Facile quantification of nanosized bioparticles in bright-field micrographs of Gires-Tournois biosensor using deep learning

Abstract

Identifying viral load profiles in infected patients, particularly those with high virulence, through decentralized equipment facilitates accurately assessing clinical severity and determining optimal therapeutic strategies. However, conventional microscope systems face challenges in accurately observing nanoscale bioparticles with low refractive index (∼ 1.5) that induce weak light-matter interactions. Here, we present a synergistic biosensing framework that addresses these limitations by utilizing a nanophotonic resonator and deep learning algorithms. We functionalize Gires-Tournois resonators as immunosensors without the need for labeling or amplification, maximizing the color difference of analytes in bright-field micrographs. By automatically refining surface detects and color deviations, our combined systems employing a convolutional neural network with a dilated back-end achieved a mean absolute error of 2.37 in particle count estimation. Our system exhibits a dynamic linear range for detection limits of 138 pg ml−1 and estimated imperceptibly small clusters (e.g., two- or three-particle clusters) across the clinical spectrum from ‘asymptomatic’ to ‘severe’ cases. Furthermore, we demonstrate high accuracy when applied to analytes smaller than the diffraction limit (< 258 nm) of the microscopy system using transfer learning. The presented system offers a valuable quantitative biosensing technique for early screening and triage of emerging viruses, reducing costs and time requirements in diagnostics.