Breast cancer image classification method based on low-frequency information attenuation and supervised contrast learning

Abstract

The invention provides a breast cancer image classification method based on combination of difficult sample generation based on low-frequency information attenuation and supervised contrast learning, belongs to the field of medical image classification, and is used for breast cancer x-ray image classification. The method comprises the following steps: carrying out initial segmentation on a breast cancer X-ray image to be processed by using an Otsu segmentation algorithm, removing irrelevant backgrounds and standardizing the size; suppressing low-frequency information of the segmented picture through a low-frequency suppression module; the features of the Otsu segmentation algorithm processed image and the features of the Otsu segmentation algorithm and the low-frequency suppression module processed image are extracted through a deep convolutional neural network; in the pre-training stage, learning loss is supervised and compared, samples of the same class are pulled in, and samples of different classes are pushed away; in a classifier training stage, features of the image processed by the Otsu segmentation algorithm are extracted through a deep convolutional neural network, and finally, the highest probability value is determined by means of a full connection layer to serve as a final diagnosis result. Experiments carried out on two public data sets (INbreast and CBIS-DDSM) and one private data set (In-home) show that the breast cancer X-ray image classification method has better performance compared with a previous breast cancer X-ray image classification method.

Description

Technical Field

[0001] This invention provides a breast cancer image classification method that combines low-frequency information attenuation to generate difficult samples with supervised contrastive learning, belonging to the field of medical image classification. Background Technology

[0002] Breast cancer is one of the most common cancers among women, with the highest incidence rate, seriously threatening women's health, especially in its late stages where the mortality rate is high. Early detection and aggressive treatment lead to a good prognosis and can significantly reduce mortality. Mammography, as a unique diagnostic method for breast cancer, has advantages such as high sensitivity, low cost, convenience, simplicity, and non-invasiveness. However, as the primary method for early breast cancer screening, current mammograms rely heavily on the radiologist's manual interpretation, and the accuracy of diagnosis depends heavily on the doctor's professional knowledge, making it highly subjective.

[0003] Achieving objective and accurate diagnoses is particularly challenging for inexperienced young radiologists. Furthermore, manual examination of medical images is a time-consuming and labor-intensive task. The high workload faced by physicians daily further increases the possibility of misdiagnosis. Therefore, developing a reliable computer-aided diagnostic (CAD) system to assist physicians in diagnosing breast cancer X-ray images is of significant research importance. A trustworthy CAD system has the potential to reduce the workload of radiologists, help them make more objective and accurate judgments, and improve overall work efficiency.

[0004] With advancements in computer vision, deep learning has been widely applied to the diagnosis of breast imaging due to its immense potential in feature extraction. The most widely used loss function in deep classification models is cross-entropy loss. However, recent research has shown that feature extractors trained using cross-entropy lack discriminative power.

[0005] In recent years, supervised contrastive learning has extended contrastive learning to labeled data. This method encourages normalized embeddings of the same class to move closer together while pushing embeddings of different classes apart. Similar to how radiologists enhance their diagnostic abilities by encountering “unusual” cases, Khosla et al. theoretically demonstrated that difficult samples can enhance the learning capabilities of supervised contrastive models. However, constructing difficult samples remains a pressing challenge. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a breast cancer image classification method (SupCon-LFA) that combines low-frequency information attenuation to generate a large number of difficult samples with supervised contrastive learning. This method generates a large number of difficult samples through a low-frequency suppression module and combines this with supervised contrastive learning to improve the model's ability to extract discriminative features. Extensive experiments have demonstrated that the classification performance of this invention is better than the most advanced breast cancer X-ray image classification methods previously available, resulting in more reliable diagnostic results.

[0007] A breast cancer image classification method combining low-frequency information attenuation to generate difficult samples and supervised contrastive learning, with the overall structure diagram shown below. Figure 2 As shown, it includes the following steps:

[0008] S1. The breast cancer X-ray image to be processed is initially segmented using the Otsu segmentation algorithm to remove irrelevant background and standardize the size;

[0009] S2. The cropped image obtained in S1 is processed by a low-frequency suppression module to suppress low-frequency information in the image;

[0010] S3. Extract features from the images obtained in S1 and S2 respectively using a deep convolutional neural network;

[0011] S4. In the pre-training stage, the low-dimensional representation of the features extracted in S3 is used with supervised contrastive learning loss to bring in samples of the same class and push away samples of different classes.

[0012] S5. The feature extractor obtained from S4 training is frozen, and a fully connected layer is added for the classifier training stage. The SOftmax activation function is used to calculate the probability value of benign or malignant, and the category with the higher probability value is used as the final diagnosis result.

[0013] Preferably, the irrelevant background cropping process described in S1 is as follows: the breast is located by edge detection and Otsu segmentation, and excess black background is removed to increase the effective pixel ratio in the image. Then, the image is cropped by selecting an appropriate bounding box based on the segmentation line or separator line, thereby generating the cropped image x. c .

[0014] Preferably, the low-frequency suppression step in S2 is as follows: First, the cropped image is converted to the frequency domain using a fast Fourier transform, and then a mask M∈R is introduced, where all values ​​except the central region are 1. r×r This allows us to obtain the frequency domain image after information attenuation. Finally, the spatial domain mammogram image after low-frequency attenuation is obtained through inverse Fourier transform. The calculation formula is as follows:

[0015]

[0016] Where ⊙ denotes element-wise multiplication, Fl (x i ) is x c The frequency domain image, where F is the Fourier transform. It is a spatial domain mammogram after low-frequency attenuation processing.

[0017] Preferably, the deep convolutional neural network feature extraction process described in S3 involves: using the same CNN backbone to perform deep feature extraction on the image obtained from Otus segmentation and the image obtained from Otus segmentation and low-frequency suppression processing, respectively. Using this feature extractor, the network can autonomously learn and capture abstract and high-level features in the image. The calculation formula is as follows:

[0018] o = f(x),

[0019] The feature vector o has a dimension of 2048.

[0020] Preferably, the supervised contrastive learning loss described in S4 can bring samples of the same class closer together and push samples of different classes further apart in the feature space. The calculation formula is as follows:

[0021]

[0022] in, Includes samples of the same category as anchor points. Includes all other samples in the batch except the anchor point. i ·z p Calculate the similarity between the anchor point and similar samples, and z i ·z a Calculate the similarity between the anchor point and all other samples except the anchor point. B represents the index of all samples in the batch. The dot symbol · represents the inner product.

[0023] Preferably, the classification step of the fully connected layer described in S5 is as follows: The malignancy probability of a single local feature is calculated using the fully connected layer; the malignancy probabilities of all regions are considered comprehensively; and the final malignancy probability is calculated using the following formula:

[0024] p = S(ω·o + b),

[0025] Where S represents the Softmax activation function, ω and b represent the weights and biases of the linear regression layer, and · represents matrix multiplication. Attached Figure Description

[0026] Figure 1 This is a technical flowchart of the present invention.

[0027] Figure 2 This is a general structural diagram of the SupCon-LFA of the present invention.

[0028] Figure 3This is a structural diagram of the low-frequency suppression module of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] A breast cancer image classification method combining low-frequency information attenuation to generate difficult samples with supervised contrastive learning includes the following steps:

[0031] S1. The breast cancer X-ray image to be processed is initially segmented using the Otsu segmentation algorithm to remove irrelevant background and standardize the size;

[0032] S2. The cropped image obtained in S1 is processed by a low-frequency suppression module to suppress low-frequency information in the image;

[0033] S3. Extract features from the images obtained in S1 and S2 respectively using a deep convolutional neural network;

[0034] S4. In the pre-training stage, the low-dimensional representation of the features extracted in S3 is used with supervised contrastive learning loss to bring in samples of the same class and push away samples of different classes.

[0035] S5. The feature extractor obtained from S4 training is frozen, and a fully connected layer is added for the classifier training stage. The highest probability value is used as the final diagnosis result.

[0036] The following explains the meanings of some of the English terms used in this invention: Otsu's method: a commonly used threshold segmentation algorithm; Softmax: a commonly used activation function; CNN: convolutional neural network; LFA: low-frequency suppression module; SupCon: supervised contrastive learning loss.

[0037] Dataset preparation:

[0038] The CBIS-DDSM dataset consists of 3071 breast cancer X-ray images with classification and segmentation annotations. It includes 753 cases of calcification and 891 cases of masses, covering normal, benign, and malignant cases, and provides validated pathological information. The dataset was obtained from a source database, with 80% designated for training and 20% for testing. The INbreast dataset contains 115 fully annotated, full-view digital breast cancer X-ray images, each a bilateral multi-view image, totaling 410 images. This database includes various lesion types, including masses, calcifications, asymmetry, and deformities. To facilitate cross-validation, all images were randomly split into 80% training and 20% testing sets. The In-house dataset includes 549 benign images and 532 malignant images, totaling 1081 images, sourced from the medical record database of the Cancer Hospital of Shandong First Medical University. This dataset was approved by the Ethics Review Committee of the Cancer Hospital of Shandong First Medical University. Due to the use of anonymized and retrospective data, informed consent from patients was waived. We used 868 images for training and 213 for testing to ensure there was no overlap between the two groups of patients.

[0039] Parameters and Experiments:

[0040] The experiments of this invention were run on a 4 NVIDIA V100 GPU equipped with the PyTorch 1.11.0 deep learning framework, compiled in Python 3.8. The ResNet50 architecture is based on code from the torchvision library. In the experiments, the CNN was initialized using pre-trained weights from ImageNet, and then fine-tuned with a relatively small learning rate. These pre-trained weights came from public packages of the PyTorch vision library. Training lasted 120 epochs, divided into two phases: a 60-epoch pre-training phase and a 60-epoch classifier training phase. In the pre-training phase, ResNet50 was used as the encoder, with supervised contrastive loss as the loss function. The optimizer used SGD, with an initial learning rate of 5e-3 decaying to 0.1 every 15 epochs, momentum set to 0.9, and weights decaying to 1e-4. During the classifier training phase, the pre-trained model was frozen during training. The cross-entropy loss function and SGD optimizer were used, with an initial learning rate of 0.005 decaying to 0.2 every 30 epochs. Momentum was set to 0.9, and weight decay was set to 0. To ensure fair comparison with previous methods and considering GPU resource limitations, all images were uniformly resized to 448x448. This paper quantitatively evaluates the model's classification performance on three datasets using five commonly used metrics: Area Under the Curve (AUC), Accuracy, Recall, Specificity, and Precision.

[0041] The SupCon-LFA classification network of this invention was tested and evaluated alongside various other classification networks on the CBIS-DDSM, INbreast, and In-house datasets. The final results are shown in Table 1. Compared with supervised contrastive learning methods that use random flipping and rotation for data augmentation, the proposed SupCon-LFA improved the accuracy (Acc) by 3.6%, 5.5%, and 7.1% on the INbreast, CBIS-DDSM, and In-house datasets, respectively; improved the AUC by 3.6%, 4%, and 8.3%; improved the recall by 33.4%, 3.4%, and 6.5%; changed the specificity by -10.8%, +3.5%, and +7.8%; and improved the precision by 13.4%, 2.2%, and 5.0%. These results demonstrate that LFA can effectively improve model performance by introducing frequency domain information to generate more challenging hard samples. This mechanism is essentially similar to the process by which radiologists continuously improve their diagnostic abilities only through repeated exposure to real "abnormal" cases. Compared to mainstream methods for mammography image classification, SupCon-LFA demonstrates superior performance, further validating its learning mechanism that effectively achieves intra-class aggregation and inter-class separation in the feature space by leveraging difficult samples. This enables the model to capture more discriminative deep feature representations. Notably, unlike other methods that perform exceptionally well on a few metrics while performing poorly on others, SupCon-LFA exhibits a relatively balanced performance across the five evaluation metrics. This indicates that our SupCon-LFA model possesses good performance stability and generalization ability, providing strong support for its future application in real-world clinical settings.

[0042] Table 1. Classification results of different classification methods on different datasets.

[0043]

[0044] The ablation study results for the three datasets are shown in Table 2. On the INbreast, CBIS-DDSM, and In-house datasets, compared to the original baseline network, using SupCon improved the model's accuracy by 5.4%, 2.5%, and 0.7%, respectively; AUC by 12.9%, 8.4%, and 0.6%; Recall by 0.1%, 0.8%, and 1%; Specificity by 23.4%, 6.4%, and 1.8%; and Precision by 3.6%, 7.6%, and 1.5%. Introducing LFA improved accuracy by 1.1%, 2.9%, and 3.0%, AUC by 2.2%, 8.2%, and 3.1%; Recall by 3.0%, 2.4%, and 3.6%; Specificity by 2.8%, 4.6%, and 3.4%; and Precision by 3.3%, 0.9%, and 3.2%. The results show that the introduction of both SupCon and LFA significantly improves model performance, indicating that: (1) LFA can generate difficult samples that effectively enhance the model's classification ability; (2) unlike cross-entropy loss which relies solely on labels for boundary division, SupCon enhances the model's discriminative ability in the feature extraction stage by explicitly constraining the similarity between similar features and the difference between dissimilar features. Furthermore, the SupCon-LFA model, constructed by combining SupCon and LFA, outperforms the use of either SupCon or LFA alone in all metrics, further verifying their complementary nature in improving model performance and their ability to synergistically enhance the model's ability to identify and classify key features.

[0045] Table 4 Ablation Experiment Results

[0046]

[0047] In terms of leveraging frequency domain information to generate challenging samples to improve model classification performance, models using low-frequency augmentation (LFA) consistently outperform those using high-frequency augmentation (HFA), regardless of whether cross-entropy loss or supervised contrastive loss is used. This indicates that LFA can generate samples of appropriate difficulty, which is beneficial for model learning, while HFA generates challenging samples by weakening the edges of lesion shapes, which may produce overly complex samples and hinder model learning.

[0048] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A breast cancer image classification method combining low-frequency information attenuation to generate difficult samples with supervised contrastive learning, characterized in that, The classification network includes: S1. The mammogram of breast cancer to be processed is initially segmented using the Otsu segmentation algorithm to remove irrelevant background and standardize the size; S2. The cropped image obtained in S1 is processed by a low-frequency suppression module to suppress low-frequency information in the image; S3. Extract features from the images obtained in S1 and S2 respectively using a deep convolutional neural network; S4. In the pre-training stage, the low-dimensional representation of the features extracted in S3 is used with supervised contrastive learning loss to bring in samples of the same class and push away samples of different classes. S5. The feature extractor obtained from S4 training is frozen, and a fully connected layer is added for the classifier training stage. The highest probability value is used as the final diagnosis result.

2. The breast cancer image classification method based on low-frequency information attenuation to generate difficult samples and supervised contrastive learning as described in claim 1, characterized in that, The irrelevant background cropping process described in S1 is as follows: the breast is located through edge detection and Otsu segmentation, and excess black background is removed to increase the effective pixel ratio in the image. Then, the image is cropped by selecting an appropriate bounding box based on the dividing line, thereby generating the cropped image x. c .

3. The breast cancer image classification method based on low-frequency information attenuation to generate difficult samples and supervised contrastive learning as described in claim 1, characterized in that, The low-frequency suppression step described in S2 is as follows: First, the cropped image is converted to the frequency domain using a Fast Fourier Transform. Then, a mask M∈R is introduced, where all values ​​except the central region are 1. r×r This allows us to obtain the frequency domain image after information attenuation. Finally, the spatial domain mammogram image after low-frequency attenuation is obtained through inverse Fourier transform. The calculation formula is as follows: Where ⊙ denotes element-wise multiplication, F l (x i ) is x c The frequency domain image, where F is the Fourier transform. It is a spatial domain mammogram after low-frequency attenuation processing.

4. The breast cancer image classification method based on low-frequency information attenuation to generate difficult samples and supervised contrastive learning as described in claim 1, characterized in that, The deep convolutional neural network feature extraction process described in S3 involves using the same CNN backbone to extract deep features from images obtained through Otus segmentation and images obtained through Otus segmentation and low-frequency suppression. Using this feature extractor, the network can autonomously learn and capture abstract and high-level features in the images. The calculation formula is as follows: o = f(x), The feature vector o has a dimension of 2048.

5. The breast cancer image classification method based on low-frequency information attenuation to generate difficult samples and supervised contrastive learning as described in claim 1, characterized in that, The supervised contrastive learning loss described in S4 can bring samples of the same class closer to the model in the feature space and push samples of different classes further away. The calculation formula is as follows: in, Includes samples of the same category as the anchor point. Includes all other samples in the batch except the anchor point. i ·z p Calculate the similarity between the anchor point and similar samples, and z i ·z a Calculate the similarity between the anchor point and all other samples except the anchor point. B represents the index of all samples in the batch. The dot symbol · represents the inner product.

6. The breast cancer image classification method based on low-frequency information attenuation to generate difficult samples and supervised contrastive learning as described in claim 1, characterized in that, The steps for classification using the fully connected layer described in S5 are as follows: Calculate the malignancy probability of a single local feature using the fully connected layer; comprehensively consider the malignancy probabilities of all regions; and calculate the final malignancy probability using the following formula: p = S(ω·o + b), Where S represents the Softmax activation function, ω and b represent the weights and biases of the linear regression layer, and · represents matrix multiplication.

Images