Methods, terminals, and media for quantitative detection of components in label-deficient serum systems

By combining a multi-task deep learning model with a credibility mask matrix, the problems of matrix competitive adsorption, signal heterogeneity, and missing labels in the quantitative detection of serum components are solved, achieving high-precision detection even with incomplete labels, and improving the robustness of the model and the credibility of the detection results.

CN121762527BActive Publication Date: 2026-05-26HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
Filing Date
2026-02-28
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

Existing technologies for quantitative detection of serum components suffer from problems such as severe matrix competition adsorption, heterogeneity of signal physical response, and low data utilization due to missing labels, resulting in poor model generalization ability.

Method used

A multi-task deep learning model is adopted, which combines a confidence mask matrix and a hierarchical specific heterogeneous path to construct a component quantitative detection method for a label-deficient serum system. By sharing the feature extraction module, the hierarchical specific heterogeneous path module and the quantitative prediction output head, the confidence mask matrix is ​​used to weight and control the loss calculation and gradient propagation, thereby shielding the interference of missing labels and improving the robustness of the model.

Benefits of technology

Even with incomplete clinical labels, the model maintains training stability, improves label utilization and data robustness, significantly enhances the scientific validity and credibility of the detection results, and is able to extract effective features in both low signal-to-noise ratio and adsorption saturation regions, thereby improving detection accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121762527B_ABST
    Figure CN121762527B_ABST
Patent Text Reader

Abstract

This invention relates to the fields of spectral analysis and deep learning technologies, and discloses a method, terminal, and medium for quantitative detection of components in serum systems lacking labels. The method acquires Raman spectral data from multiple serum samples and preprocesses them to obtain standardized one-dimensional Raman spectral input data, constructing a training set. Based on the presence status of labels in the samples of the training set and the quality confidence of the spectra, a confidence mask matrix is ​​generated. A multi-task deep learning model is constructed, including a shared feature extraction module, a hierarchical specific heterogeneous path module, and a quantitative prediction output head. During model training, the confidence mask matrix is ​​used to weight and control loss calculation and gradient propagation. The preprocessed Raman spectral data of the serum samples to be tested is input into the trained model, and the detection results are output. This invention ensures that the model can maintain training stability even with incomplete clinical labels, significantly improving label utilization and data robustness.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Chromatographic analysis method for traditional Chinese medicine compound coronary prescription sample based on artificial neural network and differential spectrum

    CN117368361A

  • Raman spectrum nonlinear generation method driven by mixed machine learning

    CN120470287A