A jitter distortion correction image processing system for ophthalmic OCT

Abstract

The present invention belongs to the field of image processing technology, and specifically relates to a jitter distortion correction image processing system for ophthalmic OCT, comprising a data acquisition module, which uses an SD-OCT acquisition device to acquire images of human eye retina, and different layers in the image are marked with different colors; a data set division module, which divides the acquired and annotated images into different subsets, and divides them into a test data set and a training data set; a data enhancement module, which performs data enhancement on the training data set; a task establishment module, which establishes training tasks, test tasks and detection tasks; a network model, which executes training tasks, test tasks and detection tasks, and uses a trained network model to process images of human eye retina that need to be detected; the system solves the problem that a large amount of volume data is needed as a reference in the current method, which takes a long time and leads to a complex diagnostic system, reduces the dependence on large-scale reference data, and improves processing efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a jitter distortion correction image processing system for ophthalmic OCT. Background Art

[0002] Fundus optical coherence tomography (OCT) is a high-resolution imaging technology widely used in ophthalmic diagnosis and research. Due to its non-invasive, rapid and high-resolution characteristics, OCT has become an imaging method for diagnosing and monitoring most retinal diseases. During the acquisition of fundus images, head movement, breathing, vascular pulsation and involuntary eye movement can cause image distortion. Correction of this movement is one of the main challenges of OCT imaging, because motion distortion not only affects the segmentation effect of volume data, but also reduces the reliability of retinal biomarkers and affects the diagnosis of eye diseases. Therefore, it is particularly important to find a fast and accurate method to align and segment OCT images.

[0003] Existing OCT motion correction methods can be divided into hardware methods and software methods. Although hardware-based methods usually work well, they are not available for all OCT systems and will increase the cost of hardware construction. Most traditional software algorithm methods require multiple OCT volume data or multimodal images as references, which brings additional burdens to the clinical environment. Deep learning methods have shown great potential in solving the problem of motion artifacts in OCT images. Deep learning models can extract more discriminative features by learning a large amount of OCT image data to generate displacement vectors to guide the registration of OCT images. However, the use of deep learning registration still faces some challenges. First, the amount of data on distorted ophthalmic OCT images is limited, and deep learning models require a large amount of data training to avoid overfitting. Second, the current data set does not have volume data after correction of distortion. The annotation process of OCT images is time-consuming and requires professional knowledge, especially for the registration annotation of three-dimensional volume data, which is more difficult. Summary of the invention

[0004] On one hand, the present invention provides a jitter distortion correction image processing system for ophthalmic OCT, which solves the problem that the current method requires a large amount of volume data as a reference, takes a long time, and leads to a complex diagnostic system.

[0005] The present invention is achieved in this way.

[0006] A jitter distortion correction image processing system for ophthalmic OCT, the system comprising:

[0007] The data acquisition module uses the SD-OCT acquisition device to acquire images of the human eye retina. The acquired images are 3D volume data, which are annotated by ophthalmologists. The annotations include marking whether there is a deformity and whether it is a diseased image, and segmenting the layer structure of the retina. The layer structure includes the retinal nerve fiber layer, ganglion cell layer, inner plexiform layer, inner nuclear layer, outer plexiform layer, outer nuclear layer, external limiting membrane and retinal pigment epithelium. Different layers are marked with different colors in the image.

[0008] A data set division module divides the collected and annotated images into different subsets, wherein the subsets include a normal morphology image data set and an abnormal morphology image data set, and the two data sets are randomly divided into a test data set and a training data set at a ratio of 1:3, respectively;

[0009] The data enhancement module performs data enhancement on the training data set, and expands the data samples of the training data set by rotating, flipping, scaling, cropping, adjusting brightness and contrast, elastic transformation, and color space transformation;

[0010] Task establishment module, to establish training tasks, test tasks and detection tasks;

[0011] The network model performs training tasks, uses the training data set for training, and obtains the trained weight file for the test task; performs the test task, uses the test data set to evaluate the performance of the network model; performs the detection task, and uses the trained network model to process the image of the human eye retina that needs to be detected;

[0012] The network model includes:

[0013] Feature extractor, which encodes the input image into a feature map;

[0014] The alignment branch processes the feature map extracted by the feature extractor through the decoder 3D convolution block of the alignment branch and outputs an alignment vector;

[0015] The spatial transformation module adjusts the position of the B-scan in the 3D volume data according to the alignment vector output by the alignment branch, and aligns the feature map output by the feature extractor to complete the alignment task;

[0016] The segmentation branch uses the segmentation branch decoder to gradually restore the feature map aligned by the spatial transformation module to the original resolution, and finally converts the feature map into a probability map of the segmentation result, outputting the surface position distribution of each A-scan.

[0017] Furthermore, the feature extractor includes four layers of downsampling modules that output downward in sequence, each layer of the downsampling module includes an encoder 2D convolution block and a mamba block, a mamba block is set at the output end of the encoder 2D convolution block in each downsampling module, and the mamba block set at the output end of the encoder 2D convolution block of the upper layer is output to the encoder 2D convolution block of the next layer.

[0018] Furthermore, the processing process of each layer of downsampling module is:

[0019] ,

[0020] ,

[0021] Where X is the output of the previous layer and the input of the current layer. At the beginning, it is the input image, and in the middle layer it is the feature map. Y ​​is the output of the previous layer and the input of the current layer. Z is the output of the current layer. Norm() is a normalization operation. ReLU() is an activation function in deep learning. Conv2d() is a 2D convolution, which is only used in the encoder 2D convolution block. Represents the Mamba operation. The initial input image is 3D volume data, expressed as [B,C,H,W,D], where B is the batch of input data, C is the dimension of input data, H, W, D are height, width and depth respectively, and the input image data is flattened into: [B*C,D,H,W].

[0022] Furthermore, the alignment branch includes a decoder 3D convolution block of four layers of alignment branches, and the decoder 3D convolution block of each layer of alignment branch is connected to the output of a Mamba block of the same layer in a downsampling, and a Mamba block is set at the output end of the decoder 3D convolution block of each layer of alignment branch. The data passes through the decoder 3D convolution block of the four layers of alignment branches from bottom to top to form an upsampling, and the alignment vector is output through the Mamba block of the top layer.

[0023] Further, the alignment vector is expressed as: , where each element represents the displacement of the B scan in the Z direction, , Represents the number of two-dimensional data contained in the 3D volume data, that is, the number of B scans of the OCT data.

[0024] Further, the spatial transformation module is arranged at the output end of the uppermost Mamba block in the alignment branch, receives the 3D volume data, and aligns the 3D volume data according to the alignment vector;

[0025] The spatial transformation module is also respectively arranged between the Mamba block of each layer of the downsampling module and the segmentation branch decoder of the layer corresponding to the segmentation branch, aligns the feature map output by each layer of the downsampling module according to the alignment vector, and outputs it to the segmentation branch decoder of the layer corresponding to the segmentation branch.

[0026] Furthermore, the spatial transformation module includes a positioning network, a parameterized transformation module, a sampling grid generation module and a resampler, wherein the positioning network predicts the spatial transformation parameters according to the input feature map θ ,

[0027] ,

[0028] The input feature map X A function that generates spatial transformation parameters, θ is the spatial transformation parameter;

[0029] The parameterized transformation module uses the spatial transformation parameters predicted by the localization network. θ To construct a transformation matrix, the transformation matrix is:

[0030] ,

[0031] The transformation parameters θ Construct a transformation matrix;

[0032] The sampling grid generator generates a sampling grid based on the size of the input feature map and the transformation matrix. The sampling grid generator outputs a sampling grid , where G is the standard grid coordinate,

[0033]

[0034] The resampler samples the input feature map according to a sampling grid and outputs a new feature map.

[0035] Furthermore, the segmentation branch includes four layers of segmentation branch decoders, the input of each layer of segmentation branch decoder corresponds to the output of a layer of spatial transformation module, a Mamba block is set for the output of each segmentation branch decoder, the data is processed from bottom to top to form upsampling, and the final data is output by the Mamba block of the top layer.

[0036] Furthermore, the feature maps output by the 2D convolutional blocks of the encoders at each layer are aligned and concatenated with the feature maps output by the previous segmentation branch decoder of the corresponding layer in the segmentation branch, and then sent to the segmentation branch decoder of the current layer for decoding.

[0037] Furthermore, the data input to the Mamba block is an input tensor, and the input tensor is reshaped, and the dimension of the tensor is adjusted to adapt the input tensor to the input requirements of the SSM algorithm; the tensor is flattened and normalized to reduce the sensitivity of the network model to changes in the scale of the input data; after normalization, it is divided into two branches, each entering a fully connected layer for linear projection, one of which is activated by a SiLU function to introduce nonlinear characteristics; the other branch performs a 1D causal convolution operation to ensure that the output at the current moment only depends on the current and previous inputs, and the output of the 1D causal convolution enters the SSM algorithm after a SiLU activation function; then the output data of the two branches are merged by linear projection through a fully connected layer, and the shape of the tensor is reshaped.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] Traditional methods are easily affected by device differences and individual differences when processing data from different devices and patients, resulting in unstable results. The present invention improves the stability and robustness of the system through training with a large amount of diversified data, and can adapt to different devices and patient data. Hardware-based OCT image processing methods usually require additional equipment and sensors, which increases the complexity and cost of the system. The present invention achieves efficient image processing through a pure software method, reduces hardware dependence, significantly reduces dependence on large-scale reference data, improves processing efficiency, and thus reduces the cost and complexity of system construction.

[0040] The method of the present invention combines the two image processing tasks by adding the alignment network to the segmentation network. An alignment branch of the deep learning network can be used to align images and features, improve the segmentation effect of OCT3d images, and accurately segment and identify the morphological changes of the fundus retina. The image alignment and segmentation tasks are combined in a deep learning model to achieve collaborative optimization and significantly improve the overall effect of image processing. This integration method breaks the limitation of traditionally processing alignment and segmentation tasks separately and provides a more efficient solution. Deep learning techniques such as data enhancement, transfer learning and multi-task learning are used in the network model training process to ensure efficient training of the network model. At the same time, the optimized network model architecture and inference algorithm ensure fast image processing capabilities and break the performance bottleneck of deep learning models in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of the functional structure of a network model in an image processing system for jitter distortion correction of ophthalmic OCT provided by an embodiment of the present invention;

[0042] Figure 2A functional schematic diagram of a Mamba block in an image processing system for jitter distortion correction of ophthalmic OCT provided by an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of a decoder 3D convolution block in an image processing system for jitter distortion correction of ophthalmic OCT provided by an embodiment of the present invention;

[0044] Figure 4 A schematic diagram of an encoder 2D convolution block in an image processing system for jitter distortion correction of ophthalmic OCT provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0046] The present invention provides an image processing system for jitter distortion correction of ophthalmic OCT, comprising:

[0047] The data acquisition module uses the SD-OCT acquisition device (the SD-OCT acquisition device belongs to frequency domain optical coherence tomography) to collect 3D volume data of the human retina. The acquisition process specifically includes patient positioning, image focus and acquisition, and image optimization. The optimized image is annotated by an ophthalmologist. The annotations here include whether there is a deformity and whether it is a diseased image. The layer structure of the retina is segmented. The layer structure here includes the retinal nerve fiber layer, ganglion cell layer, inner plexiform layer, inner nuclear layer, outer plexiform layer, outer nuclear layer, external limiting membrane and retinal pigment epithelium. Different layers are marked with different colors in the image.

[0048] A data set division module divides the collected and annotated images into different subsets, wherein the subsets include a normal state image data set and an abnormal state image data set, and the two data sets are randomly divided into a test data set and a training data set at a ratio of 1:3, respectively;

[0049] The data enhancement module performs data enhancement on the training data set, and expands the data samples of the training data set by rotating, flipping, scaling, cropping, adjusting brightness and contrast, elastic transformation, and color space transformation;

[0050] Task establishment module, to establish training tasks, test tasks and detection tasks;

[0051] The network model performs training tasks, uses the training data set for training, and obtains the trained weight file for the testing task; performs testing tasks, uses the testing data set to evaluate the performance of the network model; performs detection tasks, and uses the trained network model to process the image of the human eye retina that needs to be detected.

[0052] See also Figure 1 As shown, the network model includes:

[0053] Feature extractor, which encodes the input image into a feature map;

[0054] The alignment branch processes the feature map extracted by the feature extractor through the decoder 3D convolution block of the alignment branch and outputs an alignment vector;

[0055] The spatial transformation module adjusts the position of the B-scan in the 3D volume data according to the alignment vector output by the alignment branch, and aligns the feature map output by the feature extractor to complete the alignment task;

[0056] The segmentation branch uses the segmentation branch decoder to gradually restore the feature map aligned by the spatial transformation module to the original resolution, and finally converts the feature map into a probability map of the segmentation result, outputting the surface position distribution of each A-scan.

[0057] The feature extractor includes four layers of downsampling modules that output downward in sequence. Each layer of downsampling modules includes an encoder 2D convolution block and a Mamba block. A Mamba block is set at the output end of the encoder 2D convolution block in each downsampling module. The Mamba block set at the output end of the encoder 2D convolution block of the upper layer is output to the encoder 2D convolution block of the next layer. Figure 4 As shown, each encoder 2D convolution block includes two 3*3 2D convolution layers, the output of one 3*3 2D convolution layer is connected to the input of another 3*3 2D convolution layer, and each 3*3 2D convolution layer is provided with a normalization layer and a ReLUctant activation function.

[0058] The processing process of each layer of downsampling module is:

[0059] ,

[0060] ,

[0061] Where X is the output of the previous layer and the input of the current layer. At the beginning, it is the input image, and in the middle layer it is the feature map. Y ​​is the output of the previous layer and the input of the current layer. Z is the output of the current layer. Norm() is a normalization operation. ReLU() is an activation function in deep learning. Conv2d() is a 2D convolution, which is only used in the encoder 2D convolution block. Represents the Mamba operation. The initial input image is 3D volume data, expressed as [B,C,H,W,D], where B is the batch of input data, C is the dimension of input data, H, W, D are height, width and depth respectively, and the input image data is flattened into: [B*C,D,H,W].

[0062] The encoder 2D convolutional block effectively reduces the spatial size of the feature map and increases the receptive field, thereby capturing a wider range of contextual information. This processing method enables the network model to gradually extract more abstract features from the image, from primary edge and texture information to high-level semantic information. This hierarchical feature extraction is very beneficial for the network model to learn more representative and robust feature representations in subsequent tasks, thereby improving the image's ability to capture spatial features.

[0063] The alignment branch includes four layers of decoder 3D convolution blocks of alignment branches. The number of layers is consistent with that of the feature extractor. The decoder 3D convolution blocks of each layer of alignment branches are connected to the output of the Mamba block of the same layer in a downsampling module. A Mamba block is set at the output end of the decoder 3D convolution block of each layer of alignment branch. The data passes through the decoder 3D convolution blocks of the four layers of alignment branches from bottom to top to form upsampling, and the alignment vector is output through the Mamba block of the top layer.

[0064] The data processed by the decoder 3D convolution block of each layer of the alignment branch is the data output by the Mamba block of the same layer in the downsampling module corresponding to the current layer and the data output by the decoder 3D convolution block of the alignment branch of the previous layer via the Mamba block. These two data are concatenated and decoded.

[0065] See also Figure 3 As shown, each decoder 3D convolution block includes two 2*2*1 3D convolution layers, the output of one 2*2*1 3D convolution layer is connected to the input of another 2*2*1 3D convolution layer, and each 2*2*1 3D convolution layer is provided with a normalization layer and a ReLUctant activation function.

[0066] The alignment vector output by the alignment branch is expressed as: , where each element represents the displacement of the B scan in the Z direction, , Represents the number of 2D data contained in the 3D volume data, that is, the number of B scans of the OCT data. Since smoothness is one of the inherent properties of the retinal layer, if the B scans are correctly aligned, the surface position of the same layer should be close to the surface position of the adjacent B scans.

[0067] The alignment vector has two purposes: one is to adjust the position of the B-scan in the 3D volume data, and the other is to align the feature map in the segmentation branch.

[0068] There are multiple spatial transformation modules, one of which is set at the output end of the uppermost Mamba block in the alignment branch, receives the 3D volume data, and aligns the 3D volume data according to the alignment vector;

[0069] The spatial transformation module is also respectively arranged between the Mamba block of each layer of the downsampling module and the segmentation branch decoder of the layer corresponding to the segmentation branch, and aligns the feature map output by each layer of the downsampling module according to the alignment vector, and outputs it to the segmentation branch decoder of the layer corresponding to the segmentation branch.

[0070] The spatial transformation module uses the STM (Spatial transformer module) alignment module.

[0071] The spatial transformation module includes a localization network, a parameterized transformation module, a sampling grid generation module, and a resampler, where the localization network predicts the spatial transformation parameters based on the input feature map. θ ,

[0072] ,

[0073] The input feature map X A function that generates spatial transformation parameters, θ is the spatial transformation parameter;

[0074] The parameterized transformation module uses the spatial transformation parameters predicted by the localization network. θ To construct a transformation matrix, the transformation matrix is:

[0075] ,

[0076] The transformation parameters θ Construct a transformation matrix;

[0077] The sampling grid generator generates a sampling grid based on the size of the input feature map and the transformation matrix. The sampling grid generator outputs a sampling grid , where G is the standard grid coordinate,

[0078] ,

[0079] The resampler samples the input feature map according to a sampling grid and outputs a new aligned feature map.

[0080] The segmentation branch uses the nearest neighbor interpolation method to upsample the feature map. It includes four layers of segmentation branch decoders. The input of each layer of segmentation branch decoder corresponds to the output of a layer of spatial transformation module. A Mamba block is set at the output of each segmentation branch decoder. The data is processed from bottom to top to form upsampling, and the final data is output by the Mamba block of the top layer.

[0081] The feature maps output by the 2D convolutional blocks of each layer of the encoder are aligned and concatenated with the feature maps output by the previous layer of the segmentation branch decoder of the corresponding layer in the segmentation branch, and then sent to the segmentation branch decoder of the current layer for decoding.

[0082] The Mamba block in the above structure is a module used to process sequence data and image data, mainly used to capture long-range dependencies and context information. It can effectively enhance the feature expression ability without increasing the computational complexity, enhance the feature expression ability, and improve the discriminative performance of the network model.

[0083] See also Figure 2 As shown in the figure, the input data of the Mamba block is the input tensor, which is reshaped and adapted to the input requirements of the SSM algorithm by adjusting the dimension of the tensor; the tensor is flattened and normalized to reduce the sensitivity of the network model to changes in the scale of the input data; after normalization, it is divided into two branches, each entering a fully connected layer for linear projection, one of which is activated by the SiLU function to introduce nonlinear characteristics; the other branch performs a 1D causal convolution operation to ensure that the output at the current moment only depends on the current and previous inputs, and the output of the 1D causal convolution enters the SSM algorithm (state space sequence model) after a SiLU activation function; then the output data of the two branches are reshaped through the fully connected layer.

[0084] The SSM algorithm is the core of the Mamba block and can be expressed as the following linear ordinary differential equation (ODE):

[0085] ,

[0086] The state matrix and , is a parameter, Indicates the influence of the control quantity on the state quantity. Indicates the output, Indicates the impact of the current state quantity on the output. Indicates input, and Represent the hidden states of the previous image frame and the current image frame respectively. The Mamba block adopts an input-dependent selection mechanism based on the SSM algorithm. Unlike the traditional time- and input-invariant SSM algorithm, it allows effective information filtering from the input. It also combines a hardware-aware algorithm that can be linearly scaled with the sequence length, demonstrating significant computational efficiency during training and inference.

[0087] The network model of the present invention uses mean squared error (MSE) as the loss function in the alignment branch:

[0088] ,

[0089] The mean squared error calculates the average of the squared differences between the predicted values ​​and the actual values. is the predicted value, is the true value, is the sample size, is the index of the B-scan sample. In the alignment task, these two variables are set to two adjacent B-scans instead of the true value and predicted value in the original formula. In this way, MSE can be used to measure the similarity between the two B-scans to evaluate its effect.

[0090] In the segmentation branch, cross-entropy loss is used. Cross-entropy loss is one of the most commonly used loss functions in image segmentation, especially for binary and multi-classification segmentation tasks. The cross-entropy loss formula is as follows:

[0091] ,

[0092] in is an indicative function, indicating which category the element belongs to; is the predicted probability, the predicted probability that the observed sample belongs to category c. The predicted probability needs to be estimated and calculated in advance; is the loss function, is the total number of samples, is the total number of categories, It is The samples belong to The indicator variable of the class, The model is The sample prediction belongs to The probability of the class, and Represent the index of the category and sample respectively.

[0093] Dice Loss measures similarity by calculating the ratio of the intersection and union of two sets. The formula is as follows:

[0094] ,

[0095] is the real target mask, To predict the target mask, we hope and The intersection should be as large as possible, and the proportion should be as large as possible, but the loss needs to be gradually reduced, so a negative sign is added in front of the ratio. Represents a smaller number in order to prevent the numerator and denominator from being equal to zero.

[0096] After the system of the present invention generates a set of prediction results, these prediction results are then compared in detail with the corresponding expert annotation results to quantitatively evaluate the accuracy of the system of the present invention in performing ophthalmic OCT image segmentation. Once the accuracy of the system reaches or exceeds the established performance threshold, it is considered to be up to standard. If the accuracy does not meet the standard, the system will automatically send instructions to the task establishment module to request continued training and optimization, thereby ensuring the diagnostic accuracy of the final system.

[0097] The evaluation criteria include mean absolute distance (MAD) and NCC (normalization cross correlation) between two adjacent B-scans to ensure the multidimensionality and comprehensiveness of the evaluation results. They are defined as:

[0098] ,

[0099] ,

[0100] In the MAD formula, is the coordinate of the i-th point in the predicted segmentation result, is the coordinate of the i-th point of the true label. In this task, a retinal layer is selected and the first point of this retinal layer is taken as , the last point is used as PN, which measures the overall deviation between the predicted segmentation result and the true label.

[0101] In the NCC formula, and are the i-th pixel or sample point in image A and image B respectively, and are the average values ​​of image A and image B, respectively. In this task, A and B are two adjacent B scans to measure the similarity between the two B scans.

[0102] These two evaluation indicators can be used to evaluate the effect of the alignment task and the accuracy of the segmentation.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A jitter distortion correction image processing system for ophthalmic OCT, characterized in that: The system includes: The data acquisition module uses the SD-OCT acquisition device to acquire images of the human eye retina. The acquired images are 3D volume data, which are annotated by ophthalmologists. The annotations include marking whether there is a deformity and whether it is a diseased image, and segmenting the layer structure of the retina. The layer structure includes the retinal nerve fiber layer, ganglion cell layer, inner plexiform layer, inner nuclear layer, outer plexiform layer, outer nuclear layer, external limiting membrane and retinal pigment epithelium. Different layers are marked with different colors in the image. A data set division module divides the collected and annotated images into different subsets, wherein the subsets include a normal morphology image data set and an abnormal morphology image data set, and the two data sets are randomly divided into a test data set and a training data set at a ratio of 1:3, respectively; The data enhancement module performs data enhancement on the training data set, and expands the data samples of the training data set by rotating, flipping, scaling, cropping, adjusting brightness and contrast, elastic transformation, and color space transformation; Task establishment module, to establish training tasks, test tasks and detection tasks; The network model performs training tasks, uses the training data set for training, and obtains the trained weight file for the test task; performs the test task, uses the test data set to evaluate the performance of the network model; performs the detection task, and uses the trained network model to process the image of the human eye retina that needs to be detected; The network model includes: a feature extractor, encoding an input image into a feature map; The alignment branch processes the feature map extracted by the feature extractor through the decoder 3D convolution block of the alignment branch and outputs an alignment vector; The spatial transformation module adjusts the position of the B-scan in the 3D volume data according to the alignment vector output by the alignment branch, and aligns the feature map output by the feature extractor to complete the alignment task; The segmentation branch uses the segmentation branch decoder to gradually restore the feature map aligned by the spatial transformation module to the original resolution, and finally converts the feature map into a probability map of the segmentation result, outputting the surface position distribution of each A-scan.

2. The jitter distortion correction image processing system for ophthalmic OCT according to claim 1, characterized in that: The feature extractor includes four layers of downsampling modules that output downward in sequence, each layer of the downsampling modules includes an encoder 2D convolution block and a mamba block, a mamba block is set at the output end of the encoder 2D convolution block in each downsampling module, and the mamba block set at the output end of the encoder 2D convolution block of the upper layer is output to the encoder 2D convolution block of the next layer.

3. The image processing system for jitter distortion correction of ophthalmic OCT according to claim 2, characterized in that: The processing process of each layer of downsampling module is: ， ， Where X is the output of the previous layer and the input of the current layer. At the beginning, it is the input image, and in the middle layer it is the feature map. Y ​​is the output of the previous layer and the input of the current layer. Z is the output of the current layer. Norm() is a normalization operation. ReLU() is an activation function in deep learning. Conv2d() is a 2D convolution, which is only used in the encoder 2D convolution block. Represents the Mamba operation. The initial input image is 3D volume data, expressed as [B,C,H,W,D], where B is the batch of input data, C is the dimension of input data, H, W, D are height, width and depth respectively, and the input image data is flattened into: [B*C,D,H,W].

4. The image processing system for jitter distortion correction of ophthalmic OCT according to claim 2, characterized in that: The alignment branch includes a decoder 3D convolution block of four layers of alignment branches, and the decoder 3D convolution block of each layer of alignment branch is connected to the output of a Mamba block of the same layer in a downsampling. A Mamba block is set at the output end of the decoder 3D convolution block of each layer of alignment branch. Data passes through the decoder 3D convolution block of the four layers of alignment branches from bottom to top to form upsampling, and the alignment vector is output through the Mamba block of the top layer.

5. The image processing system for jitter distortion correction of ophthalmic OCT according to claim 4, characterized in that: The alignment vector is expressed as: , where each element represents the displacement of the B scan in the Z direction, , Represents the number of two-dimensional data contained in the 3D volume data, that is, the number of B scans of the OCT data.

6. The image processing system for jitter distortion correction of ophthalmic OCT according to claim 4, characterized in that: The spatial transformation module is arranged at the output end of the uppermost Mamba block in the alignment branch, receives the 3D volume data, and aligns the 3D volume data according to the alignment vector; The spatial transformation module is also respectively arranged between the Mamba block of each layer of the downsampling module and the segmentation branch decoder of the layer corresponding to the segmentation branch, aligns the feature map output by each layer of the downsampling module according to the alignment vector, and outputs it to the segmentation branch decoder of the layer corresponding to the segmentation branch.

7. The image processing system for jitter distortion correction of ophthalmic OCT according to claim 6, characterized in that: The spatial transformation module includes a positioning network, a parameterized transformation module, a sampling grid generation module and a resampler, wherein the positioning network predicts the spatial transformation parameters according to the input feature map. θ , ， The input feature map X A function that generates spatial transformation parameters, θ is the spatial transformation parameter; The parameterized transformation module uses the spatial transformation parameters predicted by the localization network. θ To construct a transformation matrix, the transformation matrix is: ， The transformation parameters θ Construct a transformation matrix; The sampling grid generator generates a sampling grid based on the size of the input feature map and the transformation matrix. The sampling grid generator outputs a sampling grid , where G is the standard grid coordinate, ， The resampler samples the input feature map according to a sampling grid and outputs a new feature map.

8. The image processing system for jitter distortion correction of ophthalmic OCT according to claim 1, characterized in that: The segmentation branch includes four layers of segmentation branch decoders, the input of each layer of segmentation branch decoder corresponds to the output of a layer of spatial transformation module, a Mamba block is set at the output of each segmentation branch decoder, the data is processed from bottom to top to form upsampling, and the final data is output by the Mamba block of the top layer.

9. The image processing system for jitter distortion correction of ophthalmic OCT according to claim 8, characterized in that: The feature maps output by the 2D convolutional blocks of each layer of the encoder are aligned and concatenated with the feature maps output by the previous layer of the segmentation branch decoder of the corresponding layer in the segmentation branch, and then sent to the segmentation branch decoder of the current layer for decoding.

10. The image processing system for jitter distortion correction of ophthalmic OCT according to any one of claims 2 to 9, characterized in that: The data input by the Mamba block is an input tensor, and the shape of the input tensor is reshaped, and the dimension of the tensor is adjusted to make the input tensor adapt to the input requirements of the SSM algorithm; the tensor is flattened and normalized to reduce the sensitivity of the network model to changes in the scale of the input data; After normalization, it is divided into two branches, each entering a fully connected layer for linear projection. One branch uses the SiLU activation function to introduce nonlinear characteristics; the other branch performs a 1D causal convolution operation to ensure that the output at the current moment only depends on the current and previous inputs. The output of the 1D causal convolution passes through a SiLU activation function and enters the SSM algorithm; then the output data of the two branches are merged through linear projection through the fully connected layer and the shape of the tensor is reshaped.

Images