DeepGT: Deep learning-based quantification of nanosized bioparticles in bright-field micrographs of Gires-Tournois biosensor

Abstract

Rapid and decentralized quantification of viral load profiles in infected patients is vital for assessing clinical severity and tailoring appropriate therapeutic strategies. Although microscopic imaging offers potential for labelfree and amplification-free quantitative diagnostics, the small size (similar to 100 nm in diameter) and low refractive index (n similar to 1.5) of bioparticles present challenges in achieving accurate estimations, consequently increasing the limit of detection (LoD). In this study, we present a novel synergistic biosensing approach, DeepGT, combining Gires-Tournois (GT) sensing platforms with deep learning algorithms to enhance nanoscale bioparticle counting accuracy. The GT sensing platform serves as a photonic resonator, increasing bioparticle visibility in bright-field microscopy and maximizing chromatic contrast. By employing a back-end with a dilated convolutional neural network architecture, DeepGT effectively refines artifacts and color deviations, significantly improving particle estimation accuracy (MAE similar to 2.37 across 1596 images) compared to rule-based algorithms (MAE similar to 13.47). Notably, the enhanced accuracy in detecting invisible particles (e.g., two- or three-particles) enables an LoD of 138 pg ml(-1), facilitating a dynamic linear correlation at low viral concentration ranges within the clinical spectrum of infection, from asymptomatic to severe cases. Leveraging transfer learning, DeepGT, which relies on a chromatometry-based strategy instead of a spatial resolution approach, exhibits exceptional precision when analyzing particles of diverse dimensions smaller than the microscopy system&apos;s minimum diffraction limit in visible light (< 258 nm). The DeepGT approach holds promise for early screening and triage of emerging viruses, reducing costs and time requirements in diagnostics.