Convolution Neural Networks for Point-of-Care Diagnostics of Bacterial Infections in Blood

Abstract

ABSTRACTA droplet of blood, when evaporated on a surface, leaves dried residue—the fractal patterns formed on the dried residues can act as markers for infection present in the blood. Exploiting the unique patterns found in the residues of a naturally dried droplet of blood, we propose a Point-of-Care (POC) diagnostic tool for detecting broad-spectrum of bacterial infections (such as Enterobacter aerogenes, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, and Pseudomonas aeruginosa, Salmonella Typhi) in blood. The diagnosis process we propose is straightforward and can be performed with the following steps: A droplet of blood (healthy or infected) of volume range 0.5 to 2 μl is allowed to dry on a clean glass surface and is imaged using a conventional optical microscope. A computer algorithm based on the framework of convolution neural network (CNN) is used to classify the captured images of dried blood droplets according to the bacterial infection. In total, our multiclass model reports an accuracy of 92% for detecting six bacterial species infections in the blood (with control being the uninfected or healthy blood). The high accuracy of detecting bacteria in the blood reported in this article is commensurate with the standard bacteriological tests. Thus, this article presents a proof-of-concept of a potential futuristic tool for a rapid and low-cost diagnosis of bacterial infection in the blood.