Retinal OCT full-automatic imaging and image enhancement method

Abstract

The present application relates to the technical field of optical coherence tomography, and discloses a retinal OCT full-automatic imaging and image enhancement method, which comprises the following steps: automatically positioning the three-dimensional space position of the pupil through a binocular vision system and aligning an imaging light beam; cooperatively determining an optimal refractive compensation value based on the joint evaluation of the interference signal intensity and the axial registration stability of a short sequence image; collecting multiple B-scan images, performing motion correction, removing false frames and image averaging to generate a final image; performing quality evaluation on the final image, and adaptively retriggering the acquisition process or re-focusing according to the evaluation result. The present application realizes the automation of the whole imaging process by introducing cooperative refractive compensation and closed-loop feedback control, enhances the anti-interference ability to the motion of the measured person, reduces the dependence on the experience of the operator, and can stably obtain high signal-to-noise ratio retinal tomographic images.

Description

Technical Field

[0001] This invention relates to the field of optical coherence tomography (OCT) technology, and in particular to a fully automated retinal OCT imaging and image enhancement method. Background Technology

[0002] Optical coherence tomography (OCT) is a non-invasive, high-resolution imaging technique for biological tissues. In ophthalmology, OCT devices are widely used to acquire tomographic images of the retina, providing crucial structural information for the diagnosis and monitoring of various fundus diseases such as glaucoma and macular degeneration.

[0003] Current OCT imaging procedures typically require extensive manual intervention by a professional operator. The operator must observe the subject's pupils via real-time video and manually operate levers or knobs to adjust the position of the OCT imaging beam to align it with the center of the pupil. Subsequently, the operator must manually adjust the system's refractive compensation until a clear retinal structure is observed in the preview image. This series of operations is not only time-consuming, but the image quality is also highly dependent on the operator's professional skills and clinical experience, leading to inconsistencies in imaging results between different operators or at different times within the same procedure.

[0004] In determining refractive compensation, some existing techniques rely solely on assessing the total intensity of the interference signal to find the optimal focal point. However, the point of strongest interference signal intensity does not always coincide with the point where the retinal structure is clearest. In cases of minute eye movements by the subject, simply pursuing maximum signal intensity can cause the focal plane to deviate from the target layer, thus affecting the structural resolution of the final image.

[0005] Furthermore, during the few seconds of data acquisition, the subject's eyes undergo involuntary axial and lateral movements. These movements result in displacement and deformation between the acquired image frames. Directly averaging these misaligned images to suppress noise can actually blur structural edges and reduce the effective resolution of the images. Therefore, accurate registration between images is a crucial technical problem that must be solved before improving the signal-to-noise ratio.

[0006] Current imaging workflows are mostly open-loop systems. After completing a data acquisition, the operator needs to visually inspect the generated image to determine if its quality is acceptable. If imaging fails due to factors such as significant movement of the subject or initial focusing inaccuracy, the operator must manually restart the entire process. This process lacks an automatic quality assessment and a closed-loop feedback mechanism for adaptive adjustment after failure, reducing the overall success rate and efficiency of imaging. Summary of the Invention

[0007] The purpose of this invention is to provide a fully automated retinal OCT imaging and image enhancement method, which solves the problems of low imaging efficiency and unstable image quality caused by the reliance on manual intervention, lack of precise autofocus strategy, susceptibility to eye movement interference and lack of automatic quality feedback in existing OCT imaging processes.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] Step S1: Simultaneously acquire images containing the subject's pupils using cameras positioned on both sides of the optical path of the imaging device. Then, input the acquired images into a pre-trained pupil localization model, which processes the images and outputs the two-dimensional coordinates of the pupil center in the camera coordinate system.

[0010] By combining binocular vision calibration parameters, the three-dimensional spatial position of the pupil center in the device's world coordinate system is calculated. Finally, the motor drive system is controlled to adjust the position of the optical probe of the imaging device, aligning the main axis of the emitted beam with the calculated pupil center position.

[0011] Step S2: After the beam is aligned with the center of the pupil, the initial refractive compensation procedure is initiated. This procedure is based on the intensity of the interference spectrum signal. An evaluation function S based on the intensity of the A-scan interference spectrum signal is defined. spec (d), where d represents the current refractive compensation value. The specific form of this function is...

[0012] Possible forms:

[0013]

[0014] In the formula, A p,q (d) is the signal amplitude of the q-th sampling point in the p-th A-scan when the current refractive compensation value is d, where P is the total number of A-scans in a single acquisition, and Q is the number of sampling points in each A-scan. Using a search algorithm, such as hill climbing, starting from an initial compensation value, along the path that makes S... spec (d) Iteratively adjust the current refractive compensation value d in the increasing direction until the function value converges to a local maximum. The compensation value corresponding to this convergence point is determined as a coarse optimal compensation value.

[0015] Step S3: Within the neighborhood of the coarse optimal compensation value, set multiple discrete candidate compensation locations. At each candidate compensation location, rapidly and continuously acquire a short sequence of B-scan images.

[0016] Subsequently, based on this short sequence of B-scan images, two core metrics were calculated:

[0017] 1. Signal strength index: The aforementioned evaluation function S is directly adopted. spec(d) Calculated value at the current candidate compensation position.

[0018] 2. Axial registration stability index: obtained by performing axial registration processing on this short sequence of B-scan images. Based on the above two indices, a joint evaluation function is constructed, for example:

[0019] F joint (d k ) = w spec ·S′ spec (d k )+wF joint (d k ) = w spec ·S′ spec (d k )+w stab ·S′ stab (d k );

[0020] In the formula, d k It is the k-th candidate compensation position, S′ spec (d k ) and S′ stab (d k These are the normalized values ​​of the signal strength index and the axial registration stability index, respectively. spec and w stab These are preset weighting coefficients.

[0021] By calculating and comparing the joint evaluation function values ​​of all candidate compensation locations, the one that makes F... joint (d k The largest candidate position is determined as the final optimal refractive compensation value.

[0022] Step S4: Set the system's refractive compensation to the final optimal refractive compensation value and continuously acquire N frames of B-scan raw images. Then, perform registration-guided denoising processing on these N frames of B-scan raw images. This processing flow includes:

[0023] Axial motion correction: Select one frame as the reference frame, use dynamic programming algorithm to calculate the axial offset of each column A-scan in other frames relative to the reference frame, and correct accordingly.

[0024] Lateral motion correction: Based on axial correction, calculate the global lateral offset of each frame relative to the reference frame and correct accordingly.

[0025] Artifact frame removal: After motion correction, calculate the image similarity score between each frame and the reference frame. Image frames with scores below a preset similarity threshold are identified as artifact frames and removed.

[0026] Image averaging: Pixel alignment and arithmetic averaging are performed on all the remaining corrected valid frames to generate a final image.

[0027] Step S5: Quality assessment and feedback. The quality of the generated final image is evaluated, and a quantified quality metric is calculated. This quality metric is compared with a preset threshold. If the quality metric is greater than or equal to the preset threshold, it is determined that the current imaging is qualified, the process ends, and the final image is output. If the quality metric is less than the preset threshold, it is determined that the current imaging is unqualified, and an adaptive re-trigger process is initiated.

[0028] In a specific embodiment, the method for obtaining the axial registration stability index in step S3 includes: performing axial registration on the short sequence B-scan images, and calculating a statistical value representing the structural stability of the sequence images based on the A-scan offset generated during the axial registration process, which is the axial registration stability index.

[0029] Preferably, the statistical value representing the structural stability of the sequence images is inversely proportional to the degree of dispersion of the A-scan offsets. For example, this statistical value can be defined as the reciprocal of the variance of the set of all A-scan offsets.

[0030] In a specific embodiment, the combined evaluation function in step S3 is a composite evaluation value generated by combining the signal strength index and the axial registration stability index, and this composite evaluation value is used to determine the final optimal refractive compensation value from multiple candidate compensation positions.

[0031] In a specific embodiment, the axial motion correction in step S4 includes: determining an axial offset for each column of A-scan of the image to be corrected to correct the vertical deformation of the image; the lateral motion correction includes determining a global lateral offset for the image to be corrected to correct the horizontal displacement of the image.

[0032] Preferably, the process of determining the global lateral offset in the lateral motion correction includes: using the pixel intensity correlation between images as a similarity metric, such as using the normalized cross-correlation function, and searching for the lateral offset position that makes the similarity metric reach the peak through an iterative adjustment method.

[0033] In a specific embodiment, the steps of artifact frame removal in step S4 include: removing the image frames with scores lower than a preset similarity threshold according to the image similarity scores between the image after motion correction and the reference frame.

[0034] In summary, the present invention includes at least one of the following beneficial technical effects:

[0035] 1. This invention achieves full automation of imaging front-end alignment and focusing, reducing the complexity of operation and dependence on human experience. By adopting a binocular image acquisition module and a pupil positioning model, it realizes automatic identification and tracking alignment of the three-dimensional spatial position of the pupil. Combined with the subsequent automatic refractive compensation process, it replaces the steps that require manual adjustment by the operator in the traditional process, ensuring the repeatability of imaging operation and the consistency of results.

[0036] 2. This invention improves the accuracy of refractive compensation and the signal-to-noise ratio and structural clarity of the final image. It not only evaluates the intensity of interference signals but also introduces the axial registration stability of short-sequence B-scan images as an evaluation index, constructing a joint evaluation function. This method can determine an optimal compensation value that simultaneously ensures signal quality and image structural stability. Combined with motion correction and alignment averaging processing in the main data acquisition stage, it effectively suppresses the blurring effect caused by the physiological movement of the subject.

[0037] 3. This invention establishes a closed-loop feedback and adaptive adjustment mechanism for imaging quality, which improves the success rate and efficiency of imaging. After the final image is generated, it is quantitatively evaluated by a comprehensive quality index, and the evaluation results determine whether to output a qualified image or trigger different adaptive re-triggering processes for different failure reasons. This mechanism avoids manual screening of unqualified images and can automatically adjust the acquisition strategy from failures, thereby improving the final success rate of single imaging without human intervention. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0039] Figure 2 This is a schematic diagram illustrating the automatic pupil positioning principle of the present invention;

[0040] Figure 3 This is a schematic diagram of the synergistic refractive compensation process of the present invention;

[0041] Figure 4 This is a comparison chart of the main data acquisition and registration / denoising effects of the present invention;

[0042] Figure 5 This is a flowchart of the quality assessment and adaptive feedback decision-making process of the present invention. Detailed Implementation

[0043] The following is in conjunction with the appendix Figure 1 - Appendix Figure 5 The present invention will be further described in detail below.

[0044] This invention provides a fully automated retinal OCT imaging and image enhancement method:

[0045] See attached document Figure 1 This invention provides a fully automated retinal OCT imaging and image enhancement method, which is executed on an OCT device configured with specific hardware and processing modules. The OCT device includes:

[0046] The binocular image acquisition module is configured on both sides of the OCT imaging optical path to acquire images including the pupils of the subject.

[0047] The optical probe drive module is used to adjust the position of the imaging beam according to instructions;

[0048] The central processing module is used to execute all the calculation and control logic of the method of the present invention;

[0049] And an OCT imaging module, used to emit scanning light, receive interference signals and acquire data.

[0050] The method provided by the embodiments of the present invention may generally include the following steps performed in sequence:

[0051] Step S1, Automatic Pupil Positioning. This step utilizes the binocular image acquisition module and the central processing module to determine the three-dimensional spatial position of the pupil, and the optical probe drive module completes beam alignment.

[0052] Step S2, preliminary refractive compensation. This step, based on the interference spectral signal acquired by the OCT imaging module, determines a rough optimal compensation value in the central processing module through a search algorithm.

[0053] Step S3, Fine-tuned refractive compensation. This step involves acquiring short sequences of B-scan images and, by combining signal intensity and image registration stability, constructing and optimizing a joint evaluation function in the central processing module to determine the final optimal refractive compensation value.

[0054] Step S4, Master Data Acquisition and Processing. In this step, multiple frames of B-scan images are acquired under the determined optimal compensation value, and registration, artifact removal, and averaging are performed in the central processing module to generate a final image.

[0055] Step S5, Quality Assessment and Feedback. This step involves evaluating the quality of the generated final image in the central processing module and determining whether to output the image or trigger the adaptive re-triggering process based on the evaluation results.

[0056] In the preliminary refractive compensation and fine refractive compensation steps, an evaluation function S based on the intensity of the interference spectral signal is used. spec (d) is used to quantize signal quality, and its form is:

[0057]

[0058] Where d is the current refractive compensation value; A p,q (d) is the signal amplitude of the qth sampling point in the p-th A-scan when the refractive compensation value is d; P is the total number of A-scans in a single acquisition; Q is the number of sampling points for each A-scan.

[0059] In the fine refractive compensation step, a joint evaluation function F joint (d k A method is constructed to determine the optimal solution from multiple candidate locations, and its form is:

[0060] F joint (d k ) = w spec ·S′ spec (d k )+w stab ·S′ stab (d k );

[0061] Where, d k S′ represents the k-th candidate compensation position. spec (d k S' represents the normalized signal strength index; stab (d k ) represents the axial registration stability index after normalization; w spec and w stab These are the preset weighting coefficients.

[0062] See attached document Figure 2 The execution of the method of the present invention begins with step S1, namely automatic pupil localization. This step is performed on an OCT device that integrates a binocular image acquisition module, an optical probe driving module, and a central processing module. The binocular image acquisition module includes two fixed-position cameras symmetrically arranged on both sides of the OCT imaging optical path. When the localization process is initiated, the two cameras synchronously acquire image frames containing the same pupil of the subject and transmit them to the central processing module.

[0063] In the central processing module, a pre-trained pupil localization model receives image frames from the binocular image acquisition module. This model is a convolutional neural network that processes the input 2D image and generates a 2D heatmap. In this heatmap, the point with the highest pixel intensity value corresponds to the predicted position of the pupil center. To obtain sub-pixel accuracy coordinates, a 2D Gaussian function is applied to the peak region of the heatmap to resolve the precise 2D coordinates of the pupil center in both camera views, denoted as (u... A ,v A ) and (u B ,v B ).

[0064] In a more preferred embodiment, the pupil localization model may specifically employ a convolutional neural network based on the U-Net architecture. This network is trained on a large dataset of eye images containing different lighting conditions, pupil features of different ethnicities, and partial occlusion scenarios, enabling it to robustly identify pupils in complex scenes. During training, the network parameters are optimized by minimizing the mean squared error loss function between the network's output heatmap and a pre-labeled Gaussian heatmap with a peak at the pupil center. This approach ensures the model's high accuracy and generalization ability.

[0065] After obtaining the two-dimensional coordinates from the two views, the central processing module calculates the three-dimensional spatial position of the pupil center based on a pre-calibrated binocular vision model. This calculation process employs a triangulation algorithm. The algorithm's input includes two two-dimensional coordinates (u... A ,v A ) and (u B ,v B The internal parameter matrix of each camera and the external parameter matrix representing the relative position and attitude of the two cameras are obtained. By solving, the unique three-dimensional coordinates (X,Y,Z) of the pupil center in the device world coordinate system are obtained.

[0066] Finally, the central processing module compares the calculated target 3D coordinates (X, Y, Z) with the current position of the imaging beam and generates a displacement control command. This command is sent to the optical probe drive module. The optical probe drive module includes a multi-axis motor, which drives the optical elements or the entire probe in the imaging optical path to move according to the received command until the central axis of the OCT imaging beam is aligned with the calculated pupil center 3D coordinates (X, Y, Z), with an alignment error less than a preset threshold. This step provides a stable optical path foundation for subsequent refractive compensation and data acquisition.

[0067] See attached document Figure 3 After completing the automatic pupil localization in step S1, the method proceeds to steps S2 and S3, which together constitute the collaborative refractive compensation stage.

[0068] The method first performs step S2, which is preliminary refractive compensation. The input to this step is the interference spectrum signal from the OCT imaging module. The central processing module changes the refractive compensation value d by adjusting the variable focus lens or moving the reflector in the imaging optical path.

[0069] In a preferred embodiment, the action of "adjusting the variable focal length lens in the imaging optical path" is achieved by controlling an electrically tunable lens (ETL). The ETL rapidly and without mechanical vibration changes its focal length by altering the current or voltage applied to it, thereby enabling continuous scanning of the refractive compensation value. Compared to traditional mechanically moving mirrors, this method offers advantages such as fast response speed, no noise, and long lifespan, making it particularly suitable for the initial refractive compensation stage requiring rapid iterative searching.

[0070] Evaluation function S spec (d) The physical meaning is that when the focal plane of the OCT system is precisely conjugate with the photoreceptor layer of the retina, the light flux reflected from the retina is maximized, resulting in the overall amplitude of the interference spectral signal generated in the interferometer also reaching its peak. Therefore, for S spec The peak-finding process of (d) physically corresponds to the process of finding the optimal focus.

[0071] For each refractive compensation value d, an interference spectrum signal is acquired once, and the evaluation function S is calculated. spec The value of (d):

[0072]

[0073] Where d is the current refractive compensation value; A p,q (d) is the signal amplitude of the qth sampling point in the p-th A-scan when the refractive compensation value is d; P is the total number of A-scans in a single acquisition; Q is the number of sampling points for each A-scan.

[0074] The central processing module uses a hill-climbing method to evaluate the function S. spec (d) Optimize. Starting from a preset initial compensation value d... start Initially, the search is performed in the neighborhood of the compensation value d with a fixed coarse adjustment step size Δd, and then moves towards S. spec (d) Increased directional movement.

[0075] The initial compensation value d start The diopter can be set to 0, and the search range for the compensation value can be limited to an interval that conforms to the refractive state of most people's eyes, such as -15D to +15D. Furthermore, to further improve optimization efficiency, the coarse adjustment step size Δd can also adopt an adaptive adjustment strategy. For example, a larger step size can be used in the early stages of the search to quickly approach the peak region, and the step size can be automatically reduced to perform a more refined search when the incremental change of the function value slows down. The iteration stops when the function values ​​at adjacent positions on both sides of the compensation value d are both less than the function value at the current position. The compensation value at this point is determined as the coarsely optimal compensation value d. coarseo .

[0076] Subsequently, the method performs step S3, namely fine refractive compensation. This step approximates the optimal compensation value d. coarse Within a predefined neighborhood, K candidate compensation positions {d1, d2, ..., d3} are selected with a fine-tuning step size δd smaller than Δd. K}. At each candidate compensation position d k Above, the OCT imaging module rapidly and continuously acquires a short sequence of B-scan images containing M frames, denoted as C. k ={B k,1 B k,2 ,...,B k，M}

[0077] For each short sequence image set C k The central processing module calculates two metrics. The first is the signal strength metric, whose value is S. spec (d k The second is the axial registration stability index S. stab (d k The calculation process is as follows:

[0078] With the first frame B of the sequence k，1 As the reference frame, for each other frame B in the sequence k，m (m=2,…,M), an axial offset vector s is calculated using a dynamic programming algorithm. k，m Each element of this vector corresponds to an A-scan, recording its axial displacement relative to the corresponding A-scan of the reference frame. Collecting all elements of all offset vectors in this sequence forms an offset set. Axial registration stability index S stab (d k The variance of the offset set is defined as the reciprocal of the variance of that offset set.

[0079]

[0080] The value of this index is inversely proportional to the degree of dispersion of the A-scan offset. This is a union operation. The operation will take the set of all offsets from frame 2 to frame M (s... k，2 ,s k，3 ,...,s k,M This data is then merged into a single, larger set. This final large set contains the data for the candidate compensation value d. k Below, the axial offsets of all A-scans across all non-reference frames; Var(·) is the standard variance calculation function, used to measure the dispersion or variability of all offset values ​​in the final large set mentioned above; s k,m Denotes a set containing values ​​with refractive compensation of d. kUnder the given conditions, the axial offset of the m-th frame (m ranging from 2 to M) in the acquired short sequence of images relative to all A-scans (image columns) of the first frame (reference frame); M is the total number of frames in the short sequence of B-scan images used for evaluation. For example, M can be set to 10.

[0081] The axial registration stability index S stab (d k The underlying mechanism is that when the OCT system is well focused, the edges of the various retinal structures (such as the ganglion cell layer, inner and outer plexiform layers, etc.) in the acquired Bscan image are sharp and the features are obvious.

[0082] Therefore, in continuously acquired short image sequences, even with minute ocular nystagmus, the image registration algorithm can stably identify the same structural features and accurately calculate their axial displacement in each frame, resulting in a very small ensemble variance of the displacement. Conversely, if the system is out of focus and the image structure is blurred, the registration algorithm will experience greater uncertainty in feature identification, leading to larger fluctuations in the calculated axial displacement between frames, and a corresponding increase in the ensemble variance.

[0083] Therefore, S stab (d k It can very sensitively reflect the structural clarity of an image, and is a better indicator than the simple signal strength index S. spec (d k A key and effective supplement to ).

[0084] Next, the central processing module constructs a joint evaluation function F. joint (d k First, for the set of signal strength indices {S} at all candidate locations... spec (d k )} and the set of axial registration stability indices {S stab (d k Normalize the values ​​from maximum to minimum for each index, resulting in the normalized index set {S′}. spec (d k )} and {S′ stab (d k The joint evaluation function is obtained by weighted summation of the two normalized indicators:

[0085] F joint (d k ) = w spec ·S′ spec (d k )+w stab ·S′ stab (d k );

[0086] Among them, w spec and w stab These are preset weighting coefficients that sum to 1. S′ spec (d k S′ is the normalized signal strength index. stab (d k ) is the normalized axial registration stability index.

[0087] In a specific embodiment, the weighting coefficient w spec and w stab The value of w can be adjusted according to the clinical application scenario of the device. For example, when used for screening or imaging of subjects with good cooperation, w can be appropriately increased. spec The weighting of w should be adjusted to prioritize obtaining the strongest signal response. However, when imaging subjects with poor fixation or nystagmus, the weighting should be increased. stab The weights are prioritized to ensure the structural stability of the image, even if this sacrifices some signal strength. These two weights can also be automatically adjusted by the system based on the initial eye movement assessment, thus achieving a higher level of intelligence.

[0088] The central processing module calculates F for all K candidate compensation positions. joint (d k The value is calculated, and the candidate compensation position that maximizes this function value is selected, which is then determined as the final optimal refractive compensation value d. opt This value is used in subsequent master data acquisition steps.

[0089] See attached document Figure 4 After completing step S3 and determining the final optimal refractive compensation value d opt Then, the method proceeds to step S4.

[0090] In this step, the central processing module first instructs the OCT imaging module to set the system's refractive compensation to d. opt And continuously acquire a B-scan raw image sequence containing N frames, denoted as {B1, B2, ..., B...} N The image sequence is then fed into the central processing module for registration-guided denoising.

[0091] The process first performs axial motion correction. The first frame B1 in the sequence is selected as the reference frame B. ref For any frame of the sequence, image B to be corrected i (i = 2, ..., N), a dynamic programming algorithm is applied to correct its axial motion deformation. This algorithm is for image B. i For each column of A-scan, find a value relative to the reference frame B. refThe optimal axial offset corresponding to the A-scan is determined to maximize the structural similarity between the corrected image and the reference frame. After this process, an image sequence {B′1, B′2, ..., B′} that has undergone only axial correction is obtained. N}, where B′1 is B1.

[0092] Next, lateral motion correction is performed on the axially corrected image sequence. This process aims to correct the motion of each frame B. i ′ Relative to reference frame B ref The global horizontal displacement is calculated by the normalized cross-correlation function C. ncc (h i To measure the correlation of pixel intensity between images:

[0093]

[0094] Where: C ncc (h i ) is the image B to be corrected i ′ The lateral offset being evaluated is h. i Time and reference frame B ref similarity score, h i This is the lateral offset being evaluated, B. ref (x, y) and B i ′ (xh i , y) are the intensity values ​​at pixel coordinates (x, y) of the reference frame and the offset image to be corrected, respectively. and These are the average pixel intensity values ​​of the two images within the overlapping region; B i ′ (xh i ,y) represents the lateral offset h being evaluated. i The pixel intensity value of the image to be corrected at coordinates (x, y).

[0095] This part calculates the covariance of two images (the reference image and the offset image to be corrected); This part is the product of the standard deviations of the pixel intensity of the two images (strictly speaking, the unnormalized standard deviations).

[0096] This function normalizes the covariance of image pixel intensities by dividing the product of their respective standard deviations; its range is [-1, 1], and it is used to evaluate the degree of linear correlation between two images. To determine how C... ncc (h i The optimal lateral offset h to reach the peak value i,optA hill-climbing method is used for iterative search to avoid the large amount of computation required for a global traversal search.

[0097] To further improve robustness to rotational motion during lateral motion correction, a phase correlation method can be introduced. This method utilizes the translation properties of the Fourier transform to efficiently calculate not only lateral and longitudinal translations but also the rotation angles between frames, thus providing more comprehensive motion compensation for the image. After correcting all motion parameters, calculating the normalized cross-correlation function yields a more accurate similarity score, making artifact frame removal more reliable.

[0098] Determine the optimal lateral offset h for each frame i,opt and its corresponding maximum similarity score C ncc (h i,opt After that, artifact frame removal is performed. A preset similarity threshold τ is set. ncc For each frame of an image, if its maximum similarity score C... ncc (h i,opt (less than τ) ncc If the frame is not identified as an artifact frame, it will be removed from the sequence.

[0099] Finally, image averaging is performed on all the valid frames that passed the screening. These valid frames are then aligned according to their respective axial and lateral correction parameters to obtain a fully registered set of valid frames {B}. i " |i∈V}, where V is the set of indices for all valid frames. Final image B final The following is obtained by taking the pixel-level arithmetic mean of all images in this set:

[0100]

[0101] Among them: B final (x,y) is the intensity value of the final generated image at pixel coordinates (x,y), N valid It is the total number of valid frames, B i " (x,y) is the intensity value of the i-th valid and fully registered image at pixel coordinates (x,y). This is a scaling factor of the summation result, representing the division by the total number of valid frames. This step, through averaging, significantly reduces random noise in the image, thereby generating a final image with a high signal-to-noise ratio.

[0102] See attached document Figure 5 In step S4, the final image B is generated. final Then, the method proceeds to step S5, where the central processing module performs a quality assessment on the image.

[0103] This step aims to quantify the imaging quality of the final image to determine whether it meets the output criteria. To this end, a comprehensive quality index Q is calculated. final This indicator is obtained by weighted summation of multiple sub-indicators:

[0104] Q final =w valid ·R valid +w struct ·C struct ;

[0105] Among them: Q final It is the final calculated comprehensive quality index, R. valid It is the effective frame rate metric, whose value is the number of effective frames N used to generate the final image. valid The ratio of the total number of acquisition frames N in the main data acquisition phase, i.e., R valid =N valid / N. This metric reflects the stability of the data acquisition process, C struct It is an indicator of structural continuity.

[0106] This indicator is obtained by analyzing the final image B. final Edge detection algorithms (such as the Sobel operator) are applied for calculation. In the extracted edge map, the contours of key layered structures such as the retinal pigment epithelium (RPE) are identified, and a value is calculated based on the integrity of the contour.

[0107] Regarding the structural continuity index C struct A specific calculation method is as follows:

[0108] First, for the final image B final Image segmentation is performed to automatically identify and extract the set of pixel coordinates for the retinal pigment epithelium (RPE). Then, connected component analysis is performed on this pixel set in the lateral dimension. Ideally, a healthy, high-quality RPE layer should appear as a continuous curve with almost no breaks across the entire image.

[0109] Therefore, C struct This can be quantified as the ratio of the lateral span of the largest connected component belonging to the RPE layer to the total width of the image. The closer this ratio is to 1, the more complete and continuous the RPE layer structure, and the higher the image quality. Compared to simple edge detection, this method is more robust and accurate in evaluating the structural validity of images.

[0110] For example, this value can be directly proportional to the total pixel length of the contour, or inversely proportional to the number of breakpoints in the contour. This metric reflects the structural sharpness of the image, w valid and w structThese are preset weighting coefficients assigned to the effective frame rate metric and the structural continuity metric, respectively.

[0111] The central processing module will calculate the comprehensive quality index Q final With a preset quality threshold Q thresh Compare them. If Q final ≥Q thresh If the image is deemed satisfactory, the process terminates, and the final image B is generated. final As a result, output or store. If Q final thresh If the image is not captured, the imaging is deemed unqualified, and the adaptive re-triggering process is triggered.

[0112] In one specific embodiment, when the adaptive re-trigger process is triggered, the central processing module further analyzes each sub-indicator to determine the reason for the non-compliance. If the analysis concludes that the main reason for the quality index being less than the preset threshold is the effective frame rate R... valid If the value is too low, the system determines the cause to be excessive eye movement by the subject. In this case, the system increases the total number of frames N for the next main data acquisition, for example, by setting a new acquisition frame count N. new = N + ΔN, where ΔN is a preset number of incremental frames. Then, the method returns to step S3, using the original optimal refractive compensation value to re-execute the subsequent process, aiming to obtain a sufficient number of valid frames from more acquired samples.

[0113] In another specific embodiment, when the analysis yields the effective frame rate R valid Within acceptable limits, but the structural continuity index C... struct If the overall quality index is too low, resulting in a value below the preset threshold, the system determines the cause as poor initial refractive compensation. In this case, the system will discard all currently acquired data and return to step S1 to re-execute the complete front-end alignment and focusing process, including initial and fine refractive compensation, starting from automatic pupil positioning.

[0114] In addition, there exists an intermediate state. When system analysis reveals an effective frame rate R... valid and structural continuity index C struct When all values ​​are only slightly below their respective ideal thresholds, resulting in an unqualified overall quality index, the system can determine that there is a slight deviation in the fine refractive compensation value. In this case, the system does not need to completely return to step S1, but can return to step S3, using the currently determined optimal refractive compensation value d. opt Centered on the target, use a smaller fine-tuning step size δd ′ An additional, finer refractive compensation search is performed. This more targeted feedback strategy maximizes time savings while maintaining image quality, further improving the overall efficiency of the fully automated process.​

Claims

1. A fully automated retinal OCT imaging and image enhancement method, characterized in that, Includes the following steps: S1. Images containing the pupil are acquired by cameras on both sides of the imaging device, the three-dimensional spatial position of the pupil center is determined by the pupil positioning model, and the beam of the imaging device is controlled to be aligned with the pupil center. S2. Based on the intensity of the interference spectrum signal, a rough optimal compensation value is determined through a search algorithm; S3. Select multiple candidate compensation positions within the neighborhood of the coarse optimal compensation value, and acquire short-sequence B-scan images at each candidate compensation position; perform axial registration on the short-sequence B-scan images, and calculate a statistical value characterizing the structural stability of the sequence images based on the A-scan offset generated during the axial registration process, using the statistical value as an axial registration stability index; generate a joint evaluation function as a composite evaluation value based on the signal intensity index of each candidate compensation position and the axial registration stability index, and determine the final optimal refractive compensation value according to the joint evaluation function; S4. Set the final optimal refractive compensation value and acquire N frames of B-scan raw images; The N frames of B-scan original images are subjected to registration-guided denoising processing, which includes axial motion correction, lateral motion correction, artifact frame removal, and image averaging, thereby generating a final image. S5. Perform a quality assessment on the final image to obtain a quality index, and compare the quality index with a preset threshold. If the quality index is greater than or equal to the preset threshold, the imaging is deemed qualified and the final image is output; if the quality index is less than the preset threshold, the imaging is deemed unqualified and an adaptive re-triggering process is triggered.

2. The fully automated retinal OCT imaging and image enhancement method according to claim 1, characterized in that, The statistical value representing the structural stability of the image sequence is inversely proportional to the degree of dispersion of the A-scan offset.

3. The fully automated retinal OCT imaging and image enhancement method according to claim 1, characterized in that, In step S4, the axial motion correction includes: An axial offset is determined for each column of the A-scan of the image to be corrected, thereby correcting the vertical distortion of the image; The lateral motion correction includes determining a global lateral offset for the image to be corrected, thereby correcting the horizontal displacement of the image.

4. The fully automated retinal OCT imaging and image enhancement method according to claim 3, characterized in that, The process of determining the global lateral offset in the lateral motion correction includes: The correlation of pixel intensity between images is used as a similarity metric, and the horizontal offset position that makes the similarity metric reach its peak is searched through iterative adjustment.

5. The fully automated retinal OCT imaging and image enhancement method according to claim 1, characterized in that, In step S4, the artifact frame removal process includes: Based on the similarity score between the motion-corrected image and the reference frame, image frames with scores lower than a preset similarity threshold are removed.

6. The fully automated retinal OCT imaging and image enhancement method according to claim 1, characterized in that, In step S5, the quality indicators include at least one of the following: The proportion of the effective frames used to generate the final image to the total number of acquired frames N; The continuity index of the retinal layered structure in the final image.

7. The fully automated retinal OCT imaging and image enhancement method according to claim 1, characterized in that, The adaptive re-triggering process includes: When the reason for the unqualified imaging is determined to be excessive eye movement of the subject, the total number of frames N of the next N-frame B-scan original images is increased, and the process returns to step S3.

8. The fully automated retinal OCT imaging and image enhancement method according to claim 1, characterized in that, The adaptive re-triggering process includes: If the reason for the unqualified imaging is that the initial refractive compensation is poor, return to step S1 and repeat the initial refractive compensation and fine refractive compensation.

Images