Deep Learning on chromatographic data for Segmentation and Sensitive Analysis

Abstract

Lateral flow immunoassay (LFIA) is one of the most common methods in point-of-care testing, which is widely applied in some conditions for various applications. Image segmentation is an increasingly popular experimental paradigm to efficiently test the target area in LFIA. However, due to process pollution, and problems related to the experimental operation and irregular structure of the background of the reaction, currently available tools cannot be used to extract correct signals from these images, which affects the accuracy of detection. Machine learning has significantly improved modern biochemical analysis by pushing the limits of traditional techniques for the recognition and processing of images. In this paper, the U-Net, a variant of the convolutional neural network (CNN) is used for the quantitative analysis of LFIA images for the accurate quantification of single- and multi-target images. By graying, binarizing, and labeling different concentrations of test strips, the target area of LFIA images containing the T-/C-lines is extracted and obtained. Then it provides updated trends and directions for the development of LFIA technology. Several indicators are introduced to evaluate the proposed U-Net structure to verify the feasibility and effectiveness of its image processing capability. When the trained U-Net model was used to process images, the peak signal-to-noise ratio was 22.4 dB, significantly higher than prevalent methods in the area that have reported only a 4 dB improvement in the quality of the graphics. The intersection-over-union between samples also increased to above 93%. Our results show that the proposed method has significant potential for detecting a segmented target in an LFIA area, especially weak positive signals and multichannel detection. With other modifications, this deep learning method can be applied as a powerful tool to study rapid detection devices, systems, and biological images.