A breast cancer auxiliary identification method and device based on multi-parameter MRI images

Abstract

The present invention provides a method and device for assisting breast cancer identification based on multi-parameter MRI images. The method comprises: respectively acquiring breast MRI images I1 and I2 of a DWI-ADC sequence and a T1 dynamic enhancement sequence of the same patient; extracting a 3D image block I from I1 1‑3D ; Extract 3D image blocks from I2 and stitch them together in time order to form 4D image blocks I containing time information 2‑4D ; Use 3D convolutional neural network module to respectively analyze 3D image blocks I 1‑3D , 4D image block I 2‑4D Feature extraction is performed to obtain features F1 and F2; features F1 and F2 are fused, and the fused features are input into a benign and malignant classifier to obtain the benign and malignant nature of the breast lesion. The present invention improves the recognition accuracy of breast cancer by extracting and fusing features of two different sequences of MRI images.

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging technology, and in particular relates to a breast cancer auxiliary identification method and device based on multi-parameter MRI (Magnetic Resonance Imaging) images. Background Art

[0002] At present, breast cancer has become the cancer that poses the greatest threat to women's health worldwide. In China, breast cancer presents an "urbanization" phenomenon. In relatively economically developed areas, urban women have an increased risk of cancer, and breast cancer has become a major public health issue in today's society. Medical imaging plays an important role in breast cancer diagnosis, staging, treatment design, and efficacy evaluation, and is a prerequisite for improving the treatment effect of breast cancer. Taking medical imaging examinations to assist in early diagnosis of tumors can help reduce the mortality rate of breast cancer patients. The negative predictive value of breast MRI in diagnosing breast cancer is as high as 98%, which can more intuitively reflect the morphological characteristics of the lesion and is the most sensitive examination method for detecting breast cancer. In MRI, bilateral breasts can be developed simultaneously for easy comparison and identification, and have unique value in the diagnosis of multifocal and multicenter lesions, lesion range assessment, and axillary lymph node status determination. Unfortunately, however, breast MRI is not very sensitive to the diagnosis of microcalcifications, and there is a certain overlap in the manifestations of benign and malignant lesions. Some benign lesions with rich blood supply can show enhancement patterns similar to tumors, resulting in misdiagnosis. Therefore, the empirical interpretation of MRI imaging signs is highly subjective, and its diagnostic accuracy is limited by the experience and professional ability of the physician.

[0003] With the continuous development of computer technology, image processing and pattern recognition technology, computer-aided diagnosis technology (CAD) has emerged and applied to clinical diagnosis, which has improved the diagnostic accuracy of breast cancer to a certain extent. The extraction of features in traditional CAD systems needs to be manually set based on experience, including shallow features such as grayscale, shape and texture, which can only achieve auxiliary diagnosis to a certain extent. Deep learning methods do not require prior knowledge and automatically extract the most predictive high-level semantic features from image big data; the larger the amount of data given to it, the higher the model recognition accuracy will be, and there is hope to form an efficient intelligent diagnosis system. Artificial Intelligence (AI) can extract high-dimensional information contained in medical images that is difficult for the naked eye to perceive. Using intelligent diagnosis systems to pre-diagnose MRI images can reduce a lot of film reading time, which helps to alleviate the contradiction between the growing medical needs of patients, the continuously increasing number of images and the relatively insufficient gap between imaging diagnosis experts. Using AI-assisted diagnosis technology to assist in breast cancer diagnosis and screening is a feasible way to achieve early detection and early treatment of breast cancer and improve the survival rate of breast cancer patients.

[0004] The prior art generally performs auxiliary diagnosis of benign and malignant breast cancer lesions by inputting a single-parameter MRI image into a trained neural network model, but the disadvantage is that the accuracy is not high. To this end, the present invention provides a breast cancer auxiliary diagnosis method and device based on multi-parameter MRI images. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a breast cancer auxiliary identification method and device based on multi-parameter MRI images, which utilizes multi-parameter nuclear magnetic resonance MRI images to extract and fuse lesion information on different nuclear magnetic resonance sequences, thereby effectively improving the accuracy of judging the benign and malignant nature of breast lesions.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions.

[0007] In a first aspect, the present invention provides a method for assisting breast cancer identification based on multi-parameter MRI images, comprising the following steps:

[0008] Step 101, respectively acquiring breast MRI images I1 and I2 of the same patient in a DWI-ADC sequence and a T1 dynamic enhancement sequence;

[0009] Step 102: extract 3D image block I from I1 1-3D ; Extract 3D image blocks from I2 and stitch them together in time order to form 4D image blocks I containing time information 2-4D ;

[0010] Step 103: Use a 3D convolutional neural network module to respectively process the 3D image blocks I 1-3D , 4D image block I 2-4D Perform feature extraction to obtain features F1 and F2;

[0011] Step 104, the features F1 and F2 are fused, and the fused features are input into a benign or malignant classifier to obtain the benign or malignant status of the breast lesion.

[0012] Furthermore, the step 102 further includes:

[0013] Divide the region of interest 1 from I1, and then divide the region of interest 1 into 3D image blocks I 1-3D ;

[0014] From {I2(t i )}, and then divide the region of interest 2 into 3D image blocks, and splice them together in time sequence to form a 4D image block I containing time information 2-4D .

[0015] Furthermore, features F1 and F2 are fused based on the attention mechanism.

[0016] Furthermore, the fusion of features F1 and F2 based on the attention mechanism includes:

[0017] Add features F1 and F2, perform 1*1 convolution, and then perform Sigmoid operation to obtain the weight matrix;

[0018] The features F1 and F2 are fused as follows:

[0019] F=F1×W+F2×(1-W) (1)

[0020] In the formula, F is the feature after the fusion of F1 and F2, and W is the weight matrix.

[0021] Furthermore, the benign and malignant classifier is composed of a fully connected layer, and the fused features are input into the fully connected layer to obtain the probability that the lesion is malignant; if the probability is greater than a set threshold, the lesion is malignant; otherwise, the lesion is benign.

[0022] In a second aspect, the present invention provides a breast cancer auxiliary identification device based on multi-parameter MRI images, comprising:

[0023] An image acquisition module, used to respectively acquire breast MRI images I1 and I2 of a DWI-ADC sequence and a T1 dynamic enhancement sequence of the same patient;

[0024] Image block extraction module, used to extract 3D image block I from I1 1-3D ; Extract 3D image blocks from I2 and stitch them together in time order to form 4D image blocks I containing time information 2-4D ;

[0025] The feature extraction module is used to use the 3D convolutional neural network module to extract the 3D image blocks I 1-3D , 4D image block I 2-4D Perform feature extraction to obtain features F1 and F2;

[0026] The lesion classification module is used to fuse the features F1 and F2, and input the fused features into a benign or malignant classifier to obtain the benign or malignant status of the breast lesion.

[0027] Furthermore, the feature extraction module is also used for:

[0028] Divide the region of interest 1 from I1, and then divide the region of interest 1 into 3D image blocks I 1-3D ;

[0029] The region of interest 2 is divided from I2, and then the region of interest 2 is divided into 3D image blocks, and they are spliced ​​together in time sequence to form a 4D image block I containing time information. 2-4D .

[0030] Furthermore, the lesion classification module fuses features F1 and F2 based on an attention mechanism.

[0031] Furthermore, the fusion of features F1 and F2 based on the attention mechanism includes:

[0032] Add features F1 and F2, perform 1*1 convolution, and then perform Sigmoid operation to obtain the weight matrix;

[0033] The features F1 and F2 are fused as follows:

[0034] F=F1×W+F2×(1-W) (1)

[0035] In the formula, F is the feature after the fusion of F1 and F2, and W is the weight matrix.

[0036] Furthermore, the benign and malignant classifier is composed of a fully connected layer, and the fused features are input into the fully connected layer to obtain the probability that the lesion is malignant; if the probability is greater than a set threshold, the lesion is malignant; otherwise, the lesion is benign.

[0037] Compared with the prior art, the present invention has the following beneficial effects.

[0038] The present invention obtains breast MRI images I1 and I2 of the same patient's DWI-ADC sequence and T1 dynamic enhancement sequence respectively, and extracts a 3D image block I from I1. 1-3D , extract 3D image blocks from I2 and stitch them together in time sequence to form a 4D image block I containing time information 2-4D , using 3D convolutional neural network modules to respectively 1-3D ,I 2-4D Feature extraction is performed to obtain features F1 and F2, and features F1 and F2 are fused. The fused features are input into a benign and malignant classifier to obtain the benign and malignant nature of breast lesions, thereby realizing auxiliary identification of breast cancer based on multi-parameter MRI images. Since the MRI images of the T1 dynamic enhancement sequence contain multi-angle and multi-phase information of the lesions, and the MRI images of the DWI-ADC sequence contain information about the diffusion of water molecules in breast and lesion tissues, the present invention improves the accuracy of breast cancer identification by extracting and fusing the features of these two different sequences of MRI images. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1The present invention is a flowchart of a method for assisting breast cancer identification based on multi-parameter MRI images.

[0040] Figure 2 Schematic diagram of the principle of obtaining the attention mechanism weight matrix.

[0041] Figure 3 The present invention is a block diagram of a breast cancer auxiliary identification device based on multi-parameter MRI images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention clearer and more understandable, the present invention is further described below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0043] Figure 1 The present invention is a flowchart of a method for assisting breast cancer identification based on multi-parameter MRI images, which includes the following steps:

[0044] Step 101, respectively acquiring breast MRI images I1 and I2 of the same patient in a DWI-ADC sequence and a T1 dynamic enhancement sequence;

[0045] Step 102: extract 3D image block I from I1 1-3D ; Extract 3D image blocks from I2 and stitch them together in time order to form 4D image blocks I containing time information 2-4D ;

[0046] Step 103: Use a 3D convolutional neural network module to respectively process the 3D image blocks I 1-3D , 4D image block I 2-4D Perform feature extraction to obtain features F1 and F2;

[0047] Step 104, the features F1 and F2 are fused, and the fused features are input into a benign or malignant classifier to obtain the benign or malignant status of the breast lesion.

[0048] In this embodiment, step 101 is mainly used to obtain multi-parameter breast nuclear magnetic resonance MRI images of the same patient. The multi-parameter MRI images of this embodiment include T1 dynamic enhancement sequence and DWI-ADC sequence. The T1 dynamic enhancement sequence covers imaging before contrast agent injection and imaging collected at multiple time points after contrast agent injection. For example, 5 phases are continuously collected after contrast agent injection, each phase acquisition time is 65s, and the k-space (Fourier space) center acquisition time is 52 seconds, that is, multi-phase contrast enhancement images are collected at about 46 seconds, 111 seconds, 176 seconds and 241 seconds after contrast agent injection. This timing information provides a multi-angle and multi-phase assessment of the lesion, which helps to more fully understand the dynamic changes of the lesion. In contrast, the DWI-ADC sequence provides information about the diffusion of water molecules in breast and lesion tissues.

[0049] In this embodiment, step 102 is mainly used to divide two types of MRI images into 3D image blocks. The MRI image of the DWI-ADC sequence is directly divided into multiple small 3D image blocks; for the T1 dynamic enhancement sequence MRI image, it is first divided into multiple small 3D image blocks, and then the 3D image blocks are spliced ​​together in chronological order to form a 4D image block containing time information. Dividing the image into small image blocks and then performing feature extraction has the following advantages: first, it can improve computational efficiency and reduce computational complexity and memory usage; second, it can capture local features of the image; third, the randomness of dividing small image blocks can increase the robustness of the model and prevent overfitting caused by too little data.

[0050] In this embodiment, step 103 is mainly used to extract features from the obtained 3D image blocks and 4D image blocks. In this embodiment, a 3D convolutional neural network CNN is used for feature extraction. CNN is a feedforward neural network, but unlike a general fully connected feedforward neural network, its convolution layer has the characteristics of local connection and weight sharing, so it can greatly reduce the number of weight parameters, thereby reducing the complexity of the model and improving the running speed. A typical CNN mainly includes a convolution layer and a pooling layer. The function of the convolution layer is to extract the features of a local area through the convolution operation of the convolution kernel and the input image. Different convolution kernels are equivalent to different feature extractors; the function of the pooling layer is to perform feature selection, reduce the number of features, and further reduce the number of parameters. In this embodiment, the overall design of the two CNNs used for feature extraction of 3D image blocks and 4D image blocks is basically the same, and both are based on the Residual Block of ResNet for feature extraction. The difference in details is that due to the different dimensions of the two image blocks, different numbers of channels are used in the first convolution layer.

[0051] In this embodiment, step 104 is mainly used for feature fusion and benign and malignant identification of lesions. In this embodiment, the features F1 and F2 extracted from the 3D image block and the 4D image block are first fused, and then the fused features are input into a benign and malignant classifier to obtain the benign and malignant of the breast lesion. The benign and malignant classifier is a binary classifier that outputs two benign and malignant categories of the lesion: benign and malignant.

[0052] During the model training phase, the probability distribution of benign and malignant lesions is compared with the actual benign and malignant results verified by tissue biopsy. The cross entropy loss function is used to calculate the loss value and gradient, and the back propagation mechanism is used to optimize the parameters of the above 3D convolutional network. This process helps the model to continuously learn and more accurately reflect the lesion characteristics in the MRI images. The cross entropy loss function can be expressed as:

[0053]

[0054] In the formula, L is the cross entropy loss, N is the number of samples, and y i is the value of the i-th sample label (0 or 1), p i is the probability that the i-th sample is benign.

[0055] In this embodiment, since the MRI images of the T1 dynamic enhancement sequence contain multi-angle and multi-phase information of the lesion, and the MRI images of the DWI-ADC sequence contain information about the diffusion of water molecules in the breast and lesion tissues, this embodiment improves the accuracy of identifying benign and malignant lesions by extracting and fusing features of these two MRI images and identifying benign and malignant lesions based on the fused features.

[0056] As an optional embodiment, the step 102 further includes:

[0057] Divide the region of interest 1 from I1, and then divide the region of interest 1 into 3D image blocks I 1-3D ;

[0058] From {I2(t i )}, and then divide the region of interest 2 into 3D image blocks, and splice them together in time sequence to form a 4D image block I containing time information 2-4D .

[0059] This embodiment provides an improved solution to reduce the amount of calculation. As mentioned above, dividing the MRI image into multiple small 3D image blocks and then performing feature extraction can reduce the amount of calculation and improve the data processing speed; in order to further reduce the amount of calculation, this embodiment first separates the region of interest 1 and region of interest 2 containing the lesion and the surrounding structure of the lesion from the two types of MRI images, and then only divides the region of interest 1 and region of interest 2 into multiple small 3D image blocks, and does not process other regions. The region of interest can be generated by the radiologist using software to semi-automatically outline the lesion contour; it can also be automatically generated by inputting the image into a trained lesion segmentation model.

[0060] As an optional embodiment, features F1 and F2 are fused based on an attention mechanism.

[0061] This embodiment provides a technical solution for feature fusion. This embodiment uses the attention mechanism to fuse the extracted features F1 and F2. The attention mechanism is a commonly used technique in computer science and machine learning, which allows the model to selectively focus on key parts of the input data when processing information. This mechanism originates from the study of human vision, in which humans selectively focus on part of all information while ignoring other visible information to make rational use of limited visual information processing resources. The calculation of the attention mechanism can be divided into two steps: one is to calculate the attention distribution on all input information; the other is to calculate the weighted sum of the input information based on the attention distribution. In this embodiment, features F1 and F2 are used as inputs to the attention module, and the attention module outputs the fused features of F1 and F2.

[0062] As an optional embodiment, the fusing of features F1 and F2 based on the attention mechanism includes:

[0063] Add features F1 and F2, perform 1*1 convolution, and then perform Sigmoid operation to obtain the weight matrix;

[0064] The features F1 and F2 are fused as follows:

[0065] F=F1×W+F2×(1-W) (1)

[0066] In the formula, F is the feature after the fusion of F1 and F2, and W is the weight matrix.

[0067] This embodiment provides a technical solution for feature fusion based on the attention mechanism. First, the features F1 and F2 are summed, and then a 1*1 convolution operation is performed. Through the 1*1 convolution, the feature matrix can be compressed to a single channel without changing the image size, and an eigenvalue is generated for each pixel point at each position on the plane. Then, each eigenvalue is normalized to between 0 and 1 through the Sigmoid operation to obtain the weight matrix W, as shown in Figure 2As shown. The expression of Sigmoid is as follows:

[0068]

[0069] With the weight matrix W, we can perform weighted summation of F1 and F2 according to formula (1) to obtain the fusion matrix F. In formula (1), W is a 1*h*w matrix, 1 refers to a matrix of size 1*h*w, in which the element at each position is 1, and h and w are the number of rows and columns of the feature matrix.

[0070] As an optional embodiment, the benign and malignant classifier is composed of a fully connected layer, and the fused features are input into the fully connected layer to obtain the probability that the lesion is malignant; if the probability is greater than a set threshold, the lesion is malignant; otherwise, the lesion is benign.

[0071] This embodiment provides a technical solution for a benign and malignant classifier. The benign and malignant classifier of this embodiment is composed of a fully connected layer. The fused feature F is input into the fully connected layer, and the probability that the lesion is malignant is output. The probability is a floating point number between 0 and 1, where the larger the probability value, the higher the possibility that the lesion is malignant. Therefore, the benign and malignant nature of the lesion can be judged according to the size of the probability. Generally, the probability is compared with a set threshold (such as 0.5). If the probability exceeds the set threshold, it is considered to be malignant; otherwise, it is benign.

[0072] Figure 3 The present invention is a schematic diagram of a breast cancer auxiliary identification device based on multi-parameter MRI images according to an embodiment of the present invention. The device includes:

[0073] An image acquisition module 11 is used to respectively acquire breast MRI images I1 and I2 of a DWI-ADC sequence and a T1 dynamic enhancement sequence of the same patient;

[0074] The image block extraction module 12 is used to extract the 3D image block I from I1. 1-3D ; Extract 3D image blocks from I2 and stitch them together in time order to form 4D image blocks I containing time information 2-4D ;

[0075] The feature extraction module 13 is used to respectively extract the 3D image blocks I by using the 3D convolutional neural network module. 1-3D , 4D image block I 2-4D Perform feature extraction to obtain features F1 and F2;

[0076] The lesion classification module 14 is used to fuse the features F1 and F2, and input the fused features into a benign or malignant classifier to obtain the benign or malignant status of the breast lesion.

[0077] The device of this embodiment can be used to perform Figure 1 The technical solution of the method embodiment shown has similar implementation principles and technical effects, which will not be described in detail here. The same is true for the following embodiments, which will not be described in detail.

[0078] As an optional embodiment, the feature extraction module is further used for:

[0079] Divide the region of interest 1 from I1, and then divide the region of interest 1 into 3D image blocks I 1-3D ;

[0080] The region of interest 2 is divided from I2, and then the region of interest 2 is divided into 3D image blocks, and they are spliced ​​together in time sequence to form a 4D image block I containing time information. 2-4D .

[0081] As an optional embodiment, the lesion classification module fuses features F1 and F2 based on an attention mechanism.

[0082] As an optional embodiment, the fusing of features F1 and F2 based on the attention mechanism includes:

[0083] Add features F1 and F2, perform 1*1 convolution, and then perform Sigmoid operation to obtain the weight matrix;

[0084] The features F1 and F2 are fused as follows:

[0085] F=F1×W+F2×(1-W) (1)

[0086] In the formula, F is the feature after the fusion of F1 and F2, and W is the weight matrix.

[0087] As an optional embodiment, the benign and malignant classifier is composed of a fully connected layer, and the fused features are input into the fully connected layer to obtain the probability that the lesion is malignant; if the probability is greater than a set threshold, the lesion is malignant; otherwise, the lesion is benign.

[0088] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A breast cancer auxiliary identification method based on multi-parameter MRI images, characterized in that: The following steps are involved: Step 101, respectively obtaining breast MRI images I1 and I2 of a DWI-ADC sequence and a T1 dynamic enhancement sequence of the same patient; the T1 dynamic enhancement sequence is obtained by arranging images before contrast agent injection and images acquired at multiple time points after contrast agent injection in chronological order; Step 102: divide I1 into a plurality of small 3D image blocks, denoted as I 1-3D ; Divide I2 into multiple small 3D image blocks and splice them together in time sequence to form a 4D image block I containing time information 2-4D ; Step 103: Use a 3D convolutional neural network module to respectively process the 3D image blocks I 1-3D , 4D image block I 2-4D Perform feature extraction to obtain features F1 and F2; Step 104, the features F1 and F2 are fused, and the fused features are input into a benign or malignant classifier to obtain the benign or malignant status of the breast lesion.

2. The breast cancer auxiliary identification method based on multi-parameter MRI images according to claim 1, characterized in that: The step 102 further includes: Divide the region of interest 1 from I1, and then divide the region of interest 1 into multiple small 3D image blocks I 1-3D ; From {I2(t i )}, and then divide the region of interest 2 into multiple small 3D image blocks, and splice them together in time sequence to form a 4D image block I containing time information 2-4D .

3. The breast cancer auxiliary identification method based on multi-parameter MRI images according to claim 1, characterized in that: Features F1 and F2 are fused based on the attention mechanism.

4. The breast cancer auxiliary identification method based on multi-parameter MRI images according to claim 3, characterized in that: The fusion of features F1 and F2 based on the attention mechanism includes: Add features F1 and F2, perform 1*1 convolution, and then perform Sigmoid operation to obtain the weight matrix; The features F1 and F2 are fused as follows: F=F1×W+F2×(1-W) (1) In the formula, F is the feature after the fusion of F1 and F2, and W is the weight matrix.

5. The breast cancer auxiliary identification method based on multi-parameter MRI images according to claim 1, characterized in that: The benign and malignant classifier is composed of a fully connected layer, and the fused features are input into the fully connected layer to obtain the probability that the lesion is malignant; if the probability is greater than a set threshold, the lesion is malignant; otherwise, the lesion is benign.

6. A breast cancer auxiliary identification device based on multi-parameter MRI images, characterized in that: include: An image acquisition module is used to respectively acquire breast MRI images I1 and I2 of a DWI-ADC sequence and a T1 dynamic enhancement sequence of the same patient; the T1 dynamic enhancement sequence is obtained by arranging images before contrast agent injection and images acquired at multiple time points after contrast agent injection in chronological order; The image block extraction module is used to divide I1 into multiple small 3D image blocks, denoted as I 1-3D ; Divide I2 into multiple small 3D image blocks and stitch them together in time sequence to form a 4D image block I containing time information 2-4D ; The feature extraction module is used to use the 3D convolutional neural network module to extract the 3D image blocks I 1-3D , 4D image block I 2-4D Perform feature extraction to obtain features F1 and F2; The lesion classification module is used to fuse the features F1 and F2, and input the fused features into a benign or malignant classifier to obtain the benign or malignant status of the breast lesion.

7. The breast cancer auxiliary identification device based on multi-parameter MRI images according to claim 6, characterized in that: The feature extraction module is also used for: Divide the region of interest 1 from I1, and then divide the region of interest 1 into multiple small 3D image blocks I 1-3D ; The region of interest 2 is divided from I2, and then the region of interest 2 is divided into multiple small 3D image blocks, which are spliced ​​together in time sequence to form a 4D image block I containing time information. 2-4D .

8. The breast cancer auxiliary identification device based on multi-parameter MRI images according to claim 6, characterized in that: The lesion classification module fuses features F1 and F2 based on the attention mechanism.

9. The breast cancer auxiliary identification device based on multi-parameter MRI images according to claim 8, characterized in that: The fusion of features F1 and F2 based on the attention mechanism includes: Add features F1 and F2, perform 1*1 convolution, and then perform Sigmoid operation to obtain the weight matrix; The features F1 and F2 are fused as follows: F=F1×W+F2×(1-W) (1) In the formula, F is the feature after the fusion of F1 and F2, and W is the weight matrix.

10. The breast cancer auxiliary identification device based on multi-parameter MRI images according to claim 6, characterized in that: The benign and malignant classifier is composed of a fully connected layer, and the fused features are input into the fully connected layer to obtain the probability that the lesion is malignant; if the probability is greater than a set threshold, the lesion is malignant; otherwise, the lesion is benign.

Images