A breast cancer molecular subtype prediction method based on model-driven meta-learning
By employing a model-driven meta-learning approach, utilizing a spatiotemporal recurrent attention classifier and an improved meta-learning strategy, the problem of predicting molecular subtypes of breast cancer with a limited number of medical image samples was solved, achieving high-precision prediction on small sample data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIANGNAN UNIV
- Filing Date
- 2021-12-06
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies struggle to accurately classify molecular subtypes of breast cancer with limited medical image samples, and deep learning algorithms are limited in effectiveness when medical image data is scarce, exhibiting subjectivity and overfitting issues.
A model-driven meta-learning approach is adopted. By constructing a spatiotemporal recurrent attention classifier, dynamically enhanced magnetic resonance imaging and labeled data are used, combined with an improved meta-learning strategy to optimize the classifier for predicting molecular subtypes of breast cancer.
Accurate prediction of breast cancer molecular subtypes was achieved with small sample data, avoiding the subjectivity of manual judgment, improving classification accuracy and reducing root mean square error.
Smart Images
Figure CN114187472B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of few-shot learning, and in particular to a method for predicting molecular subtypes of breast cancer based on model-driven meta-learning. Background Technology
[0002] In recent years, molecular subtypes of breast cancer have become a research hotspot because different molecular subtypes of breast cancer show significant differences in disease expression, response to treatment, prognosis, and survival outcomes. Some traditional methods for identifying breast cancer molecular subtypes have attempted to predict the molecular subtype of breast cancer through patient imaging, such as mammography, breast ultrasound, positron emission tomography (PET), and dynamic contrast-enhanced magnetic resonance imaging (MRI). However, these techniques generally rely on manual feature extraction, which is subjective and makes it difficult to objectively reflect the essential characteristics of breast cancer.
[0003] Recently, deep learning algorithms have achieved great success in medical image classification, and more research is trying to apply deep learning algorithms to the field of cancer detection and diagnosis. However, the extremely low data volume of medical images hinders deep learning algorithms from fully realizing their potential and limits the ability to evaluate results. Therefore, classifying medical images using a small number of samples is a challenging problem. Summary of the Invention
[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.
[0005] In view of the aforementioned existing problems, the present invention is proposed.
[0006] Therefore, this invention provides a model-driven meta-learning-based method for predicting molecular subtypes of breast cancer, which can avoid the subjectivity of manual judgment while solving the problem of classifying a small number of medical image samples.
[0007] To address the aforementioned technical problems, this invention provides the following technical solution: It includes obtaining dynamically enhanced magnetic resonance (MRI) images and their labels from a breast cancer database; processing the MRI images to obtain MRI volume data; matching the MRI volume data with the MRI image labels to obtain labeled MRI volume data; dividing the labeled MRI volume data into a support set and a query set; constructing a spatiotemporal recurrent attention classifier using the support set and query set; optimizing the spatiotemporal recurrent attention classifier using an improved meta-learning strategy; and predicting molecular subtypes using the spatiotemporal recurrent attention classifier.
[0008] As a preferred embodiment of the model-driven meta-learning-based molecular subtype prediction method for breast cancer described in this invention, the method further includes: the dynamically enhanced magnetic resonance image comprising three spatial dimensions and one temporal dimension; the labels of the dynamically enhanced magnetic resonance image comprising normal type, luminal epithelial type, HER-2 overexpressing type, and basal cell-like type.
[0009] As a preferred embodiment of the model-driven meta-learning-based method for predicting molecular subtypes of breast cancer described in this invention, the processing of the dynamically enhanced magnetic resonance image includes: extracting a region of interest (ROI) image of the lesion region from the dynamically enhanced magnetic resonance image according to the manually labeled area size; uniformly sampling the ROI image into images with the same pixel count and placing them into a three-dimensional matrix to obtain dynamically enhanced magnetic resonance volume data.
[0010] As a preferred embodiment of the model-driven meta-learning-based method for predicting molecular subtypes of breast cancer described in this invention, the method of dividing the labeled dynamic enhanced magnetic resonance volume data into a support set and a query set includes dividing the labeled dynamic enhanced magnetic resonance volume data into a support set and a query set using an N-way K-shot classification strategy.
[0011] As a preferred embodiment of the model-driven meta-learning-based breast cancer molecular subtype prediction method of the present invention, the spatiotemporal recurrent attention classifier includes: a recurrent neural network, an attention mechanism, a batch normalization layer, and a pooling layer; the spatiotemporal recurrent attention classifier uses a recurrent neural network with an added attention mechanism, first connecting an n×n×n convolution operation using a linear rectified activation function, then connecting a batch normalization layer to alleviate overfitting, then connecting an m×m×m max pooling operation with an added attention mechanism to connect to the next layer; finally, connecting an n×n×n convolution operation using a linear rectified activation function. Then, a batch normalization layer is connected to alleviate overfitting, followed by an m×m×m max pooling operation with an attention mechanism to connect to the next layer; next, an n×n×n convolution operation is connected using a linear rectified activation function, followed by another batch normalization layer to alleviate overfitting, and then an m×m×m max pooling operation; then a flattening operation is performed; finally, a fully connected layer is used with a normalized exponential activation function for result prediction; all convolution operations use 'a' filters; the inner loop uses gradient descent with a learning rate freely chosen by the model; the outer loop has a learning rate of 'b' and uses the Adam optimizer.
[0012] As a preferred embodiment of the model-driven meta-learning-based molecular subtype prediction method for breast cancer described in this invention, the recurrent neural network comprises three convolutional layers: one is the input to the unit from the previous layer, another is the hidden state of past and future time frames, and the last is the hidden state of the previous iteration. This represents the feature representation at layer l, time frame t, and iteration number i. This represents the feature representation at layer l-1, time frame t, and iteration number i. This represents the feature representation at layer l, time frame t, and iteration number i-1; It is the representation calculated as information propagates forward within CRAN. It is the expression calculated at time t-1 when the information propagates forward within the CRAN; It is the representation calculated when information propagates backward within CRAN. This is the representation calculated at time t+1 when the information propagates backward within the CRAN. The detailed representation of CRAN is as follows:
[0013]
[0014]
[0015]
[0016] * indicates a convolution operation. W is the linear rectification activation function. l W is the filter input to the hidden convolution. i For the hidden-to-hidden recurrent convolution that evolves during the iteration process, W t A is a time-evolving circular convolution filter, where t is time, and A l This is a deviation term.
[0017] As a preferred embodiment of the model-driven meta-learning-based breast cancer molecular subtype prediction method described in this invention, it further includes:
[0018] The steps of a spatiotemporal recurrent attention classifier in learning image morphology and pharmacokinetic feature representations are as follows: Given a dynamic MR sequence...
[0019]
[0020] Define each MRI volume V at time t t for:
[0021]
[0022] A convolutional recurrent attention network is constructed to learn morphological and pharmacokinetic characterizations, and recurrent attention is periodically used to simulate dynamic contrast-enhanced dependence, which can be represented as:
[0023] y spatio-temporal =f M (f M-1 (…(f1(x nt ))))
[0024] Where H and W are the height and width of the input image, S is the number of slices in each spatial volume, T is the time point, and V1, V2, ... V T This represents the volume data at times 1, 2, ..., T. This represents the total data with a height of H and a width of W, which has S spatial slices and T temporal data. This represents the volume data for the 1st, 2nd, ..., Sth time interval t; y spatio-temporal The prediction representing the convolutional recurrent attention, x nt denoted as a sliced downsampled image sequence, f represents a convolutional recurrent attention network, and M represents the number of iterations.
[0025] As a preferred embodiment of the model-driven meta-learning-based breast cancer molecular subtype prediction method described in this invention, the improved meta-learning strategy includes:
[0026] Perform two-step standard training on a small sample dataset, which includes inner loop update and outer loop update;
[0027] Calculate the loss of the spatiotemporal recurrent attention classifier:
[0028]
[0029] For each layer, L learning rate instances are provided for the classifier to choose from. The inner loop learning rate formula for each layer is defined as:
[0030]
[0031] Where, α n Let Φ be the learning rate of the inner loop of the nth layer of the classifier, and let Φ be the random selection function, where l1, l2, ..., l n Let L represent the learning rate for the 1st, 2nd, ..., nth layers, and let L be the total number of trainable layers in the convolutional recurrent attention network. For the loss of the i-th support set or query set in the J-th batch of tasks, x i For the i-th support set or query set data, y i For the label of the i-th support set or query set, f θ′ (x i ) is the predicted value for the support set or query set.
[0032] As a preferred embodiment of the model-driven meta-learning-based breast cancer molecular subtype prediction method of the present invention, wherein: the inner loop update and the outer loop update include,
[0033] The inner loop update: The meta-learning strategy learns a batch of tasks, defining a neural network f with meta-parameters θ. θ Randomly initialize θ = θ0, and then... j After performing a small amount of M-gradient descent on the data, θ is obtained. M Using query set Y j Evaluation Network Classification performance; j is the index of a batch of tasks, M is the number of updates in the inner loop; Supported set X j Calculate adaptive parameters using gradient descent:
[0034]
[0035] External loop update:
[0036]
[0037] in, Let the weights be the basic network parameters after M gradient descent iterations on task j. Let α be the basic network parameter weights after M-1 gradient descent iterations on task j, and α be the inner loop learning rate. For the stochastic gradient descent process of θ, For the neural network after M-1 gradient descent iterations on task j, The loss is for the support set; θ0 is the randomly initialized parameter, and β is the outer loop learning rate. For the stochastic gradient descent process of θ, The loss is for the query set.
[0038] The beneficial effects of this invention are: by using model-driven and meta-learning techniques to explore 4D spatiotemporal correlations, it is possible to accurately predict the molecular subtypes of breast cancer with a small amount of data samples. Attached Figure Description
[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:
[0040] Figure 1This is a flowchart illustrating the model-driven meta-learning-based method for predicting molecular subtypes of breast cancer as described in the first embodiment of the present invention.
[0041] Figure 2 This is a region of interest map of the model-driven meta-learning-based molecular subtype prediction method for breast cancer described in the first embodiment of the present invention.
[0042] Figure 3 This is a diagram illustrating the model-driven meta-learning architecture of the breast cancer molecular subtype prediction method based on model-driven meta-learning as described in the first embodiment of the present invention.
[0043] Figure 4 This is a diagram of the spatiotemporal recurrent attention classifier structure for the model-driven meta-learning-based molecular subtype prediction method for breast cancer as described in the first embodiment of the present invention.
[0044] Figure 5 This is a diagram of the convolutional recurrent attention network structure of the model-driven meta-learning-based breast cancer molecular subtype prediction method described in the first embodiment of the present invention.
[0045] Figure 6 This is a diagram illustrating the meta-learning strategy of the model-driven meta-learning-based molecular subtype prediction method for breast cancer as described in the first embodiment of the present invention. Detailed Implementation
[0046] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0049] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.
[0050] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0051] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0052] Example 1
[0053] Reference Figures 1-6 This is the first embodiment of the present invention, which provides a method for predicting molecular subtypes of breast cancer based on model-driven meta-learning, including:
[0054] S1: Obtain dynamic contrast-enhanced MRI images and their labels from a breast cancer database.
[0055] (1) Dynamically enhanced magnetic resonance imaging includes three spatial dimensions and one temporal dimension;
[0056] (2) The labels of dynamic contrast-enhanced magnetic resonance images include normal-like, luminal, HER-2-enriched and basal-like.
[0057] S2: Reference Figure 2 The dynamically enhanced magnetic resonance image is processed to obtain dynamically enhanced magnetic resonance volume data, and the dynamically enhanced magnetic resonance volume data is matched with the labels of the dynamically enhanced magnetic resonance image to obtain labeled dynamically enhanced magnetic resonance volume data.
[0058] (1) Based on the manually labeled area size of the lesion region, extract regions of interest (ROI) images of 1 and 1.5 times the size of the lesion region from the dynamic contrast-enhanced magnetic resonance image.
[0059] (2) In order to ensure that the number of images for each case is the same, the present invention selects the region of interest (ROI) images from the slices located at the position of the largest cross section of the tumor in the sequence and the four slices before and after it;
[0060] (3) In the 9 slices, the region of interest (ROI) images were uniformly sampled to a size of 32×32 pixels, and then placed into a 3D matrix in sequence. Finally, 147 volumetric data points of size 32×32×9 were obtained.
[0061] S3: The labeled dynamically enhanced magnetic resonance volume data is divided into a support set and a query set, and a spatiotemporal cyclic attention classifier is constructed using the support set and the query set.
[0062] (1) Construct an N-way K-shot classification task from 147 dynamic contrast-enhanced magnetic resonance (DCE-MRI) volume data;
[0063] (2) Given a support set and a query set of samples, which consist of {1,2,3,4,5} samples, each sample comes from 4 different classes, and construct classification tasks of 4-way 1-shot, 4-way 2-shot, 4-way 3-shot, 4-way 4-shot, and 4-way 5-shot;
[0064] (3)Reference Figure 4 The spatiotemporal recurrent attention classifier includes: recurrent neural network, attention mechanism, batch normalization layer and pooling layer;
[0065] Furthermore, the spatiotemporal recurrent attention classification employs a recurrent neural network (RNN) with an attention mechanism (RAN). First, it connects to a 7×7×7 convolution operation using the Rectified Linear Activation Function (ReLU), followed by a batch normalization layer to mitigate overfitting. Then, it connects to a 2×2×2 max pooling operation with an attention mechanism to connect to the next layer. Next, it connects to a 5×5×5 convolution operation using ReLU, followed by a batch normalization layer to further mitigate overfitting, and then a 2×2×2 max pooling operation. An attention mechanism is added to connect to the next layer; then a 1×1×1 convolution operation is performed, using the Rectified Linear Array (ReLU) activation function, followed by a batch normalization layer to alleviate overfitting, and then a 2×2×2 max pooling operation; next, a flattening operation is performed; finally, a fully connected layer is used, using the Softmax activation function for result prediction; all convolution operations in this invention use 64 filters; the inner loop uses a custom gradient descent process, with the learning rate freely chosen by the model; the outer loop has a learning rate of 0.001 and uses the Adam optimizer;
[0066] Among them, reference Figure 5 A recurrent neural network consists of three convolutional layers: one receives the input from the previous layer, another receives the hidden states from past and future time frames, and the last receives the hidden states from the previous iteration. This represents the feature representation at layer l, time frame t, and iteration number i. This represents the feature representation at layer l-1, time frame t, and iteration number i. This represents the feature representation at layer l, time frame t, and iteration number i-1; It is the representation calculated as information propagates forward within CRAN. It is the expression calculated at time t-1 when the information propagates forward within the CRAN; It is the representation calculated when information propagates backward within CRAN. This is the representation calculated at time t+1 when the information propagates backward within the CRAN. The detailed representation of CRAN is as follows:
[0067]
[0068]
[0069]
[0070] * indicates a convolution operation. W is the linear rectification activation function. l W is the filter input to the hidden convolution. iFor the hidden-to-hidden recurrent convolution that evolves during the iteration process, W t A is a time-evolving circular convolution filter, where t is time, and A l This is a deviation term;
[0071] Furthermore, a spatiotemporal recurrent attention classifier (STRAC) is constructed to learn morphological and pharmacokinetic feature representations from 4D images, specifically including the following steps:
[0072] The dynamic MR sequence is given as follows
[0073]
[0074] Define each MRI volume V at time t t for
[0075]
[0076] A convolutional recurrent attention network is constructed to learn morphological and pharmacokinetic characterizations, and recurrent attention is periodically used to simulate dynamic contrast-enhanced dependence, which can be represented as:
[0077] y spatio-temporal =f M (f M-1 (…(f1(x nt ))))
[0078] Where H and W are the height and width of the input image, S is the number of slices in each spatial volume, T is the time point, and V1, V2, ... V T This represents the volume data at times 1, 2, ..., T. This represents the total data with a height of H and a width of W, which has S spatial slices and T temporal data. This represents the volume data for the 1st, 2nd, ..., Sth time interval t; y spatio-temporal The prediction representing the convolutional recurrent attention, x nt denoted as a sliced downsampled image sequence, f represents a convolutional recurrent attention network, and M represents the number of iterations.
[0079] S4: Reference Figure 3 The spatiotemporal recurrent attention classifier is optimized using an improved meta-learning strategy, and molecular subtype prediction is performed through the spatiotemporal recurrent attention classifier.
[0080] (1) The improved meta-learning strategy includes performing two-step standard training on a small sample dataset, wherein the two-step standard training includes an inner loop update and an outer loop update;
[0081] (2)Reference Figure 6The inner loop update process: The meta-learning strategy learns batches of tasks: Define a neural network f with meta-parameters θ. θ Randomly initialize θ = θ0, and then... j After performing a small amount of M-gradient descent on the data, θ is obtained. M Using query set Y j Evaluation Network The classification performance. j is the index of a batch of tasks, and M is the number of updates in the inner loop. Support set X j Calculate adaptive parameters using gradient descent:
[0082]
[0083] This represents the weights of the basic network parameters after M gradient descent iterations on task j. The weights represent the basic network parameters after M-1 gradient descent iterations on task j, and α represents the inner loop learning rate. This represents the stochastic gradient descent (SGD) process for θ. This represents the neural network after M-1 gradient descent iterations on task j. This represents the loss of the support set. We assume the task batch size is j, and based on using the query set Y... j The performance of initializing θ0 is evaluated by averaging the total loss.
[0084] Outer loop update process: The process is represented as:
[0085]
[0086] θ0 represents the randomly initialized parameters, and β represents the outer loop learning rate. This represents the stochastic gradient descent (SGD) process for θ. The loss for the query set;
[0087] (3) Calculate the loss of the spatiotemporal recurrent attention classifier:
[0088]
[0089] For each layer, L learning rate instances are provided for the classifier to choose from. The inner loop learning rate formula for each layer is defined as:
[0090]
[0091] Where, α n Let Φ be the learning rate of the inner loop of the nth layer of the classifier, and let Φ be the random selection function, where l1, l2, ..., l n Let L represent the learning rate for the 1st, 2nd, ..., nth layers, and let L be the total number of trainable layers in the convolutional recurrent attention network. For the loss of the i-th support set or query set in the J-th batch of tasks, x i For the i-th support set or query set data, y i For the label of the i-th support set or query set, f θ′ (x i ) is the predicted value for the support set or query set.
[0092] This embodiment also provides a readable storage medium storing a computer program, which, when executed by a processor, is used to implement the methods provided in the various embodiments described above.
[0093] The readable storage medium can be a computer storage medium or a communication medium. The communication medium includes any medium that facilitates the transfer of computer programs from one place to another. The computer storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer. For example, the readable storage medium is coupled to a processor, thereby enabling the processor to read information from the readable storage medium and to write information to the readable storage medium.
[0094] Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application-specific integrated circuit (ASIC). In addition, the ASIC can be located in the user equipment. Alternatively, the processor and the readable storage medium can exist as discrete components in the communication equipment. The readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a CD-ROM, magnetic tape, a floppy disk, and an optical data storage device, etc.
[0095] Example 2
[0096] To verify the effectiveness of the techniques used in this method, this implementation selected the prototype network (Snell et al. 2017), the matching network (Vinyals et al. 2016), the compression-excitation network (SE-Net) (Hu et al. 2018), and the method itself for comparative testing. The experimental results were compared using scientific methods to verify the actual effectiveness of this method.
[0097] To verify the effectiveness of the techniques used in this method, this embodiment compares traditional model-driven meta-learning with the improved model-driven meta-learning method of this invention. The results are shown in the table below:
[0098] Table 1. Comparison of model-driven meta-learning accuracy.
[0099]
[0100]
[0101] Referring to Table 1, it is clear that for the same number of iterations and the same N-way K-shot, the improved model-driven meta-learning achieves greater accuracy than the unimproved model-driven meta-learning.
[0102] Traditional technical solutions require a large number of data samples, and training a high-performance neural network takes a lot of time and is not suitable for small datasets. With limited data, overfitting will occur, resulting in very poor accuracy and root mean square error.
[0103] To verify that this method can achieve better accuracy and reduce root mean square error compared to traditional methods with a small number of data samples, this embodiment will use traditional prototype networks (Snell et al. 2017), matching networks (Vinyals et al. 2016), compression-excitation networks (SE-Net) (Hu et al. 2018), and this method to conduct real-time measurement and comparison of accuracy and root mean square error.
[0104] To ensure the validity of the experimental results, the traditional prototype network (Snell et al. 2017), the matching network (Vinyals et al. 2016), the compression-excitation network (SE-Net) (Hu et al. 2018), and the proposed method were all implemented on Jupyter Notebook using the Tensorflow and Keras frameworks, and trained and tested using two 11GB Nvidia RTX 2080Ti GPUs. For 147 dynamic NMR volumetric images, 88 (approximately 60%) were used as training images and 59 (approximately 40%) as test images. The experimental results were obtained after 120 training cycles, as shown in the table below:
[0105] Table 2 compares the classification performance of this method with other methods.
[0106]
[0107] Referring to Table 2, it can be seen that the accuracy of this method is significantly higher than that of the other three methods, and the root mean square error is significantly lower than that of the other three methods, thus achieving better identification results.
[0108] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting molecular subtypes of breast cancer based on model-driven meta-learning, characterized in that: include, Dynamic contrast-enhanced MRI images and their labels were obtained from a breast cancer database. The dynamically enhanced magnetic resonance image is processed to obtain dynamically enhanced magnetic resonance volume data, and the dynamically enhanced magnetic resonance volume data is matched with the label of the dynamically enhanced magnetic resonance image to obtain labeled dynamically enhanced magnetic resonance volume data. The labeled dynamically enhanced magnetic resonance volume data is divided into a support set and a query set, and a spatiotemporal cyclic attention classifier is constructed using the support set and the query set. The spatiotemporal recurrent attention classifier is optimized using an improved meta-learning strategy, and molecular subtype prediction is performed through the spatiotemporal recurrent attention classifier; The labeled dynamically enhanced magnetic resonance volume data is divided into a support set and a query set, including: The labeled dynamic enhanced magnetic resonance volume data are divided into support set and query set using an N-way K-shot classification strategy; The spatiotemporal recurrent attention classifier includes: a recurrent neural network, an attention mechanism, a batch normalization layer, and a pooling layer; The spatiotemporal recurrent attention classifier uses a recurrent neural network with an added attention mechanism, first connecting... The convolution operation uses a linear rectified activation function, followed by a batch normalization layer to mitigate overfitting, and then connects... Max pooling is performed and an attention mechanism is added to connect to the next layer; then the connection is made. The convolution operation uses a linear rectified activation function, followed by a batch normalization layer to mitigate overfitting, and then connects... Max pooling is performed and an attention mechanism is added to connect to the next layer; then the connection continues. The convolution operation uses a linear rectified activation function, followed by a batch normalization layer to mitigate overfitting, and then connects... The process involves max pooling, followed by flattening, and finally a fully connected layer with a normalized exponential activation function for result prediction. Each convolution operation uses *a* filters. The inner loop employs gradient descent with a learning rate chosen freely by the model. The outer loop has a learning rate of *b* and uses the Adam optimizer. The improved meta-learning strategy includes, Perform two-step standard training on a small sample dataset, which includes inner loop update and outer loop update; Calculate the loss of the spatiotemporal recurrent attention classifier: For each layer, set There are 10 learning rate instances available for the classifier to choose from; the inner loop learning rate formula for each layer is defined as: in, For the classifier The learning rate of the inner loop of the layer. For random selection function, Indicates the first learning rate, This represents the total number of trainable layers in a convolutional recurrent attention network. In the first The first batch of tasks The loss of the support set or query set. For the first Groups support sets or query sets of data. For the first Supports set or query set tags. To provide predicted values for the support set or query set.
2. The breast cancer molecular subtype prediction method based on model-driven meta-learning as described in claim 1, characterized in that: It also includes, The dynamically enhanced magnetic resonance image includes three spatial dimensions and one temporal dimension; The labels for the dynamically enhanced magnetic resonance images include normal type, luminal epithelial type, HER-2 overexpressing type, and basal cell-like type.
3. The breast cancer molecular subtype prediction method based on model-driven meta-learning as described in claim 2, characterized in that: Processing the dynamically enhanced magnetic resonance image includes... Based on the manually labeled area size of the lesion region, a region of interest image of the lesion region is cropped from the dynamic contrast-enhanced magnetic resonance image; The region of interest image is uniformly sampled into images with the same pixel count and placed into a three-dimensional matrix to obtain dynamically enhanced magnetic resonance volume data.
4. The breast cancer molecular subtype prediction method based on model-driven meta-learning as described in claim 3, characterized in that: The recurrent neural network consists of three convolutional layers: one receives the input from the previous layer, another receives the hidden states from past and future time frames, and the last receives the hidden states from the previous iteration. Indicates the first Layers, time frames and number of iterations Feature representation, Indicates the first Layers, time frames and number of iterations Feature representation, Indicates the first Layers, time frames and number of iterations Feature representation; It is the representation calculated as information propagates forward within CRAN. Is The representation computed at any given moment as information propagates forward within the CRAN; It is the representation calculated when information propagates backward within CRAN. Is The representation computed at any given moment as information propagates backward within CRAN; the detailed representation in CRAN is: , , For convolution operations, For linear rectification activation function, The filter that is input into the hidden convolution. For the hidden-to-hidden recurrent convolution that evolves during the iteration process, A time-evolving circular convolution filter, For a moment, This is a deviation term.
5. The breast cancer molecular subtype prediction method based on model-driven meta-learning as described in claim 4, characterized in that: It also includes, The steps of a spatiotemporal recurrent attention classifier in learning image morphology and pharmacokinetic feature representations are as follows: Given a dynamic MR sequence... Define time Each MRI volume for: A convolutional recurrent attention network is constructed to learn morphological and pharmacokinetic characterizations, and recurrent attention is periodically used to simulate dynamic contrast-enhanced dependence, which can be represented as: in, and Given the height and width of the input image, The number of slices in each spatial volume. For a point in time, Indicates in Time-of-flight volume data, Indicates having in space Number of slices and have The height of each time series is Width is Total data; Indicates in specific Time of the first Volume data The prediction representing the convolutional recurrent attention. This represents a sequence of downsampled images with slices. This represents a convolutional recurrent attention network. Indicates the number of iterations.
6. The breast cancer molecular subtype prediction method based on model-driven meta-learning as described in claim 5, characterized in that: The inner loop update and the outer loop update include, The inner loop update: The meta-learning strategy learns a batch of tasks, defining a meta-parameter... Neural Networks Random initialization In response to the support set A small amount of data After gradient descent, we obtain Using query sets Evaluation Network Classification performance; It is an index of a batch of tasks. This refers to the number of times the inner loop updates; it supports sets. Calculate adaptive parameters using gradient descent: External loop update: in, In the mission Up past Basic network parameter weights after subgradient descent In the mission Up past Basic network parameter weights after subgradient descent For the internal loop learning rate, for The stochastic gradient descent process, In the mission Up past Neural network after subgradient descent To support the loss of the set; To initialize parameters randomly, For the outer loop learning rate, for The stochastic gradient descent process, The loss is for the query set.