Jilin University Licenses Multimodal Data Fusion-Based Quality Inspection Method for Diagnostic Kit Raw Materials to Changchun Jinzerui Medical Technology Co., Ltd. for RMB 130,000

Recently, Jilin University released a public notice on the conversion of scientific and technological achievements, proposing to transfer its“A Method for Quality Inspection of Raw Materials in Test Kits Based on Multimodal Data Fusion”The patent is licensed to Changchun Jinzerui Medical Technology Co. Ltd through agreement-based pricing, with the transaction structured as a non-exclusive license for a term ofFive Years, the transaction amount is130,000 yuan。

This patent is byThe First Hospital of Jilin University Clinical Medical CollegeofTian Runhui and Lv MengJointly completed. It involves the field of automated quality testing of raw chemical reagent materials, with the core being the application ofSpectral sensors, microscopy imaging equipment, and chemical sensorsSynchronous acquisition of multimodal data. These data undergo spatiotemporal alignment via an adaptive spatiotemporal feature calibration network, followed by redundancy removal through a hierarchical information entropy screening mechanism, and finally yield results after cross-modal consistency verification and adversarial perturbation robustness testing. This method accurately performs protein activity level classification and micro-particle uniformity defect diagnosis, effectively addressing issues such as inaccurate multimodal data alignment and cumulative redundancy interference in traditional detection, thereby enhancing detection accuracy and stability.

In the market application of quality testing for raw materials in diagnostic kits, traditional methods have long faced multiple pain points, severely constraining the reliability and efficiency of quality control.

On the one hand,Single-Modality DetectionMethodsOften relying solely on chemical indicator analysis or physical image recognition, these methods are significantly constrained by specific scenarios, making it difficult to comprehensively characterize the multidimensional properties of raw materials and prone to overlooking critical quality information.

On the other hand,Current Fusion Methods for Simple Multimodal Data Concatenation, failing to adequately model the intrinsic correlations and redundancies among multi-source data. For instance, due to differences in equipment, sampling frequency, and representation dimensions between spectral data and microscopic image data, direct fusion often leads to issues such as mismatched feature scales and temporal misalignment, resulting in the loss or distortion of critical information.

Meanwhile, multi-modal data contain both complementary features and substantial redundant or irrelevant noise. Conventional approaches, such as static weighting or simple concatenation, fail to adaptively distinguish informative signals from redundant interference, leading to feature space expansion and noise accumulation, which ultimately compromise the robustness and generalization capability of detection models.

These pain points manifest directly asHigh Fluctuation in Detection Results and High False Positive Rate in Complex Scenarios. Even with an increased number of sensors or optimized unimodal algorithms, the core contradiction in multimodal data synergy has not been fundamentally resolved, making it difficult to meet the high demands for detection accuracy and stability in actual production.

In the data acquisition phase, this solution incorporates a dynamic collaboration mechanism.This mechanism is based onSpectroscopy, Microscopic Imaging, ChemistryDynamically adjust the sampling frequency based on the differences in physical characteristics among the three types of data. The spectral sensor can switch sampling modes to adapt to the rate of compositional changes, while the chemical sensor and microscopic imaging equipment maintain strict temporal synchronization. If anomalies occur in data from a single modality, the system triggers cross-device compensatory acquisition to ensure the spatiotemporal consistency of multi-source data.

During the data processing phase, this solution pioneers an adaptive spatiotemporal feature calibration network.This network employs asymmetric convolutional kernels to differentially extract features from various modalities, and incorporates a bidirectional cross-modal attention gating mechanism to precisely address temporal phase shifts and spatial scale mismatches, thereby achieving adaptive alignment of multimodal data. Meanwhile, the proposed solution introducesHierarchical Information Entropy Screening Mechanism, through dual filtering, redundant features can be sparsified and weakly discriminative features suppressed; furthermore, the strategy can be adaptively adjusted according to actual scenarios to prevent the loss of effective complementary information.

In the result validation and output phase, this solution establishes an integrated system combining closed-loop feedback with dual verification.This system calibrates network parameters through cross-modal consistency verification, enhances noise suppression capabilities via adversarial perturbation testing, and improves robustness in complex scenarios by incorporating historical data compensation strategies. Furthermore, the solution employs a combination of specialized classification models, utilizing Softmax classifiers and Support Vector Machine (SVM) models to achieve precise classification of protein activity levels and diagnosis of particle uniformity defects, respectively. Coupled with cross-modal compensation and feature backtracking mechanisms, this approach ensures the accuracy and reliability of detection results.

The aforementioned end-to-end technological innovations not only address the core pain points of traditional detection methods in a targeted manner but also differentiate themselves from existing solutions currently available on the market. With the widespread application of multimodal data fusion technology in the field of diagnostics, a number of solutions tailored to related scenarios have gradually emerged in the market.

Beijing Zhongchi Weiye's Automatic Coagulation Analyzer Reagent Identification and Positioning SystemCommercial applications have been implemented on a small scale. The core function of this product isProvides automatic identification and positioning for reagents used in coagulation analyzers. Its core technical logic bears some relevance to this patent, but there are significant differences in application scenarios and the specificity of technical design. Its multimodal data fusion primarily integrates “reagent image microstructure data + waveform energy data.” The reagent image recognition module extracts and encodes microstructural features, which are then combined with waveform energy mapping data. A dynamic feedback fusion algorithm adjusts the fusion strategy, ultimately achieving rapid identification and localization of reagents, and indirectly enabling basic quality screening of coagulation reagent raw materials (such as reagent appearance and assessment of basic activity).

Kuoran Bio KRpath Fully Automated AI Pathology Image Analysis SystemIt is primarily used for clinical pathological diagnosis. Its core technology is also based on “multimodal data fusion + AI algorithms,” and it can be regarded as a similar competing product related to this patent. Its multimodal data fusion mainly integrates “pathological image modality (morphological features) + molecular phenotype modality (protein expression data) + clinical information modality.” Through adaptive deep learning algorithms, it achieves functions such as image annotation, cell recognition, and phenotypic feature identification.

Antu Bio Multimodal AI Diagnostic PlatformIntegratedChemiluminescent Quantitative Assay, Gene Sequencing, Mass Spectrometry AnalysisClass III technologies attempt to achieve “full-dimensional detection” through multimodal data fusion. However, their core application scenarios focus on the diagnosis of clinical samples (such as tumor marker detection and pathogen identification), rather than quality control of raw materials for test kits.

Overall, multimodal fusion technology is driving the detection field to develop in greater depth, with various innovative solutions continuously emerging. This patent introducesSystematic designs including tri-modal collaborative acquisition of “spectral + image + chemical” data, adaptive spatiotemporal feature calibration, hierarchical information entropy screening, and closed-loop feedback optimization, thereby providing a new technical pathway for quality testing of raw materials used in diagnostic kits. We look forward to collaborating with more industry peers to continuously explore the broader possibilities of technological integration and application scenarios, jointly driving the industry toward greater intelligence and precision.