A clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution

Abstract

The application discloses a clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution, comprising: carrying out pretreatment and normalization treatment on an original clinical image to obtain a time domain image; converting the time domain image into a frequency domain image through a position consistency module to generate a frequency spectrum capable of preserving local and position information; carrying out masking and supervision on the generated frequency spectrum through a high-frequency self-encoder module to obtain a self-encoding frequency spectrum; inputting the self-encoding frequency spectrum into an anti-aliasing complex convolution model for training to learn specific features of different clinical lesions in the image; and integrating and classifying the specific features learned by the anti-aliasing complex convolution model to obtain a clinical image lesion classification result. The method can better utilize frequency domain information of the image on the basis of preserving time domain features of the clinical image, solves the problems of brightness imbalance and inter-class confusion in lesion classification, and greatly improves the accuracy of automatic classification.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of clinical image recognition classification, and relates to computer vision technology, in particular to a clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution. BACKGROUND

[0002] Tumor is a big killer that harms life and health, and has a high mortality rate. Timely diagnosis of tumor is crucial for effective treatment to reduce mortality and improve prognosis. Unfortunately, the early clinical symptoms of tumor are not typical, which increases the difficulty of identifying tumors in early treatment. Despite this, the type of early lesions is closely related to the stage of cancer development. Previous clinical studies have observed an early lesion-cancer sequence, i.e., pre-existing early lesions can be transformed into tumors. Therefore, screening of early lesions plays a crucial role in the prevention of tumors, which can help doctors develop appropriate prevention and treatment programs.

[0003] Pathological biopsy is the most commonly used clinical program for early screening of clinical lesions, and it plays a key role in reducing the incidence and mortality of tumors. If early lesions can be detected in time and removed under an endoscope, the incidence of tumors can be significantly reduced. However, pathological biopsy requires surgery and is highly dependent on the experience level of physicians. If computer-aided means can be used to diagnose lesions through images, it can greatly reduce the workload of clinics and reduce the pain of patients undergoing biopsy surgery. Therefore, it is of great significance to assist clinical automatic and accurate classification through computer-aided diagnosis for early clinical lesions.

[0004] Deep learning has attracted widespread attention in the field of medical image analysis in recent years, as it has higher performance in image classification detection compared to previous state-of-the-art techniques. Moreover, the advantage of deep learning models is that they do not require pre-processing of image datasets and can use raw images for training, automatically extracting and learning relevant features. However, developing a satisfactory deep learning model for clinical lesion classification still presents challenges, as clinical diseases often have significant similarities and close distribution. Therefore, how to guide the network to extract lesion-specific features and improve the discriminability of the model is a problem that needs to be solved. However, most current deep learning architectures usually only consider general features, and cannot fully exploit the specificity information of clinical images.

[0005] The patent document with patent number 202010834609.4 discloses a method for detecting and classifying lesions in clinical images. Although this method can automatically classify lesions, it is a two-stage method. This requires first extracting local images from global images, which is very time-consuming and has a loss of precision. In addition, this method lacks exploration of image-specific features and has the problem of inter-class confusion. SUMMARY

[0006] The present application aims to overcome the deficiencies in the prior art, and provides a clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution, which can fully exploit the frequency domain information in the image, decouple the image content and irrelevant structure using the advantages of the image frequency domain, and extract lesion-specific features, explore an intelligent analysis method suitable for clinical image data classification, and successfully solve the problems of uneven brightness and inter-class confusion of clinical images using the brightness insensitivity and specific features of the frequency domain. This effective auxiliary analysis method can realize real-time classification of clinical lesions, assist clinical screening of early lesions, and improve the survival rate of patients.

[0007] The technical scheme is as follows: in order to achieve the above-mentioned purpose, the present application provides a clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution, comprising the following steps:

[0008] S1: pre-processing and normalization processing of the original clinical image to obtain a time domain image;

[0009] S2: converting the time domain image obtained in step S1 into a frequency domain image through the constructed position consistency module to generate a spectrum capable of preserving local and position information;

[0010] S3: masking and supervision of the spectrum generated in step S2 through the constructed high-frequency autoencoder module to guide subsequent network learning of useful frequency features, and obtaining an autoencoding spectrum;

[0011] S4: inputting the autoencoding spectrum obtained in step S3 into the constructed anti-aliasing complex convolution model for training to learn the specific features of different clinical lesions in the image;

[0012] S5: integrating and classifying the specific features learned by the anti-aliasing complex convolution model in step S4 to obtain a clinical image lesion classification result.

[0013] Further, the pre-processing and normalization processing process in step S1 is as follows:

[0014] A1: using a gray threshold segmentation method to eliminate white edges and transition exposure parts in the image;

[0015] A2: obtaining the rectangular frame position, length and width of the lesion area in the image according to the gray projection histogram in the vertical and horizontal directions of the image, and cutting the required area according to the obtained rectangular frame length and width to remove the redundant image parts outside the required area;

[0016] A3: using a low-pass filter to remove noise in the image signal;

[0017] A4: Standardize the image to the same size by interpolation, reduce the complexity of calculation;

[0018] A5: Normalize the standardized image to 0-1, so that the subsequent classification network can find the optimal solution more quickly.

[0019] Further, the step S2 is specifically:

[0020] B1: Cut the image into blocks, and cut the whole image into blocks according to KxK size;

[0021] B2: Fourier transform is performed on each image block obtained in step B1 to obtain a frequency spectrum block, and the specific formula is as follows:

[0022]

[0023] Wherein, M, N represent the width and height of the image block respectively, F m (x, y) represents the mth image block, F m (u, v) represents the obtained frequency spectrum block;

[0024] B3: Randomly shuffle all frequency spectrum blocks in a single image obtained in step B2;

[0025] B4: Space splicing is performed on the shuffled frequency spectrum blocks to obtain a frequency spectrum graph F of a single image, and the formula is as follows:

[0026] F j = Concat(F ij ), 1≤i≤K,

[0027] F = Concat(F j ), 1≤j≤K.

[0028] The position consistency module in step S2 firstly obtains a complex frequency spectrum graph by slicing and spatial stacking of the discrete Fourier transform time image, effectively aggregates local information and improves the learning ability of non-local information.

[0029] Further, the step S3 is specifically:

[0030] C1: Preserve the low frequency area of each frequency spectrum block of each image; for each frequency spectrum block, the low frequency area in the middle accounts for about 10% of the frequency spectrum block;

[0031] C2: For the area other than the low frequency in step C1, it is regarded as a high frequency area, and a certain proportion is randomly masked to obtain the final masked frequency spectrum graph. The purpose is to develop a self-supervised task by automatically masking a certain proportion of high frequency areas and keeping low frequency areas.

[0032] The high-frequency auto-encoding module proposed in step S3 is an effective way to keep the low-frequency signal intact and observe part of the high-frequency signal. For visual tasks, the low-frequency channel usually provides more information than the high-frequency channel. When the training data contains noise, the high-frequency component is more likely to be contaminated, thereby reducing the performance of the model. Therefore, the high-frequency auto-encoding module counteracts the interference of redundant signals in the frequency spectrum on the semantics of the image. It allows the model to develop a self-supervised task by automatically shielding a certain proportion of the high-frequency region and keeping the low-frequency region. This encourages the model to construct key information of the image from the latent representation, resulting in a more efficient optimization process.

[0033] Further, the filter is embedded in the anti-aliasing complex convolution model in step S4, and the specific feature acquisition process is:

[0034] D1: split the complex spectrum obtained in step S3 into real and imaginary parts;

[0035] D2: input the real and imaginary parts into the complex convolution model for training;

[0036] D3: low-pass filter the features obtained from each complex convolution block to obtain the anti-aliasing feature map;

[0037] D4: obtain the final feature through five anti-aliasing complex convolution blocks.

[0038] In combination with the above scheme, the reason why step S4 can obtain specific features is:

[0039] The image has unique characteristics different from the time domain after being converted from the time domain to the frequency domain, so specific features can be extracted. When the image is translated, the image spectrum remains unchanged, indicating the insensitivity of the frequency domain image to the location of the lesion. In addition, the frequency domain image is insensitive to brightness. Since most of the energy in the frequency domain is concentrated on the direct current component, which represents the average brightness of the image, the decoupling of image brightness and content can be achieved. Finally, the phase image in the frequency domain contains features that are helpful for sub-classification, such as texture, contour features, etc. If the phase image of the image is exchanged, the style of the two images remains unchanged, but their content is exchanged, because the phase image contains information that preserves the edges and overall structure of the image.

[0040] Step S4 inputs the auto-encoding spectrum obtained in step S3 into the proposed anti-aliasing complex convolution model for training to learn the specific features of different clinical lesions in the image. In addition, a filter is embedded in the complex convolution model to prevent spectral aliasing of features during down-sampling.

[0041] Further, the complex convolution model in step D2 is modified from a ResNet18 network, which is changed to a complex form.

[0042] Further, the construction of the complex convolution model replaces ResNet18 with a complex form, which specifically includes: replacing the convolution kernel and block kernel architecture with a complex form, and for each convolution block, the convolution and ReLU function form are specifically as follows:

[0043] The input is the real part a and the imaginary part b of the complex matrix P=a+bi in the spectrum graph, and the complex convolution kernel is represented as Q=c+di, and the complex convolution calculation is as follows:

[0044]

[0045] The complex ReLU satisfies the Cauchy-Riemann equation when the real part and the imaginary part are strictly positive or strictly negative, and the calculation formula is as follows:

[0046]

[0047] Further, in step D3, a Gaussian filter with a size of 3*3 is used to filter the features obtained after complex convolution, to reduce the influence of spectral aliasing on model generalization.

[0048] Further, the step S5 specifically includes:

[0049] E1: integrating the real part and the imaginary part of the network output to obtain the final classification feature, and the specific formula is as follows:

[0050]

[0051] Where F final , F real and F imag represent the final classification feature, the real part feature and the imaginary part feature respectively;

[0052] E2: obtaining the classification result according to the final feature obtained in step E1.

[0053] Further, in step E2, the classification result is obtained by a loss function Softmax, and the calculation formula of the loss function Softmax is as follows:

[0054]

[0055] Where y k is the input prediction probability value, θ k is the network training weight hyperparameter, and F final is the input.

[0056] The clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution provides a new idea for lesion classification.

[0057] Advantages: Compared with the prior art, the present application has the following advantages:

[0058] 1. The method can realize automatic classification of clinical images based on deep learning, thereby avoiding manual heuristic learning and improving the generalization ability of the model.

[0059] 2. The method fully utilizes the frequency domain information of the clinical image, decouples the image content and irrelevant structure by utilizing the advantages of the image frequency domain, extracts lesion-specific features, explores an intelligent analysis method suitable for clinical image data classification, and successfully solves the problems of uneven brightness of clinical images and inter-class confusion by utilizing the brightness insensitivity and specific features of the frequency domain, thereby improving the classification accuracy of early clinical lesions. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flowchart of the method of the present application;

[0061] Figure 2 is a structural block diagram of the application in the automatic classification of lesions in clinical endoscopic intestinal images;

[0062] Figure 3 is a result graph comparing the original image with the image after preprocessing through steps A1-A3, taking endoscopic intestinal images as an example;

[0063] Figure 4 is a result graph comparing the frequency spectrum obtained after step S2 with the time domain image obtained in step 1, taking endoscopic intestinal images as an example;

[0064] Figure 5 is a result graph comparing the frequency spectrum obtained after step S3 with the frequency spectrum obtained in step S2, taking endoscopic intestinal images as an example;

[0065] Figure 6 is an architectural diagram of the convolution block of steps D1-D4;

[0066] Figure 7 is an architectural diagram of the complex convolution and complex ReLU function in the complex network in step D2;

[0067] Figure 8 is a schematic diagram of five kinds of endoscopic intestinal images and corresponding clinical treatment methods. DETAILED DESCRIPTION

[0068] The present application is further illustrated by the following description and examples with reference to the accompanying drawings, in which the examples are presented by way of illustration and not by way of limitation by the scope of the present application, and as will become apparent to those skilled in the art upon reading, various modifications to the application in terms of the various embodiments and applications thereof can be made without departing from its scope, which is to be defined by the appended claims.

[0069] As shown in the drawings, Figure 1 The present application provides a clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution, which comprises the following steps:

[0070] S1: pre-processing and normalization processing of the original clinical image to obtain a time domain image;

[0071] S2: converting the time domain image obtained in step S1 into a frequency domain image through the constructed position consistency module to generate a spectrum capable of preserving local and position information;

[0072] S3: masking and supervising the spectrum generated in step S2 through the constructed high-frequency autoencoder module to guide subsequent network learning of useful frequency features to obtain an autoencoding spectrum;

[0073] S4: inputting the autoencoding spectrum obtained in step S3 into the constructed anti-aliasing complex convolution model for training to learn the specificity features of different clinical lesions in the image;

[0074] S5: integrating and classifying the specificity features learned by the anti-aliasing complex convolution model in step S4 to obtain a clinical image lesion classification result.

[0075] Based on the above, the above scheme is applied to automatic classification of lesions in clinical endoscopic intestinal images in this embodiment, and with reference to Figure 2 , the specific process is as follows:

[0076] Step S1: pre-processing and normalization processing:

[0077] A1: using gray threshold segmentation method to eliminate white edges and transition exposure parts in the image;

[0078] A2: obtaining the rectangular frame position, length and width of the lesion area in the image according to the gray projection histogram in the vertical and horizontal directions of the image, and cutting the required area according to the obtained rectangular frame length and width to remove the redundant image parts outside the required area;

[0079] A3: using a low-pass filter to remove noise in the image signal;

[0080] As shown in the drawings, Figure 3As shown, the embodiment is an endoscopic intestinal image as an example, a result image compared with the original image after preprocessing through steps A1-A3;

[0081] A4: Standardize the image by interpolation to the same size, reduce the complexity of calculation;

[0082] A5: Normalize the standardized image to 0-1, which is convenient for the subsequent classification network to find the optimal solution faster.

[0083] Step S2: The position consistency module first obtains a complex spectrum graph by slicing and spatial stacking of the discrete Fourier transform time image, effectively aggregating local information and improving the learning ability of non-local information, including the following steps:

[0084] B1: The image is cut into blocks, and the whole image is cut into blocks according to KxK size;

[0085] B2: Fourier transform is performed on each image block obtained in step B1 to obtain a spectrum block, and the specific formula is as follows:

[0086]

[0087] Where M, N represent the width and height of the image block, F m (x, y) represents the mth image block, F m (u, v) represents the obtained spectrum block;

[0088] B3: Randomly shuffle all spectrum blocks in a single image obtained in step B2;

[0089] B4: Spatially splice the shuffled spectrum blocks to obtain the spectrum graph F of a single image, and the formula is as follows:

[0090] F j = Concat(F ij ), 1≤i≤K,

[0091] F = Concat(F j ), 1≤j≤K.

[0092] As Figure 4 shown, the embodiment is an endoscopic intestinal image as an example, a result image compared with the original image after preprocessing through steps A1-A3;

[0093] Step S3:

[0094] C1: Keep the low-frequency area of each spectrum block of each image; for each spectrum block, the low-frequency area in the middle accounts for about ten percent of the spectrum block;

[0095] C2: For the region except low frequency in step C1, it is considered as high frequency region, and a certain proportion of it is randomly masked to get the final masked frequency spectrum. The purpose is to develop a self-supervised task by automatically masking a certain proportion of high frequency region and keeping low frequency region.

[0096] As shown in Figure 5 The embodiment is an endoscopic enteroscopy image as an example, and a result image obtained by comparing the frequency spectrum obtained after step S3 processing and the frequency spectrum obtained in step S2.

[0097] The high frequency auto-encoding module proposed in step S3 is an effective way to keep the low frequency signal complete and observe part of the high frequency signal. For visual tasks, the low frequency channel usually provides more information than the high frequency channel. When the training data contains noise, the high frequency component is more easily contaminated, thereby reducing the performance of the model. Therefore, the high frequency auto-encoding module counteracts the interference of redundant signals in the frequency spectrum on the semantics of the image. It allows the model to develop a self-supervised task by automatically masking a certain proportion of high frequency region and keeping low frequency region. This prompts the model to construct the key information of the image from the latent representation, resulting in a more efficient optimization process.

[0098] Step S4: Embedding filter in anti-aliasing complex convolution model, specific feature acquisition process is:

[0099] D1: Split the complex frequency spectrum obtained in step S3 into real and imaginary parts;

[0100] D2: Input the real and imaginary parts into the complex convolution model for training,

[0101] D3: Low-pass filter the features obtained for each complex convolution block to obtain an anti-aliasing feature map;

[0102] D4: Obtain the final feature through five anti-aliasing complex convolution blocks.

[0103] The reason why the above process can obtain specific features is that:

[0104] The image has unique characteristics different from the time domain after being converted from the time domain to the frequency domain, so specific features can be extracted. When the image is translated, the image frequency spectrum remains unchanged, indicating the insensitivity of the frequency domain graph to the location of the lesion. In addition, the frequency domain graph is insensitive to brightness. Since most of the energy in the frequency domain is concentrated on the direct current component, the direct current component represents the average brightness of the image, so the decoupling of image brightness and content can be achieved. Finally, the phase graph in the frequency domain contains features that are helpful for sub-classification, such as texture, contour features, etc. If the phase graph of the image is exchanged, the style of the two images remains unchanged, but their content is exchanged, because the phase graph contains information that preserves the edges and overall structure of the image.

[0105] Step S4 inputs the self-encoding spectrum obtained in step S3 into the proposed anti-aliasing complex convolution model for training to learn the specific features of different clinical lesions in the image. In addition, a filter is embedded in the complex convolution model to prevent spectral aliasing of features in the down-sampling process.

[0106] As shown in Figure 6 , it is the architecture diagram of the convolution block of steps D1-D4. The complex convolution model is modified from ResNet18 network. ResNet18 is changed to complex form. Refer to Figure 7 , ResNet18 is changed to complex form, which specifically includes changing the convolution kernel, block kernel architecture to complex form. For each layer of convolution block, the convolution and ReLU function form is specifically as follows:

[0107] The input is the real part a and the imaginary part b of the complex matrix P=a+bi in the spectrum diagram. The complex convolution kernel is represented as Q=c+di, and the complex convolution calculation is as follows:

[0108]

[0109] The complex ReLU satisfies the Cauchy-Riemann equation when the real part and the imaginary part are strictly positive or strictly negative. The calculation formula is as follows:

[0110]

[0111] In step D3, a Gaussian filter with a size of 3x3 is used to filter the features obtained after complex convolution to reduce the influence of spectral aliasing on model generalization.

[0112] Step S5 specifically includes:

[0113] E1: The real part and the imaginary part of the network output are integrated to obtain the final classification feature. The specific formula is as follows:

[0114]

[0115] Where F final , F real and F imag represent the final classification feature, the real part feature and the imaginary part feature, respectively.

[0116] E2: According to the final feature obtained in step E1, the classification result is obtained through the loss function Softmax. The calculation formula of the loss function Softmax is as follows:

[0117]

[0118] Where y k is the input prediction probability value, and θ kis the network training weight hyperparameter, F final is the input.

[0119] In this embodiment, the clinical images are obtained by a medical imaging system and are classified into different categories according to the patient's pathology, such as benign and malignant, etc. Among them, the colon images in the example are obtained by an endoscope system, and there are five categories of images, namely normal, polyp, adenoma, tumor and ulcerative colitis. Each disease has its specific clinical treatment plan.

[0120] As shown in Figure 8 , five kinds of colon endoscopic images and corresponding clinical treatment methods are obtained in this embodiment.

[0121] In summary, the method of the present application is suitable for automatic classification of clinical image lesions. A new deep learning method is provided, which proposes adaptive frequency domain learning and anti-aliasing complex convolutional network, fully utilizes the frequency domain information in clinical images, and realizes the rapid and accurate classification of early lesions. This automatic classification method is crucial for clinical analysis, which can help doctors identify the type of early disease and determine the most effective treatment method, thereby reducing the pain of patients.

Claims

1. A clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution, characterized in that, Includes the following steps: S1: Preprocess and normalize the original clinical images to obtain time-domain images; S2: The time-domain image obtained in step S1 is converted into a frequency-domain image by the constructed position consistency module, generating a spectrum that can retain local and positional information; S3: Mask and supervise the spectrum generated in step S2 using the constructed high-frequency autoencoder module to obtain the autoencoded spectrum; S4: Input the autoencoded spectrum obtained in step S3 into the constructed anti-aliasing complex convolution model for training, so as to learn the specific features of different clinical lesions in the image; S5: Integrate and classify the specific features learned by the anti-aliasing complex convolution model in step S4 to obtain the clinical image lesion classification results; Step S2 is as follows: B1: Cut the image into blocks, dividing the entire image into blocks of size K×K; B2: Perform a Fourier transform on each image block obtained in step B1 to obtain the spectral block. The specific formula is as follows: Where M and N represent the width and height of the image patch, respectively, and F m (x, y) represents the m-th image patch, F m (x, y) represents the obtained spectrum block; B3: Randomly shuffle all spectral blocks in the single image obtained in step B2; B4: Spatial stitching is performed on the shuffled spectral blocks to obtain the spectrum F of a single image, as shown in the following formula: F j =Concat(F ij ),1≤i≤K, F=Concat(F j ),1≤j≤K; Step S3 is as follows: C1: Preserve the low-frequency region of each spectral block in each image; for each spectral block, the middle low-frequency region occupies 10% of the spectral block; C2: For the regions other than low frequencies in step C1, they are considered as high-frequency regions and randomly masked according to the proportion to obtain the final masked spectrum. In step S4, the process of embedding filters in the anti-aliasing complex convolution model and obtaining specific features is as follows: D1: Decompose the complex spectrum obtained in step S3 into real and imaginary parts; D2: Input both the real and imaginary parts into the complex convolution model for training; D3: For the features obtained from each complex convolution block, perform low-pass filtering to obtain the anti-aliasing feature map; D4: The final feature is obtained after five anti-aliasing complex convolutional blocks.

2. The clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution according to claim 1, characterized in that, The preprocessing and normalization process in step S1 is as follows: A1: Use grayscale thresholding to eliminate white edges and overexposed areas in the image; A2: Based on the grayscale projection histograms of the image in the vertical and horizontal directions, obtain the position, length, and width of the rectangle containing the lesion area in the image. Then, based on the obtained rectangle length and width, crop the required area to remove the excess image part outside the required area. A3: Use a low-pass filter to remove noise from the image signal; A4: By interpolating, images are standardized to the same size, reducing computational complexity; A5: Normalize the standardized image to 0-1.

3. The clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution according to claim 1, characterized in that, The complex convolution model in step D2 is obtained by modifying the ResNet18 network, changing ResNet18 into a complex form.

4. The clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution according to claim 3, characterized in that, The construction of the complex convolution model involves modifying ResNet18 to a complex form, specifically by changing the convolution kernel and block kernel architecture to complex forms. For each convolutional block, including the convolution and ReLU function, the specific forms are as follows: The input is the real part 'a' and the imaginary part 'b' of the complex matrix P = a + bi in the spectrum, where the complex convolution kernel is represented as Q = c + di. The complex convolution is calculated as follows: The complex number ReLU satisfies the Cauchy-Riemann equation when both its real and imaginary parts are strictly positive or strictly negative, and the calculation formula is as follows:

5. A clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution according to claim 1, characterized in that, In step D3, a Gaussian filter is used to filter the features obtained after complex convolution.

6. A clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution according to claim 4, characterized in that, Step S5 specifically includes: E1: Integrate the real and imaginary features of the network output to obtain the final classification features. The specific formula is as follows: Among them, F final F real and F imag These represent the final classification features, real features, and imaginary features, respectively. E2: Based on the final features obtained in step E1, the classification result is obtained.

7. A clinical image lesion classification method based on adaptive frequency domain learning and anti-aliasing complex convolution according to claim 6, characterized in that, In step E2, the classification result is obtained through the loss function Softmax.

Images