Self-adaptive accumulated strain method for optical coherence elastography

Abstract

The invention discloses a self-adaptive accumulated strain method for optical coherence elastography. The method comprises the following steps: constructing and training a strain estimation network based on uncertainty learning by using a deep learning framework; building an optical coherence tomography system; a series of interference spectrums # imgabs0 # subscripts s = 1, 2,... in the deformation process of the tested sample are continuously collected through the established optical coherence tomography system, S represents the number of frames, and m and j represent horizontal coordinates and depth coordinates respectively; a corresponding wrapped phase field is obtained based on a series of collected interference spectrums # imgabs1 # in the deformation process of a tested sample, the obtained corresponding wrapped phase field is input into the trained strain estimation network based on uncertainty learning, and the accumulated strain E is finally obtained by continuously accumulating the strain of each time. According to the method, any parameter selection except a preset threshold is not needed, and the strain change can be more sensitively captured, so that the inter-frame interval is adaptively adjusted according to the change, and the accumulated strain quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computational imaging, and in particular to an adaptive cumulative strain method for optical coherence elastic imaging. Background Art

[0002] Optical coherence tomography (OCT) is a tomographic imaging method based on the principle of low-coherence interferometry. Due to its non-invasive and high-resolution characteristics, it is often used to image the internal structure of transparent and translucent objects. When combined with phase contrast technology, this method can characterize the microscopic deformation of objects under different mechanical loads and is often used to measure strain within materials. This technique is called phase-contrast optical coherence elastography (PC-OCE).

[0003] During the PC-OCE measurement process, it is first necessary to use the established OCT system to collect the interference spectrum of the measured sample before and after deformation, and then perform Fourier transform on the interference spectrum to extract the phase information. Then, a differential operation is performed to obtain the wrapped phase field containing the deformation information. Finally, the strain is directly calculated using commonly used strain estimation methods such as the vector method or deep learning method. However, during the measurement process, if the deformation of the measured sample is too large and causes pixel-level displacement, it will cause speckle decorrelation of the wrapped phase field. This will result in the loss of phase information, making it impossible for the previously mentioned strain estimation method to perform correct strain calculations. For speckle decorrelation, there are currently two methods, including spatial domain tracking and time domain tracking. Among them, the spatial domain tracking method is limited by the size of the camera format, and the tracking range is limited. In order to overcome the problem of limited range, the time domain tracking method was proposed. Specifically, an OCT system is used to continuously measure the deformation process of a sample. This captures a series of interference spectra. After Fourier transformation, adjacent frames are differentiated to obtain a series of wrapped phase fields. These wrapped phase fields represent the decomposition of the original large deformation into smaller deformations. Strain estimation methods are then used to calculate strains one by one. Finally, all the resulting strains are accumulated to obtain the strain of the sample before and after the large deformation. However, this frame-by-frame strain accumulation method still has two problems: low computational efficiency and multiple accumulation of noise. Therefore, an appropriate frame spacing is required during the strain accumulation process to obtain higher-quality strains. However, a fixed frame spacing performs poorly in the case of nonlinear deformation of the measured object. Therefore, a strain accumulation method that automatically adjusts the frame spacing according to the speed of deformation is needed. Building on previous methods, an adaptive strain accumulation method based on noise thresholding has been proposed. This method proposes that the adaptive inter-frame spacing can be determined based on phase noise quantification. Based on this, a method for locating phase noise points is introduced to define the criteria for selecting the inter-frame spacing. This method uses the noise threshold to select an appropriate frame spacing, and then uses the vector method to calculate strain. However, this method still has two problems: 1) The vector method has a limited tolerance for noise. When the number of interferometric spectral frames is large, the number of accumulations inevitably increases, which leads to multiple accumulations of noise, resulting in a decrease in the signal-to-noise ratio of the final accumulated strain result. 2) The adaptive adjustment of the frame spacing in this method essentially indirectly reflects the strain quality by using the amount of noise wrapped in the phase field as an indicator, rather than directly using the quality of the strain calculated by the strain estimation method as an indicator to guide the selection of the frame interval. Therefore, further exploration of adaptive cumulative strain methods is needed to improve the quality of strain estimation. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide an adaptive cumulative strain method for optical coherence elastography, directly evaluating the strain quality by the output uncertainty and using this as an index to guide the selection of frame spacing during the accumulation process to achieve high signal-to-noise ratio strain estimation.

[0005] To achieve the above object, the technical solution provided by the present invention is as follows:

[0006] An adaptive cumulative strain method for optical coherence elastography, including:

[0007] Using a deep learning framework to build and train a strain estimation network based on uncertainty learning;

[0008] Building an optical coherence tomography system;

[0009] Continuously collecting a series of interference spectra during the deformation process of the measured sample through the built optical coherence tomography system The subscript s = 1, 2,..., S represents the number of frames, and m and j represent the horizontal coordinate and the depth coordinate respectively;

[0010] Based on a series of interference spectra collected during the deformation process of the measured sample Obtaining the corresponding wrapped phase field, and inputting the obtained corresponding wrapped phase field into the trained strain estimation network based on uncertainty learning, and finally obtaining the cumulative strain E by continuously accumulating the strain each time.

[0011] Furthermore, the process of obtaining the cumulative strain E includes:

[0012] A1. Inputting a series of spectra The subscript s = 1, 2,..., S represents the number of frames, and a preset threshold α is set;

[0013] A2. Initializing the number of frames s = 1, the frame interval Δs = 1, and the cumulative strain E = 0;

[0014] A3. When s ≥ S, output E; when s < S, execute the following steps:

[0015] a. If s + Δs > S, then Δs = S - s, and based on a series of interference spectra collected during the deformation process of the measured sample Obtaining the corresponding wrapped phase field, and inputting the obtained corresponding wrapped phase field into the trained strain estimation network based on uncertainty learning, and calculating the corresponding strain ε and the average uncertainty U;

[0016] b. When U < α and s + Δs < S, execute the following steps:

[0017] i. Δs = Δs + 1;

[0018] ii. Based on a series of interference spectra collected during the deformation process of the measured sample Obtain the corresponding wrapped phase field, and input the obtained wrapped phase field into the trained strain estimation network based on uncertainty learning to calculate the corresponding strain ε and average uncertainty U;

[0019] c. When U > α and Δs > 1, perform the following steps:

[0020] i. Δs = Δs - 1;

[0021] ii. Based on a series of interference spectra collected during the deformation process of the measured sample Obtain the corresponding wrapped phase field, and input the obtained wrapped phase field into the trained strain estimation network based on uncertainty learning to calculate the corresponding strain ε and average uncertainty U;

[0022] dE=E+ε;

[0023] es=s+Δs;

[0024] f. Return to step A3.

[0025] Furthermore, the calculation formula of the wrapped phase field is as follows:

[0026]

[0027] Among them, Re and Im represent the real and imaginary parts of the complex number; Δs is the latest inter-frame interval selected adaptively; Δφ s→s+Δs (m,j) is the phase φ s and phase φ s+Δs The phase difference between them is the corresponding wrapped phase field.

[0028] Furthermore, the strain ε and the average uncertainty U are calculated, including:

[0029] The wrapped phase field is input into the trained strain estimation network based on uncertainty learning. The strain estimation network uses the Unet++ structure, and its output follows a Gaussian distribution. Each prediction is sampled from the distribution, which is estimated by Monte Carlo sampling.

[0030] Perform multiple forward propagations on the strain estimation network and calculate the average value to obtain the strain ε, which is as follows:

[0031]

[0032] Among them, y is the result of each sampling, and T is the total number of sampling times;

[0033] The uncertainty u is calculated simultaneously for multiple sampling to quantify the quality of strain prediction. The calculation formula is as follows:

[0034]

[0035] The calculation formula of the average uncertainty U is as follows:

[0036]

[0037] Where M and J represent the maximum number of pixels in the horizontal and vertical directions, respectively.

[0038] Compared with the existing technology, the principles and advantages of this technical solution are as follows:

[0039] This technical solution uses a strain estimation network based on uncertainty learning for strain estimation. Its tolerance to noise is superior to the existing vector method (a strain estimation method). This means that under the same number of interference spectra, the method described in this technical solution requires fewer accumulation times than the existing vector method. Excessive accumulation times will cause noise errors to accumulate and propagate, affecting imaging quality. At the same time, the noise threshold of the existing vector method is based on the wrapped phase field, which is used to obtain a low-noise wrapped phase field, and then calculate the strain using the vector method. This is essentially a method that uses the quality of the wrapped phase field to indirectly reflect the quality of the strain. The uncertainty network used in the method described in this technical solution outputs uncertainty to directly represent the quality of the strain at the same time as the strain, and then uses a preset threshold to complete the adaptive accumulation of strain. This direct characterization method can more keenly capture strain changes, thereby adaptively adjusting the inter-frame interval based on the changes, improving the quality of the accumulated strain. In addition, in addition to setting a preset noise threshold, the existing vector method itself requires window parameter selection, and the strain estimation results of different windows vary. In actual use, it is often necessary to adjust the window multiple times, which is relatively cumbersome. However, the method described in this technical solution does not require any parameter selection other than the preset threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the services required for use in the embodiments or the prior art descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 This is a principle flow chart of the adaptive cumulative strain method for optical coherence elastography according to an embodiment of the present invention;

[0042] Figure 2Flow chart for obtaining strain and uncertainty by inputting the obtained wrapped phase field into the trained strain estimation network based on uncertainty learning in the adaptive cumulative strain method for optical coherence elastography according to an embodiment of the present invention;

[0043] Figure 3 Flow chart for adaptively adjusting the frame interval in the adaptive cumulative strain method for optical coherence elastography according to an embodiment of the present invention;

[0044] Figure 4 Comparison chart of strain estimation effects under different methods;

[0045] Figure 5 Signal-to-noise ratio comparison chart between the present invention and the adaptive cumulative method based on noise threshold. Detailed implementation manners

[0046] The present invention will be further described below in conjunction with specific embodiments:

[0047] As [[ID=z1]] Figures 1 to 3 shown, the adaptive cumulative strain method for optical coherence elastography described in this embodiment includes the following steps:

[0048] S1. Build and train a strain estimation network based on uncertainty learning using a deep learning framework;

[0049] S2. Build an optical coherence tomography system;

[0050] S3. Continuously collect a series of interference spectra during the deformation process of the measured sample through the built optical coherence tomography system The subscript s = ¹, ², …, S represents the number of frames, and m and j respectively represent the horizontal coordinate and the depth coordinate;

[0051] S4. Based on the series of interference spectra collected during the deformation process of the measured sample obtain the corresponding wrapped phase field, and input the obtained corresponding wrapped phase field into the trained strain estimation network based on uncertainty learning. By continuously accumulating the strain each time, the cumulative strain E is finally obtained.

[0052] In this step, the process of obtaining the cumulative strain E includes:

[0053] A1. Input a series of spectra The subscript s = ¹, ², …, S represents the number of frames, and a preset threshold α is set;

[0054] A2. Initialize the number of frames s = ¹, the frame interval Δs = ¹, and the cumulative strain E = 0;

[0055] A3. When s ≥ S, output E; when s < S, execute the following steps:

[0056] a. If s + Δs > S, then Δs = S - s. Based on a series of interference spectra during the deformation process of the measured sample collected, obtain the corresponding wrapped phase field, and input the obtained corresponding wrapped phase field into the trained strain estimation network based on uncertainty learning to calculate the corresponding strain ε and average uncertainty U;

[0057] b. When U < α and s + Δs < S, perform the following steps:

[0058] i. Δs = Δs + 1;

[0059] ii. Based on a series of interference spectra during the deformation process of the measured sample collected, obtain the corresponding wrapped phase field, and input the obtained corresponding wrapped phase field into the trained strain estimation network based on uncertainty learning to calculate the corresponding strain ε and average uncertainty U;

[0060] c. When U > α and Δs > 1, perform the following steps:

[0061] i. Δs = Δs - 1;

[0062] ii. Based on a series of interference spectra during the deformation process of the measured sample collected, obtain the corresponding wrapped phase field, and input the obtained corresponding wrapped phase field into the trained strain estimation network based on uncertainty learning to calculate the corresponding strain ε and average uncertainty U;

[0063] d. E = E + ε;

[0064] e. s = s + Δs;

[0065] f. Return to step A3. <(

[0066] In the above, the calculation formula of the wrapped phase field is as follows:

[0067]

[0068] where Re and Im respectively represent the real part and imaginary part of the complex number; Δs is the latest frame interval adaptively selected; Δφ s→s+Δs (m, j) is the phase difference between the phase φ s and the phase φ s+Δs that is, the corresponding wrapped phase field.

[0069] In the above, calculating the strain ε and average uncertainty U includes:

[0070] The wrapped phase field is input into the trained strain estimation network based on uncertainty learning. The strain estimation network uses the Unet++ structure, and its output follows a Gaussian distribution. Each prediction is sampled from the distribution, which is estimated by Monte Carlo sampling.

[0071] Perform multiple forward propagations on the strain estimation network and calculate the average value to obtain the strain ε, which is as follows:

[0072]

[0073] Among them, y is the result of each sampling, and T is the total number of sampling times;

[0074] The uncertainty u is calculated simultaneously for multiple sampling to quantify the quality of strain prediction. The calculation formula is as follows:

[0075]

[0076] The calculation formula of the average uncertainty U is as follows:

[0077]

[0078] Where M and J represent the maximum number of pixels in the horizontal and vertical directions, respectively.

[0079] In order to verify the practicability of the method proposed in this invention, the strain estimation is performed on the decorrelation phase of the speckle pattern generated by the internal deformation of the polymer droplet during the solidification process. Figure 4 As shown, (a) is the wrapped phase field. It can be seen that as the deformation increases, speckle decorrelation occurs and the phase fringes will disappear, which will cause the original strain estimation method to fail. (b) is the strain estimation result using the vector method (a type of strain estimation method). It can be found that the strain in the phase decorrelation area cannot be calculated. Similarly, (d) is the strain estimation result using the deep learning method. It is also impossible to estimate the phase decorrelation area. (c) is an adaptive cumulative increment method based on noise threshold. It is not difficult to see that the previous decorrelation areas have been successfully estimated. However, the strain quality in some areas is still poor and the signal-to-noise ratio is low. (e) is an adaptive cumulative strain method based on uncertainty learning. In addition to correctly estimating the strain in the decorrelation area, this method can obtain higher strain quality for areas where the strain quality of the adaptive cumulative increment method based on noise threshold is poor.

[0080] Then the signal-to-noise ratios of the strain estimation results of the two adaptive cumulative strain methods are calculated, as follows: Figure 5As shown in the figure, the signal-to-noise ratio of the strain estimation results of the method of the present invention is significantly improved compared to the adaptive accumulation method based on noise threshold. Specifically, the average signal-to-noise ratio of the method of the present invention is 8.3dB, while the average signal-to-noise ratio of the comparison method is 4.8dB, an average signal-to-noise ratio improvement of 3.5dB, equivalent to a 71.5% improvement. This comparison shows that the method of the present invention has a significant improvement over the adaptive accumulation method based on noise threshold.

[0081] The embodiments described above are only preferred embodiments of the present invention and are not intended to limit the scope of implementation of the present invention. Therefore, any changes made based on the shape and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. An adaptive cumulative strain method for optical coherence elastography, characterized in that: Including: Building and training a strain estimation network based on uncertainty learning using a deep learning framework; Building an optical coherence tomography system; The optical coherence tomography system is used to continuously collect a series of interference spectra during the deformation process of the sample being tested. Subscript s = 1, 2, ..., S represents the frame number, m and j represent the horizontal coordinate and depth coordinate, respectively; Based on a series of interference spectra collected during the deformation process of the measured sample The corresponding wrapped phase field is obtained and input into the trained strain estimation network based on uncertainty learning. The accumulated strain E is finally obtained by continuously accumulating the strain each time.

2. The adaptive cumulative strain method for optical coherence elastography according to claim 1, characterized in that: The process of obtaining the cumulative strain E includes: A1. Input a series of spectra Subscript s=1, 2, ..., S represents the number of frames, and the threshold α is preset; A2. Initialize the frame number s = 1, the frame interval Δs = 1, and the cumulative strain E = 0; A3. When s ≥ S, output E; when s < S, perform the following steps: a. If s+Δs>S, then Δs=Ss, based on a series of interference spectra collected during the deformation process of the measured sample Obtain the corresponding wrapped phase field, and input the obtained wrapped phase field into the trained strain estimation network based on uncertainty learning to calculate the corresponding strain ε and average uncertainty U; b. When U < α and s + Δs < S, perform the following steps: i. Δs = Δs + 1; ii. Based on a series of interference spectra collected during the deformation process of the measured sample Obtain the corresponding wrapped phase field, and input the obtained wrapped phase field into the trained strain estimation network based on uncertainty learning to calculate the corresponding strain ε and average uncertainty U; c. When U > α and Δs > 1, perform the following steps: i. Δs = Δs - 1; ii. Based on a series of interference spectra collected during the deformation process of the measured sample Obtain the corresponding wrapped phase field, and input the obtained wrapped phase field into the trained strain estimation network based on uncertainty learning to calculate the corresponding strain ε and average uncertainty U; d. E = E + ε; e. s = s + Δs; f. Return to step A3.

3. The adaptive cumulative strain method for optical coherence elastography according to claim 1 or 2, characterized in that: The calculation formula of the wrapped phase field is as follows: Among them, Re and Im represent the real and imaginary parts of the complex number; Δs is the latest inter-frame interval selected adaptively; Δφ s→s+Δs (m,j) is the phase φ s and phase φ s+Δs The phase difference between them is the corresponding wrapped phase field.

4. The adaptive cumulative strain method for optical coherence elastography according to claim 2, characterized in that: Calculating the strain ε and the average uncertainty U includes: Inputting the wrapped phase field into the trained strain estimation network based on uncertainty learning. The strain estimation network uses the Unet++ structure, and its output follows a Gaussian distribution. Each prediction is sampled from the distribution, and the distribution is estimated through Monte Carlo sampling; Performing multiple forward propagations on the strain estimation network and calculating the mean to obtain the strain ε. The formula is as follows: where y is the result of each sampling, and T is the total number of samplings; Calculating the uncertainty u during multiple samplings to quantify the quality of strain prediction. The calculation formula is as follows: The calculation formula of the average uncertainty U is as follows: where M and J represent the maximum number of pixels in the horizontal and vertical directions respectively.

Images