CN-121980417-A - Cardiovascular disease classification system and method based on common-private characteristic multi-source vibration spectrum fusion network and application

CN121980417ACN 121980417 ACN121980417 ACN 121980417ACN-121980417-A

Abstract

The invention relates to a cardiovascular disease classification system, a cardiovascular disease classification method and application of a multisource vibration spectrum fusion network based on common-private characteristics. A cardiovascular disease classification system based on a multi-source vibration spectrum fusion network of common-private features comprises a data collection module, a data preprocessing module, a multi-scale private feature extractor module, a dynamic sharing feature generation module, a bidirectional attention fusion module and a classifier module, wherein the data collection module is used for collecting Raman spectrum data and infrared spectrum data of serum, the data preprocessing module is used for preprocessing the Raman spectrum data and the infrared spectrum data, the multi-scale private feature extractor module is used for extracting multi-scale private features of the preprocessed Raman spectrum data and the preprocessed infrared spectrum data, the dynamic sharing feature generation module is used for extracting dynamic sharing features of the preprocessed Raman spectrum data and the preprocessed infrared spectrum data, the bidirectional attention fusion module is used for carrying out deep interaction and self-adaption fusion on the multi-scale private features, and the classifier module is used for splicing and classifying the fused multi-scale private features and the dynamic sharing features. According to the technical scheme, the classification performance is good.

Inventors

YANG YINING
LV XIAOYI
CHEN CHENG
CHEN CHEN
CHANG CHENJIE
YAN LEI
TAO JING

Assignees

新疆维吾尔自治区人民医院
新疆大学

Dates

Publication Date: 20260505
Application Date: 20260116

Claims (10)

1. A cardiovascular disease classification system based on a multi-source vibration spectrum fusion network of common-private characteristics, comprising: the data collection module is used for collecting Raman spectrum data and infrared spectrum data of serum; The data preprocessing module is used for preprocessing the Raman spectrum data and the infrared spectrum data; The multi-scale private feature extractor module adopts a deep convolutional neural network structure to extract the multi-scale private features of the preprocessed Raman spectrum data and infrared spectrum data; The dynamic sharing characteristic generation module adopts a two-stage structure of dual-mode projection-dynamic gating fusion to extract the dynamic sharing characteristics of the preprocessed Raman spectrum data and the preprocessed infrared spectrum data; the bidirectional attention fusion module is used for carrying out deep interaction and self-adaptive fusion on the multi-scale private features; And the classifier module is used for splicing and classifying the fused multi-scale private features and the dynamic sharing features.
2. The cardiovascular disease classification system according to claim 1, wherein, The multi-scale private feature extractor module comprises a private feature extractor of Raman spectrum data and a private feature extractor of infrared spectrum data.
3. The cardiovascular disease classification system according to claim 2, wherein, The private feature extractor of the Raman spectrum data is a parallel multi-scale one-dimensional convolutional neural network and is composed of three parallel branches, wherein the output feature graphs of the branches are spliced into feature vectors after global average pooling, and finally the feature vectors are subjected to degradation and fusion through a full-connection layer to output Raman spectrum private feature vectors; The private feature extractor of the infrared spectrum data is a serial-parallel mixed multi-scale convolution network and comprises two main branches, the outputs of the two main branches are subjected to global average pooling to obtain 2 feature vectors, the 2 feature vectors are spliced to obtain fusion features, and finally the fusion features are projected into the infrared spectrum private feature vectors through a full-connection layer.
4. The cardiovascular disease classification system according to claim 3, wherein, In the private feature extractor of the Raman spectrum data, the sizes of branch kernels are 3, 5 and 7 respectively.
5. The cardiovascular disease classification system according to claim 3, wherein, In the private feature extractor of infrared spectrum data, the path of the first main branch firstly uses a convolution layer with the kernel size of 5 to extract basic features, then reduces the sequence length through a maximum pooling layer, and then refines semantic features through a convolution layer with the kernel size of 3; And the path of the second main branch firstly uses a large convolution kernel with the kernel size of 15, then reduces the sequence length by exciting a pooling layer, and then optimizes the pooled coarse granularity characteristic by a convolution layer with the kernel size of 5.
6. The cardiovascular disease classification system according to claim 1, wherein, The dynamic shared feature generation module maps the infrared and Raman features to feature spaces with the same dimension through an independent linear projection layer, generates sample-level weights through a gating network, performs weighted fusion on bimodal projection features, and finally outputs dynamic shared features.
7. The cardiovascular disease classification system according to claim 1, wherein, In the bidirectional attention fusion module, infrared private characteristics are used as inquiry, raman private characteristics are used as key values, and the fusion process is as follows: Attention adopts a scaling dot product mechanism, supports multi-head attention to capture multi-dimensional dependence, and for single-head attention scores Wherein Is a single head dimension; The score was then normalized to a probability distribution by Softmax, ensuring a weight sum of 1, Applying the attention weight to the value vector to obtain the attention characteristic of the focusing key information, Stitching the bi-directional attention features into Outputs gating weights through a two-layer network, Wherein , ; And then carrying out weighted fusion.
8. The cardiovascular disease classification system according to claim 7, wherein, In the bidirectional attention fusion module, a residual path and layer normalization processing is adopted, wherein the process is that an average value of original private features is adopted as a residual item, basic feature information is reserved, and then a residual addition result is subjected to layer normalization, and the formula is as follows: wherein And The mean and variance of the features are respectively, And As a result of the learnable scaling parameters and offset parameters, Preventing denominator from being 0.
9. A method for classifying cardiovascular diseases based on a multisource vibration spectrum fusion network with common-private characteristics, characterized in that the cardiovascular disease classification system according to any one of claims 1-8 comprises the following steps: (1) Collecting raman spectrum data and infrared spectrum data of serum and tears; (2) Preprocessing the Raman spectrum data and the infrared spectrum data; (3) Extracting the multiscale private features of the preprocessed Raman spectrum data and infrared spectrum data by a multiscale private feature extractor; Extracting the dynamic sharing characteristics of the preprocessed Raman spectrum data and the preprocessed infrared spectrum data by adopting a dynamic sharing characteristic generating method of a two-stage structure of dual-mode projection-dynamic gating fusion; (4) And after the multi-scale private features are subjected to deep interaction and self-adaptive fusion, the multi-scale private features are spliced with the dynamic sharing features to be classified.
10. Use of the cardiovascular disease classification system according to any one of claims 1-8, or the cardiovascular disease classification method according to claim 9, in a cardiovascular disease assisted diagnosis device.

Description

Cardiovascular disease classification system and method based on common-private characteristic multi-source vibration spectrum fusion network and application Technical Field The invention belongs to medical data processing, and particularly relates to a cardiovascular disease classification system and method based on a multisource vibration spectrum fusion network with common-private characteristics, and application. Background Aortic dissection (Aortic Dissection, AD) is a life threatening cardiovascular emergency with global high incidence. It has the advantages of rapid onset, rapid progress of disease, extremely high death rate, and 1-2% increase of death rate per hour after treatment delay. Early, rapid, accurate diagnosis is critical in reducing mortality. Computed Tomography Angiography (CTA) is currently the "gold standard" for diagnosing AD, with extremely high sensitivity and specificity. Transesophageal echocardiography (TEE) and Magnetic Resonance Angiography (MRA) are also important means. However, CTA has limitations in that it is radiation-exposed, expensive in equipment, difficult to accomplish quickly at the bedside, etc. TEE is a semi-invasive procedure and MRA is performed for a long period of time, and is not suitable for extremely critical or unstable patients. D-dimer detection is often used to exclude diagnostics, but its specificity is not high. Aortic stenosis (Aortic Stenosis, AS) refers to a type of disease in which aortic valve patency is limited, resulting in left ventricular outflow obstruction, and its severity can be classified into types such AS C0 (mild calcification) and C1 (severe calcification) according to the degree of calcification, etc. The disease usually progresses chronically, but severe stenosis can occur with acute decompensation, leading to heart failure and even sudden death. Transthoracic Echocardiography (TTE) is the preferred non-invasive method of assessing aortic stenosis, and can measure the trans-valve flow rate, pressure differential, and orifice area to determine severity and help identify the degree of calcification. The accuracy of ultrasound assessment is highly dependent on the experience of the operator, and for very mild or very severe lesions, there may be judgment deviations, affecting early intervention or surgical timing. Thus, there is a great clinical need for a new method that overcomes the limitations of the prior art, enabling a rapid, noninvasive and objective accurate assessment of AD and AS calcification. Notably, there is an overlap in the clinical manifestations of AD and severe AS, such AS chest pain, dyspnea, syncope and even cardiogenic shock, which can be misinterpreted AS acute exacerbation of aortic stenosis. Especially when AD involves the aortic root and causes acute aortic regurgitation, the resulting murmur and hemodynamic changes are easily confused with severe aortic stenosis. Although both are critical cardiovascular diseases, the treatment principles and urgency are quite different. AD requires immediate surgery or endoluminal repair and is contraindicated for thrombolysis or anticoagulation. While severe aortic stenosis (type C1) eventually requires surgical intervention, it generally allows for a certain period of drug stabilization and timing assessment, and the blood pressure management strategy also requires individual adjustment, which should avoid hypotension in severe stenoses. Thus, errors in differential diagnosis can directly result in a loss of treatment timing, even with disastrous consequences due to improper administration, such as incorrect anticoagulation in AD patients. The development of an objective diagnosis method capable of rapidly and accurately distinguishing the severity degree of AD and aortic valve stenosis (AS) has urgent and important clinical significance for optimizing clinical decisions and improving prognosis of patients. The key driving force for breaking through the performance bottleneck of a single information source and promoting the diagnosis and development of artificial intelligence auxiliary diseases is that complementary information from different sources is integrated to form a more comprehensive and accurate representation of a target object. Aiming at the multi-source spectrum fusion technology, a plurality of researches and applications exist at present, but the existing method still has obvious limitations, and the depth and the breadth of the clinical application are restricted. For example, most fusion strategies remain in low-level feature stitching or late decision weighting, and lack of explicit modeling of deep complementary relationships and difference features between multi-source spectra results in underutilization of information and susceptibility of discriminative features to be inundated with noise or redundancy. And current models generally lack an explicit separation mechanism for "common features-private features". In addition, the existing method general