Methods for using medical ultrasound images to assist in determining the severity of lesions

Abstract

This invention discloses a method for assisting in the determination of lesion severity using medical ultrasound images, comprising: preprocessing medical ultrasound images; constructing a deep neural network model, dividing the preprocessed ultrasound images into training, validation, and test sets; using the deep neural network model to extract features from the preprocessed medical ultrasound images to obtain feature maps, then performing feature enhancement on the feature maps to obtain feature-enhanced maps; training the deep neural network model using the feature-enhanced maps from the training set to obtain the parameters of the trained convolutional neural network model; validating the trained convolutional neural network model parameters using the feature-enhanced maps from the validation set, and testing the model using the test set; selecting parameters whose test accuracy meets the requirements as the final parameters of the deep neural network model; and using the deep neural network model obtained in step 2 to predict the severity of medical ultrasound images. This invention improves the accuracy of identification and reduces the workload of doctors.

Description

Technical Field

[0001] This invention belongs to the technical field of deep learning applications in medical diagnosis, specifically relating to a method for using medical ultrasound images to assist in determining the severity of lesions. Background Technology

[0002] Breast cancer has become one of the most common cancers worldwide, and also the leading cause of cancer death among women. Ultrasound imaging is one of the important early screening methods for breast diseases. Studies have shown that the 5-year survival rate for early-stage breast cancer is as high as 95%. Therefore, early detection and early treatment are the most effective means of preventing and treating breast cancer.

[0003] However, my country's medical resources are scarce and unevenly distributed, with some remote areas lacking experienced physicians, significantly impacting the detection and diagnosis of early breast lesions. Therefore, we aim to research intelligent diagnostic and treatment methods for breast diseases and construct corresponding auxiliary diagnostic systems. This will promote the downward flow of medical resources and improve the efficiency and accuracy of breast disease screening.

[0004] On the other hand, deep neural networks have become an important means of realizing artificial intelligence, achieving great success in many fields in recent years, and achieving results comparable to human doctors in the intelligent diagnosis and treatment of breast diseases. In addition to classifying breast ultrasound for benign or malignant lesions, research is also being conducted on intelligent BI-RADS grading based on breast ultrasound images. BI-RADS grading is an internationally and domestically recognized standard for measuring the severity of breast lesions, typically including nine levels: 0, 1, 2, 3, 4a, 4b, 4c, 5, and 6. The levels, from low to high, indicate progressively increasing malignancy.

[0005] However, deep neural networks still face many challenges when processing breast ultrasound images. First, feature extraction is difficult. Breast lesions exhibit various characteristics on ultrasound images, including internal features and edge features. Furthermore, the size and location of lesions vary across different images, further complicating feature extraction. Second, there is significant noise and human influence. The imaging quality of ultrasound images depends on the physician's operation and equipment precision. Typically, breast ultrasound images have a low signal-to-noise ratio, are not standardized, and lesions occupy only a small area on the image. Extracting discriminative features for malignant lesion identification is a crucial problem to solve in constructing classification algorithms. Third, there is limited data volume and data imbalance. In reality, breast cancer patients constitute only a small percentage of the population, meaning the data used to build models contains only a small number of malignant samples, while the majority are benign. Therefore, there is an urgent need to design an effective training and judgment method to learn the features of malignant samples to improve recognition accuracy. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for medical ultrasound image-assisted determination of lesion severity. This method can improve the accuracy of identification by extracting discriminative features for malignant lesion identification from a small number of malignant samples.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for determining the severity of lesions using medical ultrasound images includes the following steps:

[0009] Step 1: Preprocess the medical ultrasound images;

[0010] Step 2: Build a deep neural network model for predicting the severity of lesions. Divide the preprocessed ultrasound images into training, validation, and test sets. Use the deep neural network model to extract features from the preprocessed medical ultrasound images to obtain feature maps. Then, perform feature enhancement on the feature maps to obtain feature-enhanced maps. Use the feature-enhanced maps from the training set to train the deep neural network model and obtain the parameters of the trained convolutional neural network model. Then, use the feature-enhanced maps from the validation set to verify the parameters of the trained convolutional neural network model and use the test set to test the model. Select the parameters that meet the test accuracy requirements as the final parameters of the deep neural network model.

[0011] Step 3: Use the deep neural network model obtained in Step 2 to predict medical ultrasound images.

[0012] Furthermore, the preprocessing method in step 1 is as follows:

[0013] The medical ultrasound image is cropped, and the contrast of the cropped breast ultrasound image is enhanced. Then, the contrast-enhanced ultrasound image is cropped into a square with the original image center as the center to obtain the preprocessed ultrasound image.

[0014] Furthermore, the method for contrast enhancement processing of breast ultrasound images is as follows: the ultrasound image is sequentially processed by grayscale, binarization, dilation, restoration, image denoising, and histogram equalization.

[0015] Furthermore, the binarization conversion process in step 1 is as follows:

[0016] First, a threshold T is selected. Pixels with grayscale values ​​greater than or equal to T are set to 255, and pixels with grayscale values ​​less than T are set to 0. The processing formula is as follows:

[0017]

[0018] Where g(x,y) represents the pixel value after binarization, f(x,y) represents the pixel value before binarization, and T is the global threshold.

[0019] Furthermore, in step 1, the method for histogram equalization is as follows:

[0020] The process involves sequentially scanning each pixel of the denoised grayscale image and calculating its grayscale histogram. The cumulative distribution function of the grayscale histogram is then calculated. Based on the cumulative distribution function and histogram equalization principles, the mapping relationship between input and output is obtained. Finally, image transformation is performed based on the mapping relationship. The mapping method is as follows:

[0021]

[0022] Among them, S k This refers to the value of the current gray level after being mapped by the cumulative distribution function, where n is the sum of the pixels in the image. j L is the number of pixels at the current gray level, and L is the total number of gray levels in the image.

[0023] Through mapping, each grayscale pixel value in the ultrasound image is converted into a new grayscale value after mapping. This step improves the contrast of the image, thus obtaining an image with enhanced contrast.

[0024] Furthermore, step 2 specifically includes the following sub-steps:

[0025] S2.1: Divide the preprocessed ultrasound images from step 1 into training set, validation set, and test set.

[0026] S2.2: Construct ResNet34 as the overall architecture of the deep neural network model, and use ResNet34 as a feature extraction module to extract features from ultrasound images in the training set, validation set, and test set;

[0027] S2.3: Construct a feature enhancement module, which includes a 1×1 convolutional layer, a score generation module, and a threshold generation module. The feature map extracted by the feature extraction module in step S2.1 is input into the 1×1 convolutional layer and compressed by channel to obtain a single-channel feature map. Then, the compressed single-channel feature map is passed through the threshold generation module to generate a threshold and the score generation module to generate the position score of each pixel. Then, a mask is calculated based on the threshold and position score to divide the sample into positive or negative samples, and then the feature enhancement map is output.

[0028] S2.4: Construct a module including global max pooling and fully connected layers as a classification head, and randomly initialize the parameters. Its input is the feature map extracted in step 2.1, and its output is the category of lesion severity.

[0029] S2.5: Input the positive and negative samples from the training set into the classification head for training. Positive and negative samples are trained separately, sharing the same classification head parameters. For the same classification head, each sample incurs its own loss. The total loss during training is the sum of the losses from positive and negative samples in steps S2.2, S2.3, and S2.4. The sum is then fed back to update the parameters of the deep neural network model. After training, the model parameters are validated using a validation set, and the parameters from the round with the smallest loss are selected. Finally, the accuracy of the deep neural network model parameters is tested using a test set. Parameters meeting the accuracy requirements are selected as the final parameters of the deep neural network model for subsequent prediction.

[0030] Furthermore, the method for generating the threshold through the threshold generation module in step S2.3 is as follows:

[0031] First, the single-channel feature map obtained in step S2.2 is input into the threshold generation module. The single-channel feature map is downsampled in the threshold generation module by average pooling, and then the two-dimensional feature map is converted into a one-dimensional sequence by flattening. Then, the one-dimensional sequence is input into the fully connected layer, and finally the obtained value is processed by the sigmoid activation function to map its value to 0~1, which is used as the threshold k.

[0032] Furthermore, the method for generating the score for each pixel position using the score generation module in step S2.3 is as follows:

[0033] First, the single-channel feature map obtained in step S2.2 is input into the score generation module. In the score generation module, the single-channel feature map is downsampled by average pooling, then flattened, and then the sequence is nonlinearly transformed by two fully connected layers. After passing through two fully connected layers, the feature map is reshaped to transform the one-dimensional sequence back into a two-dimensional matrix. Then, an interpolation operation is performed to restore the feature map size to the single-channel size. Finally, the restored feature map is activated by the sigmoid activation function to map the values ​​between 0 and 1, which is used as the position score with the same size as the original feature map.

[0034] Furthermore, the method for generating positive and negative samples based on scores and thresholds in step S2.3 is as follows:

[0035] A mask is generated based on the obtained threshold k and position score. The mask is then combined with the feature map obtained in step 2.2 to generate a masked feature map. The specific formula is as follows:

[0036]

[0037] Where ω is a constant, when the score is greater than k, the mask is close to 1, and conversely, when the score is less than k, the mask is close to 0.

[0038] The feature map is denoted as image. It is combined with mask to generate positive and negative samples. Positive samples are the feature map plus the feature map multiplied by the mask, i.e., image + image × mask, and negative samples are the feature map minus the feature map multiplied by the mask, i.e., image - image × mask.

[0039] Furthermore, in step 2.5, when training the deep neural network model, the labels of positive samples are the labels of the original ultrasound images, and the loss function is the cross-entropy loss:

[0040]

[0041] In the formula, n is the number of predicted categories, which is 5 here; p(x i Let q(x) be the probability distribution of the sample labels. i ) represents the probability distribution of the prediction results;

[0042] The negative sample label is handled by minimizing the probability of predicting a positive label. The loss function used is the mean squared error loss (MSE), which calculates the mean squared error loss equal to the probability of predicting a positive sample minus zero. The implementation method is as follows:

[0043]

[0044] In the formula, Y i This indicates the probability of predicting the corresponding positive label;

[0045] The validation set and test set are generated only with positive samples after the feature enhancement module. The positive samples of the validation set are used to validate the trained deep neural network model, and the positive samples of the test set are used to test the accuracy of the deep neural network model.

[0046] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention constructs a feature extractor to extract features from ultrasound images. This feature extractor has different functions for extracting features at different depths, and can more accurately extract lesion features in ultrasound images. This overcomes the problem in existing technologies where lesions only occupy a small area in the image, making it difficult to extract malignant lesion features from ultrasound images. This invention also uses the classification head to predict and generate loss in the intermediate layers of the deep neural network model, instead of just using the deepest layer features for prediction, thus meeting the requirement of using multi-scale features. On the other hand, by using attention scores, the region with a high attention score in the image is enhanced, while the region with a low attention score is discarded, making the model's learning more effective. Thus, the model of this invention can efficiently learn malignant sample features even with only a small number of malignant samples, thereby making the prediction more accurate. Attached Figure Description

[0047] Figure 1 The original breast ultrasound image is from an embodiment of the present invention.

[0048] Figure 2 This is a preprocessed breast ultrasound image according to an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the structure of a deep neural network model according to an embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the feature enhancement module in an embodiment of the present invention;

[0051] Figure 5 This is a schematic diagram of the threshold generation module in an embodiment of the present invention;

[0052] Figure 6 This is a schematic diagram of the score generation module in an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0054] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0055] The present invention will be further described below with reference to specific embodiments, but these are not intended to limit the scope of the invention.

[0056] This invention discloses a method for determining the severity of lesions using medical ultrasound images, comprising the following steps:

[0057] Step 1: Preprocess the acquired medical ultrasound images; in this embodiment, breast ultrasound images are used for illustration, but this method is also applicable to medical ultrasound images from other departments; the specific method for Step 1 is as follows:

[0058] S1.1: Manually crop breast ultrasound images from hospitals to eliminate potential interference from boundary areas, such as... Figure 1 As shown, the black area around the effective area of ​​the breast ultrasound image is cropped;

[0059] S1.2: The cropped breast ultrasound image is subjected to contrast enhancement processing. The contrast enhancement processing method is to perform grayscale processing, binarization, dilation, image restoration, image denoising, and histogram equalization on the cropped breast ultrasound image in sequence. Specifically, the cropped image is subjected to grayscale processing. The operation steps are to perform weighted average of the RGB three-channel values ​​of each pixel position of the ultrasound image to obtain a grayscale value of 0 to 255 for each position, and the number of channels of the ultrasound image is changed from 3 to 1.

[0060] The image after grayscale processing is then binarized. The specific method is as follows: First, a threshold T is selected. Pixels with grayscale values ​​greater than or equal to T are converted to 255, and pixels with grayscale values ​​less than T are converted to 0. The processing formula is:

[0061]

[0062] Where g(x,y) represents the pixel value after binarization, f(x,y) represents the pixel value before binarization, and T is the global threshold, which is set to 100 here.

[0063] The binarized breast ultrasound images were sequentially subjected to dilation, restoration, and image denoising, followed by histogram equalization. The histogram equalization method was as follows:

[0064] The process involves sequentially scanning each pixel of the denoised grayscale image and calculating its grayscale histogram. The cumulative distribution function of the grayscale histogram is then calculated. Based on the cumulative distribution function and histogram equalization principles, the mapping relationship between input and output is obtained. Finally, image transformation is performed based on the mapping relationship. The mapping method is as follows:

[0065]

[0066] Among them, S k This refers to the value of the current gray level after being mapped by the cumulative distribution function, where n is the sum of the pixels in the image. jL is the number of pixels at the current gray level, and L is the total number of gray levels in the image.

[0067] Through mapping, each grayscale pixel value in the ultrasound image is transformed into a new grayscale value after mapping. This step improves the contrast of the image, thus obtaining an image with enhanced features.

[0068] S1.3: To avoid distortion and feature changes caused by inputting rectangular images into the neural network, the enhanced image is cropped along its shorter side to transform it into a square image, resulting in the preprocessed image, see [link to image]. Figure 2 The cutting method is as follows:

[0069] First, determine the lengths of the short and long sides of the image. Then, crop the long side using the center of the original breast ultrasound image as the center point, making the cropped square image as close as possible to the center of the original image. This yields the preprocessed breast ultrasound image, as shown below. Figure 2 As shown.

[0070] Step 2: Construct a deep neural network model for predicting lesion severity. The preprocessed ultrasound images are divided into training, validation, and test sets. The deep neural network model extracts features from the preprocessed medical ultrasound images to obtain feature maps. These feature maps are then augmented to obtain augmented feature maps. The deep neural network model is trained using the augmented feature maps from the training set to obtain the parameters of the trained convolutional neural network model. The augmented feature maps from the validation set are then used to validate the trained convolutional neural network model parameters, and the model is tested using the test set. Parameters that meet the required test accuracy are selected as the final parameters of the deep neural network model. Its structure is as follows: Figure 3 As shown;

[0071] S2.1: Divide the preprocessed ultrasound images from step 1 into a training set, a validation set, and a test set;

[0072] S2.2: Construct ResNet34 as the overall architecture of the deep neural network model, and use ResNet34 as a feature extraction module to extract features from ultrasound images in the training set, validation set, and test set;

[0073] In this step, ResNet34 is constructed as a feature extractor, and the last fully connected layer classifier is removed. Then, the weights of ResNet34 pre-trained on ImageNet (a large natural image dataset) are loaded to obtain the trained feature extraction module. In this embodiment, ultrasound images of size C0×H×W (C is the number of feature map channels, H and W are the height and width of the feature map, respectively) from the training set, validation set, and test set are input into the pre-trained feature extraction module. The sizes of these feature maps change sequentially to C1×H / 4×W / 4, C2×H / 8×W / 8, C3×H / 16×W / 16, and C4×H / 32×W / 32. The first three sizes of feature maps are then input into the subsequent step S2.3 for processing. The last layer is not enhanced and is directly input into the classification head to obtain the corresponding loss value.

[0074] S2.3: Construct the feature enhancement module. The structure of the feature enhancement module is as follows: Figure 4 As shown, the feature enhancement module includes a 1×1 convolutional layer, a threshold generation module, and a score generation module. The feature map extracted by the feature extraction module in step S2.2 is input into the 1×1 convolutional layer and compressed by channel to obtain a single-channel feature map. Then, the compressed single-channel feature map is input into the threshold generation module to generate the threshold and the score generation module to generate the position score of each pixel. Then, the mask is calculated based on the threshold and position score to divide the image sample into positive or negative samples, thereby obtaining the output feature enhancement map.

[0075] This step specifically includes the following sub-steps:

[0076] S2.3.1: Compress the feature map so that the number of channels in the feature map becomes a single channel;

[0077] The feature map extracted by the feature extraction module in step 2.3 is input into a 1×1 convolutional layer to compress the feature map by channel. The feature map size is converted from C×H×W (C is the number of feature map channels, H and W are the height and width of the feature map, respectively) to 1×H×W.

[0078] S2.3.2: Construct a threshold generation module to generate the threshold for selecting positive and negative samples in the current ultrasound image;

[0079] The compressed feature map is processed by a threshold generation module to generate a threshold, as follows: Figure 5As shown, the single-channel feature map (1×H×W) is first downsampled by a factor of 2 using average pooling, resulting in a feature map size of 1×H / 2×W / 2. Then, the two-dimensional feature map is converted into a one-dimensional sequence with a size of 1×HW / 4 using a flatten operation. Next, the one-dimensional sequence is passed through a fully connected layer with an input dimension of HW / 4 and an output dimension of 1, resulting in a final dimension of 1×1. The resulting 1×1 value is then passed through a sigmoid activation function to map its value to 0 to 1, which is used as the threshold k.

[0080] S2.3.3: Construct a score generation module to generate scores for each pixel position in the current ultrasound image;

[0081] The compressed feature map is processed by a score generation module to generate position scores, as follows: Figure 6 As shown, the single-channel feature map (1×H×W) is first downsampled by a factor of 2 using average pooling, resulting in a feature map size of 1×H / 2×W / 2. Then, a flatten operation transforms the 2D feature map into a 1D sequence with a size of 1×HW / 4. Next, two fully connected layers perform a non-linear transformation on the sequence. The first fully connected layer has an input dimension of HW / 4 and an output dimension of HW / 64; the second fully connected layer has an input dimension of HW / 64 and an output dimension of HW / 4. The first fully connected layer reduces the dimension by a factor of 16 to reduce computational cost. After two fully connected layers, the feature map size remains 1×HW / 4. Next, a reshape operation is performed on the feature map, transforming the 1D sequence back into a 2D matrix with a size of 1×H / 2×W / 2. After the reshape operation, an interpolation operation restores the feature map size to 1×H×W. Finally, the resized feature map is activated by the sigmoid function, which maps the values ​​to between 0 and 1, and serves as the position score with the same size as the original feature map.

[0082] S2.3.4: Generate a mask based on the threshold obtained in S2.3.2 and the position score obtained in S2.3.3. Combine the mask with the feature map obtained in step S2.2 to generate a masked feature map. Depending on the masking method, the masked feature map is divided into positive samples and negative samples.

[0083] The mask is generated using the threshold k and position score obtained in S2.2.2 and S2.2.3, as follows:

[0084]

[0085] Where ω is a constant, a large number, which is set to 100 in this embodiment, so that when the score is greater than k, the mask is close to 1, and conversely, when the score is less than k, the mask is close to 0.

[0086] The feature map (denoted as image) is combined with the mask to generate positive and negative samples. Positive samples are the feature map plus the feature map multiplied by the mask (image + image × mask), and negative samples are the feature map minus the feature map multiplied by the mask (image - image × mask).

[0087] S2.4: Construct a module including global max pooling and a fully connected layer as the classification head, and randomly initialize the parameters. Its input is the feature map extracted in step 2.1, and its output is the category of lesion severity. For the feature map of size C×H×W, first pass global max pooling to convert the two-dimensional image into a one-dimensional vector, and the feature map size becomes C×1×1. Then pass through the fully connected layer. The input dimension of the fully connected layer is C, and the output is 5-dimensional, so the prediction of 5 categories is obtained. These five categories are (0, 1, 2, 3), 4a, 4b, 4c, and (5, 6) in the BI-RADS classification.

[0088] S2.5: Input the positive and negative samples from the training set into the classification head for training. Positive and negative samples are trained separately, sharing the same classification head parameters. For the same classification head, each sample generates its own loss. The total loss during training is the sum of the losses from the positive and negative samples in steps S2.2, S2.3, and S2.4. The summation result is fed back to update the parameters of the deep neural network model. After training, the model parameters are validated using a validation set, and the parameters from the round with the smallest loss are selected. Finally, the accuracy of the deep neural network model parameters is tested using a test set. Parameters that meet the accuracy requirements are selected as the final parameters of the deep neural network model for subsequent prediction processes.

[0089] The positive and negative samples from the training set are input into the classification head for training, such as... Figure 4 As shown, positive and negative samples share the same classification head parameters. The label of a positive sample is the label of the ultrasound image, and the label of a negative sample is the negation of the ultrasound image label. The ultrasound images of the validation set and test set are all output as positive samples after passing through the feature enhancement module, and no negative samples are generated.

[0090] Positive and negative samples are trained separately. For the same classification head, separate losses are generated for each. The labels of positive samples are the labels of the original ultrasound images, and the loss function is the cross-entropy loss.

[0091]

[0092] In the formula, p(x)i Let q(x) be the probability distribution of the sample labels. i ) represents the probability distribution of the prediction results;

[0093] The negative sample label is handled by minimizing the probability of predicting a positive label. The loss function used is the mean squared error loss (MSE), which calculates the mean squared error loss equal to the probability of predicting a positive sample minus zero. The implementation method is as follows:

[0094]

[0095] Where, Y i This indicates the probability of predicting the corresponding positive label;

[0096] The total loss is the sum of the losses of positive and negative samples in the three stages of S2.2, S2.3 and S2.4. The summation result is fed back and the parameters of the deep neural network model are updated.

[0097] The training process uses a loss function and gradient descent to find the optimal model value. The loss function is the cross-entropy loss function, and the gradient descent method uses the Adam adaptive momentum estimation algorithm. In this implementation, the number of training epochs L is 100. The Adam optimizer is used for gradient descent calculation, with an exponentially decaying learning rate of 0.00001 initially and a batch size of 64. After training for L epochs, the parameters from the epoch with the smallest loss value are obtained in the validation set. The accuracy of the deep neural network model parameters is then tested using a test set. Parameters with satisfactory accuracy are selected as the final parameters of the deep neural network model for use in the subsequent prediction process.

[0098] S3: Use the deep neural network model obtained in step 2 to predict breast ultrasound images;

[0099] To address the problem of low diagnostic accuracy and high misdiagnosis rate of ultrasound images, which leads to delays in optimal treatment for breast cancer patients, this embodiment enables the classification of patients' BI-RADS using ultrasound images, allowing for targeted treatment based on different classifications. To illustrate the effectiveness of this embodiment, the prediction accuracy of the validation and test sets is compared with that of the ResNet34 backbone network, as shown in Table 1. Table 1 presents the comparison results between the method in this embodiment and ResNet34 on a private ultrasound dataset:

[0100] Table 1 Performance comparison between the present invention and the original resNet34

[0101]

[0102] As can be seen from Table 1, the prediction results of this embodiment are better than those of the backbone network ResNet34, which improves the accuracy of recognition, thereby providing doctors with more reference information and greatly reducing their workload.

[0103] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the protection scope of the present invention.

Claims

1. A method for determining the severity of lesions using medical ultrasound images, characterized in that, Includes the following steps: Step 1: Preprocess the medical ultrasound images; Step 2: Build a deep neural network model for predicting the severity of lesions. Divide the preprocessed ultrasound images into training, validation, and test sets. Use the deep neural network model to extract features from the preprocessed medical ultrasound images to obtain feature maps. Then, perform feature enhancement on the feature maps to obtain feature-enhanced maps. Use the feature-enhanced maps from the training set to train the deep neural network model and obtain the parameters of the trained convolutional neural network model. Then, use the feature-enhanced maps from the validation set to verify the parameters of the trained convolutional neural network model and use the test set to test the model. Select the parameters that meet the test accuracy requirements as the final parameters of the deep neural network model. Step 3: Use the deep neural network model obtained in Step 2 to predict medical ultrasound images; Step 2 specifically includes the following sub-steps: S2.1: Divide the preprocessed ultrasound images from step 1 into a training set, a validation set, and a test set; S2.2: Construct ResNet34 as the overall architecture of the deep neural network model, and use ResNet34 as a feature extraction module to extract features from ultrasound images in the training set, validation set, and test set; S2.3: Construct a feature enhancement module, which includes a 1×1 convolutional layer, a score generation module, and a threshold generation module. The feature map extracted by the feature extraction module in step S2.2 is input into the 1×1 convolutional layer and compressed by channel to obtain a single-channel feature map. Then, the compressed single-channel feature map is passed through the threshold generation module to generate a threshold and the score generation module to generate the position score of each pixel. Then, a mask is calculated based on the threshold and position score to divide the sample into positive or negative samples, and then the feature enhancement map is output. S2.4: Construct a module including global max pooling and fully connected layers as a classification head, and randomly initialize the parameters. Its input is the feature map extracted in step 2.2, and its output is the category of lesion severity. S2.5: Input the positive and negative samples from the training set into the classification head for training. Positive and negative samples are trained separately, sharing the same classification head parameters. For the same classification head, each sample incurs its own loss. The total loss during training is the sum of the losses from positive and negative samples in steps S2.2, S2.3, and S2.

4. The sum is then fed back to update the parameters of the deep neural network model. After training, the model parameters are validated using a validation set, and the parameters from the round with the smallest loss are selected. Finally, the accuracy of the deep neural network model parameters is tested using a test set. Parameters meeting the accuracy requirements are selected as the final parameters of the deep neural network model for subsequent prediction.

2. The method for determining the severity of lesions using medical ultrasound images according to claim 1, characterized in that, The preprocessing method in step 1 is as follows: The medical ultrasound image is cropped, and the contrast of the cropped breast ultrasound image is enhanced. Then, the contrast-enhanced ultrasound image is cropped into a square with the original image center as the center to obtain the preprocessed ultrasound image.

3. The method for determining the severity of lesions using medical ultrasound images according to claim 2, characterized in that, The method for contrast enhancement processing of breast ultrasound images is as follows: the ultrasound image is processed sequentially by grayscale, binarization, dilation, restoration, image denoising, and histogram equalization.

4. The method for determining the severity of lesions using medical ultrasound images according to claim 3, characterized in that, The binarization process in step 1 is as follows: First, a threshold T is selected. Pixels with grayscale values ​​greater than or equal to T are set to 255, and pixels with grayscale values ​​less than T are set to 0. The processing formula is as follows: in, This represents the pixel value after binarization. This represents the pixel value before binarization, and T is the global threshold.

5. The method for determining the severity of lesions using medical ultrasound images according to claim 3, characterized in that, In step 1, the histogram equalization process is performed as follows: The process involves sequentially scanning each pixel of the denoised grayscale image and calculating its grayscale histogram. The cumulative distribution function of the grayscale histogram is then calculated. Based on the cumulative distribution function and histogram equalization principles, the mapping relationship between input and output is obtained. Finally, image transformation is performed based on the mapping relationship. The mapping method is as follows: ； in, This refers to the value of the current gray level after being mapped by the cumulative distribution function, where n is the sum of the pixels in the image. L is the number of pixels at the current gray level, and L is the total number of gray levels in the image. Through mapping, each grayscale pixel value in the ultrasound image is converted into a new grayscale value after mapping. This step improves the contrast of the image, thus obtaining an image with enhanced contrast.

6. The method for determining the severity of lesions using medical ultrasound images according to claim 1, characterized in that, The method for generating the threshold through the threshold generation module in step S2.3 is as follows: First, the single-channel feature map obtained in step S2.2 is input into the threshold generation module. The single-channel feature map is downsampled in the threshold generation module by average pooling, and then the two-dimensional feature map is converted into a one-dimensional sequence by flattening. Then, the one-dimensional sequence is input into the fully connected layer, and finally the obtained value is processed by the sigmoid activation function to map its value to 0~1, which is used as the threshold k.

7. The method for determining the severity of lesions using medical ultrasound images according to claim 1, characterized in that, The method for generating the score for each pixel position using the score generation module in step S2.3 is as follows: First, the single-channel feature map obtained in step S2.2 is input into the score generation module. The single-channel feature map is downsampled in the score generation module by average pooling, then flattened, and then the sequence is nonlinearly transformed by two fully connected layers. After passing through two fully connected layers, the feature map is reshaped to transform the one-dimensional sequence back into a two-dimensional matrix. Then, an interpolation operation is performed to restore the feature map size to a single-channel size. Finally, the restored feature map is activated by the sigmoid function, mapping the values ​​to between 0 and 1, which serves as the position score with the same size as the original feature map.

8. The method for determining the severity of lesions using medical ultrasound images according to claim 1, characterized in that, The method for generating positive and negative samples based on scores and thresholds in step S2.3 is as follows: A mask is generated based on the obtained threshold k and position score. The mask is then combined with the feature map obtained in step 2.2 to generate a masked feature map. The specific formula is as follows: ； in, The value is a constant. When the score is greater than k, the mask is close to 1, and conversely, when the score is less than k, the mask is close to 0. The feature map is denoted as image. It is combined with mask to generate positive and negative samples. Positive samples are the feature map plus the feature map multiplied by the mask, i.e., image + image × mask. Negative samples are the feature map minus the feature map multiplied by the mask, i.e., image - image × mask.

9. The method for determining the severity of lesions using medical ultrasound images according to claim 1, characterized in that, In step 2.5, when training the deep neural network model, the labels of positive samples are the labels of the original ultrasound images, and the loss function is the cross-entropy loss. In the formula, n is the number of predicted categories; Let be the probability distribution of the sample labels. This represents the probability distribution of the predicted results; The negative sample label is handled by minimizing the probability of predicting a positive label. The loss function used is the mean squared error loss (MSE), which calculates the mean squared error loss equal to the probability of predicting a positive sample minus zero. The implementation method is as follows: In the formula, This indicates the probability of predicting the corresponding positive label; The validation set and test set are generated only with positive samples after the feature enhancement module. The positive samples of the validation set are used to validate the trained deep neural network model, and the positive samples of the test set are used to test the accuracy of the deep neural network model.

Images