A low-cost, automated parasite diagnostic system via a portable, robotic microscope and deep learning.

Abstract

Manual hand counting of parasites in fecal samples requires costly components and substantial expertise, limiting its use in resource-constrained settings and encouraging overuse of prophylactic medication. To address this issue, a cost-effective, automated parasite diagnostic system that does not require special sample preparation or a trained user was developed. It is composed of an inexpensive (~US$350), portable, robotic microscope that can scan over the size of an entire McMaster chamber (100 mm2 ) and capture high-resolution (~1 μm lateral resolution) bright field images without need for user intervention. Fecal samples prepared using the McMaster flotation method were imaged, with the imaging region comprising the entire McMaster chamber. These images are then automatically segmented and analyzed using a trained convolution neural network (CNN) to robustly separate eggs from background debris. Simple postprocessing of the CNN output yields both egg species and egg counts. The system was validated by comparing accuracy with hand-counts by a trained operator, with excellent performance. As a further demonstration of utility, the system was used to conveniently quantify drug response over time in a single animal, showing residual disease due to Anthelmintic resistance after 2 weeks.