Breast cancer pathology image classification system based on adaptive threshold wavelet denoising

Abstract

The present invention proposes a breast cancer pathology image classification system based on adaptive threshold wavelet denoising. The loss function of the classification model training of breast cancer pathology images is associated with the selection of wavelet threshold, so that the wavelet threshold is optimized during model training and gradually approaches the threshold with good denoising effect, thereby finding the wavelet threshold suitable for the image task and improving the classification accuracy of subsequent breast cancer pathology images.

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging technology, and in particular relates to a breast cancer pathology image classification system based on adaptive threshold wavelet denoising. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Cancer is the second leading cause of death worldwide, and breast cancer has now overtaken lung cancer as the world's leading cancer. Accurate diagnosis of breast cancer is crucial for successful treatment and reducing mortality. Various diagnostic techniques are used, such as biopsy, ultrasound (US), magnetic resonance imaging (MRI), and infrared thermography. Biopsy is the most reliable and widely used technique for detecting breast cancer. This requires a pathologist to observe the morphology of the tissue under a microscope and make a diagnosis. However, pathologists who can make an accurate diagnosis require years of training, and the diagnostic process is time-consuming and laborious. Targeted computer-aided diagnosis systems can help improve the efficiency of breast cancer histopathology diagnosis, making it easier to examine more patients.

[0004] In recent years, the CNN (convolutional neural network) method, which performs well in image classification tasks, has been widely used in medical image classification tasks. Currently, CNN has become the mainstream method in breast cancer histopathology image classification research and is significantly superior to traditional methods on public datasets. However, some studies have found that in image classification tasks, the output of a trained convolutional neural network can change dramatically due to slight changes in the input, and the quality of the image can significantly affect the classification accuracy. Studies have shown that histopathology images produced by optical microscopes contain noise, which mainly comes from three sources:

[0005] 1. Uneven light source or overly strong light will produce reflections or shadows, resulting in noise such as light spots or black spots in the image.

[0006] 2. Noise from the instrument itself can also affect image quality. For example, the optical components of a microscope may have chromatic aberration, distortion, and other issues, resulting in color shift, distortion, and other noise in the image.

[0007] 3. The tissue specimen preparation process may also affect image quality. For example, the slicing process may cause uneven slice thickness and cracks, resulting in discontinuous and unclear noise in the image.

[0008] Therefore, when analyzing tissue pathology images, noise needs to be processed to ensure the accuracy and reliability of the analysis results.

[0009] Wavelet transform, Gaussian filtering and other methods are used in the classification research of natural images or tissue pathology images. Among them, wavelet transform is widely used because it can well preserve the useful information in the original signal.

[0010] A key step in wavelet transform threshold denoising is selecting the threshold. Existing methods often rely on observing pixel histograms, calculating pixel medians, and using common threshold calculation formulas. These methods rely on personal experience, and determining the relationship between the impact of different thresholds on the image and the final classification accuracy is difficult. The selected threshold may not be appropriate for the current data or task. In such cases, selecting the threshold is a challenge. Summary of the Invention

[0011] To overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a breast cancer pathology image classification system based on adaptive threshold wavelet denoising. The loss function of the breast cancer pathology image classification model training is associated with the selection of the wavelet threshold, so that the wavelet threshold is optimized during model training and gradually approaches the threshold with good denoising effect, thereby finding the wavelet threshold suitable for the image task and improving the classification accuracy of subsequent breast cancer pathology images.

[0012] To achieve the above objectives, the first aspect of the present invention provides a breast cancer pathology image classification system based on adaptive threshold wavelet denoising, comprising:

[0013] A wavelet transform module is configured to: perform wavelet transform processing on the acquired original breast cancer pathology image to obtain wavelet coefficients of each layer;

[0014] A denoising module is configured to: denoise the wavelet coefficients of each layer based on a wavelet threshold; and adaptively select the wavelet threshold based on a loss function in classification model training;

[0015] A reconstruction module is configured to: reconstruct a breast cancer pathology image using the denoised wavelet coefficients of each layer;

[0016] The classification module is configured to: input the reconstructed breast cancer pathology image into the trained classification model to obtain a classification result of the breast cancer pathology image.

[0017] A second aspect of the present invention provides a computer device comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the following steps are performed:

[0018] Perform wavelet transform on the acquired original breast cancer pathology image to obtain the wavelet coefficients of each layer;

[0019] Denoising the wavelet coefficients of each layer based on the wavelet threshold; adaptively selecting the wavelet threshold based on the loss function in the classification model training;

[0020] The denoised wavelet coefficients of each layer are used to reconstruct breast cancer pathology images;

[0021] The reconstructed breast cancer pathology image is input into the trained classification model to obtain a classification result of the breast cancer pathology image.

[0022] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are performed:

[0023] Perform wavelet transform on the acquired original breast cancer pathology image to obtain the wavelet coefficients of each layer;

[0024] Denoising the wavelet coefficients of each layer based on the wavelet threshold; adaptively selecting the wavelet threshold based on the loss function in the classification model training;

[0025] The denoised wavelet coefficients of each layer are used to reconstruct breast cancer pathology images;

[0026] The reconstructed breast cancer pathology image is input into the trained classification model to obtain a classification result of the breast cancer pathology image.

[0027] One or more of the above technical solutions have the following beneficial effects:

[0028] In the present invention, the loss function of the classification model training of breast cancer pathology images is associated with the selection of the wavelet threshold, so that the wavelet threshold is optimized during model training and gradually approaches the threshold with good denoising effect, thereby finding the wavelet threshold suitable for the image task and improving the classification accuracy of subsequent breast cancer pathology images.

[0029] In the present invention, an improved ResNeXt-50 network is used as the classification model, the proportion of different steps in the original ResNeXt-50 network structure is changed, and the activation function is reduced. The improved ResNeXt-50 network improves the accuracy of breast cancer histopathology image classification results.

[0030] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0032] Figure 1 This is a schematic diagram of a pathological image of breast cancer tissue in Example 1 of the present invention;

[0033] Figure 2 Schematic diagram of the steps of wavelet transform adaptive threshold denoising and the classification network structure in Example 1 of the present invention;

[0034] Figure 3 Schematic diagram of the threshold function in the first embodiment of the present invention;

[0035] Figure 4 Schematic diagram comparing the structures of a traditional residual block and an improved residual block in the first embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0037] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0038] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0039] Example 1

[0040] This embodiment discloses a breast cancer pathology image classification system based on adaptive threshold wavelet denoising, including:

[0041] A wavelet transform module is configured to: perform wavelet transform processing on the acquired original breast cancer pathology image to obtain wavelet coefficients of each layer;

[0042] A denoising module is configured to: denoise the wavelet coefficients of each layer based on a wavelet threshold; and adaptively select the wavelet threshold based on a loss function in classification model training;

[0043] A reconstruction module is configured to: reconstruct a breast cancer pathology image using the denoised wavelet coefficients of each layer;

[0044] The classification module is configured to: input the reconstructed breast cancer pathology image into the trained classification model to obtain a classification result of the breast cancer pathology image.

[0045] like Figure 1-Figure 2 As shown, the system proposed in this embodiment mainly includes three parts. The first is discrete wavelet transform (DWT), which can decompose the original noisy image into coefficient groups at different frequency levels: CA (Approximation Coefficient), CV (Vertical Detail Coefficient), CH (Horizontal Detail Coefficient), CD (Diagonal Detail Coefficient). Among them, CA is the approximation coefficient, and CH, CV, and CD are detail coefficients. The second step is to determine the wavelet threshold and then denoise CH, CV, and CD. Since the wavelet threshold affects the image input to the classification model, which in turn affects the model's output and loss function, the selection of the wavelet threshold is crucial.

[0046] By linking the wavelet threshold with the loss function, the wavelet threshold can be optimized during training and gradually approach a threshold with good denoising results. The third step is to perform an inverse discrete wavelet transform (IDWT) on the CA and the denoised CV, CH, and CD images to reconstruct a denoised breast cancer histopathology image. This denoised breast cancer histopathology image serves as input for classification.

[0047] For discrete wavelet transform and inverse discrete wavelet transform: For a one-dimensional input signal x, DWT decomposes it into a set of approximate coefficients a j and detail coefficient d j , as shown in formula (1) and formula (2):

[0048] a j =∑ n h n ·x j,n (1)

[0049] d j =∑ n g n ·x j,n (2)

[0050] Among them, h and g are the low-pass and high-pass filters of the orthogonal wavelet, respectively, and x j,n is the discrete sample in the input signal, j is the number of decomposition levels, and n is the displacement.

[0051] Formula (1) and formula (2) can be used to recursively decompose the approximate coefficients of the previous level, where the original signal x is the level j and the level j is 0, as shown in formula (4) and formula (5):

[0052] a0=x (3)

[0053] a j =∑ n h n ·a j-1,n (4)

[0054] d j =∑ n g n ·a j-1,n (5)

[0055] IDWT can reconstruct the approximate coefficients and detail coefficients into the original signal, as shown in formula (6):

[0056]

[0057] Where J is the maximum number of series and n is the displacement.

[0058] The processing of 2D images is similar to that of one-dimensional signals. It only requires performing DWT or IDWT in the horizontal and vertical directions respectively, as shown in the following formula:

[0059] CA j =∑ h ∑ v h h ·h v ·a j-1,h,v (7)

[0060] CV j =∑ h ∑ v h h ·g v ·a j-1,h,v (8)

[0061] CH j =∑ h ∑ v g h ·h v ·a j-1,h,v (9)

[0062] CD j =∑ h ∑ v g h ·g v ·a j-1,h,v (10)

[0063] The subscripts h and v refer to the horizontal and vertical displacements. h 、h v The horizontal displacement of the low-pass filter of the orthogonal wavelet and the vertical displacement of the low-pass filter of the orthogonal wavelet are respectively; gv 、g h The horizontal displacement of the high-pass filter of the orthogonal wavelet and the vertical displacement of the high-pass filter of the orthogonal wavelet are respectively; since the common practice for wavelet transform threshold denoising is to perform DWT and IDWT only once, this embodiment experiments on the CA, CV, CH, and CD coefficients when the level j = 2.

[0064] In the denoising module of this embodiment, an appropriate wavelet threshold is optimized. The wavelet threshold is optimized as a parameter through backpropagation. Specifically, a four-dimensional tensor with a single element of 1 is initialized and input into a convolutional layer with input channel = 1, output channel = 9, kernel size = 1, stride = 1, and padding = 0. The output channel = 9 is because color breast cancer pathology images have three channels: R, G, and B. Each channel's image produces three detail coefficients: CH, CV, and CD after wavelet transform. The nine elements output by the convolutional layer are used to denoise these nine coefficients. Using this method, changes in the wavelet threshold will alter image quality and affect the final loss value.

[0065]

[0066] The method for optimizing parameters is shown in formula (11), where L is the loss function, ∈ t is the learning rate, τ t It is the parameter at the tth iteration, namely the wavelet threshold. The learning rate is the most important optimization parameter. When the learning rate is too high, the model may skip the optimal parameters during training, and when it is too low, it is easy to fall into the local optimum. During training, the parameters that determine the wavelet threshold and the parameters of the classification model are optimized together. However, the range of the wavelet threshold (usually dozens) is much larger than the range of variation of the parameters in the classification model. Therefore, this embodiment uses a larger learning rate for the convolution layer that determines the threshold. The wavelet threshold is used as a parameter in the CNN model to train together with the model. This can link the threshold with the classification result of the model, and find the threshold suitable for the image and task through back propagation during training.

[0067] like Figure 3 As shown in Figure 1, after obtaining the wavelet threshold, a threshold function is still required to generate a filter for denoising. Experiments have found that using the sgn function to make the value suddenly change from 0 to 1, or similar threshold functions that change from 0 to 1 in a very small range, has a very poor effect. This embodiment uses sigmoid to construct the threshold function, as shown in formula (12):

[0068] x denoised =x original*(sigmoid((x original -abs(threshold))*10)) (12)

[0070] Among them, x original It is the original detail coefficient, such as CH, CV, CD, x denoised is the detail coefficient after denoising, and threshold is the learned wavelet threshold. original -abs(threshold) is used to compare the detail value with the threshold value to determine whether the value of the point should be filtered out. Since the optimized parameter may be positive or negative, abs is needed to keep the threshold positive, which is convenient for comparison with x. original Subtraction.

[0071] After x original -abs(threshold), x original Pixels with values ​​greater than the threshold will have values ​​greater than 0, while pixels with values ​​less than the threshold will have values ​​less than 0. The result is then multiplied by 10 and used as the input to $sigmoid$. ×10 causes the function value to change from 0.006 to 0.993 within the interval [-0.5, 0.5] centered on the threshold, making the filtering slightly steeper. A value too large, such as several hundred, produces an overly steep waveform, which severely impairs the threshold's optimization effect. A value too small results in a flat filter waveform, impairing the filtering effect. Experiments have shown that 10 is a relatively suitable value.

[0072] In the classification module of this embodiment, the classification model adopts an improved ResNeXt network with two main improvements: changing the ratio of different steps and using fewer activation functions.

[0073] Table 1 shows the details of the improved ResNeXt-50 architecture used in this example. The output column indicates the output sizes H and W, and the convolution column indicates the convolution shape. The first parameter in the convolution refers to the kernel size, and the second refers to the number of channel dimensions. "C = 32" indicates that there are 32 convolution groups. The number of residual blocks is indicated outside the parentheses. The typical ratio is [3, 4, 6, 3]. Based on ConvNeXt research, this example uses [3, 3, 9, 3].

[0074] Table 1

[0075]

[0076]

[0077] Specifically, the classification model adopts an improved ResNeXt-50 network. The reconstructed breast cancer pathology image is input into the improved ResNeXt-50 network. First, a convolution operation is performed on the convolution layer with a convolution kernel of 7×7, a stride of 2, and a channel dimension of 64, and a 128×128 feature map is output. Then, three Res1 block operations with an input of 128 dimensions and an output of 256 dimensions are performed in sequence to output a 64×64 feature map. Three Res2 block operations with an input of 256 dimensions and an output of 512 dimensions are performed to output a 32×32 feature map. Nine Res3 block operations with an input of 512 dimensions and an output of 1024 dimensions are performed to output a 16×16 feature map. Three Res4 block operations with an input of 1024 dimensions and an output of 2048 dimensions are performed to output an 8×8 feature map. Finally, a fully connected layer is used to output a 1×1 feature map.

[0078] The designs of the traditional residual block and the improved residual block used in this embodiment are as follows: Figure 4 As shown in FIG, compared with the original ResNeXt, this embodiment cancels the ReLU layer below the 1×1conv layer in the residual block.

[0079] This embodiment verifies the technology of this embodiment on the BreaKHis and BACH breast cancer histopathology image datasets. The BreaKHis dataset is the earliest public large-scale non-full-field breast cancer histopathology image dataset, and the BACH database is a representative multi-classification dataset.

[0080] BreakHis is an eight-class classification dataset containing four benign tumor categories: adenosis, fibroadenoma, dermatoma, and tubular adenoma, and four malignant tumors: ductal carcinoma, lobular carcinoma, mucinous carcinoma, and papillary carcinoma. The dataset includes 7,909 images collected from 82 patients using four magnification factors (40×, 100×, 200×, and 400×). The specific image distribution is shown in Table 2. The images are in RGB three-channel format with a resolution of 700×460.

[0081] Table 2 Image distribution of BreakHis

[0082] category 40× 100× 200× 400× Adenopathy 114 113 111 106 Fibroadenoma 253 260 264 237 skin graft tumor 109 121 108 115 Tubular adenoma 149 150 140 130 Ductal carcinoma 864 903 896 788 Lobular carcinoma 156 170 163 137 mucinous carcinoma 205 222 196 169 Papillary carcinoma 145 142 135 138 total 1995 2081 2013 1820

[0083] BACH is a four-class classification dataset consisting of 400 breast cancer histopathology images, classified as follows: 100 normal tissues, 100 benign abnormalities, 100 carcinoma in situ, and 100 invasive carcinomas. These images are RGB images, each 2048x1536 pixels, with a pixel ratio of 0.42um x 0.42um.

[0084] Preprocessing: To reduce memory and computational costs, the original images in the BreaKHis dataset were downscaled from 700×460 pixels to 256×256 pixels, while the original images in the BACH dataset were downscaled from 2048×1536 pixels to 512×512 pixels. To reduce overfitting due to the small number of images, data augmentation methods such as vertical and horizontal mirroring, random rotation, and random cropping were used. The data was also normalized.

[0085] To evaluate the proposed method, this example used 4-fold cross-validation. The dataset was divided into four parts, each containing 25% of the total training samples. During training, three of the parts were used as training sets, and the remaining part was used for validation. Because the proposed method searches for appropriate thresholds during the training process, the thresholds obtained in each experiment vary slightly. To facilitate comparison with existing threshold calculation methods, a training and validation run was first performed on the training set to obtain a set of thresholds. This set of thresholds was then used for 4-fold cross-validation.

[0086] Training: During training, the Adam optimizer was used, the batch size was 60, and the learning rate was controlled using cosine annealing. The epoch number was set to 700 for the BreakHis dataset, and 2000 for the BACH dataset due to the small number of data points. This was to avoid incomplete optimization due to too few iterations. As mentioned previously, parameters controlling the wavelet threshold require a higher learning rate. For parameters other than the convolutional layers used to control the wavelet threshold in the model, the learning rate used for optimization was: a maximum of 5e-5 and a minimum of 5e-7. For the convolutional layers used to control the threshold, the learning rate used for optimization was: a maximum of 2e-1 and a minimum of 2e-3. Experiments were conducted in a PyTorch environment using an NVIDIA Titan X GPU.

[0087] Evaluation metrics: Accuracy (ACC) is used to evaluate the performance of the classification model. As shown in formula (12), accuracy refers to the total number of samples divided by the number of correctly predicted samples.

[0088]

[0089] Where T p and T n are respectively the positive samples and negative samples that are accurately predicted. p and F n are the positive and negative samples that are predicted incorrectly.

[0090] Comparative experiments used two representative traditional threshold selection methods: a fixed threshold method and a minimum-maximum threshold method, as shown in Equations 13 and 14, respectively. The fixed threshold method provides more thorough denoising, but it can also easily misidentify useful signals as noise. The minimum-maximum threshold method is more conservative and works better when noise is less concentrated in the high-frequency range of the signal.

[0091]

[0092]

[0093] Where N is the signal length, which is the number of pixels in a two-dimensional image. The threshold used for wavelet transform threshold denoising is obtained by multiplying λ by the image's noise variance estimate σ. The noise variance estimate is shown in Formula 15:

[0094] σ=median(image) / 0.6745 (15)

[0095] In the experiment, threshold denoising was performed on each detail coefficient of each RGB channel. Since there are two datasets, each with four magnifications and three channels per image, the number of thresholds is too large to be fully displayed. This article uses a 40× image from the BreaKHis dataset as an example. Table 3 shows the thresholds calculated by the two traditional methods and the thresholds learned by the proposed model through training.

[0096] Table 3 Thresholds obtained using different methods

[0097]

[0098] To verify the effectiveness of this scheme, experiments were conducted on the BreaKHis and BACH datasets. The results of denoising using the adaptive threshold method, the fixed threshold method, the maximum and minimum threshold method, and no threshold denoising were compared.

[0099] Table 4 Experimental results on the BreaKHis dataset

[0100]

[0101] As shown in Table 4, the results of the adaptive threshold method + improved ResNeXt method of this embodiment are the best on images of different magnifications. And using the traditional threshold selection method, whether it is a relatively large fixed threshold method or a relatively small maximum and minimum threshold method, the results obtained are not only worse than the method of this embodiment, but also worse than the results of not performing wavelet transform threshold denoising. This shows that the thresholds selected by these two methods of selecting thresholds based on image size and median may be inappropriate and lead to poor classification results. The adaptive threshold technology of this embodiment can find a more suitable wavelet threshold by integrating the threshold selection step into deep learning, testing a large number of thresholds during training, and continuously adjusting the wavelet threshold through loss function and back propagation.

[0102] This method not only selects an appropriate wavelet threshold, but also does so efficiently. Traditional methods often require trying multiple thresholds and then selecting the appropriate one based on the results. Because deep learning models are sensitive to changes in input, testing the effects of different wavelet thresholds on classification results on an already trained model, using different wavelet thresholds for filtering will inherently reduce classification accuracy. Therefore, wavelet threshold filtering must be performed on the input image during training to effectively compare different wavelet thresholds. However, it is difficult for humans to determine which wavelet threshold is best and adjust it during training, and comparisons are often made only after training is complete. This consumes a significant amount of resources and time, resulting in high costs. This method allows for continuous experimentation and adjustment of thresholds during training, allowing the appropriate threshold to be found with only a very small number of training sessions.

[0103] Table 4 also compares the results of the original ResNeXt and the improved ResNeXt. We can see that the two improvements from the ConNeXt research slightly improve the classification results. The improved ResNeXt is also used as the baseline on BACH.

[0104] Table 5 Experimental results on the BACH dataset

[0105]

[0106] Table 5 shows the results of different methods on the BACH dataset. Similar to the results on the BreakHis dataset, the adaptive thresholding method achieves the best results. The fixed thresholding method and the minimum-maximum thresholding method still perform worse than no denoising. The maximum-maximum thresholding method produces significantly worse results, and its variance is significantly greater than that of the adaptive thresholding method. This may be because the BACH dataset has far less data than the BreakHis dataset, resulting in more unstable results.

[0107] Example 2

[0108] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the following steps are implemented:

[0109] Perform wavelet transform on the acquired original breast cancer pathology image to obtain the wavelet coefficients of each layer;

[0110] Denoising the wavelet coefficients of each layer based on the wavelet threshold; adaptively selecting the wavelet threshold based on the loss function in the classification model training;

[0111] The denoised wavelet coefficients of each layer are used to reconstruct breast cancer pathology images;

[0112] The reconstructed breast cancer pathology image is input into the trained classification model to obtain a classification result of the breast cancer pathology image.

[0113] Example 3

[0114] The purpose of this embodiment is to provide a computer-readable storage medium.

[0115] A computer-readable storage medium stores a computer program, which, when executed by a processor, performs the following steps:

[0116] Perform wavelet transform on the acquired original breast cancer pathology image to obtain the wavelet coefficients of each layer;

[0117] Denoising the wavelet coefficients of each layer based on the wavelet threshold; adaptively selecting the wavelet threshold based on the loss function in the classification model training;

[0118] The denoised wavelet coefficients of each layer are used to reconstruct breast cancer pathology images;

[0119] The reconstructed breast cancer pathology image is input into the trained classification model to obtain a classification result of the breast cancer pathology image.

[0120] The steps involved in the above-mentioned embodiments 2 and 3 correspond to those in the method embodiment 1. For detailed implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media that includes one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and cause the processor to perform any method of the present invention.

[0121] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0122] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A breast cancer pathology image classification system based on adaptive threshold wavelet denoising, characterized by: include: A wavelet transform module is configured to: perform wavelet transform processing on the acquired original breast cancer pathology image to obtain wavelet coefficients of each layer; A denoising module is configured to: construct a threshold function using a sigmoid function and a wavelet threshold, and perform denoising on the wavelet coefficients of each layer using the threshold function to obtain denoised wavelet coefficients of each layer; The appropriate wavelet threshold is obtained through optimization. This threshold is then used as a parameter of the classification model through backpropagation. Specifically, a four-dimensional tensor with a single element of 1 is initialized and fed into a convolutional layer with 1 input channel, 9 output channels, kernel size 1, stride 1, and padding 0. The output channel 9 is because color breast cancer pathology images have three channels: R, G, and B. After wavelet transformation, each channel's image produces three detail coefficients: CH, CV, and CD. The nine output elements of the convolutional layer are used to denoise these nine coefficients. The wavelet threshold is adaptively selected based on the loss function in the classification model training. Specifically, the parameters that determine the wavelet threshold are trained together with the parameters of the classification model in the classification model training. The formula for optimizing the parameters is: ;in, is the loss function, is the learning rate, is the wavelet threshold at the tth iteration; The threshold function is: in, is the original detail coefficient, namely CH, CV, CD, is the detail coefficient after denoising, is the learned wavelet threshold, It is to compare the detail value with the wavelet threshold to decide whether the value of the point should be filtered out; A reconstruction module is configured to: reconstruct a breast cancer pathology image using the denoised wavelet coefficients of each layer; The classification module is configured to: input the reconstructed breast cancer pathology image into the trained classification model to obtain a classification result of the breast cancer pathology image.

2. The breast cancer pathology image classification system based on adaptive threshold wavelet denoising according to claim 1, characterized in that: In the wavelet transform module, the acquired original breast cancer pathology image is decomposed into coefficient groups of different frequency levels, including approximate coefficients and detail coefficients, and the detail coefficients are denoised.

3. The breast cancer pathology image classification system based on adaptive threshold wavelet denoising according to claim 1, characterized in that: In the classification module, the classification model adopts an improved ResNeXt-50 network, specifically: The reconstructed breast cancer pathology image is input into the improved ResNeXt-50 network. It is first convolved with a 7×7 convolution kernel, a stride of 2, and 64 channels, and outputs a 128×128 feature map. Then, the Res1 block operation with an input of 128 dimensions and an output of 256 dimensions is performed three times in sequence; the Res2 block operation with an input of 256 dimensions and an output of 512 dimensions is performed three times; the Res3 block operation with an input of 512 dimensions and an output of 1024 dimensions is performed nine times; the Res4 block operation with an input of 1024 dimensions and an output of 2048 dimensions is performed three times, and finally the output is output through the fully connected layer.

4. The breast cancer pathology image classification system based on adaptive threshold wavelet denoising according to claim 3, characterized in that: In the classification module, in the improved ResNeXt-50 network, the Relu layer below the 1×1 convolution layer in the original residual block is cancelled.

5. A computer device, characterized in that: include: A processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate via the bus. When the machine-readable instructions are executed by the processor, the following steps are performed: Perform wavelet transform on the acquired original breast cancer pathology image to obtain the wavelet coefficients of each layer; A threshold function is constructed using a sigmoid function and a wavelet threshold, and the wavelet coefficients of each layer are denoised using the threshold function to obtain denoised wavelet coefficients of each layer; The appropriate wavelet threshold is obtained through optimization. This threshold is then used as a parameter of the classification model through backpropagation. Specifically, a four-dimensional tensor with a single element of 1 is initialized and fed into a convolutional layer with 1 input channel, 9 output channels, kernel size 1, stride 1, and padding 0. The output channel 9 is because color breast cancer pathology images have three channels: R, G, and B. After wavelet transformation, each channel's image produces three detail coefficients: CH, CV, and CD. The nine output elements of the convolutional layer are used to denoise these nine coefficients. The wavelet threshold is adaptively selected based on the loss function in the classification model training. Specifically, the parameters that determine the wavelet threshold are trained together with the parameters of the classification model in the classification model training. The formula for optimizing the parameters is: ;in, is the loss function, is the learning rate, is the wavelet threshold at the tth iteration; The threshold function is: in, is the original detail coefficient, namely CH, CV, CD, is the detail coefficient after denoising, is the learned wavelet threshold, It is to compare the detail value with the wavelet threshold to decide whether the value of the point should be filtered out; The denoised wavelet coefficients of each layer are used to reconstruct breast cancer pathology images; The reconstructed breast cancer pathology image is input into the trained classification model to obtain a classification result of the breast cancer pathology image.

6. A computer-readable storage medium, characterized in that The computer readable storage medium stores a computer program, which, when executed by a processor, performs the following steps: Perform wavelet transform on the acquired original breast cancer pathology image to obtain the wavelet coefficients of each layer; A threshold function is constructed using a sigmoid function and a wavelet threshold, and the wavelet coefficients of each layer are denoised using the threshold function to obtain denoised wavelet coefficients of each layer; The appropriate wavelet threshold is obtained through optimization. This threshold is then used as a parameter of the classification model through backpropagation. Specifically, a four-dimensional tensor with a single element of 1 is initialized and fed into a convolutional layer with 1 input channel, 9 output channels, kernel size 1, stride 1, and padding 0. The output channel 9 is because color breast cancer pathology images have three channels: R, G, and B. After wavelet transformation, each channel's image produces three detail coefficients: CH, CV, and CD. The nine output elements of the convolutional layer are used to denoise these nine coefficients. The wavelet threshold is adaptively selected based on the loss function in the classification model training. Specifically, the parameters that determine the wavelet threshold are trained together with the parameters of the classification model in the classification model training. The formula for optimizing the parameters is: ;in, is the loss function, is the learning rate, is the wavelet threshold at the tth iteration; the threshold function is: in, is the original detail coefficient, namely CH, CV, CD, is the detail coefficient after denoising, is the learned wavelet threshold, It is to compare the detail value with the wavelet threshold to decide whether the value of the point should be filtered out; The denoised wavelet coefficients of each layer are used to reconstruct breast cancer pathology images; The reconstructed breast cancer pathology image is input into the trained classification model to obtain a classification result of the breast cancer pathology image.

Images