Method for improving OCT (Optical Coherence Tomography) axial and transverse resolution

Abstract

The invention discloses a method for improving OCT (optical coherence tomography) axial and transverse resolution, and belongs to the field of deep learning and signal processing. The method comprises the following steps: step S31, acquiring narrowband interference spectrums of different wavebands of the same object by using a variable waveband light source module, and filling a gap region by using a trained spectrum gap filling neural network to obtain a broadband wave number-depth domain interference spectrum; performing short-time inverse Fourier transform to obtain a broadband wave number domain interference spectrum, and finally performing fast Fourier transform to obtain an axial high-resolution A-Line; and S32, transversely splicing the axial high-resolution A-Line obtained in the step S31 to obtain a transverse low-resolution cross section, inputting the trained conditional diffusion model to generate a transverse high-resolution cross section, and obtaining a high-resolution three-dimensional OCT image. According to the invention, the imaging quality can be improved without hardware upgrading, and the axial and transverse resolution can be obviously improved in a low-cost manner.

Description

Technical Field

[0001] The present invention belongs to the fields of deep learning and signal processing, and particularly relates to a method for improving the axial and transverse resolutions of OCT. Background Art

[0002] Optical Coherence Tomography (OCT) is a non-contact, non-invasive, and high-resolution imaging technology, which is widely used in the diagnosis of fields such as ophthalmology, cardiovascular diseases, and dermatology. The OCT technology obtains the microstructural information inside biological tissues or materials by using the principle of light interference. Its imaging principle is to use the light emitted by a broadband light source to irradiate the sample, and the reflected light is interfered with the reference light through an interferometer to obtain an interference signal, thereby reconstructing the cross-sectional image of the sample.

[0003] It obtains the microscopic structural information inside biological tissues or materials through the interference principle of a broadband light source: the broadband light is reflected by the sample and then interfered with the reference light, and the cross-sectional image of the sample is reconstructed by analyzing the interference signal. However, there are dual bottlenecks in the resolution of traditional OCT systems - the axial resolution is limited by the spectral bandwidth of the light source, and the transverse resolution is limited by the beam diameter. Specifically, the axial resolution is inversely proportional to the light source bandwidth, and the transverse resolution is directly related to the size of the focused spot. Most current OCT systems use a single-band light source, with limited spectral bandwidth and a relatively wide beam, making it difficult to meet the imaging requirements of micron-scale fine structures.

[0004] To break through the resolution limitation, the existing technologies mainly rely on hardware improvements: for example, using an ultra-wideband light source to expand the axial bandwidth, or using a high numerical aperture objective lens to reduce the spot size to improve the transverse resolution. However, there are significant defects in hardware upgrades: ultra-wideband light sources are costly and lack stability, while high numerical aperture objective lenses sacrifice imaging depth and field of view. In addition, hardware methods are often costly and technically complex, while software methods only require minor modifications to the system application layer. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for improving the axial and transverse resolutions of OCT to solve the technical problems of high cost and complexity in improving the resolution through hardware in the prior art.

[0006] To solve the above technical problems, the specific technical solution of the present invention is as follows:

[0007] A method for improving the axial and transverse resolutions of OCT, the method comprising the following steps:

[0008] Step S31: Use a variable-band light source module to collect narrowband interference spectra of the same object in different bands, and use the trained spectral gap filling neural network to fill the gap regions to obtain a broadband wavenumber-depth domain interference spectrum; then perform an inverse short-time Fourier transform to obtain a broadband wavenumber domain interference spectrum, and finally perform a fast Fourier transform to obtain an axially high-resolution A-Line;

[0009] Step S32: Horizontally splice the axially high-resolution A-Line obtained in Step S31 to obtain a horizontally low-resolution cross-section, input it into the trained conditional diffusion model to generate a horizontally high-resolution cross-section, and obtain a high-resolution three-dimensional OCT image.

[0010] Further, the broadband wavenumber-depth domain interference spectrum in Step S31 is obtained in the following manner:

[0011] Use a variable-band light source module to collect narrowband raw interference spectra of the same object in more than two different bands. After preprocessing and short-time Fourier transform, obtain wavenumber-depth domain interference spectra of different bands, horizontally splice them or the overlapping regions are occluded to obtain a broadband wavenumber-depth domain interference spectrum with gaps; use the trained spectral gap filling neural network to fill the gap regions to obtain a broadband wavenumber-depth domain interference spectrum.

[0012] Further, the spectral gap filling neural network is trained in the following manner:

[0013] Step S11: Collect raw interference spectra through OCT, perform short-time Fourier transform after preprocessing, and artificially occlude some bands to construct a training dataset;

[0014] Step S12: Combine the constraint loss of the phase continuity prior to construct and train a spectral gap filling neural network with an encoder-decoder architecture.

[0015] Further, Step S11 includes the following steps:

[0016] Step S111: Collect raw interference spectra through OCT and obtain a wavenumber domain interference spectrum after preprocessing;

[0017] Step S112: Perform an inverse short-time Fourier transform on the preprocessed wavenumber domain interference spectrum to generate a wavenumber-depth domain interference spectrum containing depth information;

[0018] Step S113: Artificially occlude some bands of the wavenumber-depth domain interference spectrum obtained in Step S112. The occluded region is called the gap region, and the wavenumber-depth domain interference spectrum before occlusion is used as its ground truth. Construct a training dataset from the wavenumber domain interference spectrum, the wavenumber-depth domain interference spectrum with gaps, and the ground truth of the wavenumber-depth domain interference spectrum before occlusion.

[0019] Further, in step S12, the loss function of the spectral gap filling neural network of the encoder-decoder architecture is the sum of the wavenumber domain loss function and the wavenumber-depth domain loss function. Among them, the average absolute error is used for the wavenumber domain loss function, and the wavenumber-depth domain loss function is composed of the spectral amplitude reconstruction error L amp and the phase continuity error L phase weighted. The loss function in the wavenumber-depth domain is expressed as follows:

[0020] L k-dep = λ1·L amp + λ2L phase

[0021] where, L k-depth represents the loss function in the wavenumber-depth domain, λ1 represents the weighting coefficient of the spectral amplitude reconstruction error L amp and λ2 represents the weighting coefficient of the phase continuity error L phase ;

[0022] The spectral amplitude reconstruction error L amp is the mean square error loss and is expressed as follows:

[0023]

[0024] where, is the complex spectral amplitude predicted by the network, S i is the true spectral amplitude, and N is the number of sampling points;

[0025] The phase continuity error L phase is expressed as follows:

[0026]

[0027] where, is the true phase, is the predicted phase, is the gradient operator, is the second-order difference, and α is the second-order regularization weight coefficient.

[0028] Further, the conditional diffusion model is trained in the following manner:

[0029] Step S21: Adjust the spot size through the dynamic beam control system, and collect the transverse high-resolution and transverse low-resolution cross-section data sets within the same field of view;

[0030] Step S22: Construct and train a lateral low-resolution cross-section-guided conditional diffusion model, which includes a conditional model and a diffusion model with an encoder-decoder structure. The conditional model extracts the feature map of the low-resolution cross-section and performs feature fusion with the feature map of the noise image at each time step of the diffusion model. After denoising generation through multiple time steps, a lateral high-resolution cross-section is obtained.

[0031] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention uses a neural network to generate a wide-band interference spectrum from multiple discrete narrow-band spectra, significantly improving the axial resolution; at the same time, based on a dynamic beam control system, the spot size is dynamically adjusted, a high- and low-resolution cross-section dataset is constructed, and a conditional diffusion model is trained to achieve an improvement in lateral resolution. In axial processing, an encoder-decoder network combines a phase continuity prior constraint to accurately recover the spectral information in the gap region, breaking through the light source bandwidth limitation. This method can improve the imaging quality without hardware upgrade, generate high-resolution microstructural images, and provide a high-precision and low-cost solution for applications such as clinical diagnosis and defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments of the present invention. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 It is a schematic flowchart of the method for improving the axial and lateral resolutions of OCT of the present invention.

[0034] Figure 2 It is a schematic diagram of the training stage of the axial super-resolution model of the present invention.

[0035] Figure 3 It is a schematic diagram of the encoder-decoder model for spectral signal filling of the present invention.

[0036] Figure 4 It is a schematic diagram of the lateral super-resolution model of the present invention.

[0037] Figure 5 It is a schematic diagram of the verification stage of the axial super-resolution model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the protection scope of the present invention.

[0039] The present invention proposes a method for improving the axial and lateral resolutions of OCT. For the axial resolution, the original interference spectrum collected by OCT is collected, and after processing, a part of the wavelength band is artificially blocked to construct a data set composed of the interference spectrum with gaps and the true value of the gap region. Then, in combination with the phase continuity prior constraint loss function, a spectral gap filling neural network with an encoder-decoder structure is trained. For the lateral resolution, based on the dynamic beam control system to regulate the spot size, high-resolution and low-resolution cross-sectional images are alternately collected within the same field of view (equivalent imaging range) to construct a training data set, and the mapping relationship from low resolution to high resolution is learned through a conditional diffusion model. In the application stage, interference spectra of multiple different wavelength bands are collected by a variable wavelength band light source device, and the trained spectral gap filling neural network is used to fill the missing or phase-discontinuous regions to generate a broadband interference spectrum, which is then Fourier-transformed to obtain an axial high-resolution A-Line. The A-Lines are horizontally stitched into a low-resolution horizontal image, and the conditional diffusion model is further used to generate a horizontally high-resolution cross-section.

[0040] This embodiment is described by taking the dual-band interference spectrum and dynamic spot control as examples, but the present invention can be extended to multi-band interference spectra and multi-level spot size control scenarios, and there are multiple implementation methods by changing the neural network model. This embodiment uses a supercontinuum laser in combination with a tunable filter to achieve adjustable central wavelength, and a variable phase lens to achieve dynamic adjustment of the light source spot size. Those that collect discrete-band interference spectra or adjust the spot size in other ways also fall within the method and device for improving the axial and lateral resolutions of OCT described in the present invention.

[0041] Next, in conjunction with the accompanying drawings and specific embodiments, a detailed description will be given from aspects such as data acquisition, model architecture, optimization strategy, and system implementation.

[0042] A device for improving the axial and lateral resolutions of OCT proposed by the present invention, the device includes: a variable wavelength band light source module configured to dynamically switch the central wavelength to collect interference spectra of two or more wavelength bands. Optionally, the variable wavelength band light source module uses a supercontinuum laser in combination with a tunable filter to achieve adjustable central wavelength. A dynamic beam control system for adjusting the spot size and synchronously collecting high-resolution and low-resolution cross-sectional images in the same field of view. Optionally, the dynamic beam control system uses a variable phase lens to achieve dynamic adjustment of the light source spot size.

[0043] Based on the above device for improving the axial and lateral resolutions of OCT, the present invention also proposes a method for improving the axial and lateral resolutions of OCT, as Figure 1 shown, the method includes a training stage and a verification stage;

[0044] The training stage includes the following steps:

[0045] The training for axial super-resolution includes the following steps:

[0046] Step S11: Collect the original interference spectrum through OCT, perform short-time Fourier transform after preprocessing, artificially block part of the waveband, and construct a training dataset.

[0047] Specifically, step S1 includes the following steps:

[0048] Step S111: Collect the original interference spectrum through OCT, and obtain the interference spectrum in the wavenumber domain after preprocessing. The preprocessing includes operations such as DC removal and k-linearization. Specifically, DC removal means removing the DC component to eliminate autocorrelation noise; k-linearization means realizing wavenumber linearization based on k-space resampling of the reference spectrum.

[0049] Step S112: Perform short-time Fourier transform (STFT, Short-Time Fourier Transform) on the preprocessed interference spectrum in the wavenumber domain to generate an interference spectrum in the wavenumber-depth domain containing depth information.

[0050] Step S113: Artificially block part of the waveband of the interference spectrum in the wavenumber-depth domain obtained in step S112. The blocked area is called the gap area, and the interference spectrum in the wavenumber-depth domain before being blocked is used as its true value. Construct a training dataset from the interference spectrum in the wavenumber domain, the interference spectrum in the wavenumber-depth domain with gaps, and the true value of the interference spectrum in the wavenumber-depth domain before being blocked.

[0051] In this embodiment, as Figure 2 shown, use OCT to collect the original interference spectrum in the λ0~λ3 waveband (λ represents wavelength), obtain the interference spectrum in the wavenumber domain in the k0~k3 waveband (k represents wavenumber, k = 2π / λ) after preprocessing such as DC removal and k-linearization, and convert it into an interference spectrum in the wavenumber-depth domain through short-time Fourier transform. Artificially block the interference spectrum in the wavenumber-depth domain in the k1~k2 waveband to obtain an interference spectrum in the wavenumber-depth domain with gaps.

[0052] Step S12: Combine the constraint loss of the phase continuity prior, construct and train a spectral gap-filling neural network with an encoder-decoder architecture, and obtain a trained spectral gap-filling neural network.

[0053] Specifically, an encoder-decoder neural network model is constructed and trained, where the encoder operation extracts the context features of the non-gap region in the wavenumber-depth domain interferometric spectrum with gaps, and the decoder restores the spectral information of the gap region. A constraint loss function based on the prior of phase continuity is designed, which includes two parts: the spectral amplitude reconstruction error and the phase continuity error. The phase continuity error constrains the physical consistency of the model output by calculating the gradient difference between the predicted spectrum and the true spectrum in the phase domain.

[0054] In this embodiment, as Figure 2 shown, for the wavenumber-depth domain interferometric spectrum with gaps in the wavenumber range of k0 to k3, a neural network model based on encoder-decoder is used to fill the signal in the gap region and restore the wavenumber-depth domain interferometric spectrum before occlusion. The loss function for training the spectral gap filling neural network is the sum of the wavenumber domain loss function and the wavenumber-depth domain loss function. The spectral amplitude reconstruction error in the wavenumber domain uses the L1 loss (mean absolute error). The loss function in the wavenumber-depth domain is composed of the spectral amplitude reconstruction error L amp and the phase continuity error L phase weighted, and the loss function in the wavenumber-depth domain is expressed as follows:

[0055] L k-depth = λ1·L amp + λ2l phase

[0056] where, l k-dep represents the loss function in the wavenumber-depth domain, λ1 represents the weighting coefficient of the spectral amplitude reconstruction error L amp , and λ2 represents the weighting coefficient of the phase continuity error L phase .

[0057] The spectral amplitude reconstruction error L amp is the mean square error loss and is expressed as follows:

[0058]

[0059] where, is the complex spectral amplitude predicted by the network, S i is the true spectral amplitude, and N is the number of sampling points.

[0060] The phase continuity error L phase is calculated based on the principle of instantaneous Phase Corrected Total Variation (iPCTV):

[0061]

[0062] where, is the true phase, is the predicted phase, is the gradient operator (such as the first-order difference), is the second-order difference, and α is the second-order regularization weight coefficient. By minimizing the gradient difference between the predicted phase and the true phase the phase generated by the network is forced to change smoothly in the local area, avoiding phase mutations caused by spectral gap filling. In addition, a second-order difference term is introduced to constrain the continuity of the phase curvature, suppressing high-frequency oscillation artifacts. The parameter α balances the first-order and second-order constraint strengths to adapt to different spectral characteristics.

[0063] In this embodiment, as Figure 3 shown, two wavenumber-depth domain interferometric spectra (k0~k1, k2~k3) with different wavenumber ranges are horizontally stitched into a broadband wavenumber-depth domain interferometric spectrum with a gap. Each column with signals in this spectral map is used as an embedding vector and input into a Transformer-based encoder, which encodes it into a discrete token sequence with gaps (embedding vector). Then, randomly initialized gap tokens are added to supplement it into a complete token sequence. Finally, a decoder stacked with Transformers is used to recover the spectral signals in the gap region.

[0064] The training for horizontal super-resolution includes the following steps:

[0065] Step S21: Adjust the spot size through a dynamic beam control system, and collect a horizontal high-resolution and a horizontal low-resolution cross-section dataset within the same field of view.

[0066] Specifically, the dynamic beam control system is used to drive the detection spot to actively switch between the high-resolution diameter and the low-resolution diameter, and alternately collect horizontal high-resolution cross-sections and horizontal low-resolution cross-sections of the same object within the same field of view (the same imaging range), constructing a horizontal high- and low-resolution cross-section training set.

[0067] In this embodiment, a variable-phase lens is used to dynamically adjust the size of the light source spot, obtaining horizontal high- and low-resolution cross-sections with different image sizes within the same field of view.

[0068] Step S22: Construct and train a conditional diffusion model guided by the horizontal low-resolution cross-section to obtain a trained conditional diffusion model.

[0069] Specifically, the conditional diffusion model includes a conditional model and a diffusion model with an encoder-decoder structure, and the conditional diffusion model is constructed and trained. The conditional model extracts the feature map of the low-resolution cross-section, and in each time step of the diffusion model, it performs feature fusion with the feature map of the noisy image. After denoising generation through multiple time steps, a high-resolution cross-section in the transverse direction is obtained.

[0070] In this embodiment, as Figure 4 shown, the conditional model and the diffusion model adopt the UNet series model with an encoder-decoder structure (such as the UNet model based on a convolutional neural network, which includes two parts: a decoder and an encoder). First, the transverse low-resolution cross-section image is resampled to the same size (H, W) as the transverse high-resolution image, where H represents the height of the cross-section and W represents the width of the cross-section. The feature map of the transverse low-resolution cross-section is extracted through the encoder and decoder of the conditional model. The generation process starts from the noisy image at time x T and denoising generation is performed using the diffusion model at each time step. For time step t, the diffusion model is used to extract the feature map of the noisy image at time x t and perform pointwise addition with the feature map of the transverse low-resolution cross-section extracted by the conditional model for feature fusion. Then, through the decoder of the diffusion model, the cross-section image at time x t-1 is obtained. After T time steps, a high-resolution cross-section in the transverse direction is finally obtained.

[0071] In this embodiment, the construction method of the training dataset of the conditional diffusion model is as follows: The high-resolution image is gradually noise-added to obtain the cross-section before noise addition, the cross-section after noise addition, and the noise ground truth at each time step.

[0072] In this embodiment, the training process of the conditional diffusion model is as follows: Given any time step, map the time step t to a vector representation and then add it to each encoder-decoder block to let the model know the time step information, and inversely generate the noise used for noise addition according to the input cross-section after noise addition, and calculate the mean absolute error with the noise ground truth to train the conditional diffusion model. The cross-section after noise addition minus the noise predicted by the conditional diffusion model is the cross-section before noise addition predicted by the conditional diffusion model.

[0073] The verification stage includes the following steps:

[0074] Step S31: Use the variable-band light source module to collect the narrowband interference spectra of the same object in different bands, use the trained spectral gap-filling neural network to fill the gap area to obtain the broadband wavenumber-depth domain interference spectrum. Then, through the inverse short-time Fourier transform, obtain the broadband wavenumber domain interference spectrum, and finally, through the fast Fourier transform, obtain the high-resolution A-Line in the axial direction.

[0075] Specifically, a variable-band light source module is used to collect narrowband raw interference spectra of N (N≥2) different bands of the same object (the same object to be imaged). After preprocessing and short-time Fourier transform, the wavenumber-depth domain interference spectra of different bands are obtained. They are horizontally spliced or the overlapping regions are blocked to obtain a broadband wavenumber-depth domain interference spectrum with gaps. A trained spectral gap-filling neural network is used to fill the gap regions, and after subsequent processing, a high-resolution A-Line is obtained.

[0076] In this embodiment, as Figure 5 shown, a supercontinuum laser is used in cooperation with a tunable filter to collect narrowband raw interference spectra of two different bands (k0~k1, k2~k3) in real time. For the case of k2≤k1, after horizontal splicing, the overlapping regions can be artificially blocked to obtain a wavenumber-depth domain interference spectrum with gaps; for the case of k2>k1, a wavenumber-depth domain interference spectrum with gaps is naturally obtained after horizontal splicing. Then, a trained spectral gap-filling neural network is used to fill the gap regions to obtain a broadband wavenumber-depth domain interference spectrum. Through inverse short-time Fourier transform, a broadband wavenumber domain interference spectrum is obtained, and finally, an axial high-resolution A-Line is obtained through fast Fourier transform (FFT).

[0077] It should be emphasized that although k0, k1, k2, and k3 are used in the above description, the spectral bands used in the training set and the validation set are not fixed in the same band. During training, sample pairs of different band ranges can be collected, and during validation, bands not included in the training set can also be filled.

[0078] Step S32: Horizontally splice the axial high-resolution A-Line obtained in step S31 to obtain a horizontally low-resolution cross-section, input it into the trained conditional diffusion model, generate a horizontally high-resolution cross-section, and obtain a high-resolution three-dimensional OCT image.

[0079] Specifically, a light source module with a low-resolution diameter but variable bands is used. According to the axial high-resolution A-Line obtained in step S31, it is horizontally spliced into a cross-section with axial high resolution but horizontal low resolution. Through resampling operations such as cropping and stretching, N (N≥1) low-resolution cross-sections with a width of W can be obtained. Using the trained conditional diffusion model, N high-resolution cross-sections with a width of W are generated, and through resampling operations such as splicing and averaging, horizontal high-resolution reconstruction of any multiple of the low-resolution cross-section with any width is achieved.

[0080] The present invention combines deep learning methods to identify the context information in the non-gap region of broadband interference spectra through a neural network, and at the same time uses the neural network to analyze the organizational structure topology law in the low-resolution cross-sectional image. The gap-filling neural network is based on the short-time Fourier transform wavenumber-depth domain interference spectrum, which can achieve phase continuity constraints compared with the wavenumber domain interference spectrum. Based on the dual-modal data of the wavenumber domain and the wavenumber-depth domain, this method can obtain a broadband interference spectrum from a narrowband interference spectrum, use a neural network to extract the anatomical structure features of the photographed object, and achieve high-precision reconstruction of sub-pixel-level microstructures. Compared with the traditional hardware upgrade scheme, this technology only requires the configuration of a conventional OCT system to generate high-quality images, breaking through the physical limit while maintaining the hardware cost unchanged.

[0081] It can be understood that the present invention is described through some embodiments. As is known to those skilled in the art, without departing from the spirit and scope of the present invention, various changes or equivalent replacements can be made to these features and embodiments. Additionally, under the teaching of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.

Claims

1. A method for improving the axial and transverse resolutions of OCT, characterized in that, The method includes the following steps: Step S31: Use a variable-band light source module to collect narrowband interference spectra of the same object in different bands, and use a trained spectral gap-filling neural network to fill the gap regions to obtain a broadband wavenumber-depth domain interference spectrum; Then, perform an inverse short-time Fourier transform to obtain a broadband wavenumber domain interference spectrum, and finally perform a fast Fourier transform to obtain an axially high-resolution A-Line; Step S32: Horizontally splice the axially high-resolution A-Line obtained in Step S31 to obtain a horizontally low-resolution cross-section, input it into the trained conditional diffusion model, generate a horizontally high-resolution cross-section, and obtain a high-resolution three-dimensional OCT image.

2. The method for improving the axial and transverse resolutions of OCT according to claim 1, wherein In Step S31, the broadband wavenumber-depth domain interference spectrum is obtained in the following manner: Use a variable-band light source module to collect narrowband raw interference spectra of the same object in two or more different bands. After preprocessing and short-time Fourier transform, obtain the wavenumber-depth domain interference spectra of different bands, horizontally splice them or cover the overlapping regions to obtain a broadband wavenumber-depth domain interference spectrum with gaps; use a trained spectral gap-filling neural network to fill the gap regions to obtain a broadband wavenumber-depth domain interference spectrum.

3. The method for improving the axial and transverse resolutions of OCT according to claim 1, wherein The spectral gap-filling neural network is trained in the following manner: Step S11: Collect raw interference spectra through OCT, perform short-time Fourier transform after preprocessing, and artificially cover some bands to construct a training dataset; Step S12: Combine the constraint loss of the phase continuity prior to construct and train a spectral gap-filling neural network with an encoder-decoder architecture.

4. The method for improving the axial and transverse resolutions of OCT according to claim 3, characterized in that, Step S11 includes the following steps: Step S111: Collect raw interference spectra through OCT and obtain the wavenumber domain interference spectrum after preprocessing; Step S112: Perform an inverse short-time Fourier transform on the preprocessed wavenumber domain interference spectrum to generate a wavenumber-depth domain interference spectrum containing depth information; Step S113: Artificially cover some bands of the wavenumber-depth domain interference spectrum obtained in Step S112; the covered region is called the gap region, and the wavenumber-depth domain interference spectrum before being covered is used as its ground truth; construct a training dataset from the wavenumber domain interference spectrum, the wavenumber-depth domain interference spectrum with gaps, and the ground truth of the wavenumber-depth domain interference spectrum before being covered.

5. The method for improving the axial and lateral resolutions of OCT according to claim 4, wherein In step S12, the loss function of the spectral gap filling neural network with an encoder-decoder architecture is the sum of the wavenumber domain loss function and the wavenumber-depth domain loss function; the spectral amplitude reconstruction error in the wavenumber domain uses the L1 loss, and the loss function in the wavenumber-depth domain is composed of the weighted sum of the spectral amplitude reconstruction error L amp and the phase continuity error L pha as shown in the following formula: L k-depth = λ1·L amp + λ2L phase Among them, L k-dep represents the loss function in the wavenumber-depth domain, and λ1 represents the weighting coefficient of the spectral amplitude reconstruction error L amp ; λ2 represents the weighting coefficient of the phase continuity error L phase . Spectral amplitude reconstruction error L amp Is the mean square error loss and is expressed as follows: Among them, is the complex spectral amplitude predicted by the network, S i is the true spectral amplitude, and N is the number of sampling points; Phase continuity error L phase is expressed as follows: Among them, is the true phase, is the predicted phase, is the gradient operator, is the second-order difference, α is the second-order regularization weight coefficient, and k0 to k3 represent the wavenumber range of the interference spectrum.

6. The method for improving the axial and transverse resolutions of OCT according to claim 1, wherein, The conditional diffusion model is trained in the following manner: Step S21: Adjust the spot size through a dynamic beam control system, and collect horizontally high-resolution and horizontally low-resolution cross-section datasets within the same field of view; Step S22: Construct and train a conditional diffusion model guided by horizontally low-resolution cross-sections. The conditional diffusion model includes a conditional model and a diffusion model with an encoder-decoder structure; the conditional model extracts the feature map of the low-resolution cross-section, and in each time step of the diffusion model, performs feature fusion with the feature map of the noise image, and through denoising generation in multiple time steps, obtains a horizontally high-resolution cross-section.

Images