Data processing method based on photoacoustic imaging
The fast preprocessing model established through neural networks simplifies the preprocessing process of photoacoustic imaging, solves the complex and cumbersome problems in existing technologies, and achieves efficient photoacoustic data processing and image reconstruction.
Patent Information
- Application Number
- CN202511375017.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-25
- Publication Date
- 2026-02-13
Smart Images
Figure CN121512441A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image processing, and particularly relates to a data processing method based on photoacoustic imaging. BACKGROUND
[0002] Photoacoustic imaging (PAI) is a new biomedical imaging technology, which combines the high contrast of optical imaging and the high penetration depth of ultrasonic imaging. Its physical basis is the photoacoustic effect: when pulsed laser irradiates biological tissue, internal light absorbers (such as hemoglobin and melanin) absorb light energy and produce transient thermal expansion, exciting broadband ultrasonic waves. By detecting these photoacoustic data with an ultrasonic transducer, the data is processed and reconstructed into an image, which can realize high-resolution imaging of the internal optical absorption distribution (photoacoustic imaging).
[0003] At present, a photoacoustic data preprocessing process is required before image reconstruction, which usually includes light power compensation, filtering and peak detection. These preprocessing operations need to be performed one by one, and the process is complex and tedious, making it difficult to achieve the efficiency of preprocessing. SUMMARY
[0004] The purpose of the present application is to provide a data processing method based on photoacoustic imaging to solve the technical problem that the preprocessing operations need to be performed one by one in the prior art, the process is complex and tedious, and it is difficult to achieve the efficiency of preprocessing.
[0005] To solve the above technical problems, the present application specifically provides the following technical solutions: A data processing method based on photoacoustic imaging, characterized in that it comprises the following steps: Obtaining original photoacoustic data of photoacoustic imaging; Performing light power compensation, filtering and peak detection preprocessing on the original photoacoustic data in sequence to obtain preprocessed photoacoustic data; Using a neural network, based on the original photoacoustic data, data generated during preprocessing, and preprocessed photoacoustic data, a fast preprocessing model is established for predicting preprocessed photoacoustic data from original photoacoustic data; Based on the preprocessed photoacoustic data output by the fast preprocessing model, image reconstruction is performed to obtain photoacoustic imaging, and post-processing is performed on the photoacoustic imaging.
[0006] As a preferred scheme of the present application, the method of light power compensation comprises: Obtaining original photoacoustic data formed by each laser pulse , and the energy of the laser pulse ; Taking the rated output energy of the laser as the reference energy and the compensation coefficient of is calculated ; The original photoacoustic data is multiplied by the compensation coefficient to obtain photo power compensated data ; ; wherein, is the original photoacoustic data formed by the i-th laser pulse, is the position coordinate of the r-th ultrasonic transducer used to record the photoacoustic data, and t is time, is the energy of the i-th laser pulse.
[0007] As a preferred scheme of the present application, the filtering method comprises: selecting a filter and setting filter parameters according to the center frequency and bandwidth of the transducer; convolving the photo power compensated data with the impulse response of the filter to obtain filtered data .
[0008] As a preferred scheme of the present application, the packet peak detection method comprises: performing Hilbert transform on the filtered data to obtain quadrature signals .
[0009] constructing an analytic signal based on and calculating the modulus of the analytic signal to obtain packet peak detected data , wherein, is the imaginary unit, is the absolute value operation.
[0010] As a preferred scheme of the present application, the construction method of the fast preprocessing model comprises: taking the original photoacoustic data f, the photo power compensated data F1, and the filtered data F2 as input items of a convolutional neural network, taking the packet peak detected data F3 as an output item of the convolutional neural network, and training the neural network to obtain a fast preprocessing model for packet peak detection ; taking the original photoacoustic data f and the photo power compensated data F1 as input items of the fast preprocessing model for packet peak detection, taking the packet peak detected data F3 as an output item of the fast preprocessing model for packet peak detection, and retraining the fast preprocessing model for packet peak detection to obtain a fast preprocessing model for filtering and packet peak detection ; The original photoacoustic data f is an input item of the fast processing model of filtering and packet peak detection, and the data F3 after packet peak detection is an output item of the fast processing model of filtering and packet peak detection, and the fast preprocessing model is obtained by retraining the fast processing model of filtering and packet peak detection ; wherein, is the preprocessed photoacoustic data output by the fast processing model of packet peak detection, is the preprocessed photoacoustic data output by the fast processing model of filtering and packet peak detection, is the preprocessed photoacoustic data output by the fast preprocessing model, and the CNN is a neural network.
[0011] As a preferred scheme of the present application, the loss function of the fast preprocessing model ; The loss function of the fast processing model of filtering and packet peak detection ; The loss function of the fast processing model of packet peak detection ; wherein, is an L2 norm operation.
[0012] As a preferred scheme of the present application, the image reconstruction method of the photoacoustic data is a delay-and-sum algorithm.
[0013] As a preferred scheme of the present application, in the preprocessing sequence process of optical power compensation, filtering and packet peak detection, the data F3 after packet peak detection is the preprocessed photoacoustic data.
[0014] As a preferred scheme of the present application, the post-processing of photoacoustic imaging includes logarithmic compression and contrast, brightness adjustment.
[0015] As a preferred scheme of the present application, the original photoacoustic data f, the data F1 after optical power compensation, the data F2 after filtering, and the data F3 after packet peak detection are normalized.
[0016] Compared with the prior art, the present application has the following beneficial effects: The present application uses a neural network to establish an end-to-end prediction model for outputting preprocessed photoacoustic data from original photoacoustic data, realizes fast preprocessing of photoacoustic data, does not need to perform various preprocessing operations one by one, and improves preprocessing efficiency. BRIEF DESCRIPTION OF DRAWINGS
[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other drawings can also be obtained from the provided drawings without creative labor.
[0018] Figure 1 The data processing method flowchart based on photoacoustic imaging provided for the embodiments of the present application; Figure 2 The delay sum algorithm provided for the embodiments of the present application is based on the comparison between the photoacoustic imaging reconstructed based on the preprocessing procedure and the photoacoustic imaging reconstructed based on the fast preprocessing model. DETAILED DESCRIPTION
[0019] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.
[0020] As shown in Figure 1 The present application provides a data processing method based on photoacoustic imaging, comprising the following steps: obtaining original photoacoustic data of photoacoustic imaging; sequentially performing preprocessing of light power compensation, filtering and packet peak detection on the original photoacoustic data to obtain preprocessed photoacoustic data; using a neural network, based on the original photoacoustic data, the data generated in the preprocessing process and the preprocessed photoacoustic data, establishing a fast preprocessing model for predicting the preprocessed photoacoustic data from the original photoacoustic data; performing image reconstruction based on the preprocessed photoacoustic data output by the fast preprocessing model to obtain photoacoustic imaging, and performing post-processing on the photoacoustic imaging.
[0021] The present application uses a neural network to establish an end-to-end model inputting original photoacoustic data and outputting preprocessed photoacoustic data, so as to realize the simplification of the preprocessing procedure into a prediction process of an end-to-end neural network model, thereby improving the preprocessing efficiency.
[0022] When the present application establishes an end-to-end model inputting original photoacoustic data and outputting preprocessed photoacoustic data, the data samples are obtained through a preprocessing procedure containing various preprocessing operations, and the accuracy of the data samples is high, thereby ensuring the training accuracy of the end-to-end model at the data level.
[0023] The application adopts a step-by-step transfer learning method to gradually simplify each preprocessing operation in the preprocessing process during the construction of an end-to-end model, and finally the robustness, generalization ability of the model and the stability and efficiency of the training process are improved.
[0024] The method for optical power compensation comprises: acquiring original photoacoustic data formed by each laser pulse , and energy of the laser pulse ; taking the rated output energy of the laser as reference energy , and calculating a compensation coefficient of the original photoacoustic data ; multiplying the original photoacoustic data by the compensation coefficient to obtain optical power compensation data ; wherein, is the original photoacoustic data formed by the i-th laser pulse, is a position coordinate of the r-th ultrasonic transducer used for recording photoacoustic data, t is time, is the energy of the i-th laser pulse.
[0025] In the application, the original photoacoustic signal is scaled in amplitude by optical power compensation to eliminate signal intensity changes caused by laser pulse energy fluctuations, that is, to eliminate systematic deviations of data collected at different time points and between different pulses, and finally the image uniformity in photoacoustic imaging can be improved.
[0026] The method for filtering comprises: selecting a filter and setting filter parameters according to the center frequency and bandwidth of the transducer; (for example, a 5MHz transducer may use a 1-10MHz bandpass filter) convolving the optical power compensation data with an impulse response of the filter to obtain filtered data .
[0027] In the application, the time-domain signal after power compensation is processed by filtering preprocessing, effective signals are retained, noise is suppressed, and the subsequent envelope detection (Hilbert transform) result is more accurate, and noise is avoided from being misdetected as a signal envelope.
[0028] The method for peak detection comprises: performing Hilbert transform on the filtered data to obtain quadrature signals .
[0029] based on and constructing a resolution signal , and calculating the modulus of the resolution signal to obtain data after peak detection wherein, is an imaginary unit, is an absolute value operation.
[0030] In the present application, destructive interference is avoided, so that the reconstructed image has a darker background and higher contrast, i.e. background noise and artifacts are significantly suppressed, the energy core of the signal is highlighted, so that the boundary of the reconstructed absorber is clearer and the details are easier to distinguish, i.e. the image clarity is improved.
[0031] The construction method of the fast preprocessing model comprises: taking the original photoacoustic data f, the data F1 compensated by light power, and the filtered data F2 as input items of a convolutional neural network, taking the data F3 after peak detection as an output item of the convolutional neural network, and training the neural network to obtain a fast processing model for peak detection ; taking the original photoacoustic data f and the data F1 compensated by light power as input items of the fast processing model for peak detection, taking the data F3 after peak detection as an output item of the fast processing model for peak detection, retraining the fast processing model for peak detection to obtain a fast processing model for filtering and peak detection ; taking the original photoacoustic data f as an input item of the fast processing model for filtering and peak detection, taking the data F3 after peak detection as an output item of the fast processing model for filtering and peak detection, retraining the fast processing model for filtering and peak detection to obtain a fast preprocessing model ; wherein, is the preprocessed photoacoustic data output by the fast processing model for peak detection, is the preprocessed photoacoustic data output by the fast processing model for filtering and peak detection, is the preprocessed photoacoustic data output by the fast preprocessing model, and CNN is a neural network.
[0032] Loss function of the fast preprocessing model ; Loss function of the fast processing model for filtering and peak detection ; Loss function of the fast processing model for peak detection ; wherein, is an L2 norm operation.
[0033] The traditional end-to-end model directly learns a complex mapping from the original data f to the final output F3, which is too large to jump, difficult to train, and prone to local optimum or unexplainable error. The present application designs three progressive sub-models to build a clear learning path: The first stage establishes an end-to-end model to predict F3 from f, F1, and F2, i.e. The neural network first focuses on learning the accurate mapping of the envelope detection (or peak detection) task under the condition of having the most abundant information (original data + two intermediate results), that is, the neural network masters the preprocessing ability of peak detection. At this time, the loss function of the neural network is , the loss between the neural network prediction value and the true value, the smaller the loss, the stronger the preprocessing ability of the neural network to master the peak detection. The model of this stage is the easiest to train to convergence, so that the neural network masters the preprocessing ability of peak detection, and provides a high-precision learning target F3' for the subsequent stage.
[0034] The second stage establishes an end-to-end model to predict F3 from f and F1, i.e. , so that the neural network further learns the preprocessing ability of filtering based on the preprocessing ability of peak detection it has mastered, and focuses on learning the accurate mapping of the filtering task, and takes as the loss term, the smaller the loss, the stronger the preprocessing ability of the current neural network to master the filtering, and also takes the output F3' of the first stage as the learning target to form the loss term , avoiding the decrease of the preprocessing ability of peak detection in the process of the neural network focusing on learning the preprocessing ability of filtering, so that the preprocessing ability of peak detection of the neural network in the second stage is maintained with the first stage. Therefore, the neural network of this stage masters the preprocessing ability of filtering and peak detection, and provides a high-precision learning target F3'' for the subsequent stage.
[0035] The third stage establishes an end-to-end model to predict F3 from f and F1, i.e. , so that the neural network further learns the preprocessing ability of optical power compensation based on the preprocessing ability of filtering and peak detection it has mastered, and focuses on learning the accurate mapping of the optical power compensation task, and takes as the loss term, the smaller the loss, the stronger the preprocessing ability of the current neural network to master the optical power compensation, and also takes the output F3' of the first stage as the learning target to form the loss term , avoid causing the decrease of the pre-processing ability of the packet peak detection in the process that the neural network focuses on learning the pre-processing ability of the optical power compensation, so that the pre-processing ability of the packet peak detection of the neural network in the third stage is maintained flat with that in the first stage , avoid causing the decrease of the pre-processing ability of the packet peak detection in the process that the neural network focuses on learning the pre-processing ability of the optical power compensation, so that the pre-processing ability of the packet peak detection of the neural network in the third stage is maintained flat with that in the first stage The present application realizes step-by-step migration learning, step-by-step precipitation and maximum utilization of knowledge, greatly improves the precision and stability of the model, avoids the technical defects of direct learning from the original data f to the final output F3, the large jump, the high training difficulty, and the local optimum or the difficult-to-explain error.
[0036] The step-by-step migration method of the present application is equivalent to injecting an explicit physical process priori into this end-to-end black box model. Although the final model receives the original signal, its optimization target is explicitly constrained to simultaneously approximate the results of each stage after being processed by the traditional physical method. This means that the internal hierarchical structure of the trained model is likely to have learned the feature extraction function corresponding to the physical operations such as "power compensation", "filtering" and "Hilbert transform" spontaneously. This makes the model not only superior in performance, but also has a certain interpretability.
[0037] The training process of the final model in the present application is strongly constrained by the "pure" signals (F1, F2, F3) of each stage, and it learns how to reconstruct the clean signal after standard processing from the original data f full of noise and distortion. This learning method not only enables the model to learn the mapping function, but also improves the model performance against noise and outliers. Therefore, when facing noise patterns or slight system parameter changes that have not been seen in the training set, the model shows stronger robustness and generalization ability than the directly trained end-to-end model.
[0038] Therefore, by using this step-by-step progressive migration learning framework, the present application successfully combines the accuracy and interpretability of traditional physical algorithms with the high speed and strong robustness of deep learning models, not only improving the efficiency of photoacoustic data preprocessing, but also ensuring the reliability of the processing results.
[0039] The image reconstruction method of photoacoustic data is a delay-and-sum algorithm or other algorithms with the same function. Taking the delay-and-sum algorithm as an example, as shown in Figure 2 , the photoacoustic imaging of a tungsten filament phantom is performed Figure 2Photoacoustic imaging in section A is reconstructed from photoacoustic data obtained through traditional preprocessing procedures. Figure 2 Photoacoustic imaging (B) in this paper is reconstructed from photoacoustic data obtained by the fast preprocessing model in this invention. Although the sharpness of B is slightly lower than that of A, the reconstruction results in B are basically consistent with those in A, as follows: Discretizing the image space yields a pixel grid. And calculate the sound wave from spread to The required time, the delay time. ; In pixel grid place, from Data after peak detection at the location Extract the envelope signal at the delay time τ. The short-time reconstruction results were obtained by weighted averaging. , , It is typically used to compensate for factors such as sound beam diffusion; in, pixel grid Reconstruction results at the location (photoacoustic imaging). For the speed of sound, for The weight of m is the total number of ultrasonic transducers.
[0040] In the preprocessing sequence of optical power compensation, filtering, and peak detection, the data F3 after peak detection is the preprocessed photoacoustic data.
[0041] Post-processing of photoacoustic imaging includes logarithmic compression and contrast and brightness adjustment.
[0042] The original photoacoustic data f, the data F1 after optical power compensation, the filtered data F2, and the data F3 after peak detection are normalized.
[0043] This invention utilizes a neural network to establish an end-to-end prediction model that outputs preprocessed photoacoustic data based on the original photoacoustic data f, thereby achieving rapid preprocessing of photoacoustic data without the need for various preprocessing operations one by one, thus improving preprocessing efficiency.
[0044] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.
Claims
1. A data processing method based on photoacoustic imaging, characterized in that, The method comprises the following steps: obtaining original photoacoustic data of photoacoustic imaging; sequentially performing preprocessing of optical power compensation, filtering and envelope peak detection on the original photoacoustic data to obtain preprocessed photoacoustic data; using a neural network, based on the original photoacoustic data, data generated in the preprocessing process and the preprocessed photoacoustic data, establishing a fast preprocessing model for predicting the preprocessed photoacoustic data from the original photoacoustic data; performing image reconstruction based on the preprocessed photoacoustic data output by the fast preprocessing model to obtain photoacoustic imaging, and performing post-processing on the photoacoustic imaging.
2. The data processing method based on photoacoustic imaging according to claim 1, characterized in that: The method of optical power compensation comprises: acquiring raw photoacoustic data formed by each laser pulse , and the energy of the laser pulse ; The rated output energy of the laser is taken as the reference energy and the compensation coefficient of is calculated ; Utilizing raw photoacoustic data Multiplying by compensation factor , obtaining light power compensated data ; wherein is the raw photoacoustic data formed for the i-th laser pulse, is the position coordinate of the r-th ultrasound transducer used for recording the photoacoustic data, t is the time, is the energy of the i-th laser pulse.
3. The method of claim 2, wherein: The method of filtering comprises: selecting a filter and setting filter parameters according to the center frequency and bandwidth of the transducer; Data with optical power compensation Impulse response of the filter Performing convolution operation to obtain filtered data .
4. The data processing method based on photoacoustic imaging according to claim 3, characterized in that: The method of envelope peak detection comprises: on the filtered data performing a hilbert transform to obtain quadrature signals ; Based on and constructing a resolution signal , and calculating the module of the resolution signal to obtain the data after the packet peak detection wherein, is the imaginary unit, is the absolute value operation.
5. The method of claim 4, wherein: The method of constructing the fast preprocessing model comprises: The original photoacoustic data f, the data F1 compensated by light power, and the data F2 filtered are taken as input items of a convolutional neural network, and the data F3 detected by a packet peak is taken as an output item of the convolutional neural network, so that a fast processing model for packet peak detection is obtained by training the neural network ; The original photoacoustic data f and the data F1 compensated by light power are input items of the fast processing model of packet peak detection, and the data F3 after packet peak detection is an output item of the fast processing model of packet peak detection, and the fast processing model of packet peak detection is retrained to obtain the fast processing model of filtering and packet peak detection ; The original photoacoustic data f is taken as an input item of the fast processing model of filtering and packet peak detection, the data F3 after packet peak detection is taken as an output item of the fast processing model of filtering and packet peak detection, and the fast preprocessing model is obtained by retraining the fast processing model of filtering and packet peak detection ; wherein, pre-processed photoacoustic data output by a fast processing model for envelope detection, pre-processed photoacoustic data output by a fast processing model for filtering and envelope detection, pre-processed photoacoustic data output by a fast pre-processing model, CNN being a neural network.
6. The data processing method based on photoacoustic imaging according to claim 5, characterized in that: loss function of the fast pre-processing model ; Loss function of the fast processing model of the filtering and packet peak detection ; Loss function of the fast processing model of the peak detection ; wherein is an L2 norm operation.
7. The data processing method based on photoacoustic imaging according to claim 1, characterized in that: The image reconstruction method of the photoacoustic data is a delay-and-sum algorithm.
8. The data processing method based on photoacoustic imaging according to claim 4, characterized in that: In the preprocessing sequence of optical power compensation, filtering and envelope peak detection, the data F3 after envelope peak detection is the preprocessed photoacoustic data.
9. The method of claim 1, wherein: The post-processing of the photoacoustic imaging comprises logarithmic compression and contrast and brightness adjustment.
10. The data processing method based on photoacoustic imaging according to claim 5, characterized in that: The original photoacoustic data f, the data F1 after optical power compensation, the data F2 after filtering and the data F3 after envelope peak detection are subjected to normalization processing.