of the keratitis infiltrate was calculated using the following method: CT scan of the deepest point of the keratitis infiltrate. low Let D be the distance between the two intersection points of the polar coordinate system reference line passing through the deepest point D of the keratitis lesion and the upper and lower boundaries of the cornea. The method for determining the deepest point D of the keratitis lesion is as follows: Let N be the intersection point of the reference line in the polar coordinate system and the lower boundary of the keratitis lesion. Then, the deepest point D of the keratitis lesion is point N when the ON value is the smallest. The lesion invasion depth is calculated as follows: Lesion invasion depth IT low The distance between the polar coordinate system reference line passing through the deepest point D of the keratitis infiltrate and the two intersection points of the upper boundary of the cornea and the lower boundary of the keratitis lesion; The lesion width is calculated as follows: the lesion width IW is the length of the curve segment between the intersection of ray OL passing through the left edge point L of the lesion and ray OR passing through the right edge point R of the lesion with the upper boundary of the cornea. The lesion invasion angle is calculated as follows: the lesion invasion angle IR is the angle between the ray OL passing through the left edge point L of the lesion and the ray OR passing through the right edge point R of the lesion. The lesion eccentricity angle is calculated as follows: the right eccentricity angle EAR of the lesion is the angle between the ray OR passing through the right edge point R of the lesion and OB; the left eccentricity angle EAL of the lesion is the angle between the ray OL passing through the left edge point L of the lesion and OB.

2. The device for evaluating the three-dimensional topological morphology of keratitis lesions based on deep learning according to claim 1, characterized in that, The segmentation model includes two segmentation networks: the upper segmentation network includes a U-NET network, a multi-level feature enhancement module, and a first pseudo-label repair module. The input of the multi-level feature enhancement module is connected to the output of the downsampling stage of the U-NET network, and its output is connected to the input of the upsampling stage of the U-NET network. The U-NET network and the multi-level feature enhancement module combine to generate a first uncertainty map and two feature maps corresponding to the image sample pairs based on image sample pairs containing both labeled and unlabeled images, using multi-level feature enhancement processing. The first uncertainty map and the feature maps corresponding to the unlabeled image samples are concatenated and then input into the first pseudo-label repair module for pseudo-label repair, generating a first segmentation result. The lower segmentation network includes at least one Transformer network consisting of an encoder and a decoder, and a second pseudo-label repair module. The input image sample pairs are processed by the Transformer network to generate a second uncertainty map and two other feature maps. The second uncertainty map and the feature maps corresponding to the unlabeled image samples in the other two feature maps are concatenated and then input into the second pseudo-label repair module for pseudo-label repair, generating a second segmentation result. During training, the following four dicelosses are constructed: the diceloss between the feature map of the labeled image sample generated by the upper segmentation network and the ground truth label, which serves as the first segmentation result under supervision label, and the diceloss between the feature map of the unlabeled image sample in the lower segmentation network; the diceloss between the feature map of the labeled image sample generated by the lower segmentation network and the ground truth label, which serves as the second segmentation result under supervision label, and the diceloss between the feature map of the unlabeled image sample in the upper segmentation network. These four dicelosses together constitute the loss function of the segmentation model, which is used to optimize the parameters of the segmentation model.

3. The deep learning-based three-dimensional topological morphological feature evaluation device for keratitis lesions according to claim 2, characterized in that, In each sampling convolutional layer of the U-NET network, the number of downsampling convolutional layers is equal to the number of upsampling convolutional layers. The features generated by the downsampling convolutional layers are input to a multi-level feature enhancement module. The enhanced features output after feature enhancement are then input to the corresponding upsampling convolutional layer. That is, the UNET hop connections of the U-NET network undergo feature enhancement through a multi-level feature enhancement module. The multi-level feature enhancement module includes a semantic feature module, a spatial attention module, and a channel attention module. The semantic feature module is used to obtain the contextual multi-scale receptive field of the input features by dilated convolution with different dilation rates, thereby obtaining features at different scales. Then, the features at multiple scales are fused and input into the spatial attention module and the channel attention module respectively. The feature maps output by large-scale downsampling convolutional layers focus more on spatial features, which are extracted by the spatial attention module in the second pseudo-label repair module. The feature maps output by small-scale downsampling convolutional layers focus more on channel features, which are extracted by the channel attention module in the second pseudo-label repair module.

4. A device for evaluating the three-dimensional topological morphology of keratitis lesions based on deep learning, characterized in that, It includes sample construction unit, model training unit, segmentation unit, and feature evaluation unit. The sample construction unit is used to acquire optical coherence tomography images of the anterior segment of patients with keratitis, and to annotate the keratitis lesions, corneal contours and bilateral scleral ridges to construct an image sample set; The model training unit is used to construct a segmentation model based on a convolutional neural network for segmenting keratitis lesions, the cornea, and the scleral protrusion, and to train the segmentation model using an image sample set; The segmentation unit is used to segment the anterior segment optical coherence tomography image of the subject using a trained segmentation model to obtain the segmented keratitis lesion, cornea and two scleral processes. The feature evaluation unit is used to establish a polar coordinate system and reference lines based on the boundary of the cornea and the two scleral ridges, including: setting the two scleral ridges as points A and C, drawing the perpendicular bisector of AC intersecting the lower boundary of the cornea at point B, drawing the perpendicular bisector of AB intersecting the perpendicular bisector of AC at point O, establishing a polar coordinate system with point O as the origin, and drawing rays between OA and OC at certain intervals as reference lines. The feature evaluation unit is also used to calculate the keratitis lesion and cornea based on the polar coordinate system and reference line to determine the three-dimensional topological morphological features of the keratitis lesion. The three-dimensional topological morphological features of the keratitis lesion include: corneal thickness at the deepest point of the keratitis infiltrate, lesion infiltration depth, lesion width, lesion infiltration angle, and lesion eccentricity angle. The corneal thickness at the deepest point of the keratitis infiltrate was calculated using the following method: CT scan of the deepest point of the keratitis infiltrate. low Let D be the distance between the two intersection points of the polar coordinate system reference line passing through the deepest point D of the keratitis lesion and the upper and lower boundaries of the cornea. The method for determining the deepest point D of the keratitis lesion is as follows: Let N be the intersection point of the reference line in the polar coordinate system and the lower boundary of the keratitis lesion. Then, the deepest point D of the keratitis lesion is point N when the ON value is the smallest. The lesion invasion depth is calculated as follows: Lesion invasion depth IT low The distance between the polar coordinate system reference line passing through the deepest point D of the keratitis infiltrate and the two intersection points of the upper boundary of the cornea and the lower boundary of the keratitis lesion; The lesion width is calculated as follows: the lesion width IW is the length of the curve segment between the intersection of ray OL passing through the left edge point L of the lesion and ray OR passing through the right edge point R of the lesion with the upper boundary of the cornea. The lesion invasion angle is calculated as follows: the lesion invasion angle IR is the angle between the ray OL passing through the left edge point L of the lesion and the ray OR passing through the right edge point R of the lesion. The lesion eccentricity angle is calculated as follows: the right eccentricity angle EAR of the lesion is the angle between the ray OR passing through the right edge point R of the lesion and OB; the left eccentricity angle EAL of the lesion is the angle between the ray OL passing through the left edge point L of the lesion and OB.

Images