Organ-specific physiological age prediction system for breast cancer diagnosis

Abstract

The invention discloses an organ-specific physiological age prediction system for breast cancer diagnosis. The system comprises an image acquisition module, an age prediction network module, a sample division and dynamic confidence module, a lightweight benign and malignant prediction module, a benign and malignant discrimination module and an output module. Through a dynamic sample screening strategy based on prediction age and registration age difference value distribution and in combination with an age prediction network integrating a detail enhancement convolution module and a pixel-channel joint attention module, the problems of label noise and sample imbalance are effectively solved; the model performance is indirectly evaluated by using a lightweight network, the network can accurately deduce the organ specific physiological age of the breast tissue through noise robust training, and finally, the predicted organ specific physiological age is fused with image features to perform benign and malignant discrimination. According to the method, the accuracy, robustness and interpretability of organ-specific physiological age prediction and benign and malignant judgment of the mammary gland molybdenum target image are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image intelligent analysis and computer-aided diagnosis, and particularly relates to an organ-specific physiological age prediction system for breast cancer diagnosis. BACKGROUND

[0002] Breast cancer is one of the major malignant tumors threatening women's health. Mammography (also known as molybdenum target photography) is an important imaging means for clinical screening and early detection. With the development of deep learning, computer-aided diagnosis (CAD) technology based on molybdenum target images has made certain progress in lesion detection and benign and malignant discrimination, but it is still limited by factors such as data quality, label reliability, and model interpretability in complex clinical scenarios.

[0003] Age is not only an important risk factor for breast cancer epidemiology, but also a key biological phenotype reflecting the degenerative changes and density structure of breast tissue. Traditional methods often treat age as a demographic covariate or simply concatenate it with image features as input into a classification model, failing to fully utilize the potential prior information of "image-characterizable organ-specific physiological age". Organ-specific physiological age is not completely consistent with registered age, and the former can be comprehensively reflected by image texture and structural patterns, which has auxiliary value and interpretive significance for tumor benign and malignant discrimination.

[0004] In actual medical data, the real age label registered or recorded does not always truly reflect the physiological state of breast tissue. Due to factors such as individual disease status, work and life intensity, hormone level, and lifestyle, there may be significant differences in the degenerative changes and density structure of breast tissue among women of the same age, leading to a deviation between registered age and the "organ-specific physiological age" characterizable by breast images. Therefore, if such "real age" is directly used as a model training label, it essentially belongs to a noisy label, which can easily make the model learn irrelevant mixed information about the breast state, reducing the accuracy and robustness of organ-specific physiological age prediction and subsequent benign and malignant discrimination.

[0005] Existing noise-robust learning methods (such as loss correction, sample selection / small loss assumption, co-teaching, semi-supervised hybrid modeling, etc.) perform well on natural image tasks, but have obvious limitations when migrating to breast molybdenum target images. On the one hand, the uneven distribution of different age samples in breast molybdenum target data is not fully considered, leading to bias in model learning of age-related features; on the other hand, there is a general difference between individual registered age and organ-specific physiological age characterizable by breast images in real life, and existing multi-task learning frameworks do not explicitly model this difference and its prior dependency, thus limiting the stability and generalization ability of the model in actual clinical scenarios.

[0006] Therefore, there is an urgent need for a technical solution that combines noise-robust learning and multi-task collaborative optimization. SUMMARY

[0007] In view of the above problems, the purpose of the present application is to provide an organ-specific physiological age prediction system for breast cancer diagnosis.

[0008] The specific technical solution is as follows:

[0009] An organ-specific physiological age prediction system for breast cancer diagnosis, comprising:

[0010] An image acquisition module: used for acquiring full-field digital mammography (FFDM) image data, and synchronously extracting registration age labels and lesion benignity labels, and performing standardization, cropping and quality control on the images to form a basic data set.

[0011] An age prediction network module: used for extracting detail and attention features from the FFDM images through a detail enhancement convolution (DEConv) module and a pixel-channel joint attention (CPCA) module, and outputting an age probability distribution through a classification head to calculate an expected age value; the module is trained in the initial stage with the registration age as a supervision signal to learn the mapping relationship between the image and the registration age, and then combined with a sample division and dynamic confidence module to screen and weight the samples, so that the model gradually changes from fitting the registration age to modeling the organ-specific physiological age of the breast tissue, and obtains a prediction model that can reflect the actual physiological state of the tissue.

[0012] A sample division and dynamic confidence module: used for calculating the confidence according to the difference between the model predicted age and the registration age, preferentially learning the features of high confidence samples, and using a semi-supervised learning strategy for low confidence samples to identify samples that truly reflect the physiological state of the breast tissue from the noisy registration age labels, so as to guide the model to learn the true relationship between the image features and the organ-specific physiological age.

[0013] A lightweight benignity prediction module: connected with the age prediction network module and the sample division and dynamic confidence module, used for predicting the organ-specific physiological age of the validation set using the current model at each training cycle epoch during the training process of the age prediction network module, and combining the predicted organ-specific physiological age value with the validation set images to train a lightweight benignity classification network, and indirectly evaluating the performance of the current organ-specific physiological age prediction model through the AUC index of the lightweight network on the validation set, so as to dynamically monitor the training process and select the optimal organ-specific physiological age prediction model.

[0014] Benign / Malignant Discrimination Module: This module integrates the original image features with the final determined predicted organ-specific physiological age value to construct a joint feature benign / malignant discrimination model, enabling more accurate, robust, and interpretable predictions of the benign / malignant nature of breast tumors.

[0015] Output module: Used to output organ-specific physiological age values ​​and benign / malignant probability indicators predicted by the model, presented in a structured or visual manner, to provide doctors with auxiliary diagnostic and risk assessment support.

[0016] Furthermore, the age prediction network module introduces the DEConv module into the model network to preserve and enhance fine-grained texture and structural information, and adds the CPCA module to enhance the spatial-channel feature coupling expression. It predicts the age category distribution of breast tissue from FFDM images and calculates the expected age value based on the probability distribution. Specifically, this includes the following steps:

[0017] S101: DEConv module adaptive differential feature enhancement multi-directional differential convolution combination:

[0018] , In the formula, For the input feature map, For the first Attention weights for each branch, For the first One convolutional kernel, This is a convolution operation.

[0019] Definition of central difference convolution kernel:

[0020] , In the formula, The weights of the c-th channel in the center-difference convolution kernel. The weight of the c-th channel in the standard convolution kernel. It is the sum of the weights at all positions of the standard convolution kernel.

[0021] S102: CPCA Module Channel-Spatial Coupled Attention Feature Extraction Phase 1: Channel Attention Mechanism , In the formula, For the input feature map, For global average pooling, For global max pooling, For element-wise multiplication, This is the Sigmoid function.

[0022] Phase 2: Multi-scale spatial feature fusion , wherein, is the multi-scale spatial fusion feature, is the channel attention output feature, is the empty convolution.

[0023] Stage 3: Spatial attention: , wherein, is the final feature map containing spatial attention, is the 1x1 convolution.

[0024] S103: Final age prediction feature fusion and classification: After the fusion of the output by S102 and the output by S101, a vector is obtained by 1x1 convolution and global average pooling and sent to the classification head, and the of each age category is output. The probability distribution of the predicted age is obtained, and the expected age value can be calculated according to the probability distribution.

[0025] , , wherein, is the probability of the th age category, is the number of categories; represents the actual age value corresponding to the th category; represents the final predicted age.

[0026] Further, the sample division and dynamic confidence module adopts a probabilistic sample screening and retraining mechanism based on age difference value distribution, specifically including the following steps: S201: calculating the difference between the model predicted age and the registration age of each case: , wherein, is the model predicted age of the th case, is the registration age of the th case.

[0027] S202: based on the preset age difference value distribution parameters, calculating the selection probability of each sample, i.e. the probability of selecting the image as the training set: , wherein, P is the probability that the i-th sample is selected into the training set, P is the probability that the i-th sample is selected into the training set, P is the probability that the i-th sample is selected into the training set, P is the probability that the i-th sample is selected into the training set, P is the probability that the i-th sample is selected into the training set, P is the probability that the i-th sample is selected into the training set, P is the probability that the i-th sample is selected into the training set, P is the probability that the i-th sample is selected into the training set, P is the probability that the i-th sample is selected into the training set,

[0028] S203: According to the calculated selection probability, the training sample is dynamically selected by using Bernoulli sampling method: for each sample , a random number is generated. If , the sample is selected into the training set of the current round. This process is performed at the beginning of each training round to achieve dynamic adjustment of sample weights.

[0029] After several rounds of dynamic sampling and training, the model gradually tends to learn the feature distribution of samples with small age difference between predicted age and registered age, i.e. high confidence samples. Since the image features of these samples can better reflect the true physiological state of breast tissue, the model gradually transforms from the initial registered age fitting model to the prediction model representing the organ-specific physiological age of the breast during the continuous iteration process, thereby realizing the natural transition from "registered age prediction" to "organ-specific physiological age prediction".

[0030] Further, the lightweight benign and malignant prediction module trains a lightweight benign and malignant prediction network by combining the validation set data with the organ-specific physiological age value predicted by the current model, indirectly evaluates the organ-specific physiological age prediction model through AUC index, and selects the optimal model, which specifically includes the following steps:

[0031] 1) Age prediction: using the model trained in the current epoch to predict the organ-specific physiological age of the validation set.

[0032] 2) Lightweight network retraining: combining the original images of the validation set and their predicted organ-specific physiological age, retraining a lightweight network.

[0033] 3) AUC collection: collecting the AUC index value of the lightweight network on the validation set.

[0034] 4) Optimal model selection: statistics the change rule of AUC value with the training process, thereby finding and selecting the model with the best performance.

[0035] The design of the module is based on the following technical assumptions: accurate organ-specific physiological age prediction can capture the true physiological state of breast tissue, and the organ-specific physiological age of breast tissue is closely related to its lesion risk, so more accurate organ-specific physiological age prediction will provide more valuable features for benign and malignant discrimination, thereby improving classification performance. Based on this assumption, by monitoring the change of AUC index of the lightweight network on the validation set, the quality of the organ-specific physiological age prediction model can be indirectly reflected: the more accurate the organ-specific physiological age prediction is, the higher the AUC of benign and malignant classification is; otherwise, the lower the AUC is. Accordingly, the organ-specific physiological age prediction model that makes the AUC of the lightweight network reach the highest is selected as the final model. The advantage of this method is that it does not require additional organ-specific physiological age true value labeling, but uses the easily obtained benign and malignant labels in the clinic to indirectly optimize the organ-specific physiological age prediction.

[0036] Further, the benign and malignant discrimination module fuses the original image features and the finally determined predicted organ-specific physiological age value to train a benign and malignant discrimination model, specifically including: S401: Age feature enhancement processing: standardizing and multi-dimensionally encoding the predicted organ-specific physiological age: , In the formula, is the predicted organ-specific physiological age value, is the standardization mean, is the standardization standard deviation.

[0037] Construct an age feature encoder: , In the formula, is the age feature after the first layer of linear transformation and activation, is a linear layer that maps the input to 256 dimensions, is a batch normalization layer, is a ReLU activation function.

[0038] , In the formula, is the finally encoded age feature, is a linear layer that maps the input to dimensions, is a Dropout layer.

[0039] S402: Multi-modal feature fusion: extracting image deep features: , In the formula, To extract the deep image features from the original image ResNet101 network.

[0040] Fusion of image features and age features: In the formula, is the fused multi-modal feature vector, is the feature splicing.

[0041] S403: benign and malignant classification prediction: output the benign and malignant probability through the fusion classifier: In the formula, is the fusion feature after dimension linear layer and activation.

[0042] In the formula, is the output benign and malignant probability.

[0043] The beneficial effects of the present application are:

[0044] The present application introduces a pixel-channel joint attention (CPCA) module and a detail enhancement convolution (DEConv) module into the organ-specific physiological age prediction network to strengthen the fine-grained texture and spatial feature expression of breast images; introduces a dynamic confidence strategy based on age difference distribution in the sample division process to realize adaptive screening and iterative optimization of high-quality samples; and uses a lightweight network to dynamically monitor the performance of the training process, thereby realizing the collaborative optimization of organ-specific physiological age prediction and benign and malignant discrimination of breast molybdenum target images under the conditions of noisy labels and uneven sample distribution. This method significantly improves the prediction accuracy, stability and explainability of the model, and has good clinical application potential. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is the relationship diagram between the modules of the present application;

[0046] Figure 2 is the organ-specific physiological age prediction module flowchart of the present application. DETAILED DESCRIPTION

[0047] As shown in Figure 1 and Figure 2 , an organ-specific physiological age prediction system for breast cancer diagnosis includes:

[0048] ​​​​Image acquisition module: Used to acquire full-field mammography (FFDM) image data, and simultaneously extract registration age labels and benign / malignant lesion labels. The images are standardized, cropped, and quality-controlled to form a basic dataset. The specific process is as follows: Full-field mammography (FFDM) images and their corresponding registration age and benign / malignant labels were acquired to establish the raw dataset. Next, the acquired image data underwent standardization processing, including pixel intensity normalization, size cropping, and alignment, to eliminate the influence of imaging equipment and individual differences. Then, grayscale normalization and quality control steps were performed to remove invalid samples with artifacts, exposure abnormalities, or missing tissue boundaries, ensuring the quality of the input data. Finally, a structured, high-quality basic dataset was obtained, denoted as ______ for each image sample. The corresponding registration age is benign or malignant label This serves as the input for subsequent sample partitioning and model training.

[0049] Age prediction network module: First, a detail enhancement convolutional module is added at the network entry point to enhance input features and improve the ability to express edge and texture information; then, a pixel-channel joint attention module is integrated into the network to enhance the expression of key features; the age of breast tissue is predicted through a combination of classification and regression, specifically including the following steps:

[0050] S101: DEConv module adaptive differential feature enhancement multi-directional differential convolution combination: , In the formula, For the input feature map, For the first Attention weights for each branch, For the first One convolutional kernel, This is a convolution operation.

[0051] Definition of central difference convolution kernel: , In the formula, The weights of the c-th channel in the center-difference convolution kernel. The weight of the c-th channel in the standard convolution kernel. It is the sum of the weights at all positions of the standard convolution kernel.

[0052] S102: CPCA Module Channel-Spatial Coupled Attention Feature Extraction Phase 1: Channel Attention Mechanism , In the formula, is the input feature map, is the global average pooling, is the global max pooling, is the element-wise multiplication, is the Sigmoid function.

[0053] Stage 2: Multi-scale spatial feature fusion: , wherein, is the multi-scale spatial fusion feature, is the channel attention output feature, is the dilated convolution.

[0054] Stage 3: Spatial attention: , wherein, is the final feature map containing spatial attention, is convolution.

[0055] S103: Final age prediction feature fusion and classification: After fusing the output by S102 with the output by S101, the vector is obtained by convolution and global average pooling and sent to the classification head, and the of each age category is output. The probability distribution of the predicted age is obtained, and the expected age value can be calculated according to the probability distribution.

[0056] , , wherein, is the probability of the predicted age category, is the number of categories; represents the actual age value corresponding to the category; represents the final predicted age.

[0057] Sample division and dynamic confidence module: used for automatically dividing high-confidence training set and noisy set according to the registered age and the organ-specific physiological age difference distribution of the breast image representation, and dynamically adjusting the confidence threshold and sample weight according to the sample number of each age stage, using semi-supervised mechanism for pseudo-label or consistency regularization learning of the noisy set, which specifically includes the following steps: S201: Calculate the model predicted age of each case and the registered age The difference: , In the formula, For the first The model predicts the age of each case. For the first The registered age of each case.

[0058] S202: Based on the preset age difference distribution parameters, calculate the selection probability of each sample, that is, the probability of selecting this image as the training set: , In the formula, For the first The probability that a sample is selected into the training set. For the first The difference between the model-predicted age and the registered age for each case. and These are the mean and standard deviation of the difference distribution, respectively, obtained in this embodiment based on statistics. , , To screen for intensity coefficients.

[0059] S203: Based on the calculated selection probability, dynamically select training samples using Bernoulli sampling: for each sample Generate random numbers .like If the sample is selected, it is included in the training set for the current round. This process is performed at the beginning of each training round to achieve dynamic adjustment of sample weights.

[0060] Through multiple rounds of dynamic sampling and training iterations, the model gradually focuses on high-confidence samples with small differences between the predicted results and the registered age, thereby weakening the influence of noise labels and enabling the network to learn the real organ-specific physiological age characteristics of breast tissue contained in the images, achieving an adaptive transformation from "registered age prediction" to "organ-specific physiological age representation".

[0061] Lightweight benign / malignant prediction module: This module combines validation set data with the organ-specific physiological age values ​​predicted by the current model to train a lightweight benign / malignant prediction network. The AUC metric is used to indirectly evaluate the performance of the organ-specific physiological age prediction model and select the optimal model. The specific steps include:

[0062] 1) Age prediction: Use the model trained at the current epoch to predict organ-specific physiological age on the validation set.

[0063] 2) Lightweight network retraining: Combine the original images of the validation set with their predicted organ-specific physiological ages to retrain a lightweight network.

[0064] 3) AUC collection: collect the AUC indicator value of the lightweight network on the validation set.

[0065] 4) Optimal model selection: the change rule of statistical AUC value with the training process is found, and the model with the best performance is selected.

[0066] The above process is based on the following principle: accurate organ-specific physiological age prediction can reflect the true physiological state of breast tissue, and organ-specific physiological age is related to lesion risk, so the level of benign and malignant classification AUC can indirectly evaluate the accuracy of organ-specific physiological age prediction.

[0067] Benign and malignant discrimination module: fuse the original image features and the finally determined predicted organ-specific physiological age value to input the joint feature benign and malignant discrimination model, realize more accurate, robust and interpretable prediction of breast tumor benign and malignant, specifically including: S401: age feature enhancement processing: standardize and multi-dimensionally encode the predicted organ-specific physiological age: , In the formula, is the predicted organ-specific physiological age value, is the standardized mean, is the standardized standard deviation.

[0068] Build an age feature encoder: , In the formula, is the age feature after the first layer linear transformation and activation, is a linear layer that maps the input to 256 dimensions, is a batch normalization layer, is a ReLU activation function.

[0069] , In the formula, is the finally encoded age feature, is a linear layer that maps the input to dimensions, is a Dropout layer.

[0070] S402: multi-modal feature fusion: extract image deep features: , In the formula, is the deep image feature extracted from the original image , ResNet101 network.

[0071] Fusion of image features and age features: , wherein, is the fused multi-modal feature vector, is feature concatenation.

[0072] S403: benign and malignant classification prediction: output the benign and malignant probability through the fusion classifier:

[0073] , wherein, is the fusion feature after linear layer and activation.

[0074] , wherein, is the output benign and malignant probability.

[0075] Output module: used for outputting the organ-specific physiological age value and the benign and malignant probability index predicted by the model, and presenting in a structured or visualized manner, to provide support for doctors in auxiliary diagnosis and risk assessment.

Claims

1. An organ-specific physiological age prediction system for breast cancer diagnosis, characterized in that, include The image acquisition module is used to acquire and preprocess full-view mammography FFDM image data to form a basic dataset containing image samples, registration age labels, and benign / malignant labels. The age prediction network module, connected to the image acquisition module, is used to extract detail and attention features from FFDM images through the detail enhancement convolutional DEConv module and the pixel-channel joint attention module, and outputs the age probability distribution through the classification head to calculate the expected age value, i.e. the model predicts the age. The sample partitioning and dynamic confidence module, connected to the age prediction network module, is used to calculate the confidence level based on the difference between the model's predicted age and the registered age. It prioritizes learning the features of high-confidence samples and adopts a semi-supervised learning strategy for low-confidence samples to guide the model to learn the true relationship between organ-specific physiological age and imaging features, rather than just fitting the registered age. The output of this module is used to dynamically adjust the training sample set of the age prediction network module. A lightweight benign / malignant prediction module, connected to the age prediction network module and the sample partitioning and dynamic confidence module, is used to predict organ-specific physiological age on the validation set using the current model in each training epoch during the training of the age prediction network module. The predicted organ-specific physiological age value is combined with the validation set images to train a lightweight benign / malignant classification network. The performance of the current organ-specific physiological age prediction model is indirectly evaluated by the AUC index of this lightweight network on the validation set. This method is based on the correlation between organ-specific physiological age and disease risk, and indirectly reflects the age prediction quality through the benign / malignant classification AUC, so as to dynamically monitor the training process and select the optimal organ-specific physiological age prediction model. The benign / malignant discrimination module is used to fuse the original image features with the finally determined predicted organ-specific physiological age value to construct a joint feature benign / malignant discrimination model; The output module is used to output the predicted organ-specific physiological age value and the probability of benign or malignant transformation.

2. The organ-specific physiological age prediction system for breast cancer diagnosis as described in claim 1, characterized in that, The age prediction network module fuses the output features of the DEConv and CPCA modules, performs global average pooling and classification head processing to obtain the probability distribution of age, and then uses the classification head to output the age probability distribution to calculate the expected age value. ， In the formula, For the first The probability of each age category Number of categories; Indicates the first The actual age value corresponding to each category; This indicates the final predicted age.

3. The organ-specific physiological age prediction system for breast cancer diagnosis as described in claim 1, characterized in that, The sample splitting and dynamic confidence module adopts a probabilistic sample selection method based on the age difference distribution, including the following steps: S201: Calculate the model-predicted age for each case. With registered age The difference: ， In the formula, For the first The model predicts the age of each case. For the first The registered age of each case; S202: Based on the preset age difference distribution parameters, calculate the selection probability of each sample, that is, the probability of selecting this image as the training set: ， In the formula, For the first The probability that a sample is selected into the training set. For the first The difference between the model-predicted age and the registered age for each case. and These are the mean and standard deviation of the difference distribution, respectively, obtained in this embodiment based on statistics. , , To screen for intensity coefficients; S203: Based on the calculated probabilities, Bernoulli sampling is used to dynamically select training samples: for each sample Generate random numbers ,like If the sample is selected, it is included in the training set for the current round. This process is performed at the beginning of each training round to achieve dynamic adjustment of sample weights. Through multiple rounds of iterative screening and retraining, the model gradually focuses on high-confidence samples with a small gap between the predicted results and the registered age, thereby weakening the interference of noisy labels and enabling the network to learn the true mapping relationship between breast images and organ-specific physiological age.

4. The organ-specific physiological age prediction system for breast cancer diagnosis as described in claim 1, characterized in that, The lightweight benign / malignant prediction module selects the optimal organ-specific physiological age prediction model through the following process: 1) Use the organ-specific physiological age prediction model in the current training phase to predict the validation set and obtain the predicted organ-specific physiological age value; 2) Combine the validation set images and the predicted organ-specific physiological age values ​​to retrain a lightweight benign / malignant prediction network; 3) Collect the AUC metric values ​​of this lightweight network on the validation set; 4) Based on the variation of AUC index during the training process, select the organ-specific physiological age prediction model with the best performance.

5. The organ-specific physiological age prediction system for breast cancer diagnosis as described in claim 1, characterized in that, The benign / malignant discrimination module is specifically used to predict organ-specific physiological age values. Standardization and multi-dimensional coding are performed to obtain age characteristics. Extracting deep image features from the original image X The age features are then stitched and fused with the image features. ， In the formula, concat represents feature concatenation, which inputs the fused features into the classifier and outputs the probability of benign or malignant features. 。

Images