A registration method based on three-dimensional retinal OCT images

Abstract

The present invention discloses a registration method based on three-dimensional retinal OCT images, which relates to the technical field of OCT image registration, and includes the following steps: in the registration process of OCT images, firstly, the anatomical feature points of the retina are used as the key reference for basic alignment to perform preliminary registration to eliminate global mismatch errors; after the preliminary registration is completed, the OCT image is divided into a number of sub-areas with equal areas, so that each sub-area is used as an independent analysis unit, which is convenient for subsequent accurate analysis of the local optical characteristics of each analysis unit. The present invention eliminates global errors through preliminary registration of retinal anatomical feature points, combines image segmentation with abnormal signal extraction, uses a machine learning model to intelligently identify healthy and scar tissue areas, and dynamically adjusts the weight to prioritize healthy areas, thereby reducing the interference of scars on registration, improving registration reliability, and reducing clinical misjudgments, providing technical support for accurate diagnosis and treatment of retinal diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of OCT image registration, and in particular to a registration method based on three-dimensional retinal OCT images. Background Art

[0002] Three-dimensional retinal OCT image registration refers to spatially aligning three-dimensional optical coherence tomography (OCT) images acquired at different times or angles so that the anatomical structures of these images overlap accurately, thereby accurately monitoring the dynamic changes of retinal tissue, lesion progression or treatment effects. Since the acquisition of retinal images may be affected by eye movement, acquisition angle deviation and patient posture changes, there is a spatial misalignment between different scan data.

[0003] To solve these problems, the registration algorithm extracts key feature points or texture information of the retinal layer, and uses rigid transformation, non-rigid transformation or deep learning-based registration technology to align OCT images at different time points in three-dimensional space, thereby accurately comparing important indicators such as changes in retinal thickness, changes in the location of lesion areas, and changes in vascular morphology. This process plays an important role in the diagnosis, progression tracking and efficacy evaluation of retinal diseases, and is particularly suitable for the long-term tracking and management of lesions such as diabetic retinopathy and macular degeneration.

[0004] The prior art has the following deficiencies:

[0005] During the registration of three-dimensional OCT images, when scar tissue forms in the retinal lesion area (such as epiretinal membrane, fibrosis caused by diabetic retinopathy, or scars after retinal tear repair), the registration algorithm may be disturbed by the abnormal reflection signal of the scar tissue, resulting in scar position drift or shape deformation in the images collected at different time points (the reflection signal of the scar tissue is abnormally strong, and it is easy to be mistaken by the algorithm as a key feature point during the registration process, resulting in scar position drift or shape deformation in the registration results). This situation can easily cause doctors to make serious misjudgments when evaluating the progression of the lesion or judging the treatment effect. For example, when the registration algorithm mistakenly marks the scar tissue as a new location, the doctor may mistakenly believe that the scar tissue has proliferated or the condition has worsened, and thus arrange for the patient to undergo unnecessary treatment (such as laser surgery or anti-VEGF injection), which may cause permanent damage to healthy retinal tissue. On the other hand, if the reflection signal of the scar tissue is mistakenly weakened in a certain scan, the registration result shows that the scar has shrunk or disappeared, and the doctor may underestimate the condition and delay the treatment, which will eventually lead to the patient's condition worsening, or even serious consequences such as retinal detachment or blindness.

[0006] The following are examples:

[0007] For example, a patient with diabetic retinopathy undergoes OCT scans at different time points, and the doctor hopes to track changes in scar tissue in the macula by registering the images. In the first scan, the scar tissue is 2.3 mm² and is located to the left of the center of the macula; in the second scan (3 months later), the registered image shows that the scar is 3.1 mm² and has changed its position to the lower right corner. Based on the registered images, the doctor may judge that the patient's scar area has increased and drifted, which is usually considered to be a sign of worsening disease. However, the actual situation may be that the registration algorithm fails to accurately identify the image deviation caused by the patient's slight eye movement or light reflection changes between the two scans, causing the scar tissue to drift and become larger in the registration result. This misregistration may mislead the doctor to schedule the patient for unnecessary laser treatment, which may further damage healthy retinal tissue, causing permanent vision loss or even serious retinal damage.

[0008] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0009] The purpose of the present invention is to provide a registration method based on three-dimensional retinal OCT images, which uses the anatomical feature points of the retina for preliminary registration, eliminates global mismatch errors, and provides a stable basis for subsequent local analysis; by dividing the image into equal-area sub-areas and extracting abnormal reflection signals, combined with a machine learning model to achieve intelligent classification of healthy tissue and scar tissue, thereby laying a scientific basis for dynamic weight allocation; in the registration process, the weights of healthy tissue and scar tissue are dynamically adjusted, so that the registration algorithm relies on healthy areas first, reduces the drift and deformation caused by scar tissue, and ensures that the scar area will not become an erroneous reference point, which not only improves the reliability of the registration results, but also reduces the risk of clinical misjudgment, and provides a solid technical guarantee for the accurate diagnosis and treatment of retinal lesions to solve the problems in the above-mentioned background technology.

[0010] In order to achieve the above object, the present invention provides the following technical solution: a registration method based on three-dimensional retinal OCT images, comprising the following steps:

[0011] In the registration process of OCT images, the anatomical feature points of the retina are first used as the key reference for basic alignment to perform preliminary registration and eliminate global mismatch errors;

[0012] After the initial registration is completed, the OCT image is divided into several sub-areas of equal area, so that each sub-area is used as an independent analysis unit, which facilitates the subsequent accurate analysis of the local optical characteristics of each analysis unit;

[0013] For each divided sub-region, obtain the optical reflection characteristic information inside it, and pre-process the acquired data to extract the abnormal reflection signal. Then, quantify the abnormal signal and convert the abnormal reflection characteristics of each sub-region into numerical features to generate a standardized data set as the input basis of the machine learning model.

[0014] The quantitative features extracted from each sub-region are input into a pre-trained machine learning model, and each sub-region is intelligently identified through the model. According to the recognition results of the machine learning model, each sub-region is automatically divided into two categories: healthy tissue area and scar tissue area;

[0015] During the registration process, dynamic weights are assigned to different sub-regions. For the healthy tissue area, a high registration weight is set as the reference area for registration; for the scar tissue area, a dynamically variable low weight is set to avoid the registration algorithm from over-relying on abnormal reflection signals and reduce the interference of scar tissue on the registration results.

[0016] Preferably, the anatomical feature points of the retina are used as key references for basic alignment to perform preliminary registration, and the specific steps are as follows:

[0017] In the preliminary registration of OCT images, the anatomical feature points of the retina are first automatically identified by an algorithm;

[0018] After identifying the feature points, the spatial relationship between the anatomical feature points is calculated, including position coordinates, distance, angle and shape characteristics. The translation deviation is adjusted by comparing the optic disc position, the rotation deviation is corrected by the direction of the vascular intersection, and the scaling ratio is adjusted by the length of the retinal layer.

[0019] Finally, the calculated spatial transformation parameters are applied to preliminarily align the images at different time points, and the pixel coordinates of the images are remapped through the affine transformation matrix to ensure that the anatomical feature points overlap the mismatch errors in the images, providing a basis for subsequent refined registration.

[0020] Preferably, after the preliminary registration is completed, a fixed grid division method is used to divide the OCT image into several sub-areas with equal areas.

[0021] Preferably, an abnormal reflection signal is extracted from the acquired optical reflection characteristic information, wherein the extracted abnormal reflection signal includes a high-contrast reflection peak and an irregular spectral response. For each sub-region, the extracted high-contrast reflection peak and the irregular spectral response are quantized to generate a reflection peak abnormal sharpness factor and a spectral nonlinearity factor, respectively. A standardized data set is generated through the reflection peak abnormal sharpness factor and the spectral nonlinearity factor to provide accurate data input for the machine learning model for identifying scar areas.

[0022] Preferably, the abnormal sharpness factor of the reflection peak and the spectral nonlinear factor extracted from each sub-region are input into a pre-trained machine learning model, and a scar risk assessment index is output based on the machine learning model. The scar risk in the sub-region is intelligently assessed through the scar risk assessment index, and then the scar tissue in each sub-region is intelligently identified.

[0023] Preferably, the scar risk assessment index generated by intelligently assessing the scar risk in the sub-regions through the pre-trained machine learning model is compared and analyzed with the pre-set scar risk reference threshold, and each sub-region is automatically divided. The specific division steps are as follows:

[0024] If the scar risk assessment index is greater than the scar risk reference threshold, the sub-region is divided into a scar tissue region;

[0025] If the scar risk assessment index is less than or equal to the scar risk reference threshold, the sub-region is divided into a healthy tissue region.

[0026] Preferably, for each sub-region, the specific steps of quantizing the extracted high-contrast reflection peak to generate the reflection peak abnormal sharpness factor are as follows:

[0027] The peak value of the optical reflection signal in the sub-area is detected to identify the significant peak in the reflection signal. The local maximum detection algorithm is used to extract all the peak points in the reflection signal. Then, the sharpness is measured by calculating the local contrast difference around each peak. The local contrast difference reflects the prominence of the peak relative to the adjacent area and is quantified by the following formula: ,in: The peak point p The local contrast difference, The peak point p The reflection intensity value, The peak point p Radius on both sides r The minimum reflection intensity value within the range, The peak point p Radius on both sides r The maximum reflection intensity value within the range, p Indicates the location of the peak point, that is, the specific pixel position identified as the reflection peak in a given sub-region;

[0028] After obtaining the local contrast difference of each peak, the reflection peak abnormal sharpness factor is further calculated by the local contrast difference of each peak to quantify the abnormal sharpness of the peak. The calculation expression of the reflection peak abnormal sharpness factor is: ,in: is the reflection peak abnormal sharpness factor,P represents all the peak points detected in the sub-region, It is the median of the local contrast differences of all reflection peaks in the sub-region, reducing the impact of outliers or extreme values. The absolute difference in the denominator quantifies the degree of deviation of the reflection peak from the median value of the sub-region, enhancing the sensitivity to extreme sharp peaks.

[0029] Preferably, for each sub-region, the specific steps of quantizing the extracted irregular spectral response to generate the spectral nonlinear factor are as follows:

[0030] Get the spectral response data in the sub-region, where Represents the wavelength of the spectrum, and uses the high-order difference method to calculate the rate of change of the spectral signal to quantify its nonlinear characteristics. The wavelength The spectral response at Higher-order differences , the calculation expression is: ,in, n is the total length of the spectral data collected in the sub-region, i Represents the index in the spectral data, indicating the position of the current wavelength point in the sequence. For the i The position of the wavelength, k represents the order of higher-order differences, Indicated in wavelength The spectral response value at Indicated in wavelength The spectral response value at u is the offset between wavelength points, indicating the interval between wavelength points used to calculate the difference;

[0031] The calculated high-order difference values ​​are synthesized, and the spectral nonlinear factor is generated by introducing a relative nonlinear metric. The spectral nonlinear factor quantifies the nonlinear characteristics of the spectral response by comparing the change relationship between the high-order difference and the original spectral response. The generated expression is: ,in, represents the spectral nonlinear factor, Representative i The square of the high-order difference at a wavelength reflects the nonlinear change intensity of the spectral signal at that wavelength position, and the denominator is the energy of the original signal, ensuring that the generated spectral nonlinear factor is consistent under different spectral signal intensities.

[0032] Preferably, in the registration process, dynamic weights are assigned to different sub-regions, and the specific steps are as follows:

[0033] First, initialize the regional dynamic weight distribution to ensure that the healthy tissue area obtains a high weight. The calculation expression for the initial weight distribution is: ,in, is the registration weight of the f-th sub-region, reflecting the importance of the sub-region in the registration process, is the scar risk assessment index of the f-th sub-region, generated by the machine learning model. is the weight reduction control factor, which is used to adjust the attenuation degree of scar risk on weight. It is a weight smoothing control factor to avoid registration instability caused by too low or too high weight;

[0034] The registration objective function is adjusted based on the regional dynamic weight. The dynamic weight is introduced into the registration objective function to emphasize the healthy tissue area and weaken the influence of the scar area. The adjustment expression of the registration objective function is: ,in, is the weighted error of the registration objective function, which is used to measure the registration deviation between the template image and the reference image. N is the total number of sub-regions in the image, For the f The difference metric function of the sub-region represents the registration error between the template image and the reference image in the sub-region;

[0035] The weights are continuously updated during the iterative registration process, and the weight distribution is further optimized according to the real-time registration error. Dynamic iterative optimization and weight update are performed. The specific expression is: ,in, Indicates f The sub-region is The weight value at the iteration, that is, the weight value after dynamic adjustment during the image registration iteration process, For the f The weight value of the sub-region at the tth iteration, To dynamically update the step size factor, it is used to control the magnitude of the weight adjustment. For the f The registration error gradient of each sub-region reflects the sensitivity of the current weight to the registration objective function.

[0036] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0037] The present invention uses the anatomical feature points of the retina for preliminary alignment, eliminating global mismatch errors and providing a stable foundation for subsequent local analysis; by dividing the image into equal-area sub-areas and extracting abnormal reflection signals, the intelligent classification of healthy tissue and scar tissue is realized in combination with a machine learning model, thereby laying a scientific basis for dynamic weight allocation; in the alignment process, the weights of healthy tissue and scar tissue are dynamically adjusted so that the alignment algorithm relies on healthy areas first, reducing the drift and deformation caused by scar tissue, and ensuring that the scar area does not become an erroneous reference point, which not only improves the reliability of the alignment results, but also reduces the risk of clinical misjudgment, providing a solid technical guarantee for the accurate diagnosis and treatment of retinal diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0039] Figure 1 The present invention is a method flow chart of a registration method based on three-dimensional retinal OCT images. DETAILED DESCRIPTION

[0040] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0041] The present invention provides Figure 1 A registration method based on three-dimensional retinal OCT images is shown, comprising the following steps:

[0042] In the registration process of OCT images, the anatomical feature points of the retina (such as retinal layer boundaries, vascular intersections, optic disc positions, etc.) are first used as key references for basic alignment to perform preliminary registration and eliminate global mismatch errors;

[0043] These anatomical feature points usually remain relatively stable in healthy and diseased retinas and are therefore reliable registration reference points. The purpose of preliminary registration is to eliminate global mismatch errors, such as overall position offsets caused by patient eye movement, scanning angle changes, etc.

[0044] Use the anatomical feature points of the retina (such as retinal layer boundaries, vascular intersections, optic disc locations, etc.) as key references for basic alignment for preliminary registration. The specific steps are as follows:

[0045] In the initial registration of OCT images, the algorithm is first used to automatically identify the anatomical feature points of the retina, such as the boundaries of retinal layers, vascular intersections, and the position of the optic disc. These feature points usually remain stable in images at different time points and are reliable reference points in the registration process. For example, the retinal pigment epithelium (RPE) is a clearly identifiable boundary layer, the vascular intersection has a fixed position and unique morphology, and the optic disc is the most prominent anchor point.

[0046] The algorithm extracts these feature points through edge detection, shape recognition and other methods to provide a stable initial framework for registration, effectively avoiding mismatching caused by interference from local lesions in the early stages.

[0047] After identifying the feature points, the spatial relationship between the anatomical feature points is calculated, including position coordinates, distances, angles, and shape features. Since images at different time points may undergo translation, rotation, and scale changes due to eye movement or changes in scanning angles, the algorithm needs to analyze these geometric transformation parameters. The translation deviation is adjusted by comparing the optic disc position, the rotation deviation is corrected by the direction of the vascular intersection, and the scaling is adjusted by the length of the retinal layer. This step uses a rigid registration algorithm or an affine transformation model to ensure that the images remain aligned and consistent on a global scale.

[0048] Finally, the calculated spatial transformation parameters (such as translation, rotation, and scaling) are applied to preliminarily align the images at different time points. The pixel coordinates of the image are remapped through the affine transformation matrix to ensure that the anatomical feature points overlap in the image. This process maintains the clarity and continuity of the image through image resampling (such as bilinear interpolation). Preliminary registration can effectively reduce the overall mismatch error and provide a basis for subsequent refined registration.

[0049] After the initial registration is completed, the OCT image is divided into several sub-areas of equal area, so that each sub-area is used as an independent analysis unit, which facilitates the subsequent accurate analysis of the local optical characteristics of each analysis unit;

[0050] The purpose of dividing the sub-regions is to refine the registration units and reduce the local drift and mismatch that may occur during the overall registration, especially in areas with large changes in scar tissue. The size of the sub-regions can be adjusted according to actual needs, and a fixed grid division method is usually used.

[0051] The fixed grid division method is used to divide the image into rectangular or square sub-regions of uniform size. This method is simple and intuitive, easy to implement, and can perform independent analysis and processing in each sub-region of the image. For example, the image is divided into a 10×10 grid, each grid unit is a sub-region, and the algorithm will extract and analyze the optical reflection characteristics in each grid. The advantage of fixed grid division is that it evenly covers the entire image, ensuring that each area can be processed to avoid omissions.

[0052] It should be noted that the area of ​​the sub-region is small. The core function of dividing the OCT image into smaller sub-regions is to improve the recognition accuracy of the scar tissue area, thereby reducing the error caused by abnormal reflection signals during the registration process. Scar tissue usually manifests as high reflection signals, irregular boundaries and dynamic changes, and these features are often concentrated in local areas. If the sub-region area is too large, the characteristics of the scar tissue may be averaged or diluted by the signals of the surrounding healthy tissue, resulting in a decrease in recognition accuracy. Smaller sub-regions can more accurately capture the boundary changes and signal intensity fluctuations of scar tissue, thereby avoiding mismatches or missed judgments. At the same time, this refined division can help the algorithm identify local abnormal signals more sensitively, improve the classification accuracy of scar areas by machine learning models, and ensure that the changing trends of lesion areas are more reliably tracked and analyzed.

[0053] For each divided sub-region, obtain the optical reflection characteristic information inside it, and pre-process the acquired data to extract the abnormal reflection signal. Then, quantify the abnormal signal and convert the abnormal reflection characteristics of each sub-region into numerical features to generate a standardized data set as the input basis of the machine learning model.

[0054] For each divided sub-region, the optical reflection characteristic information inside it can be obtained in the following ways:

[0055] First, the signal intensity analysis calculates the intensity of the reflected light in the sub-region to obtain the intensity distribution of the optical reflection signal in the region as an indicator of the reflection characteristics. Secondly, spectral analysis is used to capture the changes in the reflection signal at different wavelengths by extracting the spectral data of the sub-region (such as through frequency domain or time domain analysis) to identify the optical characteristics of different tissues or lesion areas. In addition, image texture analysis methods such as grayscale co-occurrence matrix or local binary pattern (LBP) can be used to extract subtle changes in the reflection signal from the texture characteristics of the image to further understand the tissue characteristics of the region.

[0056] The process of preprocessing the acquired optical reflection characteristic information mainly includes the following steps: denoising, by applying methods such as Gaussian filtering and wavelet transform, to remove noise and unnecessary interference signals in the image to ensure the purity of the signal; normalization, converting the optical reflection characteristic data of each sub-area into a standardized numerical range to eliminate the influence of factors such as lighting and equipment differences between different image scans, which is convenient for subsequent analysis and comparison; smoothing, by using mean or median filtering methods, smoothing out high-frequency fluctuations and reducing the interference of unstable factors on data analysis. The main function of preprocessing is to provide clean, stable, and standardized data for subsequent abnormal reflection signal analysis, so that the subsequent machine learning model can accurately identify the difference between scar tissue and healthy tissue, and improve the registration accuracy and reliability.

[0057] Abnormal reflection signals are extracted from the acquired optical reflection characteristic information, wherein the extracted abnormal reflection signals include high-contrast reflection peaks and irregular spectral responses. For each sub-region, the extracted high-contrast reflection peaks and irregular spectral responses are quantified to generate reflection peak abnormal sharpness factors and spectral nonlinear factors, respectively. A standardized data set is generated through the reflection peak abnormal sharpness factors and spectral nonlinear factors to provide accurate data input for the machine learning model for identifying scar areas.

[0058] When the reflection peak contrast in the optical reflection characteristics within a sub-region is high, it usually indicates that the sub-region has a higher risk of scar tissue. Scar tissue has significantly different optical properties from normal retinal tissue due to fibrosis or collagen deposition. The reflection signal of scar tissue usually appears as abnormally strong reflection peaks, and the contrast of these peaks is significantly higher than that of the surrounding healthy tissue. This is because scar tissue is usually denser and more rigid than normal retinal tissue, resulting in a significant increase in light reflection. By comparing the intensity and morphology of the reflection peaks, scar tissue areas can be effectively identified. Therefore, an abnormal increase in the reflection peak contrast is often one of the indicators of the presence of scar tissue, indicating that the area may be at risk of scar tissue or lesions.

[0059] For each sub-region, the specific steps of quantizing the extracted high-contrast reflection peak to generate the reflection peak abnormal sharpness factor are as follows:

[0060] The peak value of the optical reflection signal in the sub-area is detected to identify the significant peak in the reflection signal. In order to accurately describe the abnormal sharpness of the peak, a local maximum detection algorithm (such as maximum value detection) is used to extract all peak points in the reflection signal; then, the local contrast difference around each peak is calculated to measure its sharpness. The local contrast difference reflects the prominence of the peak relative to the adjacent area, which is quantified by the following formula: ,in: The peak point p The local contrast difference, The peak point p The reflection intensity value, The peak point p Radius on both sides r The minimum reflection intensity value within the range, The peak point p Radius on both sides r The maximum reflection intensity value within the range, p Indicates the location of the peak point, that is, the specific pixel position identified as the reflection peak in a given sub-region;

[0061] The prominence of the reflection peak is described by calculating the local contrast difference of each peak. Prominent peaks usually correspond to abnormal pathological tissues, such as scar tissue. A higher contrast difference indicates that the sharpness of the reflection signal in this sub-region is higher, which may suggest that this sub-region is scar tissue.

[0062] After obtaining the local contrast difference of each peak, the reflection peak abnormal sharpness factor is further calculated by the local contrast difference of each peak to quantify the abnormal sharpness of the peak. The calculation expression of the reflection peak abnormal sharpness factor is: ,in: is the reflection peak abnormal sharpness factor, P represents all the peak points detected in the sub-region, It is the median of the local contrast differences of all reflection peaks in the sub-region, reducing the impact of outliers or extreme values. The absolute difference in the denominator quantifies the degree of deviation of the reflection peak from the median value of the sub-region, enhancing the sensitivity to extreme sharp peaks.

[0063] The core of this step is to quantify the abnormal sharpness of the peak. By calculating the abnormal sharpness factor of the reflection peak, areas with extreme reflection characteristics can be more clearly identified, which are usually associated with the formation of scar tissue or diseased tissue. A higher reflection peak abnormal sharpness factor indicates that the reflection signal of the sub-area is abnormally sharp, which may represent scar tissue or fibrosis areas.

[0064] It can be seen from the abnormal sharpness factor of the reflection peak that for each sub-region, the extracted high-contrast reflection peak is quantified to generate the abnormal sharpness factor of the reflection peak. The larger the performance value, the more it usually indicates that the reflection peak of the sub-region is abnormally sharp, that is, there is a higher local contrast difference. This sharp reflection peak is usually associated with scar tissue, because scar tissue usually has strong optical reflection characteristics, and its reflection signal often appears more prominent than normal retinal tissue. By quantifying the sharpness of the reflection peak, these abnormal reflection characteristics can be effectively identified, thereby inferring that the area may be at risk of scar tissue. On the contrary, if the abnormal sharpness factor of the reflection peak is small, it means that the difference between the reflection peak of the sub-region and the surrounding tissue is small, which usually indicates that there is no obvious risk of scar or lesion in the sub-region.

[0065] When the degree of irregularity of the spectral response in a sub-region is high, it usually indicates that the sub-region is at a high risk of scar tissue. Scar tissue, due to its characteristics of fibrosis and collagen deposition, often causes significant changes in the optical properties of the tissue, and these changes are usually manifested as irregularities in the spectral response. The spectral response of normal retinal tissue is usually smooth and has a certain regularity, while scar tissue may cause complex fluctuations in the spectral signal due to its structural heterogeneity and uneven reflection characteristics. Therefore, by analyzing the degree of irregularity of the spectral response in a sub-region, potential scar areas can be effectively identified and whether there is fibrosis or other pathological changes in the area can be determined. A highly irregular spectral response is an important sign of scar tissue.

[0066] For each sub-region, the specific steps of quantizing the extracted irregular spectral response to generate the spectral nonlinear factor are as follows:

[0067] The irregularity of the spectral response can be quantified by calculating the nonlinearity of the spectral response. First, the spectral response data in the sub-region is obtained, where Represents the wavelength of the spectrum. On this basis, the high-order difference method is used to calculate the rate of change of the spectral signal, especially the high-order derivative, to quantify its nonlinear characteristics. The wavelength The spectral response at Higher-order differences (in ), the calculation expression is: ,in, n is the total length of the spectral data collected in the sub-region, i Represents the index in the spectral data, indicating the position of the current wavelength point in the sequence. For the i The position of the wavelength, k represents the order of higher-order differences, Indicated in wavelength The spectral response value at Indicated in wavelength The spectral response value at u is the offset between wavelength points, indicating the interval between wavelength points used to calculate the difference;

[0068] This high-order difference reflects the local fluctuation degree of the signal, that is, the rate of change of the spectral signal. k The value of can capture higher-order nonlinear changes in spectral response, especially in scar tissue areas, where the changes in spectral response are often more complex and have larger mutations;

[0069] By calculating high-order differences, high-frequency fluctuations in the spectral response can be captured, thus reflecting the irregularity of the spectral signal. Scar tissue usually has more significant spectral nonlinear characteristics, especially in areas with more significant fibrosis or collagen deposition, the spectral response will show a complex change pattern. This method can effectively quantify the irregularity of the spectral response and provide a basis for the subsequent calculation of spectral nonlinear factors.

[0070] The calculated high-order difference values ​​are synthesized, and the spectral nonlinear factor is generated by introducing a relative nonlinear metric. The spectral nonlinear factor quantifies the nonlinear characteristics of the spectral response by comparing the change relationship between the high-order difference and the original spectral response. The generated expression is: ,in, represents the spectral nonlinear factor, Representative i The square of the high-order difference at a wavelength reflects the nonlinear change intensity of the spectral signal at that wavelength. The denominator is the energy of the original signal, ensuring that the generated spectral nonlinear factor is consistent under different spectral signal intensities.

[0071] In this way, the spectral nonlinear factor can reflect the overall nonlinearity of the spectral signal. Scar tissue usually has a more complex optical structure and reflection characteristics, resulting in significant nonlinear characteristics in the spectral response. Therefore, a higher spectral nonlinear factor can effectively identify scar tissue areas and help distinguish normal tissue from scar tissue during the registration process.

[0072] It can be seen from the spectral nonlinear factor that for each sub-region, the larger the performance value of the spectral nonlinear factor generated by quantifying the extracted irregular spectral response, the higher the degree of irregularity of the spectral response in the sub-region, which is usually one of the typical characteristics of scar tissue. Scar tissue often causes the spectral signal to present more complex nonlinear fluctuation characteristics due to its fibrosis, collagen deposition or tissue structure heterogeneity, which is manifested as significant changes in reflection intensity, irregular distribution or increased spectral fluctuations. Therefore, when the value of the spectral nonlinear factor is high, it means that the spectral characteristics of the sub-region deviate from the smooth and regular pattern of normal tissue, indicating that the sub-region may have a higher risk of scar tissue. On the contrary, if the value of the spectral nonlinear factor is low, it means that the spectral response is relatively stable and smooth, and is closer to the spectral characteristics of healthy tissue, indicating that the risk of scar tissue in the sub-region is low. Therefore, the spectral nonlinear factor is an important indicator for quantitatively assessing the risk of scar tissue.

[0073] The quantitative features extracted from each sub-region are input into a pre-trained machine learning model, and each sub-region is intelligently identified through the model. According to the recognition results of the machine learning model, each sub-region is automatically divided into two categories: healthy tissue area and scar tissue area;

[0074] The abnormal sharpness factor of the reflection peak and the spectral nonlinear factor extracted from each sub-region are input into a pre-trained machine learning model, and a scar risk assessment index is output based on the machine learning model. The scar risk in the sub-region is intelligently assessed by the scar risk assessment index, and then the scar tissue in each sub-region is intelligently identified.

[0075] A pre-trained machine learning model refers to a machine learning model that has been trained using a large number of labeled spectral data sets before realizing the intelligent identification of scar tissue in each sub-region. These data sets contain the optical property information of different sub-regions and their corresponding scar tissue risk labels. By learning the patterns and rules in these data, the model can accurately predict scar risk in new spectral data. During the training process of this model, supervised learning methods are used to establish a mapping relationship between input features (for example, abnormal sharpness factors of reflection peaks and spectral nonlinear factors) and target outputs (scar risk assessment values ​​or classification labels). Through repeated iterative optimization, the model can effectively capture the complex nonlinear relationship between optical reflectance characteristics and scar risk, and ultimately achieve a high prediction accuracy.

[0076] In practical applications, such pre-trained machine learning models are usually built based on advanced algorithmic frameworks, such as support vector machines (SVM), random forests (RF) or deep learning methods (such as convolutional neural networks, CNN). The training process includes the following key steps: the first step is to collect enough high-quality spectral data samples, which need to cover a variety of scar tissue types and healthy tissues to ensure the generalization ability of the model; the second step is to extract features from the data, such as calculating the abnormal sharpness factor of the reflection peak and the spectral nonlinear factor of each sample, and standardizing these features; the third step is to divide the data into a training set and a validation set, and train the model with the training set, and the validation set is used to evaluate the model performance; the last step is to use optimization algorithms (such as gradient descent) to continuously adjust the model parameters to minimize the prediction error. Through this series of training processes, the model can learn the intrinsic relationship between the reflection peak and spectral nonlinear characteristics and the risk of scar tissue, laying the foundation for the subsequent prediction of unknown data.

[0077] During the application phase, the pre-trained machine learning model takes the characteristics of each sub-region (abnormal sharpness factor of reflection peak and spectral nonlinear factor) as input, and outputs the scar risk assessment index after model calculation. This assessment index is usually a quantitative value, reflecting the level of scar tissue risk in the sub-region. A high risk assessment index indicates that the spectral characteristics of the region are highly consistent with the optical characteristics of scar tissue, while a low risk assessment index means that the region is closer to healthy tissue. Using the scar risk assessment index generated by the model, the system can intelligently assess the scar risk of each sub-region and further determine whether there is scar tissue in the sub-region. In this way, the pre-trained model can not only process large-scale spectral data quickly and efficiently, but also significantly improve the accuracy and reliability of scar tissue identification, thereby providing a scientific basis for subsequent medical diagnosis and treatment plans.

[0078] The advantage of this pre-trained machine learning model is that it can extract potential patterns from complex optical properties and automatically adapt to differences between different patients or scanning conditions. For example, scar tissue in different patients may exhibit different spectral characteristics, or the spectral data may have certain noise or deviations due to changes in scanning conditions (such as light intensity, angle, etc.). By introducing diverse sample data during the training phase, the model can have high robustness and generalization capabilities, and can still accurately identify scar tissue in practical applications. In addition, the machine learning model can be continuously updated and optimized as data accumulates, so that it remains efficient and reliable when dealing with new scar tissue or more complex spectral characteristics. This dynamic optimization capability makes the machine learning model an indispensable tool in complex optical analysis and provides strong technical support for scar risk assessment.

[0079] The machine learning model is not specifically limited here, and can achieve the purpose of abnormal sharpness factor of reflection peak. and spectral nonlinear factor Comprehensive analysis to generate scar risk assessment index In order to realize the technical solution of the present invention, the present invention provides a specific implementation method; scar risk assessment index The resulting calculation formula is: , where , Reflection peak abnormal sharpness factor and spectral nonlinear factor The weight factor of , All are greater than 0. Weight factor , It indicates the relative importance of the abnormal sharpness factor of the reflection peak and the spectral nonlinear factor in the comprehensive analysis when calculating the scar risk assessment index. In other words, the weight factor is used to adjust the contribution ratio of the two indexes in risk assessment. Specifically, the weight factor represent The weight of represents the influence of the abnormal sharpness of the reflection peak on the scar risk; represent The weight of spectral nonlinearity indicates the influence of spectral nonlinearity on scar risk. When the value of a certain weight factor is large, it means that it contributes more to the final risk assessment result. Therefore, by reasonably setting , The size of the weight factor can adjust the algorithm's attention to different characteristic indicators according to the actual application scenario, thereby improving the accuracy and robustness of scar risk assessment. The overall design goal of the weight factor is to allow the scar risk assessment index to comprehensively reflect the importance of the two parameters and flexibly adapt to the needs of different data characteristics in practical applications.

[0080] It can be seen from the scar risk assessment index that, for each sub-region, the larger the expression value of the abnormal sharp reflection peak factor generated by quantizing the extracted high-contrast reflection peak, the larger the expression value of the spectral nonlinear factor generated by quantizing the extracted irregular spectral response. That is, the larger the expression value of the scar risk assessment index generated when the scar risk in the sub-region is intelligently assessed by the pre-trained machine learning model, the higher the risk of scar tissue in the sub-region, and vice versa.

[0081] The scar risk assessment index generated by the pre-trained machine learning model when intelligently assessing the scar risk in the sub-region is compared and analyzed with the pre-set scar risk reference threshold, and each sub-region is automatically divided. The specific division steps are as follows:

[0082] If the scar risk assessment index is greater than the scar risk reference threshold, the sub-region is divided into a scar tissue region;

[0083] If the scar risk assessment index is less than or equal to the scar risk reference threshold, the sub-region is divided into a healthy tissue region.

[0084] In the registration process, dynamic weights are assigned to different sub-regions. For healthy tissue areas, high registration weights are set as the main reference areas for registration. For scar tissue areas, dynamically variable low weights are set to avoid excessive reliance on abnormal reflection signals by the registration algorithm, reduce the interference of scar tissue on the registration results, and reduce scar drift and deformation.

[0085] During the registration process, dynamic weights are assigned to different sub-regions. The specific steps are as follows:

[0086] First, initialize the regional dynamic weight distribution to ensure that the healthy tissue area obtains a high weight. The calculation expression for the initial weight distribution is: ,in, For the f The registration weight of a sub-region reflects the importance of the sub-region in the registration process. For the f The scar risk assessment index for each sub-region is generated by a machine learning model. is the weight reduction control factor, which is used to adjust the attenuation degree of scar risk on weight. It is a weight smoothing control factor to avoid registration instability caused by too low or too high weight;

[0087] Through this step, healthy tissue areas (scar risk assessment index is low, ) will be given a higher weight , while scar tissue areas (scar risk assessment index is higher ) will be given a lower weight. This step ensures that the healthy tissue area becomes the main reference point for the registration and the interference of the scar tissue area is suppressed.

[0088] The registration objective function is adjusted based on the regional dynamic weight. The dynamic weight is introduced into the registration objective function to emphasize the healthy tissue area and weaken the influence of the scar area. The adjustment expression of the registration objective function is: ,in, is the weighted error of the registration objective function, which is used to measure the registration deviation between the template image and the reference image. N is the total number of sub-regions in the image, For the fThe difference metric function of the sub-region represents the registration error between the template image and the reference image in the sub-region (such as mutual information or gradient correlation, etc.);

[0089] The difference metric function is a mathematical function used to measure the degree of match or difference between the template image and the reference image in a certain sub-region during the registration process. f The local features of the sub-regions are compared to generate a numerical value to describe the difference or similarity between the two. Common difference measurement methods include intensity-based differences (such as mutual information, square error), edge-based gradient correlation, and Euclidean distance based on feature points. The goal of this function is to provide a quantitative way to evaluate whether the current registration state is ideal: the smaller the difference value, the higher the match between the template image and the reference image in the sub-region. By incorporating the difference measurement function into the registration objective function, the algorithm can be guided to continuously adjust the image's pose parameters to achieve the optimal registration effect.

[0090] By introducing dynamic weights , the registration algorithm will focus on optimizing the registration error in healthy tissue areas (high weight) and reduce over-reliance on abnormal signals in scar tissue areas (low weight). This step reduces the impact of scar drift and deformation on the overall registration result.

[0091] The weights are continuously updated during the iterative registration process, and the weight distribution is further optimized according to the real-time registration error. Dynamic iterative optimization and weight update are performed. The specific expression is: ,in, Indicates f The sub-region is The weight value at the iteration, that is, the weight value after dynamic adjustment during the image registration iteration process, For the f The weight value of the sub-region at the tth iteration, To dynamically update the step size factor, it is used to control the magnitude of the weight adjustment. For the f The registration error gradient of each sub-region reflects the sensitivity of the current weight to the registration objective function;

[0092] By dynamically updating the weights, the algorithm can adjust the registration strategy in real time. For areas with large registration errors (such as scar tissue areas), the weights will be further reduced to reduce their interference; for areas with small registration errors (such as healthy tissue areas), the weights will remain stable or even slightly increased to further optimize the registration accuracy.

[0093] In the registration process, the role of assigning dynamic weights to different sub-regions is to improve the robustness and accuracy of the registration, especially in complex scenes containing scar tissue. By setting high registration weights for healthy tissue areas, the registration algorithm prioritizes areas with stable structures and normal optical properties, thereby ensuring the accuracy and reliability of the overall registration; and setting dynamically variable low weights for scar tissue areas can effectively reduce the interference of abnormal reflection signals on the registration results, and avoid drift or deformation caused by excessive or irregular reflection characteristics of scar tissue. This dynamic weight allocation mechanism can better balance the impact of normal and abnormal areas on the registration results, and ultimately improve the adaptability and robustness of the registration algorithm under complex pathological conditions.

[0094] The above registration method based on three-dimensional retinal OCT images effectively solves the problem of abnormal reflection signals of scar tissue interfering with the registration results, and significantly improves the accuracy and robustness of registration. The anatomical feature points of the retina are used for preliminary registration to eliminate global mismatch errors and provide a stable foundation for subsequent local analysis; by dividing the image into equal-area sub-regions and extracting abnormal reflection signals, the intelligent classification of healthy tissue and scar tissue is achieved by combining the machine learning model, thus laying a scientific basis for dynamic weight allocation; the weights of healthy tissue and scar tissue are dynamically adjusted during the registration process, so that the registration algorithm relies on healthy areas first, reducing the drift and deformation caused by scar tissue, and ensuring that the scar area will not become an erroneous reference point. This method not only improves the reliability of the registration results, but also reduces the risk of clinical misjudgment, providing a solid technical guarantee for the accurate diagnosis and treatment of retinal diseases.

[0095] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A registration method based on three-dimensional retinal OCT images, characterized in that: The following steps are involved: In the registration process of OCT images, the anatomical feature points of the retina are first used as the key reference for basic alignment to perform preliminary registration and eliminate global mismatch errors; After the initial registration is completed, the OCT image is divided into several sub-areas of equal area, so that each sub-area is used as an independent analysis unit, which facilitates the subsequent accurate analysis of the local optical characteristics of each analysis unit; For each divided sub-region, obtain the optical reflection characteristic information inside it, and pre-process the acquired data to extract the abnormal reflection signal. Then, quantify the abnormal signal and convert the abnormal reflection characteristics of each sub-region into numerical features to generate a standardized data set as the input basis of the machine learning model. The quantitative features extracted from each sub-region are input into a pre-trained machine learning model, and each sub-region is intelligently identified through the model. According to the recognition results of the machine learning model, each sub-region is automatically divided into two categories: healthy tissue area and scar tissue area; During the registration process, dynamic weights are assigned to different sub-regions. For the healthy tissue area, a high registration weight is set as the reference area for registration; for the scar tissue area, a dynamically variable low weight is set to avoid the registration algorithm from over-relying on abnormal reflection signals and reduce the interference of scar tissue on the registration results.

2. A registration method based on three-dimensional retinal OCT images according to claim 1, characterized in that: Use the anatomical feature points of the retina as the key reference for basic alignment for preliminary registration. The specific steps are as follows: In the preliminary registration of OCT images, the anatomical feature points of the retina are first automatically identified by an algorithm; After identifying the feature points, the spatial relationship between the anatomical feature points is calculated, including position coordinates, distance, angle and shape characteristics. The translation deviation is adjusted by comparing the optic disc position, the rotation deviation is corrected by the direction of the vascular intersection, and the scaling ratio is adjusted by the length of the retinal layer. Finally, the calculated spatial transformation parameters are applied to preliminarily align the images at different time points, and the pixel coordinates of the images are remapped through the affine transformation matrix to ensure that the anatomical feature points overlap the mismatch errors in the images, providing a basis for subsequent refined registration.

3. A registration method based on three-dimensional retinal OCT images according to claim 1, characterized in that: After the initial registration is completed, the fixed grid partitioning method is used to divide the OCT image into several sub-areas with equal areas.

4. A registration method based on three-dimensional retinal OCT images according to claim 1, characterized in that: Abnormal reflection signals are extracted from the acquired optical reflection characteristic information, wherein the extracted abnormal reflection signals include high-contrast reflection peaks and irregular spectral responses. For each sub-region, the extracted high-contrast reflection peaks and irregular spectral responses are quantified to generate reflection peak abnormal sharpness factors and spectral nonlinear factors, respectively. A standardized data set is generated through the reflection peak abnormal sharpness factors and spectral nonlinear factors to provide accurate data input for the machine learning model for identifying scar areas.

5. A registration method based on three-dimensional retinal OCT images according to claim 4, characterized in that: The abnormal sharpness factor of the reflection peak and the spectral nonlinear factor extracted from each sub-region are input into a pre-trained machine learning model, and a scar risk assessment index is output based on the machine learning model. The scar risk in the sub-region is intelligently assessed by the scar risk assessment index, and then the scar tissue in each sub-region is intelligently identified.

6. A registration method based on three-dimensional retinal OCT images according to claim 5, characterized in that: The scar risk assessment index generated by the pre-trained machine learning model when intelligently assessing the scar risk in the sub-region is compared and analyzed with the pre-set scar risk reference threshold, and each sub-region is automatically divided. The specific division steps are as follows: If the scar risk assessment index is greater than the scar risk reference threshold, the sub-region is divided into a scar tissue region; If the scar risk assessment index is less than or equal to the scar risk reference threshold, the sub-region is divided into a healthy tissue region.

7. A registration method based on three-dimensional retinal OCT images according to claim 4, characterized in that: For each sub-region, the specific steps of quantizing the extracted high-contrast reflection peak to generate the reflection peak abnormal sharpness factor are as follows: The peak value of the optical reflection signal in the sub-area is detected to identify the significant peak in the reflection signal. The local maximum detection algorithm is used to extract all the peak points in the reflection signal. Then, the sharpness is measured by calculating the local contrast difference around each peak. The local contrast difference reflects the prominence of the peak relative to the adjacent area and is quantified by the following formula: ,in: The peak point p The local contrast difference, The peak point p The reflection intensity value, The peak point p Radius on both sides r The minimum reflection intensity value within the range, The peak point p Radius on both sides r The maximum reflection intensity value within the range, p Indicates the location of the peak point, that is, the specific pixel position identified as the reflection peak in a given sub-region; After obtaining the local contrast difference of each peak, the reflection peak abnormal sharpness factor is further calculated by the local contrast difference of each peak to quantify the abnormal sharpness of the peak. The calculation expression of the reflection peak abnormal sharpness factor is: ,in: is the reflection peak abnormal sharpness factor, P represents all the peak points detected in the sub-region, It is the median of the local contrast differences of all reflection peaks in the sub-region, reducing the impact of outliers or extreme values. The absolute difference in the denominator quantifies the degree of deviation of the reflection peak from the median value of the sub-region, enhancing the sensitivity to extreme sharp peaks.

8. The registration method based on three-dimensional retinal OCT images according to claim 4, characterized in that: For each sub-region, the specific steps of quantizing the extracted irregular spectral response to generate the spectral nonlinear factor are as follows: Get the spectral response data in the sub-region, where Represents the wavelength of the spectrum, and uses the high-order difference method to calculate the rate of change of the spectral signal to quantify its nonlinear characteristics. The wavelength The spectral response at High-order differences , the calculation expression is: ,in, n is the total length of the spectral data collected in the sub-region, i Represents the index in the spectral data, indicating the position of the current wavelength point in the sequence. For the i The position of the wavelength, k represents the order of higher-order differences, Indicated in wavelength The spectral response value at Indicated in wavelength The spectral response value at u is the offset between wavelength points, indicating the interval between wavelength points used to calculate the difference; The calculated high-order difference values ​​are synthesized, and the spectral nonlinear factor is generated by introducing a relative nonlinear metric. The spectral nonlinear factor quantifies the nonlinear characteristics of the spectral response by comparing the change relationship between the high-order difference and the original spectral response. The generated expression is: ,in, represents the spectral nonlinear factor, Representative i The square of the high-order difference at a wavelength reflects the nonlinear change intensity of the spectral signal at that wavelength position, and the denominator is the energy of the original signal, ensuring that the generated spectral nonlinear factor is consistent under different spectral signal intensities.

9. The registration method based on three-dimensional retinal OCT images according to claim 5, characterized in that: During the registration process, dynamic weights are assigned to different sub-regions. The specific steps are as follows: First, initialize the regional dynamic weight distribution to ensure that the healthy tissue area obtains a high weight. The calculation expression for the initial weight distribution is: ,in, For the f The registration weight of a sub-region reflects the importance of the sub-region in the registration process. For the f The scar risk assessment index for each sub-region is generated by a machine learning model. is the weight reduction control factor, which is used to adjust the attenuation degree of scar risk on weight. It is a weight smoothing control factor to avoid registration instability caused by too low or too high weight; The registration objective function is adjusted based on the regional dynamic weight. The dynamic weight is introduced into the registration objective function to emphasize the healthy tissue area and weaken the influence of the scar area. The adjustment expression of the registration objective function is: ,in, is the weighted error of the registration objective function, which is used to measure the registration deviation between the template image and the reference image. N is the total number of sub-regions in the image, For the f The difference measurement function of the sub-region represents the registration error between the template image and the reference image in the sub-region; The weights are continuously updated during the iterative registration process, and the weight distribution is further optimized according to the real-time registration error. Dynamic iterative optimization and weight update are performed. The specific expression is: ,in, Indicates f The sub-region is The weight value at the iteration, that is, the weight value after dynamic adjustment during the image registration iteration process, For the f The weight value of the sub-region at the tth iteration, To dynamically update the step size factor, it is used to control the magnitude of the weight adjustment. For the f The registration error gradient of each sub-region reflects the sensitivity of the current weight to the registration objective function.

Images