System for prognosis risk prediction of colorectal cancer

Abstract

The invention discloses a system for prognostic risk prediction of colorectal cancer, and relates to the technical field of medical artificial intelligence, and the system is technically characterized by comprising a data acquisition and preprocessing module, a multi-modal feature extraction module, a multi-modal feature fusion module, a prognostic risk prediction module and a visualization and interpretation module. The system extracts key features of a patient by collecting pathological images, genome data and clinical information of the patient, deep fusion of multi-modal data is realized by using an attention mechanism, and a comprehensive feature vector is generated. Based on a deep learning model, the system efficiently predicts the prognosis risk of a patient, and the contribution of key features to a prediction result is displayed by adopting explanatory technologies such as SHAP and the like. The system has the characteristics of multi-modal data integration, efficient prediction and transparent result, and can provide intuitive risk stratification results and personalized treatment suggestions for doctors. Through multi-center data verification, the system is excellent in accuracy, generalization ability and clinical applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of medical artificial intelligence, and particularly to a system for predicting the prognosis risk of colorectal cancer. Background Art

[0002] Colorectal Cancer (CRC) is one of the most common malignant tumors globally, with significant geographical and age-related characteristics. According to the data of the International Agency for Research on Cancer (IARC) in 2020, CRC is the third most common cancer globally and the second leading cause of cancer death, with approximately 1.9 million new cases and 930,000 deaths each year. With the aging of the population and changes in lifestyle, the incidence of CRC is on the rise in many countries.

[0003] Currently, the treatment of CRC mainly relies on early screening, surgical resection, and adjuvant chemotherapy and radiotherapy. However, the clinical prognosis of patients varies greatly, and even among patients at the same stage, there are significant individual differences in survival time and recurrence rate. This difference is mainly affected by the following factors:

[0004] Heterogeneity of tumors: CRC exhibits complex molecular characteristics at the genomic level. The presence or absence of gene mutations such as KRAS and TP53 has an important impact on patient prognosis and treatment response.

[0005] Complexity of environmental and clinical characteristics: Factors such as patient age, gender, lesion size, and TNM stage all have varying degrees of influence on prognosis.

[0006] Diversity of treatment regimens: The choice of postoperative chemotherapy, targeted therapy, or immunotherapy needs to be based on the specific risk level of the patient.

[0007] Therefore, how to accurately evaluate the prognosis risk of CRC patients and provide a scientific basis for personalized treatment is an important challenge in current clinical treatment.

[0008] In clinical practice, the prognosis assessment of CRC patients usually depends on the following aspects:

[0009] TNM staging system: TNM staging is graded based on the pathological characteristics of three aspects: tumor, node, and metastasis. Although TNM staging is a standardized prognosis assessment tool, it is difficult to capture the molecular and biological differences between patients based solely on tumor invasion depth and metastasis, resulting in limited prediction accuracy.

[0010] Single molecular marker: In recent years, gene mutations such as KRAS, APC, TP53, and the MSI (microsatellite instability) status have been widely used to assist in predicting the treatment response and prognosis risk of patients. However, the predictive ability of these single markers is limited and cannot comprehensively reflect the disease characteristics of patients.

[0011] Imaging and pathological evaluation: Imaging examinations (CT, MRI) and pathological section analysis are important means for judging tumor invasiveness. However, these methods rely on artificial experience, are highly subjective, and are difficult to quantify in complex data patterns.

[0012] Generally speaking, traditional prognostic evaluation methods are mostly based on single-modal data, ignoring the potential complementary relationships between multi-modal data. This limitation may lead to the missed diagnosis of high-risk patients or the over-treatment of low-risk patients, which is not conducive to the optimization of personalized treatment plans.

[0013] Therefore, we urgently need to design a system for predicting the prognosis risk of colorectal cancer to solve the above problems. Summary of the Invention

[0014] The purpose of the present invention is to provide a system for predicting the prognosis risk of colorectal cancer in view of the deficiencies of the prior art to solve the problems raised in the background art.

[0015] To achieve the above object, the present invention provides the following technical solution: A system for predicting the prognosis risk of colorectal cancer, comprising:

[0016] Data acquisition and preprocessing module: Used to collect multi-modal data of colorectal cancer patients, including pathological images, genomic data, and clinical data, and perform standardized processing on the collected multi-modal data to clean outliers;

[0017] Multi-modal feature extraction module: Used to extract features from pathological images, genomic data, and clinical data respectively;

[0018] Multi-modal feature fusion module: Used to deeply fuse the feature vectors extracted from each modal data to generate a comprehensive feature vector;

[0019] Prognosis risk prediction module: Used to calculate the prognosis risk of patients based on the comprehensive feature vector and output the prediction result;

[0020] Visualization and interpretation module: Used to intuitively display the prognosis risk prediction result and provide feature contribution analysis through an interpretive algorithm;

[0021] The prediction process of the prognosis risk prediction module includes:

[0022] R = f(F image ,F gene,F clinical )

[0023] Among them, R is the patient's prognostic risk prediction value, F image 、F gene 、F clinical are the fusion features of pathological images, genomic data and clinical data, respectively, and f is the risk prediction function based on deep learning.

[0024] As a preferred technical solution of the present invention, the data acquisition and preprocessing module includes:

[0025] Pathological image processing unit: used to enhance, cut and standardize pathological images to remove noise and artifacts;

[0026] Genomic data encoding unit: used to process genomic data into a specific numerical vector form;

[0027] Clinical data cleaning unit: used to process the patient's basic information (such as age, gender, stage) and disease course data (such as lesion size, treatment history), and fill in missing values ​​or remove outliers.

[0028] As a preferred technical solution of the present invention, the multimodal feature extraction module extracts features in the following manner: The pathological image feature extraction uses a convolutional neural network (CNN) to obtain image features F through multi-layer convolution, pooling and feature mapping. image ; Genomic data feature extraction Through the fully connected network, high-dimensional genetic data is reduced in dimension to generate genetic features F gene ; Clinical data feature extraction is normalized and generates clinical features F through another fully connected network clinicalo

[0029] As a preferred technical solution of the present invention, the calculation process of pathological image feature extraction includes:

[0030] F image =CNN(X image )

[0031] Among them, X image is the input pathological image, CNN(·) represents the feature extraction function of the convolutional neural network, and its network structure includes multiple convolutional layers, ReLU activation function and maximum pooling layer.

[0032] As a preferred technical solution of the present invention, the calculation formula for extracting features of genomic data and clinical data is as follows:

[0033] F gene =σ(W g X gene +b g ),F clinical =σ(Wc X clinical +b c )

[0034] Among them, X gene and X clinical respectively represent the input matrices of genomic data and clinical data, W g , W c are the learned weight matrices respectively, b g , b c are the bias vectors respectively, and σ(·) represents the activation function.

[0035] As a preferred technical solution of the present invention, the multimodal feature fusion module realizes weighted fusion of multimodal features through an attention mechanism, and its formula is as follows:

[0036] F = Attention(F image , F gene , F clinical )

[0037] Among them, F is the fused feature vector, and Attention(·) is a fusion method based on the attention mechanism, which is used to dynamically allocate the weights of each modality data.

[0038] As a preferred technical solution of the present invention, the calculation method of the attention mechanism is:

[0039]

[0040] Among them, F i represents the i-th modality feature vector, α i represents the attention weight of the i-th modality, W i represents the learned weight of the i-th modality, The transpose of the weight vector W i means to perform a dot product calculation on W i and the feature vector F i .

[0041] As a preferred technical solution of the present invention, the output of the prognosis risk prediction module adopts a classification method based on the softmax function, and its calculation formula is:

[0042] R = softmax(WF + b)

[0043] Among them, F is the fused feature vector, W is the weight matrix of the classifier, and b is the bias term.

[0044] As a preferred technical solution of the present invention, the prediction model is trained by minimizing the following cross-entropy loss function:

[0045]

[0046] Among them, N is the number of samples, and y i is the true label of sample i, and R i is the predicted result of sample i.

[0047] As a preferred technical solution of the present invention, the visualization and interpretation module calculates the contribution degree of each feature in the prognosis risk prediction through the SHAP (SHapley Additive exPlanations) method, and its calculation formula is:

[0048]

[0049] Among them, φ i represents the Shapley value of the i-th feature, f(S) represents the contribution of the feature subset S to the model output, and |N| represents the total number of features.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows: The colorectal cancer prognosis risk prediction system of the present invention overcomes the limitations of traditional single-modal data evaluation by integrating multi-modal information such as pathological images, genomic data, and clinical data. Through multi-modal feature extraction and deep fusion, the system can comprehensively capture the complex features of patients, improve the accuracy of prognosis risk prediction, and is particularly more applicable in high-heterogeneity patient populations.

[0051] The system adopts a model structure based on the attention mechanism and SHAP interpretability technology, which can clearly show the contribution degree of each modal data and feature to the prediction result. This highly transparent prediction process significantly enhances the credibility of the model, enables doctors to understand and trust the prediction result of the system, and provides a guarantee for clinical applications.

[0052] The present invention designs an efficient system architecture that can quickly process patient data and output risk stratification and personalized treatment recommendations in real time. The system has good generalization ability and a user-friendly interface, can be directly deployed in clinical practice, significantly improves the decision-making efficiency of doctors, optimizes the allocation of medical resources, and provides personalized treatment plans for colorectal cancer patients, thereby improving the survival rate and quality of life of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a system diagram of a system for colorectal cancer prognosis risk prediction proposed by the present invention;

[0054] Figure 2 is a composition diagram of the data acquisition and preprocessing module of a system for colorectal cancer prognosis risk prediction proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0056] The following combines the attached Figure 1 and Figure 2 and multiple embodiments to describe the specific implementation manners of the present invention in detail.

[0057] Embodiment 1: Data collection and preprocessing;

[0058] Pathological image data processing: Input the pathological image X image , with the format of high-resolution HE-stained sections. After gray level equalization and noise reduction processing, the image is scaled to the standard size of 512×512 and converted into an RGB three-channel input.

[0059] The following processing formula is used to standardize each image:

[0060]

[0061] where μ is the mean value of image pixels and σ is the pixel standard deviation.

[0062] Genomic data processing: Input the genomic data X gene , including common mutant genes (such as KRAS, TP53) and expression levels, etc.

[0063] The high-dimensional gene data is processed by dimensionality reduction, and the following feature encoding formula is adopted:

[0064] X′ gene = log(1 + X gene )

[0065] Here, logarithmic transformation is used to reduce the skewness of the data distribution, and X gene represents the original expression value of the gene data.

[0066] Clinical data processing:

[0067] Input the clinical data X clinical , including continuous variables (such as CEA level) and discrete variables (such as TNM stage).

[0068] The following standardization formula is used to process the continuous variables:

[0069]

[0070] Perform one-hot encoding on discrete variables.

[0071] Embodiment 2: Multimodal feature extraction module;

[0072] Pathological image feature extraction: Use a convolutional neural network (CNN) to extract image features. The network includes multiple convolutional layers, ReLU activation functions, and max pooling layers.

[0073] The feature extraction formula is:

[0074] F image = CNN(X image )

[0075] where CNN(·) represents the computational function of the convolutional neural network, the input is the preprocessed pathological image X image , and the output is a feature vector F of length 256 image ;

[0076] Genomic feature extraction: Use a fully connected network to process gene data and reduce the dimension to generate gene feature F gene The feature extraction formula is:

[0077] F gene = σ(W g X′ gene + b g )

[0078] where W g is the weight matrix, b g is the bias term, and σ(·) is the ReLU activation function.

[0079] Clinical feature extraction: Use another fully connected network to extract features from clinical data.

[0080] The formula is:

[0081] F clinical = σ(W c X clinical ′ + b c )

[0082] The output feature vector F clinical is of length 64.

[0083] Embodiment 3: Multimodal feature fusion module;

[0084] Use a weighted fusion method based on the attention mechanism to deeply fuse the feature vectors F image , F gene , F clinical of each modality to generate a comprehensive feature vector F.

[0085] Attention mechanism calculation formula: First, calculate the attention weights for each modality: Fusion method:

[0086]

[0087] Among them, F i represents the feature vector of the i-th modality (such as F image , F gene , F clinical ), W i is the weight vector, and α i is the attention weight of modality i.

[0088] Perform weighted fusion according to the attention weights:

[0089]

[0090] The dimension of the output comprehensive feature vector F is 512.

[0091] Embodiment 4: Prognostic risk prediction module;

[0092] Risk prediction model: Input the fused feature vector F, and predict the prognostic risk distribution through a fully connected layer and the softmax function.

[0093] The formula is:

[0094] R = softmax(WF + b)

[0095] Among them, W is the classification weight matrix, b is the bias vector, R is the prediction result, and the output is the probability distribution of three categories (high risk, medium risk, low risk).

[0096] Model training: The training objective is to minimize the following cross-entropy loss function:

[0097]

[0098] Among them, N is the number of samples, y i is the true label, and R i is the predicted value.

[0099] Embodiment 5: Visualization and interpretation module;

[0100] Visualization of risk prediction results: The prediction results generated by the system include the risk stratification (high risk / medium risk / low risk) of the patient and the corresponding probability values, which are displayed on the user interface.

[0101] An example output is: "Prediction result: High risk (85%), Medium risk (10%), Low risk (5%)".

[0102] Explanation of feature contribution: The SHAP algorithm is used to calculate the contribution of each modality data, and the formula is:

[0103]

[0104] where φ i represents the contribution value of the i-th feature, and f(S) represents the contribution of the feature subset S to the model output. The feature contribution is displayed in the form of a bar chart or a heat map.

[0105] For example, the pathological image features contribute 70%, the gene data contributes 20%, and the clinical data contributes 10%.

[0106] Embodiment 6: System deployment;

[0107] Deployment method: The system is deployed in the high-performance server of the hospital in the form of Docker containers, supporting multi-user concurrent access. A friendly interface is designed for the front end for doctors to input patient data and view the risk prediction results.

[0108] Workflow:

[0109] Doctors upload the patient's pathological images, gene data, and clinical information.

[0110] The system automatically completes data preprocessing, feature extraction, risk prediction, and visual output.

[0111] Doctors formulate treatment plans based on the prediction results combined with clinical experience.

[0112] The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A system for predicting the prognostic risk of colorectal cancer, characterized in that, include: Data acquisition and preprocessing module: used to collect multimodal data of colorectal cancer patients, including pathological images, genomic data and clinical data, and to standardize the collected multimodal data and clean up abnormal values; Multimodal feature extraction module: used to extract features from pathological images, genomic data and clinical data respectively; Multimodal feature fusion module: used to deeply fuse the feature vectors extracted from each modality data to generate a comprehensive feature vector; Prognostic risk prediction module: used to calculate the patient's prognostic risk based on the comprehensive feature vector and output the prediction result; Visualization and interpretation module: used to intuitively display the prognostic risk prediction results and provide feature contribution analysis through explanatory algorithms; The prediction process of the prognostic risk prediction module includes: R = f(F image , F gene , F clinical ) Among them, R is the prognostic risk prediction value of the patient, and F image , F gene , F clinical are the fusion features of pathological images, genomic data, and clinical data respectively, and f is a risk prediction function based on deep learning.

2. The system for predicting the prognosis risk of colorectal cancer according to claim 1, wherein, The data acquisition and preprocessing module includes: Pathological image processing unit: used to enhance, cut and standardize pathological images to remove noise and artifacts; Genomic data encoding unit: used to process genomic data into a specific numerical vector form; Clinical data cleaning unit: used to process the patient's basic information and medical history data, and fill in missing values ​​or remove abnormal values.

3. The system for predicting the prognosis risk of colorectal cancer according to claim 1, wherein The multi-modal feature extraction module extracts features in the following way. For pathological image feature extraction, a convolutional neural network is used to obtain image feature F through multiple layers of convolution, pooling, and feature mapping. image ; Genomic data feature extraction is performed through a fully connected network to reduce the dimensionality of high-dimensional gene data and generate gene feature F gene ; Clinical data feature extraction is performed through another fully connected network, which normalizes and generates clinical feature F clinicalo .

4. The system for predicting the prognosis risk of colorectal cancer according to claim 3, wherein The calculation process of the pathological image feature extraction includes: F image = CNN(X image ) Among them, X image is the input pathological image, and CNN(·) represents the feature extraction function of the convolutional neural network, whose network structure includes multiple convolutional layers, ReLU activation functions, and max pooling layers.

5. The system for predicting the prognosis risk of colorectal cancer according to claim 3, wherein The calculation formula for extracting the features of the genomic data and clinical data is as follows: F gene = σ(W g X gene + b g ), F clinical = σ(W c X clinical + b c ) Among them, X gene and X clinical represent the input matrices of genomic data and clinical data respectively, W g and W c are the learned weight matrices respectively, b g and b c are the bias vectors respectively, and σ(·) represents the activation function.

6. The system for predicting the prognosis risk of colorectal cancer according to claim 1, characterized in that, The multimodal feature fusion module realizes weighted fusion of multimodal features through the attention mechanism, and its formula is as follows: F = Attention(F image , F gene , F clinical ) Among them, F is the fused feature vector, and Attention(·) is a fusion method based on the attention mechanism, which is used to dynamically assign the weight of each modality data.

7. The system for predicting the prognosis risk of colorectal cancer according to claim 6, wherein The calculation method of the attention mechanism is: Among them, F i represents the i-th modal feature vector, and α i represents the attention weight of the i-th modality, and W i represents the learning weight of the i-th modality, represents the transpose of the weight vector W i , indicating the transpose of W i and the dot product calculation is performed between W i and the feature vector F 8. The system for predicting the prognosis risk of colorectal cancer according to claim 1, characterized in that, The output of the prognostic risk prediction module adopts a classification method based on the softmax function, and its calculation formula is: R = softmax(WF+b) Among them, F is the fusion feature vector, W is the weight matrix of the classifier, and b is the bias term.

9. The system for predicting the prognosis risk of colorectal cancer according to claim 8, wherein The prediction model is trained by minimizing the following cross entropy loss function: Among them, N is the number of samples, and y i is the true label of sample i, and R i is the predicted result of sample i.

10. The system for predicting the prognosis risk of colorectal cancer according to claim 1, wherein The visualization and interpretation module calculates the contribution of each feature in the prognostic risk prediction by the SHAP method, and the calculation formula is: Among them, φ i represents the Shapley value of the i-th feature, f(S) represents the contribution of the feature subset S to the model output, and |N| represents the total number of features.

Images