Multi-bandwidth photoacoustic imaging system and method based on deep learning cascade enhancement

Abstract

The invention discloses a multi-bandwidth photoacoustic imaging system and method based on deep learning cascade enhancement, the system is composed of a pulse laser, two semicircular array probes, a multi-channel data acquisition card and a central computer, the pulse laser is used for exciting a light absorber in a target area to generate a photoacoustic signal; the two semicircular array probes are respectively provided with a high-frequency probe with the center frequency of 5.5 MHz and a low-frequency probe with the center frequency of 2.5 MHz, the high-frequency probe is used for acquiring a high-resolution image, and the low-frequency probe is used for detecting structural characteristics of a relatively large light absorber and a low signal-to-noise ratio; and the multichannel data acquisition card is used for performing digital processing on the photoacoustic signals received by the two semicircular array probes, transmitting the processed data to the central computer, and reconstructing a photoacoustic image based on a filtering back projection reconstruction algorithm. The multi-bandwidth photoacoustic imaging system disclosed by the invention shows extremely high image reconstruction capability, and provides a more accurate solution for biomedical imaging.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology and relates to a multi-bandwidth photoacoustic imaging system and method, and specifically relates to a multi-bandwidth photoacoustic imaging system and method based on deep learning cascade enhancement. Background Art

[0002] In recent years, photoacoustic imaging (PAI) has made significant progress in the biomedical field, especially showing great potential in tumor detection and functional imaging. However, traditional photoacoustic imaging methods are limited by the bandwidth and response characteristics of the detector, resulting in the loss of some spectral information, thereby limiting the resolution and information integrity of the imaging. Although multi-bandwidth imaging technology alleviates this problem by using detectors with different center frequencies, there are still challenges such as low information integration efficiency and low imaging quality. With the rapid development of deep learning technology, especially in the fields of image reconstruction and signal processing, new solutions have been provided. Summary of the invention

[0003] In order to overcome the shortcomings of limited signal bandwidth, insufficient image resolution and quality in existing photoacoustic imaging technology, the present invention provides a multi-bandwidth photoacoustic imaging system and method based on deep learning cascade enhancement. The present invention introduces a cascaded deep learning network architecture to effectively integrate multi-frequency photoacoustic signals, achieve image reconstruction from coarse to fine, significantly improve imaging accuracy and quality, and reduce the complexity of system hardware to achieve higher quality photoacoustic image reconstruction, which is suitable for high-resolution and multi-bandwidth imaging requirements in biomedical imaging.

[0004] The objective of the present invention is achieved through the following technical solutions:

[0005] A multi-bandwidth photoacoustic imaging system based on deep learning cascade enhancement consists of a pulsed laser, two semi-annular array probes with different center frequencies, a multi-channel data acquisition card and a central computer, wherein:

[0006] The pulse laser is used to excite the light absorber in the target area to generate a photoacoustic signal;

[0007] The two semi-circular array probes respectively have a high-frequency probe with a center frequency of 5.5 MHz and a low-frequency probe with a center frequency of 2.5 MHz, the high-frequency probe is used to obtain high-resolution images, and the low-frequency probe is used to detect larger light absorbers and structural features with low signal-to-noise ratio;

[0008] The multi-channel data acquisition card is used to digitally process the photoacoustic signals received by the two semi-circular array probes, and transmit the processed data to a central computer to reconstruct the photoacoustic image based on a filtered back projection reconstruction algorithm.

[0009] A multi-bandwidth photoacoustic imaging method based on deep learning cascade enhancement comprises the following steps:

[0010] Step 1: Multi-bandwidth photoacoustic imaging system hardware composition and configuration

[0011] The multi-bandwidth photoacoustic imaging system consists of a pulsed laser, two semi-circular array probes with different center frequencies, a multi-channel data acquisition card and a central computer, among which:

[0012] The pulse laser is used to excite the light absorber in the target area to generate a photoacoustic signal;

[0013] The two semi-circular array probes respectively have a high-frequency probe with a center frequency of 5.5 MHz and a low-frequency probe with a center frequency of 2.5 MHz, the high-frequency probe is used to obtain high-resolution images, and the low-frequency probe is used to detect larger light absorbers and structural features with low signal-to-noise ratio;

[0014] The multi-channel data acquisition card is used to digitally process the photoacoustic signals received by the two semi-circular array probes, and transmit the processed data to the central computer to reconstruct the photoacoustic image based on the filtered back projection reconstruction algorithm;

[0015] Step 2: Signal acquisition and preprocessing

[0016] Step 2.1: The pulse laser emits short pulses of light, which are irradiated onto the tissue within the target imaging area to stimulate photoacoustic signals;

[0017] Step 2.2: The two semi-circular array probes receive photoacoustic signals of different frequencies respectively, and these photoacoustic signals are collected by a multi-channel data acquisition card and stored in a central computer;

[0018] Step 2.3: The central computer pre-processes the collected photoacoustic signals;

[0019] Step 2.4: Use the reconstruction algorithm to generate high-frequency photoacoustic image x after preprocessing the photoacoustic signal h and low-frequency photoacoustic images x l , and divided into training set, validation set and test set for subsequent deep learning cascade network training and performance verification;

[0020] Step 3: Construction and training of deep learning cascade network

[0021] Step 3.1: The cascaded deep learning network consists of two U-Net modules in series. First, the high-frequency photoacoustic image x generated by preprocessing is h and low-frequency photoacoustic images x lShallow feature extraction is performed through convolution kernels of size 3×3, and then the extracted features are fused along the channel direction, and the fused features are input into the first U-Net module of the encoding-decoding structure. During the encoding process, this module extracts features through a double-layer Double Conv3×3 convolution layer and down-samples through a four-level maximum pooling layer. During the decoding process, up-sampling is performed through a Double Deconv3×3 deconvolution layer, and finally a rough reconstruction is completed through an output layer with a convolution kernel size of 1×1 to generate a preliminary full-ring image y 1 ; Next, the generated full ring image y 1 The high-frequency photoacoustic image x h and low-frequency photoacoustic image x l The features are fused along the channel direction, and the fused features are input into the second U-Net module of the encoding-decoding structure. During the encoding process, the module extracts features through a double-layer Double Conv3×3 convolution layer and implements downsampling through a four-level maximum pooling layer. During the decoding process, upsampling is performed through a Double Deconv3×3 deconvolution layer, and finally a fine reconstruction is completed through an output layer with a convolution kernel size of 1×1 to generate the final full ring image y 2 ;

[0022] Step 3.2: During the training process, a joint loss function is used to measure the difference between the reconstructed image and the real image. The loss function consists of two parts:

[0023] L=αL 1 +βL 2

[0024]

[0025]

[0026] Among them, L 1 y is the output of the first-level rough reconstruction model 1 The MSE loss between the real image y, L 2 y is the output of the second-level fine reconstruction model 2 The MSE loss between the real image y and α is L 1 Loss weight, β is L 2 Loss weight;

[0027] Step 3.3: Use Adam optimizer to optimize the network parameters;

[0028] Step 4: Multi-bandwidth photoacoustic reconstruction:

[0029] The photoacoustic images collected by the 2.5MHz half-ring array and the photoacoustic images collected by the 5.5MHz half-ring array were used as the input of the deep learning cascade network. The trained deep learning cascade network was used to perform multi-bandwidth photoacoustic reconstruction to obtain 2.5MHz and 5.5MHz full-ring photoacoustic images.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] 1. The present invention realizes a two-stage image reconstruction process from coarse to fine through a cascaded deep learning network (CU-Net), which significantly improves the resolution and information integrity of photoacoustic images, especially in the fusion and processing of low-frequency and high-frequency signals.

[0032] 2. The present invention adopts semi-circular array probes with different central frequencies, which can capture a wider range of photoacoustic signals, realize multi-scale and multi-bandwidth imaging, and take into account the detection requirements of large light absorbers and fine structures.

[0033] 3. This invention uses deep learning technology to achieve automatic image reconstruction, reduces reliance on manual adjustments, greatly improves processing efficiency, and is suitable for real-time imaging and large-scale application scenarios.

[0034] 4. The present invention avoids complex hardware design requirements and reduces system complexity and cost through deep learning-driven signal integration and optimization, while retaining efficient multi-angle and multi-frequency imaging capabilities.

[0035] 5. In the in vivo experiment, the multi-bandwidth photoacoustic imaging system was successfully applied to photoacoustic imaging of the mouse abdomen. CU-Net performed well in image reconstruction indicators such as PSNR (peak signal-to-noise ratio), MAE (mean absolute error) and SSIM (structural similarity index), verifying its reliability and wide applicability in practical applications.

[0036] 6. The multi-bandwidth photoacoustic imaging system of the present invention demonstrates extremely high image reconstruction capabilities, providing a more accurate solution for biomedical imaging. The system integrates high- and low-frequency probes, and combines the wide coverage capability of low-frequency probes in overall structural imaging and the high-resolution characteristics of high-frequency probes in detail capture through the deep learning cascade enhancement model CU-Net, thus showing significant advantages and huge application potential in multi-bandwidth photoacoustic imaging tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the working process of the multi-bandwidth photoacoustic tomography system;

[0038] Figure 2 This is a schematic diagram of the deep learning cascade enhancement model CU-Net proposed in the present invention;

[0039] Figure 3 This is an example of the application effect of the method proposed in the present invention in multi-bandwidth photoacoustic reconstruction. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0041] The present invention provides a multi-bandwidth photoacoustic imaging method based on deep learning cascade enhancement, the method comprising the following steps:

[0042] Step 1: Multi-bandwidth photoacoustic imaging system hardware composition and configuration

[0043] like Figure 1 As shown, the multi-bandwidth photoacoustic imaging system proposed in the present invention is composed of the following main components: a pulsed laser, two semi-annular array probes with different center frequencies, a multi-channel data acquisition card and a central computer. The pulsed laser is used to excite the light absorber in the target area to generate a photoacoustic signal. The two semi-annular array probes respectively have a high-frequency probe with a center frequency of 5.5MHz and a low-frequency probe with a center frequency of 2.5MHz. The high-frequency probe is mainly used to obtain high-resolution images, while the low-frequency probe is used to detect larger light absorbers and structural features with low signal-to-noise ratio. The multi-channel data acquisition card is used to digitize the photoacoustic signals received by the probe, and transmit the processed data to the central computer to reconstruct the photoacoustic image based on the filtered back projection reconstruction algorithm.

[0044] Step 2: Signal acquisition and preprocessing

[0045] During the experimental operation, the pulsed laser emits short pulses of light, which are irradiated on the tissue within the target imaging area to stimulate photoacoustic signals. The generation of photoacoustic signals can be expressed by the following formula:

[0046] p 0 (r) = Γ·μ a (r)·F(r)

[0047] Among them, p 0 (r) represents the initial sound pressure distribution at position r, Γ is the Grüneisen coefficient, μ a (r) is the light absorption coefficient of the tissue, and F(r) represents the local light energy density.

[0048] The propagation of photoacoustic signals in tissue can be described by the following wave equation:

[0049]

[0050] Here, p(r,t) is the acoustic pressure at position r at time t, c is the speed of sound in the tissue, and δ(t) is the instantaneous excitation of the laser pulse.

[0051] The two semi-circular array probes receive photoacoustic signals of different frequencies respectively, which are collected by a multi-channel data acquisition card and stored in a central computer. In order to ensure the accuracy and efficiency of subsequent image reconstruction, the collected photoacoustic signals must first be preprocessed. The preprocessing steps include removing noise, enhancing the effective part of the signal, and normalizing the data. The preprocessed photoacoustic signals use traditional reconstruction algorithms such as delayed sum (DAS) to generate photoacoustic image data, and are divided into training sets, validation sets, and test sets according to a certain ratio for subsequent deep learning cascade network training and performance verification.

[0052] Step 3: Construction and training of deep learning cascade network

[0053] In order to achieve high-quality image reconstruction from coarse to fine, the present invention constructs a cascade deep learning network (Cascade U-Net, CU-Net). The network consists of two U-Net modules connected in series. First, the high-frequency photoacoustic image x generated by preprocessing is h and low-frequency photoacoustic image x l Shallow feature extraction (Conv3×3) is performed through convolution kernels of size 3×3, and then the extracted features are fused along the channel direction and input into the first U-Net module of the encoding-decoding structure. During the encoding process, this module extracts features through a double-layer Double Conv3×3 convolution layer and downsamples through a four-level maximum pooling layer. During the decoding process, upsampling is performed through a Double Deconv3×3 deconvolution layer, and finally a rough reconstruction is completed through an output layer with a convolution kernel size of 1×1 to generate a preliminary full-ring image y 1 Next, the generated intermediate image y 1 The high-frequency image x of the original input h and the low frequency image x l The features are fused along the channel direction, and the fused features are input into the second U-Net module of the encoder-decoder structure. This module also performs fine reconstruction through the encoder-decoder structure to generate the final full ring image y 2 , thereby improving the quality and resolution of the image.

[0054] During the training process, the mean square error (MSE) loss function is used to measure the difference between the reconstructed image and the real image. To ensure that both stages of the cascaded U-Net play a key role in improving the overall quality and accuracy of the reconstructed image, the present invention designs a joint loss function, which consists of two parts:

[0055]

[0056] Among them, L 1 y is the output of the first-level rough reconstruction model 1 The MSE loss between the real image y.

[0057]

[0058] Among them, L 2 y is the output of the second-level fine reconstruction model 2 The MSE loss between the real image y and the final joint loss function L is expressed as:

[0059] L=αL 1 +βL 2

[0060] Where α is the coarse reconstruction L 1 Loss weight, β is the fine reconstruction L 2 Loss weights, in this embodiment, both α and β are set to 1.

[0061] The Adam optimizer is used to optimize and adjust the network parameters. The learning rate during training is set to an initial value of 0.0001, and a step-by-step reduction method is used for optimization to ensure the stability and convergence of the training.

[0062] Step 4: Multi-bandwidth photoacoustic reconstruction:

[0063] The photoacoustic images collected by the 2.5MHz half-ring array and the photoacoustic images collected by the 5.5MHz half-ring array were used as the input of the deep learning cascade network. The trained deep learning cascade network was used to perform multi-bandwidth photoacoustic reconstruction to obtain 2.5MHz and 5.5MHz full-ring photoacoustic images.

[0064] Step 5: System Verification and Application

[0065] In order to verify the effectiveness of the present invention, in vivo experiments were conducted to verify its reliability and applicability in practical applications.

[0066] The following are examples of reconstruction indicators and visualization effects of the model of the present invention, demonstrating its excellent performance in practical applications. This system was applied to photoacoustic imaging of the abdomen of mice. The low-frequency 2.5MHz photoacoustic signal and high-frequency 5.5MHz photoacoustic signal were collected, and 240 photoacoustic image data were generated using a delay-sum algorithm. The data were divided into training set, validation set, and test set in a ratio of 7:1:2. The low-frequency 2.5MHz semi-circular (180°) array photoacoustic image x l and high frequency half ring (180°) array photoacoustic image x h As the input of the deep learning cascade enhancement model CU-Net, a low-frequency 2.5MHz full-view (360°) photoacoustic image and a high-frequency 2.5MHz full-view (360°) photoacoustic image were used as the imaging target y to test the model reconstruction performance. The experimental results show that the cascade deep learning model CU-Net proposed in the present invention shows good reconstruction performance in the application of multi-bandwidth photoacoustic imaging, and performs well in image quality indicators such as PSNR, MAE and SSIM, as shown in Table 1.

[0067] Table 1 Abdominal reconstruction indexes of mice

[0068]

[0069] Figure 3 The visualization results of CU-Net in multi-bandwidth photoacoustic imaging tasks are shown, which effectively captures fine structural details, significantly improves imaging quality, and demonstrates its excellent performance in practical applications. This system integrates high- and low-frequency probes, and through the deep learning cascade enhancement model CU-Net, it combines the wide coverage capability of low-frequency probes in overall structural imaging and the high-resolution characteristics of high-frequency probes in detail capture, thus showing significant advantages and huge application potential in multi-bandwidth photoacoustic imaging tasks.

Claims

1. A multi-bandwidth photoacoustic imaging system based on deep learning cascade enhancement, characterized in that The multi-bandwidth photoacoustic imaging system is composed of a pulsed laser, two semi-circular array probes with different center frequencies, a multi-channel data acquisition card and a central computer, wherein: The pulse laser is used to excite the light absorber in the target area to generate a photoacoustic signal; The two semi-circular array probes respectively have a high-frequency probe with a center frequency of 5.5 MHz and a low-frequency probe with a center frequency of 2.5 MHz, the high-frequency probe is used to obtain high-resolution images, and the low-frequency probe is used to detect larger light absorbers and structural features with low signal-to-noise ratio; The multi-channel data acquisition card is used to digitally process the photoacoustic signals received by the two semi-circular array probes, and transmit the processed data to a central computer to reconstruct the photoacoustic image based on a filtered back projection reconstruction algorithm.

2. A multi-bandwidth photoacoustic imaging method based on deep learning cascade enhancement, characterized in that The method comprises the following steps: Step 1: Multi-bandwidth photoacoustic imaging system hardware composition and configuration The multi-bandwidth photoacoustic imaging system consists of a pulsed laser, two semi-circular array probes with different center frequencies, a multi-channel data acquisition card and a central computer, among which: The pulse laser is used to excite the light absorber in the target area to generate a photoacoustic signal; The two semi-circular array probes respectively have a high-frequency probe with a center frequency of 5.5 MHz and a low-frequency probe with a center frequency of 2.5 MHz, the high-frequency probe is used to obtain high-resolution images, and the low-frequency probe is used to detect larger light absorbers and structural features with low signal-to-noise ratio; The multi-channel data acquisition card is used to digitally process the photoacoustic signals received by the two semi-circular array probes, and transmit the processed data to the central computer to reconstruct the photoacoustic image based on the filtered back projection reconstruction algorithm; Step 2: Signal acquisition and preprocessing Step 2.1: The pulse laser emits short pulses of light, which are irradiated onto the tissue within the target imaging area to stimulate photoacoustic signals; Step 2.2: The two semi-circular array probes receive photoacoustic signals of different frequencies respectively, and these photoacoustic signals are collected by a multi-channel data acquisition card and stored in a central computer; Step 2.3: The central computer pre-processes the collected photoacoustic signals; Step 2.4: Use the reconstruction algorithm to generate high-frequency photoacoustic image x after preprocessing the photoacoustic signal h and low-frequency photoacoustic images x l , and divided into training set, validation set and test set for subsequent deep learning cascade network training and performance verification; Step 3: Construction and training of deep learning cascade network Step 3.1: The cascaded deep learning network consists of two U-Net modules in series. First, the high-frequency photoacoustic image x generated by preprocessing is h and low-frequency photoacoustic images x l Shallow feature extraction is performed through convolution kernels of size 3×3, and then the extracted features are fused along the channel direction, and the fused features are input into the first U-Net module of the encoding-decoding structure. In the encoding process, this module extracts features through a double-layer Double Conv3×3 convolution layer and implements downsampling through a four-level maximum pooling layer. In the decoding process, upsampling is performed through a Double Deconv3×3 deconvolution layer, and finally a rough reconstruction is completed through an output layer with a convolution kernel size of 1×1 to generate a preliminary full-ring image y1; then, the generated full-ring image y1 is compared with the original input high-frequency photoacoustic image x h and low-frequency photoacoustic images x l The features are fused along the channel direction, and the fused features are input into the second U-Net module of the encoding-decoding structure. During the encoding process, the module extracts features through a double-layer Double Conv3×3 convolution layer and implements downsampling through a four-level maximum pooling layer. During the decoding process, upsampling is performed through a Double Deconv3×3 deconvolution layer, and finally fine reconstruction is completed through an output layer with a convolution kernel size of 1×1 to generate the final full ring image y2; Step 3.2: During the training process, a joint loss function is used to measure the difference between the reconstructed image and the real image; Step 3.3: Use Adam optimizer to optimize the network parameters; Step 4: Multi-bandwidth photoacoustic reconstruction: The photoacoustic images collected by the 2.5MHz half-ring array and the photoacoustic images collected by the 5.5MHz half-ring array were used as the input of the deep learning cascade network. The trained deep learning cascade network was used to perform multi-bandwidth photoacoustic reconstruction to obtain 2.5MHz and 5.5MHz full-ring photoacoustic images.

3. The multi-bandwidth photoacoustic imaging method based on deep learning cascade enhancement according to claim 2 is characterized in that The joint loss function consists of two parts: L=αL1+βL2 Among them, L1 is the MSE loss between y1 output by the first-level coarse reconstruction model and the real image y, L2 is the MSE loss between y2 output by the second-level fine reconstruction model and the real image y, α is the L1 loss weight, and β is the L2 loss weight.

Images