Deep learning on lateral flow immunoassay for the analysis of detection data

Area of Science:

• Biomedical Engineering
• Medical Diagnostics
• Artificial Intelligence in Healthcare

Background:

• Lateral flow immunoassays (LFIA) are crucial for in vitro diagnostics but present challenges in peak shape analysis.
• Classical peak-finding methods struggle with complex LFIA data, failing to distinguish true peaks from noise or interference, and missing weak signals.

Purpose of the Study:

• To develop a novel deep learning-based method for accurate data processing in LFIA.
• To overcome limitations of classical methods in analyzing complex LFIA peak shapes and improve reliability for point-of-care testing (POCT) instruments.

Main Methods:

• A two-step deep learning approach: a classification model to screen double-peaks data and an improved U-Net segmentation model for integral region segmentation.
• Implementation within a custom-designed hand-held fluorescence immunochromatography analyzer.

Main Results:

• Classification model achieved 99.59% accuracy; segmentation model obtained an IoU of 0.9680.
• Demonstrated accurate Ferritin quantification (0-500 ng/ml, R² = 0.9986) with low CVs (≤1.37%) and good recovery (96.37-105.07%).
• Significantly reduced peak-finding errors caused by interference or noise, crucial for field usability of hand-held devices.

Conclusions:

• The proposed deep learning method offers a robust solution for complex LFIA data analysis.
• This advancement enhances the reliability and user-friendliness of hand-held POCT instruments, reducing the need for technical support.
• Presents a new direction for data processing in LFIA-based POCT devices.