Terahertz medical image denoising method based on SE module and DnCNN neural network

Abstract

The invention relates to a terahertz medical image denoising method based on an SE module and a DnCNN neural network, and the method comprises the steps: carrying out the simulation generation of a simulation noise which is similar to the statistics of an observation noise for an observed first noise-containing image, defining a second noise-containing image as a noise-containing image added with the simulation noise, inputting the noise-containing image into a DnCNN-SE model, carrying out the noise reduction processing, and training an NAC strategy of a self-supervision network, thereby achieving the denoising of a terahertz medical image. And training a DnCNN-SE model by taking a mean square error MSE between the denoised image and the first noisy image as a loss function, and directly carrying out denoising processing on the original noisy image y to obtain a final denoised image. An external clean sample is not needed in the process, an original noise image is regarded as a training target, and additional noise is superposed on the original noise image to serve as network input. By simulating the noise process, the model learns the general denoising capability, and the noise level does not need to be estimated in advance. And the SE module adaptively emphasizes channels with rich information and inhibits channels dominated by noise through extrusion and excitation operations, so that key information of the medical image is better reserved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a terahertz medical image denoising method based on an SE module and a DnCNN neural network. BACKGROUND

[0002] Terahertz waves refer to electromagnetic waves with a frequency of 0.1 THz to 10 THz (corresponding to a wavelength of 30 μm to 300 μm). Terahertz medical imaging has two core characteristics: non-ionizing and molecular fingerprint spectrum. Its energy is far below the ionizing radiation threshold that destroys biological molecules, and it is highly safe; at the same time, its specific frequency can resonate with the molecular vibration, rotation and collective mode of water molecules and various biological macromolecules (such as proteins, DNA, lipids, etc.) in biological tissues, forming a highly specific molecular fingerprint. This makes terahertz imaging not only provide tissue morphological structure information, but also reflect its biochemical and pathological state without labeling.

[0003] However, terahertz imaging also faces significant challenges in biomedical applications: the high water content of biological tissues causes terahertz waves to attenuate strongly in tissues, and the imaging process is easily disturbed by environmental noise. Coupled with its complex coherent detection signal characteristics, the original image obtained often accompanies significant speckle noise, random noise and information blur. At present, the research on terahertz medical image denoising is relatively scarce, which is a key problem to be solved.

[0004] In the field of medical image denoising, DnCNN is a classic supervised denoising algorithm based on deep learning, which effectively removes Gaussian noise through an end-to-end trainable deep convolutional neural network. However, when applied to real-world scenarios (including terahertz medical image denoising), there are obvious limitations: first, as a supervised network, DnCNN training requires a large number of noise-clean image pairs, which limits the model's generalization ability and contradicts the reality of the difficulty of obtaining clean terahertz image samples; second, DnCNN is usually trained for Gaussian noise with a specific noise level, and its performance decreases when dealing with unknown noise levels or non-Gaussian noise (especially the complex noise in real medical images); in addition, when its convolutional layer fuses spatial and channel information, it defaults to equal weights for each channel, ignoring the different contributions of information in different frequency bands in medical images.

[0005] To solve the above problems, the application intends to improve the DnCNN, integrate the NAC (Noise As Clean) self-supervised strategy and the SE (Squeeze-and-Excitation) channel attention module. The NAC strategy does not need external clean samples: it regards the original noisy (i.e. damaged) image as the training target, and adds additional noise on it as the network input. By simulating the noise process, the model learns the general denoising ability without needing to estimate the noise level in advance. The SE module adaptively emphasizes the channels with rich information and suppresses the channels dominated by noise through the Squeeze (global average pooling to aggregate spatial information) and Excitation (fully connected layer to learn channel weights) operations, so as to better preserve the key information of the medical image. By integrating the NAC strategy and the SE module into the DnCNN network, the shortcomings of the original DnCNN in the application of terahertz medical image denoising are made up. SUMMARY

[0006] The application aims to provide a terahertz medical image denoising method based on the SE module and the DnCNN neural network, to solve the technical problems in the prior art that the DnCNN training needs a large number of noise-clean image pairs, which limits the model generalization ability and is contrary to the reality that the terahertz image clean samples are difficult to obtain; secondly, the DnCNN is usually trained for a specific noise level of Gaussian noise, and the performance is reduced when dealing with unknown noise level or non-Gaussian noise (especially the complex noise in real medical images); in addition, when the convolutional layer fuses the spatial and channel information, it defaults that the weights of each channel are equal, ignoring the different contribution of the information in different frequency bands in the medical image.

[0007] To solve the above technical problems, the technical scheme adopted by the application is as follows:

[0008] The terahertz medical image denoising method based on the SE module and the DnCNN neural network comprises the following steps:

[0009] S1: For the observed first noisy image y, i.e. the image containing observation noise, simulate to generate simulated noise similar to the observation noise statistics, and define the second noisy image as the noisy image added with the simulated noise;

[0010] S2: input the second noisy image into the DnCNN-SE model for denoising processing to obtain the denoised image;

[0011] S3: train the DnCNN-SE model with the mean square error (MSE) between the denoised image and the first noisy image as the loss function, and after the training is completed, directly perform denoising processing on the original noisy image y to obtain the final denoised image.

[0012] Preferably, the specific process of step S1 is as follows:

[0013] S11: represent the observed first noisy image as y = x + n o , where x is the clean image, n o is the observation noise;

[0014] S12: simulate to generate a noise n o statistically similar to n s , i.e. E[n o ] ≈ E[n s ], Var[n o ] ≈ Var[n s ], define the second noisy image as z = y + n o + n s .

[0015] Preferably, the specific process of step S2 is as follows:

[0016] S21: input the second noisy image z into a 20-layer DnCNN-SE model for processing, the first layer generates C feature maps using C 3x3xnc convolution kernels, nc=1 for grayscale images and nc=3 for color images, C=64 is taken in this network, and then the ReLU activation function is used for nonlinear activation;

[0017] S22: each layer from the second layer to the 19th layer of the DnCNN-SE model uses C 3x3xnc convolution kernels to generate C feature maps, then performs batch normalization, i.e. normalizes the input to a standard normal distribution with a mean of 0 and a variance of 1, and finally inputs to the output layer after nonlinear activation by the ReLU activation function;

[0018] S23: the output layer uses nc 3x3x64 convolution kernels to reconstruct the output and obtain the denoised image f θ (z), f θ denotes the image processing function of the DnCNN-SE model; θ is the parameter of the model.

[0019] Preferably, the specific process of step S3 is as follows:

[0020] S31: use mean square error MSE as the loss function for model training, i.e.

[0021]

[0022] where N is the number of image blocks, denotes the loss function; ‖·‖ is the l2 norm; z i denotes the second noisy image block; y idenotes the first noisy image block (i = 1, 2,..., N); the DnCNN-SE neural network defines an image processing function f θ when the model parameters are θ, which functions to learn a mapping from the second noisy image z to the noise estimate f θ (z), and the ultimate goal is to obtain a denoised image by subtracting the predicted noise from the first noisy image y;

[0023] S32: training the network with the training target of minimizing the loss function to obtain the most suitable parameters;

[0024] S33: performing denoising processing on the original noisy image y through the trained DnCNN-SE model to obtain a denoised image

[0025] Preferably, the step S22 further comprises the following process:

[0026] After processing through each of the two intermediate layers, a feature map of HxWxC is generated, where H and W are the width and height of the image, C = 64 in this network structure, and the feature map is denoted as U = [u1, u2,..., u C ], which is input to the SE module, and first squeezed, i.e., through a global pooling layer to generate a channel-level statistic v ∈ R, where the c-th element of v is c = 1, 2, 3,..., C.

[0027] Excitation processing, i.e., through two fully connected layers to learn channel weights and achieve nonlinear calibration, is denoted as s = σ(g(v, W)) = σ(W2δ(W1v)), where σ is a Sigmoid activation function δ is a ReLU activation function max(0, ·), The output of the SE module is normalized by U and s, i.e. Output c = 1, 2, 3,..., C.

[0028] The feature map is repeatedly processed through the above-mentioned two intermediate layers 9 times, and after processing through the SE module once, it is input to the output layer.

[0029] Preferably, the specific execution process of the SE module on the image is as follows:

[0030] Squeezing operation: global average pooling is used, i.e., all pixel values of the feature map are added to obtain an average, spatial information is aggregated, and a channel descriptor is generated; after the input Z of the intermediate layer is processed through convolution, batch normalization, and ReLU activation, a feature map of HxWxC is generated, where H and W are the width and height of the image, C = 64 in this network structure, and the feature map is denoted as U = [u1, u2,..., uC ] by a global pooling layer, to generate a channel-level statistic v e R, where the c-th element of v is c=1,2,,3,...,C;

[0031] Excitation operation: learn the channel weights through two fully connected layers to achieve nonlinear calibration;Express the excitation operation as s=σ(g(v,W))=σ(W2δ(W1v)),where δ is the Sigmoid activation function, δ is the ReLU activation function, The output of the SE module is normalized by U and s, that is, Finally output

[0032] Preferably, the network depth nb of the DnCNN-SE model is 20, including an input layer, i.e., the first layer;Middle layer, the second layer to the nb-1 layer, i.e., the second layer to the 19th layer;Output layer, i.e., the 20th layer;

[0033] Input layer: input noise image z e H x W x nc, where H and W are the width and height of the image, nc is the input channel number, nc=1 for gray image and nc=3 for color image;First use 64 3 x 3 x C convolution kernels to generate 64 feature maps, and then use the ReLU activation function max(0, ·) for nonlinear activation;

[0034] Middle layer including Conv+BN+ReLU: use 64 3 x 3 x nc convolution kernels to generate 64 feature maps, where nc=1 for gray image and nc=3 for color image;Then, batch normalization, BN, that is, normalize the input to a standard normal distribution with mean 0 and variance 1, and finally pass through the ReLU activation function max(0, ·) for nonlinear activation;After repeating this operation twice, an SE module is added;

[0035] Output layer: the last layer uses nc 3 x 3 x 64 convolution kernels to reconstruct the output.

[0036] The beneficial effects of the present application include:

[0037] The application provides a terahertz medical image denoising method based on an SE module and a DnCNN neural network. First, in view of the fact that clean samples are difficult to obtain in a terahertz medical image, terahertz imaging is affected by tissue attenuation and environmental interference, and it is almost impossible to obtain a pure medical image without noise, the NAC strategy innovatively directly uses an original noisy image (namely, a first noisy image y) as a training target, generates a second noisy image z as network input by superimposing additional simulated noise with similar statistics on the original noisy image. This design completely abandons the strong dependence of the traditional DnCNN on paired noise-clean samples, does not require additional acquisition or construction of a pure terahertz medical image dataset, greatly reduces the data acquisition cost and data preprocessing complexity of the terahertz medical image denoising task, and solves the technical problem caused by the data scarcity in the terahertz medical imaging field.

[0038] Secondly, the traditional DnCNN needs to be trained for a specific noise level, and its performance significantly decreases when facing unknown noise levels (noise intensity differences of different acquisition devices and different tissue samples) and non-Gaussian noise (such as coherent speckle noise and pulse noise mixed scenes) in terahertz medical images. The NAC strategy dynamically superimposes simulated noise on the original noisy image during training, so that the model does not need to estimate the noise level in advance and can learn the general denoising rule covering a wider noise distribution. This feature perfectly adapts to the dynamics and complexity of terahertz medical image noise, and even when facing unknown noise intensity or noise type, it can still maintain stable denoising effect, greatly improving the generalization ability of the model to the terahertz medical imaging scene.

[0039] Thirdly, the convolutional layer of the traditional DnCNN defaults to equal weights of each channel, and cannot distinguish between the effective information channels containing tissue texture and lesion edges and the redundant information channels dominated by noise in the terahertz medical image, which easily leads to the contradiction between deleting effective information and retaining noise details during denoising. The SE module realizes channel-level attention optimization through the squeeze-excitation mechanism: the squeeze operation aggregates the spatial information of the feature map to generate a channel-level statistic, allowing the model to accurately perceive the contribution of different channels to denoising and detail retention; the excitation operation gives high weights to information-rich effective channels and low weights to noise-dominated redundant channels through nonlinear mapping, realizing adaptive calibration of channel features. The model can specifically suppress noise channels in the terahertz image while avoiding excessive suppression of effective information channels, and the thoroughness of denoising is significantly better than that of the traditional DnCNN.

[0040] Thirdly, the core value of terahertz medical images lies in the biochemical information and morphological details contained therein, which directly determine the accuracy of subsequent medical diagnosis. The SE module calibrates the channel weight and prioritizes the preservation of key details related to diagnosis during the noise reduction process: high weight is given to high-frequency feature channels that reflect lesion boundaries to ensure that the lesion edges are clear and identifiable after noise reduction; low weight is given to meaningless noise high-frequency channels to achieve precise noise filtering. This effectively addresses the shortcomings of traditional DnCNN noise reduction, which blurs details, making the denoised terahertz medical images both free of obvious noise interference and retaining all the key information needed for diagnosis, thus truly having clinical application value.

[0041] Finally, the training phase only needs to use the input images generated by the original noisy images + simulated noise to complete model optimization, without the need for complex sample pairing or noise level labeling; in the inference phase, the original noisy terahertz medical images are directly input into the trained model, and the denoising results can be quickly output. The simplified training-inference process reduces the model's dependence on hardware devices, while avoiding additional data preprocessing operations during inference, improving the efficiency of image processing in clinical scenarios. Combined with the 20-layer deep structure of DnCNN and the intermediate layer Conv+BN+ReLU combination: the deep network can fully capture the complex nonlinear relationship between noise and signal in terahertz images, and the BN layer solves the gradient vanishing / explosion problem of deep network training, accelerating model convergence and improving training stability. This design enables the model to be trained stably with limited terahertz medical image data, avoiding training fluctuations caused by insufficient data or complex noise, and ensuring the consistency and reliability of the denoising effect. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The flowchart of the terahertz medical image denoising method based on the SE module and the DnCNN neural network of the present application.

[0043] Figure 2 The architecture diagram of the SE module of the present application.

[0044] Figure 3 The DnCNN-SE network architecture diagram of the present application.

[0045] Figure 4 The NAC strategy diagram of the present application. DETAILED DESCRIPTION

[0046] The following will be described in detail with reference to the accompanying drawings Figures 1-4 Further detailed description of the present application:

[0047] The existing DnCNN is an end-to-end image denoising deep learning model, which adopts a 17-layer convolutional architecture. The core breakthrough lies in the collaborative design of residual learning and batch normalization, which greatly promotes the learning ability of the network, not only accelerating the training process, but also improving the denoising performance. However, when facing real image denoising tasks, the image prior and noise statistics during DnCNN training are different from real test images.

[0048] The present application is a noisy image y=x+n o , wherein x is a clean image, n o is an observed noise, a noise n o similar in statistics is simulated, and z=y+n s is defined o +n s As the input of the 20-layer DnCNN-SE (i.e. f θ ), after 3x3 convolution and ReLU activation function processing in the input layer (layer 1), it is processed through 3x3 convolution, batch normalization (BN, i.e. normalizing the input of this layer to a standard normal distribution with mean 0 and variance 1) and ReLU activation function (i.e. max(0,·)) in the middle layer (layer 2 to layer 19), and after every two layers, a Squeeze-and-Excitation (SE) module is added for processing. Finally, the output layer (layer 20) is output through 3x3 convolution to obtain f θ (z), and the Noise As Clean (NAC) strategy is applied to output the mean square error (MSE) between f θ (z) and the noisy image y as the loss function, and the network is trained to minimize the loss function. The trained network is used to process the noisy image y to obtain the denoised image.

[0049] The present application improves the DnCNN and integrates the NAC (Noise As Clean) self-supervised strategy and the SE (Squeeze-and-Excitation) channel attention module. The NAC strategy does not require external clean samples: it regards the original noisy (i.e. damaged) image as the training target, and superimposes additional noise on it as the network input. By simulating the noise process, the model learns the general denoising ability without the need to estimate the noise level in advance. The SE module adaptively emphasizes the channels with rich information and suppresses the channels dominated by noise through Squeeze (global average pooling to aggregate spatial information) and Excitation (fully connected layer to learn channel weights), thereby better preserving the key information of medical images. By integrating the NAC strategy and the SE module into the DnCNN network, the shortcomings of the original DnCNN in terahertz medical image denoising applications are effectively overcome.

[0050] Example 1

[0051] See appendix Figure 1 As shown, the terahertz medical image denoising method based on the SE module and DnCNN neural network includes the following steps:

[0052] S1: For the first noisy image y observed, i.e. the image containing observed noise, simulate noise that is statistically similar to the observed noise is generated, and the second noisy image is defined as the noisy image with added simulated noise;

[0053] S2: Input the second noisy image into the DnCNN-SE model for noise reduction to obtain the denoised image;

[0054] S3: Train the DnCNN-SE model using the mean square error (MSE) between the denoised image and the first noisy image as the loss function. After training, directly denoise the original noisy image y to obtain the final denoised image.

[0055] In this process, a novel approach is taken: using a first noisy image y (a real-world noisy image) as a base, a second noisy image is generated solely by simulating noise statistically similar to the observed noise (matching key statistical characteristics such as mean and variance), without requiring any clean samples for training. This breaks through the data bottleneck of terahertz medical image denoising, eliminating the need for additional acquisition or construction of clean datasets, significantly reducing experimental costs and data preprocessing complexity. The generated simulated noise is not random noise, but highly consistent with the statistical characteristics of the observed noise in the first noisy image, such as the distribution characteristics of speckle noise and Gaussian mixture noise commonly found in terahertz images. This noise simulation closely resembles the real-world scenario, ensuring a high degree of match between the noise distribution of the second noisy image and the actual noise distribution of terahertz imaging, preventing the model from overfitting to a single noise type due to significant differences between the training data and the real noise. Even when faced with terahertz images from different acquisition devices and tissue samples (with slight differences in noise intensity or type), the model can still stably denoise, demonstrating significantly better generalization ability than traditional methods trained with fixed noise.

[0056] The second noisy image is essentially a statistically similar simulated noise superimposed on a real noisy image y. This is equivalent to providing the model with training samples that enhance noise diversity without increasing the amount of original images acquired. For example, the same noisy image y can generate multiple different second noisy images by adjusting subtle parameters of the simulated noise within a statistically similar range. This allows the model to learn richer noise feature patterns, rather than being limited to a single noise distribution of a small number of samples, further improving its ability to process complex noise in terahertz images.

[0057] The mean square error MSE of the denoised image and the first noisy image y is taken as the loss function, and the pixel-level error is directly minimized. On the one hand, the optimization target of MSE is clear, which can make the denoised image consistent with the effective structure of the original noisy image at the global pixel level, and avoid local distortion, such as shape and position offset of the lesion area. On the other hand, the mathematical property of MSE is simple, which facilitates gradient calculation and model optimization, reduces training complexity, and ensures stable convergence of the model under limited data.

[0058] The specific process of step S1 is as follows:

[0059] S11: The observed first noisy image is represented as: y = x + n o , wherein x is a clean image, and n o is an observation noise;

[0060] S12: Simulate to generate a noise n o statistically similar to n s , that is, E[n o ] ≈ E[n s ], Var[n o ] ≈ Var[n s ], and define the second noisy image as: z = y + n o + n s .

[0061] The specific process of step S2 is as follows:

[0062] S21: The second noisy image z is input into the 20-layer DnCNN-SE model for processing, and the first layer generates C feature maps using C 3x3xnc convolution kernels (C = 64 is taken in this network), and then performs nonlinear activation through the ReLU activation function;

[0063] S22: Each layer from the second layer to the 19th layer of the DnCNN-SE model uses C 3x3xnc convolution kernels to generate C feature maps, and then performs batch normalization processing, that is, normalizes the input to a standard normal distribution with a mean of 0 and a variance of 1, and finally inputs to the output layer after nonlinear activation through the ReLU activation function;

[0064] S23: The output layer uses nc 3x3x64 convolution kernels to reconstruct the output, and obtains the denoised image f θ (z).

[0065] The specific process of step S3 is as follows:

[0066] S31: Take the mean square error MSE as the loss function of model training, that is:

[0067]

[0068] wherein N is the number of image blocks;

[0069] S32: training the network with the minimum loss function as the training target to obtain the most suitable parameters;

[0070] S33: performing denoising processing on the original noisy image y through the trained DnCNN-SE model to obtain a denoised image

[0071] Embodiment 2

[0072] On the basis of Embodiment 1, the following process is further included in step S22:

[0073] After each two intermediate layers are processed, a feature map of HxWxC is generated, and the feature map is denoted as U=[u1, u2,..., u C ], which is input to the SE module. First, squeezing processing is performed, that is, a global pooling layer is used to generate a channel-level statistic v e R, wherein the c-th element of v is c=1, 2, 3,..., C;

[0074] Excitation processing is performed, that is, two fully connected layers are used to learn channel weights, and non-linear calibration is realized, which is denoted as s=σ(g(v,W))=σ(W2δ(W1v)), wherein σ is a Sigmoid activation function, δ is a ReLU activation function, The output of the SE module is obtained by normalizing U and s, that is, Output

[0075] The feature map is repeatedly processed by the above two intermediate layers for 9 times, and after the SE module is processed once, the output layer is input.

[0076] Referring to Figure 2 , the specific execution process of the SE module on the image is as follows:

[0077] Squeezing operation: a global average pooling is used, that is, all pixel values of the feature map are added to obtain an average, spatial information is aggregated, and a channel descriptor is generated; after the input Z of the intermediate layer is processed by convolution, batch normalization, and ReLU activation, a feature map of HxWxC is generated, wherein H and W are the width and height of the image, and C=64 is taken in this network structure, and the feature map is denoted as U=[u1, u2,..., u C ], a channel-level statistic v e R is generated through a global pooling layer, wherein the c-th element of v is

[0078] The excitation operation: learn the channel weight through two full connection layers to realize the nonlinear calibration; the excitation operation is expressed as s = σ (g (v, w)) = σ (W2δ (W1v)), wherein σ is a Sigmoid activation function, and δ is a ReLU activation function, The output of the SE module is normalized by U and s, that is The final output

[0079] Embodiment 3

[0080] Based on the embodiment 1 or the embodiment 2, refer to Figure 3 The network depth nb of the DnCNN-SE model is 20, including an input layer, i.e. the first layer; a middle layer, the second layer to the (nb-1)th layer, i.e. the second layer to the 19th layer; and an output layer, i.e. the 20th layer;

[0081] The input layer: input a noise image z ∈ H × W × C, wherein H and W are the width and height of the image, and nc is the input channel number (nc = 1 for a gray image, and nc = 3 for a color image); 64 3 × 3 × C convolution kernels are used to generate 64 feature maps, and then a ReLU activation function max (0, ·) is used for nonlinear activation;

[0082] The middle layer including Conv+BN+ReLU: 64 3 × 3 × nc convolution kernels are used to generate 64 feature maps, wherein nc = 1 for a gray image, and nc = 3 for a color image; then batch normalization processing, BN, is performed, i.e. the input is normalized to a standard normal distribution with a mean of 0 and a variance of 1; finally, a ReLU activation function max (0, ·) is used for nonlinear activation; after repeating this operation twice, an SE module is added;

[0083] The output layer: the last layer uses nc 3 × 3 × 64 convolution kernels to reconstruct the output.

[0084] Refer to Figure 4 The NAC (Noise As Clean) strategy diagram for training the self-supervised network, wherein the network f θ is the DnCNN-SE, wherein θ refers to the parameters in the network. For an observed noisy image y = x + n o , wherein x is a clean image, and n o is an observation noise, a noise n o similar to n s is simulated, i.e. E[n o ] ≈ E[n s ], Var[n o ] ≈ Var[n s ], and a noisy image z = y + n o +ns The z is input to the DnCNN-SE for denoising to obtain f θ (z), and a loss function mean square error (MSE) is given, that is Wherein N is the number of image blocks, and the network is trained with the training target of minimizing the loss function to obtain the most suitable parameters, and after the network training is completed, the original noisy image y is directly denoised to obtain the denoised image The network trained according to the strategy is different from a traditional neural network which needs a large number of paired training sets, and an exclusive model is independently trained for each image.

[0085] In summary, the terahertz medical image denoising method based on the SE module and the DnCNN neural network provided by the application simulates and generates simulated noise similar to the statistical noise of the observed first noisy image, defines the second noisy image as the noisy image added with the simulated noise, inputs the DnCNN-SE model for denoising, trains the NAC strategy of the self-supervised network, trains the DnCNN-SE model with the mean square error (MSE) between the denoised image and the first noisy image as the loss function, directly denoises the original noisy image y, and obtains the final denoised image. The process does not need external clean samples: it regards the original noisy image as the training target, and superimposes additional noise on it as the network input. Through the simulated noise process, the model learns the general denoising ability without needing to estimate the noise level in advance. The SE module adaptively emphasizes the information-rich channels and suppresses the noise-dominant channels through squeezing and excitation operations, so that the key information of the medical image can be better preserved.

Claims

1. A terahertz medical image denoising method based on SE module and DnCNN neural network, characterized in that, Includes the following steps: S1: For the first noisy image y observed, i.e. the image containing observed noise, simulate noise that is statistically similar to the observed noise is generated, and the second noisy image is defined as the noisy image with added simulated noise; S2: Input the second noisy image into the DnCNN-SE model for noise reduction to obtain the denoised image; S3: By training the self-supervised network using the NAC strategy, the DnCNN-SE model is trained with the mean square error (MSE) between the denoised image and the first noisy image as the loss function. After training, the original noisy image y is directly denoised to obtain the final denoised image.

2. The terahertz medical image denoising method based on the SE module and DnCNN neural network according to claim 1, characterized in that, The specific process of step S1 is as follows: S11: The first noisy image observed is represented as: y = x + n o Where x is a clean image and n o To observe noise; S12: Simulation generation and n o Statistically similar noise n s That is, E[n o ]≈E[n s ], Var[n o ]≈Var[n s The second noisy image is defined as: z = y + n o +n s .

3. The terahertz medical image denoising method based on the SE module and DnCNN neural network according to claim 2, characterized in that, The specific process of step S2 is as follows: S21: Input the second noisy image z into the 20-layer DnCNN-SE model for processing. The first layer uses C 3×3×nc convolutional kernels to generate C feature maps. For gray images, nc=1, and for color images, nc=E. In this network, C=64 is used. Then, non-linear activation is performed through the ReLU activation function. S22: In the DnCNN-SE model, each of layers 2 to 19 uses C 3×3×nc convolutional kernels to generate C feature maps. After batch normalization, the input is normalized to a standard normal distribution with a mean of 0 and a variance of 1. Finally, it is non-linearly activated by the ReLU activation function and then input to the output layer. S23: The output layer uses nc 3×3×64 convolutional kernels to reconstruct the output, obtaining the denoised image f. θ (z), f θ Refers to the image processing function of the DnCNN-SE model; θ is the model parameter.

4. The terahertz medical image denoising method based on the SE module and DnCNN neural network according to claim 3, characterized in that, The specific process of step S3 is as follows: S31: The loss function for model training is the mean squared error (MSE), i.e.: Where N is the number of image patches, Refers to the loss function; ||·|| is the l2 norm; z i Refers to the second noisy image patch; y i The first noisy image patch (i = 1, 2, ..., N) is defined; when the model parameter is θ, the DnCNN-SE neural network defines the image processing function f. θ The function learns a noise estimate f from the second noisy image z. θ The mapping of (z) has the ultimate goal of obtaining the denoised image by subtracting the predicted noise from the first noisy image y; S32: Train the network with the goal of minimizing the loss function to obtain the most suitable parameters; S33: The original noisy image y is denoised using the trained DnCNN-SE model to obtain the denoised image.

5. The terahertz medical image denoising method based on the SE module and DnCNN neural network according to claim 3, characterized in that, Step S22 also includes the following process: After each two intermediate layers, a feature map of H×W×C is generated, where H and W are the width and height of the image. In this network, C = 64. Let this feature map be U = [u1, u2, ..., u...]. C The input is fed into the SE module, where it undergoes compression processing, specifically through a global pooling layer, to generate channel-level statistics v∈R, where the c-th element of v is... The activation process involves learning the channel weights through two fully connected layers to achieve nonlinear calibration, expressed as s = σ(g(v, W)) = σ(W2δ(W1v)), where σ is the Sigmoid activation function. δ is the ReLU activation function max(0,·). The output of the SE module is obtained by normalizing U and s, i.e. Output The feature map is repeated 9 times in the two intermediate layers mentioned above. After one SE module processing, it is input to the output layer.

6. The terahertz medical image denoising method based on the SE module and DnCNN neural network according to claim 5, characterized in that, The specific execution process of the SE module for images is as follows: The compression operation uses global average pooling, which sums and averages all pixel values ​​in the feature map to aggregate spatial information and generate channel descriptors. The input Z in the intermediate layer, after convolution, batch normalization, and ReLU activation, generates an H×W×C feature map, where H and W are the width and height of the image. In this network structure, C = 64, and the feature map is denoted as U = [u1, u2, ..., u...]. C Through a global pooling layer, channel-level statistics v∈R are generated, where the c-th element of v is... c = 1, 2, ..., C; Activation operation: Channel weights are learned through two fully connected layers to achieve nonlinear calibration; the activation operation is expressed as s=σ(g(v,W))=σ(W2δ(W1v)), where σ is the Sigmoid activation function and δ is the ReLU activation function. The output of the SE module is obtained by normalizing U and s, i.e. Final output 7. The terahertz medical image denoising method based on the SE module and DnCNN neural network according to claim 6, characterized in that, The network depth nb of the DnCNN-SE model is 20, including the input layer, which is the first layer; The intermediate layers are layers 2 through nb-1, i.e., layers 2 through 19; the output layer is layer 20. Input layer: Input noisy image z∈H×W×nc, where H and W are the width and height of the image, nc is the number of input channels, nc=1 for gray image and nc=3 for color image. First, 64 feature maps are generated using 64 3×3×C convolutional kernels, and then non-linear activation is performed using the ReLU activation function max(0,·). The intermediate layer includes Conv+BN+ReLU: 64 feature maps are generated using 64 3×3×nc convolutional kernels, where nc=1 for grayscale images and nc=3 for color images. After batch normalization (BN, which normalizes the input to a standard normal distribution with a mean of 0 and a variance of 1), the ReLU activation function max(0,·) is used for non-linear activation. This operation is repeated twice, and then an SE module is added. Output layer: The last layer uses nc 3×3×64 convolutional kernels to reconstruct the output.

Images