Deep learning based recognition of shape-coded microparticles

Abstract

Encoded particles have been used for multiplexed diagnostics, drugs testing, and anti-counterfeiting applications. Recently, shape-coded hydrogel particles with amphiphilic properties have enabled an amplified duplexed bioassay. However, a limitation to read multiple particle shape-codes in an automated manner and within a reasonable time prevents a widespread adaptation of such potent diagnostic platforms. In this work, we applied established deep learning based multi-class segmentation models, such as U-Net, Attention U-Net, and UNet3+, to detect five or more particle shape-codes within a single image in an automated fashion within seconds. We demonstrated that the tested models provided prosaic results, when implemented on an imbalanced and limited raw dataset, with the best intersection over union (IoU) scores of 0.76 and 0.46 for six- and eleven-class segmentation, respectively. We introduced augmentation by translocation (ABT) technique to enhance the performances of the tested models significantly, where the best IoU scores for the six and eleven classes increased to 0.92 and 0.74, respectively. These initial findings to detect multiple shapes of the particles in an automated manner underscore the potential of shape-coded particles to be used in multiplexed bioassays. The code is available at: github.com/destgeerlab/shape-coded-particles.