An EfficientNet optimization method for mammography image detection

Abstract

This invention belongs to the field of computer vision technology, specifically relating to an optimized EfficientNet method for mammography image detection. The purpose of this invention is to address the problem of low accuracy in identifying benign and malignant images due to the large size of mammography images and the small size of lesion areas. This invention selects the EfficientNet-B0 network for mammography image classification and recognition, replacing the SENet module in the EfficientNet-B0 model with a CBAM module and the swish activation function with an h-swish function. This invention combines the high speed and accuracy of EfficientNet-B0 with the channel feature sensitivity of the CBAM module, improving the recognition of mammography images while reducing model parameters. The optimized EfficientNet-B0 model requires fewer parameters, has higher accuracy, and faster recognition speed. Using the optimized EfficientNet-B0 model to identify mammography image datasets can improve the accuracy of identifying benign and malignant breast tumors.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to an EfficientNet optimization method for mammography image detection. Background Technology

[0002] Breast cancer is the most common cancer among women, accounting for 24.2% of all female cancers worldwide. Furthermore, breast cancer has the highest mortality rate among all cancers. These patients represent a large proportion of all breast cancer patients globally. Early detection of breast cancer is crucial for subsequent treatment. Mammography offers advantages such as low cost and ease of operation, making it one of the most reliable screening methods currently available. Moreover, mammography has high sensitivity, capable of detecting small breast cancer lesions and some occult lesions, thus reducing radiation dose. Mammography can accurately identify calcified lesions that cannot be distinguished by ultrasound or MRI. In addition, mammography allows for observation of the breast from multiple angles; therefore, mammography is the preferred method for breast cancer screening. However, several factors may affect the effectiveness of mammography in diagnosing cancer. First, the large number of patients increases the workload, potentially reducing the accuracy of doctors' work. Second, some radiologists may not reach consistent conclusions due to the difficulty in interpreting unimportant lesions and the subjectivity of the information.

[0003] In recent years, the widespread success of deep learning in the field of computer vision has become commonplace. Deep learning is primarily applied to mammography images in three areas: classification, object detection, and segmentation. Deep learning can automatically extract features from raw data, which is its biggest difference from traditional methods. This improves the feature extraction capabilities of algorithms, thereby enhancing the efficiency and accuracy of disease diagnosis. Currently, simply using convolutional neural network backbones for feature extraction does not significantly improve the performance in diagnosing benign or malignant tumors.

[0004] The EfficientNet model, proposed by Google at the 2019 ICML conference, is a new fast and refined model. EfficientNet uses a fixed set of scaling factors to uniformly scale each dimension (such as network width, network depth, and image resolution), employing a simple yet efficient composite scaling method. This allows for a more principled and easy scaling of benchmark CNNs to any target resource while maintaining model efficiency. Google has developed a series of versions using this technique, currently ranging from EfficientNet-B0 to EfficientNet-B8, plus EfficientNet-L2 and NoisyStudent, totaling 11 versions. As the model's accuracy and scale increase, so too does its memory requirement. Considering the performance limitations of the experimental system, this invention selects the EfficientNet-B0 model for improvement, performing benign and malignant classification of molybdenum target images. Summary of the Invention

[0005] The purpose of this invention is to provide an EfficientNet optimization method for mammography image detection.

[0006] An EfficientNet optimization method for mammography image detection includes the following steps:

[0007] Step 1: Obtain the mammogram image dataset and preprocess the samples in the dataset;

[0008] Step 2: Construct an optimized EfficientNet-B0 classification model;

[0009] Step 2.1: Construct an EfficientNet-B0 network model based on the PyTorch framework, and then use a set of fixed scaling factors to uniformly scale each dimension based on EfficientNet's adaptive model scaling technique, thereby searching for the network structure most suitable for the dataset constructed in Step 1.

[0010] Step 2.2: Optimize the MBConv module of the EfficientNet-B0 model by replacing the original SENet module with the CBAM module. The CBAM module is a module that combines channel and spatial information. It contains two independent sub-modules: the channel attention module CAM and the spatial attention module SAM. It is used to obtain feature attention maps in the channel and spatial dimensions of the given intermediate feature map, and then multiply them with the original feature map for optimization.

[0011] In the Channel Attention (CAM) module, the image features F are first input, followed by global average pooling and global maximum pooling modules. Then, a shared Multiprocessing Pooling (MLP) module is used to obtain two output features. The two output features are then added together, and finally, the sigmoid activation function is used to obtain the feature weight coefficients M. c Finally, the obtained weight coefficients are multiplied by the image features output by the MLP module to obtain new image features F';

[0012] The F' features obtained by the channel attention module (CAM) are used as the input to the spatial attention module (SAM). In the SAM, after inputting the features, pooling operations are performed along the channel dimension, using both global average pooling and global maximum pooling. The resulting features are then concatenated, followed by a convolutional module, where the sigmoid activation function is used to calculate the weight coefficients M. s The final features are obtained by multiplying the weight coefficients and the input features;

[0013] Step 2.3: Optimize the swish activation function block used in the EfficientNet-B0 model by replacing the swish activation function with the h-swish activation function;

[0014] The formula for the h-swish function is:

[0015]

[0016] Among them, ReLU6(x) is the core function of the h-swish activation function;

[0017] Step 3: Train the optimized EfficientNet-B0 classification model using the training set constructed in Step 1;

[0018] Step 4: Input the mammogram image to be detected into the trained and optimized EfficientNet-B0 classification model to obtain the benign or malignant identification result of the input image.

[0019] Furthermore, in step 1, obtaining the mammography image dataset includes obtaining publicly available mammography datasets CBISDDSM, MIAS, and datasets provided by local hospitals. Due to the different dataset formats, the image formats, sizes, and resolutions of the datasets are standardized and preprocessed to uniformly convert the format of all images in the datasets and unify the pixels.

[0020] Furthermore, the preprocessing method for the molybdenum target image samples in the dataset in step 1 includes cropping feature centers, data augmentation, and histogram equalization; the data augmentation includes performing geometric transformations such as rotation, horizontal and vertical flipping on the images to expand the dataset; adding Gaussian noise to the images improves the model's generalization ability and doubles the original data size; the Gaussian noise is... x represents the grayscale value, u represents the expected value of x, and σ represents the standard deviation of x. 2 This represents the variance of x.

[0021] The beneficial effects of this invention are as follows:

[0022] This invention combines the high speed and accuracy of EfficientNet-B0 with the channel feature sensitivity of the CBAM module, thereby improving the recognition of liver mammograms while reducing model parameters. The optimized EfficientNet-B0 model in this invention requires fewer parameters, has higher recognition accuracy, and faster recognition speed. Attached Figure Description

[0023] Figure 1 This is a network structure diagram of the optimized EfficientNet-B0 model in an embodiment of the present invention.

[0024] Figure 2 This is a structural diagram of CBAM in an embodiment of the present invention.

[0025] Figure 3 This is a structural diagram of CBAM-MBConv in an embodiment of the present invention. Detailed Implementation

[0026] The present invention will now be further described with reference to the accompanying drawings.

[0027] An optimized EfficientNet method for mammography image detection is disclosed, relating to the field of computer vision technology. The purpose of this invention is to address the problem of low accuracy in identifying benign and malignant images due to the large image size and small lesion area of ​​mammography images. The EfficientNET-B0 network is selected for mammography image classification, demonstrating high classification performance. Furthermore, the SENet module in the EfficientNET-B0 model is replaced with a CBAM module, and the swish activation function is replaced with an h-swish function. Training a mammography dataset with the improved EfficientNET-B0 network further enhances the accuracy in identifying benign and malignant breast tumors.

[0028] S1. Dataset Construction: Obtain publicly available mammography datasets MIAS and CBIS DDSM and local hospital data to construct a dataset.

[0029] S2. Data preprocessing: Preprocessing the dataset, including geometric transformations, data augmentation, and histogram equalization.

[0030] S3. Construct the EfficientNet-B0 classification model, train the model, and explore a composite scaling strategy suitable for the molybdenum target dataset.

[0031] S4. Optimize the EfficientNet-B0 model by replacing the original SENet module with the CBAM module and the swish activation function with the h-swish activation function. Train the improved EfficientNet-B0 classification model and save the training model.

[0032] S5. Breast cancer benign / malignant identification: Input mammograms from the test set and use the improved EfficientNet-B0 model.

[0033] In step S1, the dataset is constructed by acquiring publicly available mammography datasets CBIS DDSM, MIAS, and datasets provided by local hospitals. Due to the different dataset formats, the image format, size, resolution, and other parameters of the datasets are standardized and preprocessed. The dataset images are uniformly converted to JPG format, and OpenCV is used to unify the pixel size to 1024×1024.

[0034] In step S2, data preprocessing is performed on the dataset, including cropping feature centers, data augmentation, and histogram equalization. The specific steps are as follows:

[0035] Data augmentation includes performing geometric transformations on images, such as rotation and horizontal / vertical flipping, to expand the dataset; and adding Gaussian noise to images to improve the model's generalization ability. Gaussian noise refers to noise whose probability density function distribution follows a Gaussian distribution, as shown in the following equation: Where x represents the gray value, u represents the expected value of x, and σ represents the standard deviation of x. 2 Let x represent the variance. Adding Gaussian noise doubles the original data size.

[0036] In step S3, an EfficientNet-B0 classification model needs to be constructed, trained, and a composite scaling strategy suitable for the molybdenum target dataset needs to be explored. First, an EfficientNet-B0 network model is built based on the PyTorch framework. Then, based on EfficientNet's adaptive model scaling technique, each dimension is scaled uniformly using a set of fixed scaling factors to search for the network structure most suitable for the dataset.

[0037] The EfficientNet-B0 model has a depth and width scaling factor of 1 and a resolution of 224×224. The specific implementation method is to modify the depth, width and image resolution parameters of the network to train the network model. First, the depth and width scaling factors are modified to 4 and 2 respectively, and the image resolution is modified to twice the original. Finally, a set of optimal parameter scaling adjustments is explored to achieve the highest model accuracy.

[0038] In step S4, the EfficientNet-B0 model is optimized by replacing the original SENet module with the CBAM module and the swish activation function with the h-swish activation function. The improved EfficientNet-B0 classification model is then trained and the training model is saved.

[0039] like Figure 1 As shown, the improved model's main module consists of 16 MBConv modules, 2 Conv modules, 1 global average pooling layer, and 1 FC layer. This invention primarily improves the MBConv module, with the structure as follows: Figure 3 As shown, the attention mechanism in the MBConv module is first improved; secondly, the swish activation function is replaced with the h-swish activation function; and finally, the 1×1 convolutional layer in the module is changed to grouped convolution.

[0040] 1) An improvement was made to the core module of EfficientNet, the mobile inverted bottleneck convolution (MBConv). The original SENet module was replaced with a Convolutional Block Attention Module (CBAM) that combines channel and spatial information. The CBAM module is a commonly used functional module in computer vision, mainly consisting of two independent sub-modules: a Channel Attention Module (CAM) and a Spartial Attention Module (SAM). Its purpose is to sequentially derive feature attention maps from the given intermediate feature map along both the channel and spatial dimensions, and then multiply them with the original feature map for optimization, reducing the loss of original image information and thus improving the model's classification accuracy. For example... Figure 2 As shown.

[0041] In the CAM module structure, the image features F are first input, followed by two global pooling modules: global average pooling and global maximum pooling. Then, a shared MLP module is used to obtain two output features. The two output features are added together, and finally, the sigmoid activation function is used to obtain the feature weight coefficients M.c Finally, the obtained weight coefficients are multiplied by the image features output by the MLP module to obtain the new image features F'. The channel attention calculation formula is as follows:

[0042]

[0043] The F' features obtained from the CAM module are used as input to the SAM module. In the spatial attention module, after inputting the features, pooling operations are performed along the channel dimension, using both global average pooling and global maximum pooling. The resulting features are then concatenated, followed by a 7×7 convolutional module. The sigmoid activation function is used to calculate the weight coefficients M. s The final feature is obtained by multiplying the weight coefficients and the input features. The formula for calculating spatial attention is as follows:

[0044]

[0045] After replacing the SENet module with the CBAM module, the overall structure of the EfficientNet-B0 model remains unchanged. The structure of the combined MBConv and CBAM modules is as follows: Figure 3 As shown.

[0046] 2) The activation function used in the EfficientNet-B0 model is replaced and optimized. The EfficientNet-B0 model uses the swish activation function. While the swish activation function can improve classification accuracy, its derivative calculation is complex and computationally expensive, leading to slower convergence during training. When training the model using large molybdenum target images, computational speed must be considered. Therefore, to balance training speed with the improved classification performance of the swish activation function, this invention replaces the swish activation function with the h-swish activation function. The expression for the swish function is as follows: swish(x) = s·σ(x), where σ(x) = (1 + exp(-x)). -1 The h-swish activation function mainly improves upon the σ(x) function in the swish activation function, replacing it with... The formula for the h-swish function is as follows: The expression ReLU6(x) is the core function of the h-swish activation function. In fact, the h-swish activation function is a variant of the ReLU6(x) activation function, and its specific expansion is shown below:

[0047]

[0048] When performing non-linear activation of neurons in a neural network, h-swish has the same effect as the swish activation function. Compared to the original swish activation function, the h-swish activation function is easier to differentiate, which can improve the model's convergence speed.

[0049] 3) Model training: The dataset constructed in step S2 is divided into training, validation, and test sets. The improved EfficientNET-B0 is used as the base network. The corresponding hyperparameters are set, the Adam optimizer is used, and the model is trained using the training set. The optimal model weights are saved.

[0050] In step S5, the benign or malignant nature of breast cancer is identified by inputting mammogram images from the test set and using the improved EfficientNet-B0 model to identify the benign or malignant nature of the input images.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An EfficientNet optimization method for mammography molybdenum target image detection, characterized in that, The method comprises the following steps: Step 1: obtaining a breast molybdenum target image data set, and preprocessing samples in the data set; Step 2: constructing an optimized EfficientNet-B0 classification model; Step 2.1: constructing an EfficientNet-B0 network model based on a pytorch framework, and then uniformly scaling each dimension by using a set of fixed scaling coefficients based on the adaptive model scaling technology of the EfficientNet, so as to search for a network structure most suitable for the data set constructed in step 1; Step 2.2: optimizing the MBConv module of the EfficientNet-B0 model, and replacing the original SENet module with a CBAM module; The CBAM module is a module combining channel and spatial two-dimensional information, and comprises two independent sub-modules, namely a channel attention module CAM and a spatial attention module SAM, which are used to obtain feature attention maps in the channel and spatial two-dimensional directions in sequence, and then multiply the original feature map to perform optimization; In the channel attention module CAM, first input the image feature F, then the global average pooling and global maximum pooling module, and then through a shared MLP module to obtain two output features respectively, add the two output features, and then use the activation function sigmoid function to obtain the feature weight coefficient M c , and finally use the obtained weight coefficient M c and the image feature output by the MLP module to obtain a new image feature F'; The F' feature obtained by the channel attention module CAM is taken as an input of a spatial attention module SAM; in the spatial attention module SAM, the input feature is subjected to a pooling operation in a channel dimension, still global average pooling and global maximum pooling, the obtained features are connected, and a convolution module is connected after the pooling module to calculate a weight coefficient M using a sigmoid activation function s The final feature is obtained by multiplying the weight coefficient M s and the input feature F'. Step 2.3: optimizing the swish activation function block used in the EfficientNet-B0 model, and replacing the swish activation function with a h-swish activation function; The formula of the h-swish function is: Wherein, ReLU6(x) is the core function of the h-swish activation function; Step 3: learning and training the optimized EfficientNet-B0 classification model by using the training set constructed in step 1; Step 4: inputting the breast molybdenum target image to be detected into the trained optimized EfficientNet-B0 classification model, so as to obtain a benign or malignant identification result of the input image.

2. The EfficientNet optimization method for mammography detection according to claim 1, wherein: In the step 1, the breast molybdenum target image data set is obtained by obtaining a public molybdenum target data set CBISDDSM, a MIAS and a data set provided by a local hospital, and since the data set formats are different, the image formats, sizes and resolutions of the data sets are standardized and preprocessed, all the image formats in the data set are uniformly converted, and the pixels are unified.

3. The EfficientNet optimization method for mammography detection according to claim 1, wherein: The method for preprocessing the molybdenum target image samples in the data set in step 1 includes several cropping feature centers, data enhancement and histogram equalization; the data enhancement includes geometric transformation such as rotation, horizontal and vertical flipping of the image, expansion of the data set, addition of Gaussian noise to the image, improvement of the generalization ability of the model, expansion of the original data by one time, and Gaussian noise x represents a gray value, u represents an expected value of x, sigma represents a standard deviation of x, and sigma 2 represents the variance of x.

Images