Mammogram segmentation method and system based on multi-view self-supervised deep learning

Abstract

The present application belongs to the field of medical image analysis, and provides a breast molybdenum target image segmentation method and system based on multi-view self-supervised deep learning, which comprises: acquiring a breast molybdenum target image and performing pretreatment; based on the pretreated breast molybdenum target image, a pre-trained multi-view self-supervised image segmentation network is used for breast molybdenum target image mass segmentation; wherein the process of training the multi-view self-supervised image segmentation network comprises a self-supervised pre-training stage and a downstream task transfer training stage, and the self-supervised pre-training stage specifically comprises: acquiring breast molybdenum target images of different views of the same patient's ipsilateral breast molybdenum target as self-supervised positive samples; acquiring non-ipsilateral breast molybdenum target images of the same patient and breast molybdenum target images of different patients as self-supervised negative samples; performing self-supervised pre-training based on the self-supervised positive samples and the self-supervised negative samples to obtain a pre-training model; the method has good robustness to dataset differences and downstream task differences.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image analysis, and particularly relates to a breast mammogram image segmentation method and system based on multi-view self-supervised deep learning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Breast mammography has the advantages of low dose, high sensitivity, simplicity and convenience, and has been widely used as an economical and effective method for early detection of breast cancer. Breast cancer is the most common cancer in women, and research shows that breast cancer accounts for 29% of female cancer incidence and 15% of cancer mortality. Early diagnosis of breast cancer is crucial for patient survival. In order to reduce the problem of tissue overlap and performance loss in dense breasts, standard digital mammography requires two low-dose x-ray projection views for each breast, one cranio-caudal (CC) view and one mediolateral oblique (MLO) view. Radiologists usually use all views when diagnosing breast cancer, and can cross-check the lesion site by combining CC and MLO views for ipsilateral analysis, and use the symmetry information of bilateral ipsilateral view analysis to improve the accuracy of decision-making. Mass is one of the important signs of breast cancer, and determining the mass location based on breast mammography x-ray film and further classifying breast mass (BMC) is considered an effective method for early diagnosis of breast cancer, and detailed delineation of mass contour is also of great significance for subsequent treatment.

[0004] The research of computer-aided diagnosis (CAD) based on mammography has never stopped, especially the application of deep learning technology, which has improved the overall efficiency and accuracy of breast cancer diagnosis. In order to improve the ability of CAD to identify and evaluate lesions, just like the traditional doctor's diagnosis process, the information contained in multiple views (CC and MLO) has been included in the process of developing CAD algorithms. For example, Engeland and Karssemeijer developed an algorithm to detect and evaluate lesions in two views of the same breast, and Wang et al. proposed an algorithm to detect the asymmetry of two corresponding views of the breast, which improved the performance of the CAD algorithm as expected. Since Krizhevsky et al. used convolutional neural networks (CNN) to achieve a decisive advantage in the ImageNet Large Scale Visual Recognition Challenge in 2012, CNN has been widely used in various fields of visual analysis in the past few years, including many applications on medical images such as mammography. For example, Lotter et al. developed a two-stage algorithm, which has two different multi-scale CNNs, one for masses and the other for calcifications, for scanning and analyzing image blocks. There are also works that directly classify and segment masses at the image level, which do not involve analyzing images separately to search for soft tissue lesions and calcification areas, thereby achieving good performance. However, it is worth noting that the supervised training method based on CNN requires a large amount of pixel-level or image block-level labeled training set, which requires professional doctors to finely outline the lesion area or crop the image block containing the lesion, which undoubtedly increases the difficulty of constructing the training data set, so it is meaningful to minimize the amount of training data set required by the training algorithm. An effective method to achieve this goal is transfer learning, which involves using a deep learning model that has already been trained, keeping most of the model's internal parameter values unchanged, and only fine-tuning the network's parameters for new applications, which is particularly suitable for medical small data scenarios, and has the effect of accelerating convergence and improving accuracy. As a paradigm, it has been widely used in various scenarios of visual analysis.

[0005] The pre-trained model of transfer learning can be trained by a labeled large upstream data set through supervised learning, or can be obtained by designing a semi-supervised or self-supervised learning task. The supervised pre-training such as Samala et al. successfully fine-tunes the pre-trained deep learning AlexNet model for the classification task of only 1500 breast molybdenum target lesion image blocks, considering that the original training data set of AlexNet includes the use of more than 1.2 million natural (non-medical) images, and the effect of transfer learning is very significant. And the related research constructs a self-supervised learning method (SSL) on the breast molybdenum target mass detection task, which reduces the dependence of the downstream task on labels. At the same time, the effectiveness of the self-supervised pre-training method has gradually surpassed the supervised pre-training method, and considering that self-supervision does not require human labels, it has gradually been applied in the field of medical images, providing efficient and sustainable assistance for the development of medical CAD algorithms.

[0006] The multi-view (CC and MLO) analysis of the breast molybdenum target conforms to the diagnosis standard and process of the doctor, and the current deep learning research using both views often lacks a reasonable and effective model pre-training method, which cannot fully utilize the great advantage of transfer learning on small medical data sets, and the current deep learning model based on self-supervised transfer learning on the breast molybdenum target also ignores the information distribution of the multi-view of the breast molybdenum target. Therefore, in view of the characteristics of the multi-view of the breast molybdenum target, it is particularly important to design a reasonable self-supervised pre-training method to realize transfer learning, to reduce the demand for fine annotation quantity of the downstream task, and to further improve the effectiveness of the CAD algorithm based on the breast molybdenum target of the deep learning cancer screening.

[0007] The existing deep learning algorithm applied to the breast molybdenum target fails to design a reasonable self-supervised pre-training method for the characteristics of the multi-view of the breast molybdenum target, and in the tasks such as breast mass detection segmentation and breast benign and malignant grading, the current method often uses a random or large labeled data set to supervise the pre-training of the model parameters, resulting in slow training convergence, serious dependence on upstream annotation data, and poor performance on small data sets, thereby causing the demand for fine annotation quantity in the downstream task of breast molybdenum target mass segmentation to be unable to meet the requirements, and the breast mass cannot be accurately segmented. SUMMARY

[0008] In order to solve the above problems, the present application provides a multi-view self-supervised deep learning breast molybdenum target image segmentation method and system, which can effectively reduce the demand for label data quantity of the downstream task, i.e. achieve similar algorithm performance under the condition of a small amount of label data as under the condition of the same amount of label data.

[0009] According to some embodiments, the first aspect of the present application provides a multi-view self-supervised deep learning breast mammogram image segmentation method, which adopts the following technical solutions:

[0010] The multi-view self-supervised deep learning breast mammogram image segmentation method comprises the following steps:

[0011] Obtaining a breast mammogram image and performing preprocessing;

[0012] Based on the preprocessed breast mammogram image, a pre-trained multi-view self-supervised image segmentation network is used to perform breast mammogram image mass segmentation;

[0013] The process of training the multi-view self-supervised image segmentation network comprises a self-supervised pre-training stage and a downstream task transfer training stage, and the self-supervised pre-training stage specifically comprises:

[0014] Obtaining breast mammogram images of different views of the same patient's ipsilateral breast as self-supervised positive samples;

[0015] Obtaining breast mammogram images of the same patient's non-ipsilateral breast and breast mammogram images of different patients as self-supervised negative samples;

[0016] Based on the self-supervised positive samples and the self-supervised negative samples, a self-supervised pre-training is performed to obtain a pre-training model.

[0017] Further, the step of obtaining a breast mammogram image and performing preprocessing comprises:

[0018] Obtaining breast mammogram images of different views of the patient;

[0019] The breast mammogram images of different views of the patient are normalized by using a Min-Max normalization method to obtain preprocessed breast mammogram images.

[0020] Further, the step of performing self-supervised pre-training based on the self-supervised positive samples and the self-supervised negative samples to obtain a pre-training model comprises:

[0021] Selecting a backbone network;

[0022] In the training process, the proportion of positive sample pairs and negative sample pairs in each batch is controlled to train the backbone network;

[0023] Obtaining the pre-training model.

[0024] Further, in the process of self-supervised training, when the inputs of the batch all pass through the forward process of the backbone network, an InfoNCE loss function is used, and specifically:

[0025]

[0026] where q*k is the model predicted probability, t is a temperature hyperparameter, k is the number of classes in the dataset, and i represents the i-th sample in each batch.

[0027] Further, the backbone network adopts a 50-layer residual convolutional neural network ResNet-50, which is composed of a convolutional layer, a batch normalization layer, an activation layer and a residual connection unit.

[0028] Further, the downstream task transfer training stage specifically includes:

[0029] According to the breast molybdenum target image mass segmentation requirements, a segmentation decoder and a skip layer connection are added to the pre-trained model to obtain a multi-view self-supervised image segmentation network.

[0030] The Min-Max normalization method is used to normalize the breast molybdenum target images of different views of the patient to obtain preprocessed breast molybdenum target image samples.

[0031] Based on the preprocessed breast molybdenum target image samples, the multi-view self-supervised image segmentation network is trained to obtain the multi-view self-supervised image segmentation network.

[0032] Further, the segmentation decoder is composed of a convolutional layer, an up-sampling layer, a batch normalization layer and an activation layer.

[0033] The skip layer connection cross-layer connection combines the features of the encoder backbone network and the segmentation decoder at the same level to increase the low-level semantic information.

[0034] According to some embodiments, the second aspect of the present application provides a multi-view self-supervised deep learning breast molybdenum target image segmentation system based on the following technical solutions:

[0035] The multi-view self-supervised deep learning breast molybdenum target image segmentation system comprises:

[0036] The image acquisition module is configured to acquire breast molybdenum target images and perform preprocessing;

[0037] The image recognition module is configured to perform breast molybdenum target image mass segmentation based on the preprocessed breast molybdenum target images using the pre-trained multi-view self-supervised image recognition network.

[0038] The process of training the multi-view self-supervised image segmentation network includes a self-supervised pre-training stage and a downstream task transfer training stage, and the self-supervised pre-training stage specifically includes:

[0039] The breast molybdenum target images of different views of the same patient are obtained as self-supervised positive samples.

[0040] The non-same-side breast molybdenum target image of the same patient and the breast molybdenum target image of different patients are acquired as self-supervised negative samples.

[0041] The self-supervised pre-training is performed based on the self-supervised positive samples and the self-supervised negative samples, and a pre-training model is obtained.

[0042] According to some embodiments, a third aspect of the present application provides a computer-readable storage medium.

[0043] A computer-readable storage medium, which stores a computer program, the program being executed by a processor to implement the steps of the breast molybdenum target image segmentation method based on multi-view self-supervised deep learning according to the first aspect.

[0044] According to some embodiments, a fourth aspect of the present application provides a computer device.

[0045] A computer device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the breast molybdenum target image segmentation method based on multi-view self-supervised deep learning when executing the program.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] In the process of developing the CAD algorithm for breast molybdenum target mass segmentation, the multi-view self-supervised pre-training method is adopted, which can accelerate the model convergence and improve the performance of the model on small sample data compared with the existing mass segmentation method using a large amount of labeled data training, and the pre-trained model can be quickly migrated to different medical institution data sets, and has good robustness to data set differences and downstream task differences. BRIEF DESCRIPTION OF DRAWINGS

[0048] The drawings constituting a part of the specification of the present application are used to provide a further understanding of the present application, and the illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application.

[0049] Figure 1 is the breast molybdenum target image segmentation method flowchart based on multi-view self-supervised deep learning according to the embodiments of the present application. DETAILED DESCRIPTION

[0050] The present application will be further described below in conjunction with the drawings and embodiments.

[0051] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0052] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0053] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0054] Example 1

[0055] like Figure 1 As shown, this embodiment provides a mammogram image segmentation method based on multi-view self-supervised deep learning. This embodiment uses the application of this method to a server as an example for illustration. It is understood that this method can also be applied to terminals, and can also be applied to systems including terminals, servers, and other components, and implemented through interaction between the terminal and the server. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, CDN security services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein. In this embodiment, the method includes the following steps:

[0056] Acquire mammograms and preprocess them;

[0057] Based on the preprocessed mammogram images, a pre-trained multi-view self-supervised image segmentation network was used to segment the mammogram mass.

[0058] The process of training a multi-view self-supervised image segmentation network includes a self-supervised pre-training stage and a downstream task transfer training stage. Specifically, the self-supervised pre-training stage comprises:

[0059] Mammograms of the same patient's ipsilateral breast from different angles were obtained as self-supervised positive samples.

[0060] obtaining mammography images of the same patient and mammography images of different patients as self-supervised negative samples;

[0061] performing self-supervised pre-training based on the self-supervised positive samples and the self-supervised negative samples to obtain a pre-trained model.

[0062] Specifically, the method of obtaining and preprocessing the mammography images comprises:

[0063] obtaining mammography images of the same patient and mammography images of different patients as self-supervised negative samples;

[0064] obtaining mammography images of the same patient and mammography images of different patients as self-supervised negative samples;

[0065] The whole process of the method of the embodiment is divided into two stages, i.e., a multi-view unlabeled self-supervised learning stage and a downstream task transfer learning stage.

[0066] The multi-view unlabeled self-supervised learning stage comprises the following steps.

[0067] Step 1: For the mammography images, the X-ray images of different views (CC and MLO) of the same patient are marked as and wherein i = 0, 1, 2, …, N-1, and N is the number of patients in the data set.

[0068] Step 2: All the mammography images are normalized to the interval [0, 1] by using the Min-Max normalization method, and the normalization formula is as follows:

[0069]

[0070] wherein x represents the mammography image, min represents the minimum value of the image observation, and max represents the maximum value of the image observation.

[0071] obtaining mammography images of the same patient and mammography images of different patients as self-supervised negative samples;

[0072] obtaining mammography images of the same patient and mammography images of different patients as self-supervised negative samples;

[0073] Step 3: A backbone network (such as ResNet-50) is selected, and in the training process, the proportion of positive sample pairs (i.e., the CC and MLO images of the same patient are positive sample pairs) and negative sample pairs in each batch is controlled (0.2 is used in the experiment).

[0074] The best parameters selected in the multi-group comparison experiment are because the number of positive sample pairs is far less than the number of negative sample pairs, and in order to prevent the optimization process from selecting trivial solutions, the proportion of positive sample pairs and negative sample pairs in each batch in the training process is appropriately balanced.

[0075] The process of contrastive learning is specifically that after each batch of positive and negative samples uses InfoNCE to calculate the loss, the network parameters are updated by gradient backpropagation. The proportion of control samples is because the number of positive samples is much smaller than that of negative samples. In order to prevent the optimization process from selecting trivial solutions, the proportion of positive and negative samples in each batch during the training process is appropriately balanced.

[0076] The backbone network adopts a 50-layer residual convolutional neural network ResNet-50, which is composed of a unit stacked by a convolutional layer, a batch normalization layer, an activation layer and a residual connection.

[0077] Step 4: When the input of this batch passes through the forward process of the network, the InfoNCE loss function is used to calculate as follows:

[0078]

[0079] Where q*k is the probability predicted by the model, t is a temperature hyperparameter, which is a scalar (0.9 is used in the experiment), k is the number of classes in the dataset, and i represents the i-th sample in each batch. According to the calculated loss L q Gradient backpropagation is used to update the parameters.

[0080] Step 5: During the training process, the initial learning rate is 0.01. Every 1 / 4 of the iteration number, the learning rate of the optimizer is reduced by a factor of 0.5. Finally, when the loss function no longer decreases, the training process is terminated, and the pre-trained model is obtained.

[0081] Where the downstream task transfer learning stage

[0082] The downstream task only uses the parameters of the pre-trained network in the upstream. The data in the second stage is not related to the data in the first stage, and the data in the first stage does not need to be labeled. The second stage uses specific data sets according to specific tasks (tumor segmentation).

[0083] Step 1: Use the same backbone network (such as ResNet-50) as the multi-view unlabeled self-supervised learning stage, initialize with the pre-trained model parameters, and add a segmentation decoder and cross-layer connection to the pre-trained model according to the requirements of breast molybdenum target image tumor segmentation, to obtain a multi-view self-supervised image segmentation network;

[0084] The segmentation decoder is a U-Net-like decoder structure composed of convolutional layers, up-sampling layers, batch normalization layers and activation layers, and has a characteristic cross-layer connection with the encoder backbone network. The cross-layer (skip-layer) connection combines the features of the encoder backbone network with the features of the decoder at the same level, which increases the low-level semantic information.

[0085] According to the different task types, the backbone network is adjusted, such as downstream benign and malignant classification tasks, the last linear layer of the backbone network is adjusted (random initialization); downstream breast mass segmentation tasks, add segmentation decoder part and skip layer connection (similar to the encoding decoding structure of UNet), the added part is also randomly initialized.

[0086] Step 2: The downstream task input image adopts the same normalization method as the pre-training stage.

[0087] Specifically, the Min-Max normalization method is used for the breast molybdenum target image to normalize the breast molybdenum target images of different views of the patient, and the preprocessed normalized breast molybdenum target image sample is obtained.

[0088] Based on the preprocessed normalized breast molybdenum target image sample, a multi-view self-supervised image segmentation network is trained, and the multi-view self-supervised image segmentation network is obtained.

[0089] Step 3: During the training process, different loss functions suitable for classification or segmentation are used according to the different tasks.

[0090] Step 4: During the downstream training process, a lower initial learning rate such as 0.001 is used, and the number of iterations for training is also reduced to 1 / 4 of the direct training (without pre-training parameter initialization), until the model converges.

[0091] The same side molybdenum target x-ray films of the same patient of different views (CC and MLO) are taken as paired instance data, i.e. positive samples of self-supervised learning, and the x-ray films of the same patient of different sides and different patients are taken as negative samples of self-supervised learning, a contrast learning task is designed, and the backbone network is supervised and trained using the InfoNCE loss function. The trained backbone network is used for model initialization in downstream tasks such as breast mass segmentation and breast benign and malignant classification tasks. After fine-tuning, the DICE of the mass segmentation task on the public dataset DDSM is 3.2 and 2.6 percentage points higher than that of random initialization and ImageNet supervised pre-training, respectively, and the benign and malignant classification on INBreast is 5.7 and 4.9 percentage points higher than that of random initialization and ImageNet supervised pre-training, respectively.

[0092] The backbone network pre-trained by the multi-view self-supervised method described in the embodiment is used for model initialization in downstream tasks such as breast mass segmentation and breast benign and malignant classification. After fine-tuning, the DICE of the mass segmentation task on the public dataset DDSM is 3.2 and 2.6 percentage points higher than that of random initialization and ImageNet supervised pre-training, respectively, and the benign and malignant classification on INBreast is 5.7 and 4.9 percentage points higher than that of random initialization and ImageNet supervised pre-training, respectively. At the same time, it can effectively reduce the demand for label data of downstream tasks. The results on DDSM show that only half of the label data is needed to achieve the same algorithm performance as the original amount of label data.

[0093] Embodiment two

[0094] The embodiment provides a multi-view self-supervised deep learning breast mammogram image segmentation system, comprising:

[0095] An image acquisition module configured to acquire breast mammogram images and perform preprocessing;

[0096] An image recognition module configured to perform breast mass segmentation on the breast mammogram images based on the preprocessed breast mammogram images using a pre-trained multi-view self-supervised image recognition network.

[0097] The process of training the multi-view self-supervised image segmentation network includes a self-supervised pre-training phase and a downstream task transfer training phase. The self-supervised pre-training phase is specifically:

[0098] Obtaining breast mammogram images of different views of the same patient's ipsilateral breast mammogram as self-supervised positive samples;

[0099] Obtaining breast mammogram images of the same patient's non-ipsilateral breast and breast mammogram images of different patients as self-supervised negative samples;

[0100] Self-supervised pre-training based on self-supervised positive samples and self-supervised negative samples to obtain a pre-training model.

[0101] The above modules and corresponding steps achieve the same examples and application scenarios as the above embodiment one, but are not limited to the content disclosed in the above embodiment one. It should be noted that the above modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0102] The descriptions of various embodiments in the above embodiments have different focuses, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0103] The proposed system can be implemented in other ways. For example, the system embodiments described above are only illustrative, and for example, the division of the above modules is only a logical functional division, and in actual implementation, there can be another division manner, for example, a plurality of modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0104] Embodiment three

[0105] The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize steps in a multi-view self-supervised deep learning breast mammogram segmentation method according to the above embodiment one.

[0106] Embodiment four

[0107] The embodiment provides a computer device, which comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor realizes steps in a multi-view self-supervised deep learning breast mammogram segmentation method according to the above embodiment one when executing the program.

[0108] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer usable program code.

[0109] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The function of one flow or multiple flows and / or blocks Figure 1 The function of one flow or multiple flows and / or blocks

[0110] These computer program instructions can also be stored in a computer readable storage medium capable of guiding a computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a manufactured product including instruction devices, which implement the functions specified in the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0111] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate a computer implemented process, so that the instructions executed on the computer or other programmable data processing devices provide steps for implementing the processes described in the flowcharts Figure 1 one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0112] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiment methods. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

[0113] Although the specific embodiments of the present application are described above with reference to the accompanying drawings, the description is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A multi-view self-supervised deep learning breast mammogram segmentation method, characterized in that, The method comprises the following steps: acquire mammography images and preprocess them; based on the preprocessed mammography images, use a pre-trained multi-view self-supervised image segmentation network to segment the mammography images; wherein the process of training the multi-view self-supervised image segmentation network includes a self-supervised pre-training stage and a downstream task transfer training stage, and the self-supervised pre-training stage specifically comprises: acquire mammography images of different views of the same breast of the same patient as self-supervised positive samples; acquire mammography images of non-same breast of the same patient and mammography images of different patients as self-supervised negative samples; based on the self-supervised positive samples and the self-supervised negative samples, perform self-supervised pre-training to obtain a pre-training model, which comprises: select a backbone network; during the training process, control the proportion of positive sample pairs and negative sample pairs in each batch, and train the backbone network; obtain the pre-training model; during the self-supervised pre-training process, when the input of each batch passes through the forward process of the backbone network, an InfoNCE loss function is used, specifically: where, is the model predicted probability, is a temperature hyperparameter, is the number of classes in the dataset, i denotes the ith sample in each batch; the downstream task transfer training stage specifically comprises: according to the requirements of mammography image mass segmentation, add a segmentation decoder and a cross-layer connection to the pre-training model to obtain a multi-view self-supervised image segmentation network; use a Min-Max normalization method to normalize the mammography images of different views of the patient to obtain preprocessed mammography image samples; based on the preprocessed mammography image samples, train the multi-view self-supervised image segmentation network to obtain the multi-view self-supervised image segmentation network.

2. The multi-view self-supervised deep learning-based breast molybdenum target image segmentation method of claim 1, wherein, the process of acquiring mammography images and preprocessing them comprises: acquire mammography images of different views of the patient; use a Min-Max normalization method to normalize the mammography images of different views of the patient to obtain preprocessed mammography images.

3. The multi-view self-supervised deep learning-based breast molybdenum target image segmentation method of claim 1, wherein, the backbone network uses a 50-layer residual convolutional neural network ResNet-50, which is composed of a convolutional layer, a batch normalization layer, an activation layer, and a residual connection.

4. The multi-view self-supervised deep learning-based breast molybdenum target image segmentation method of claim 1, wherein, the segmentation decoder is composed of a convolutional layer, an up-sampling layer, a batch normalization layer, and an activation layer; the cross-layer connection combines the features of the encoder backbone network and the features of the same level of the segmentation decoder together, which plays a role in increasing low-level semantic information.

5. The multi-view self-supervised deep learning breast mammogram segmentation system based on, characterized in that, The method comprises the following steps: an image acquisition module configured to acquire mammography images and preprocess them; an image recognition module configured to use a pre-trained multi-view self-supervised image recognition network to segment the mammography images based on the preprocessed mammography images; wherein the process of training the multi-view self-supervised image segmentation network includes a self-supervised pre-training stage and a downstream task transfer training stage, and the self-supervised pre-training stage specifically comprises: acquire mammography images of different views of the same breast of the same patient as self-supervised positive samples; acquire mammography images of non-same breast of the same patient and mammography images of different patients as self-supervised negative samples; based on the self-supervised positive samples and the self-supervised negative samples, perform self-supervised pre-training to obtain a pre-training model, which comprises: Select the backbone network; in the training process, control the proportion of positive sample pairs and negative sample pairs of each batch, train the backbone network; obtain a pre-training model; In the process of self-supervised pre-training, when each batch of inputs passes through the forward process of the backbone network, an InfoNCE loss function is adopted, specifically: where, is the model predicted probability, is a temperature hyperparameter, is the number of classes in the dataset, i indicates the ith sample in each batch; In the downstream task transfer training phase, specifically: According to the breast molybdenum target image mass segmentation requirements, a segmentation decoder and a cross-layer connection are added to the pre-training model to obtain a multi-view self-supervised image segmentation network; a Min-Max normalization method is used to normalize the breast molybdenum target images of different views of the patient to obtain preprocessed breast molybdenum target image samples; based on the preprocessed breast molybdenum target image samples, the multi-view self-supervised image segmentation network is trained to obtain a multi-view self-supervised image segmentation network.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the multi-view self-supervised deep learning breast molybdenum target image segmentation method according to any one of claims 1-4.

7. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the multi-view self-supervised deep learning breast molybdenum target image segmentation method according to any one of claims 1-4.

Images