The Potential of Deep Learning Object Detection in Citizen-Driven Snail Host Monitoring to Map Putative Disease Transmission Sites

Abstract

Schistosomiasis is a neglected tropical disease caused by parasitic flukes transmitted by freshwater snails. Despite increasing efforts of mass drug administration, schistosomiasis remains a public health concern and the World Health Organization recommends complementary snail control. To address the need for broad-scale and actual snail distribution data to guide snail control, we adopted a citizen science approach and recruited citizen scientists (CSs) to perform weekly snail sampling in an endemic setting in Uganda. Snails were identified, sorted and counted according to genus, and photographed; and the photos were uploaded for expert-led validation and feedback. However, expert validation is time-consuming and introduces a delay in verified data output. Thus, artificial intelligence could provide a solution by means of automated detection and counting of multiple snails collected from the field. Trained on approximately 2,500 citizen-collected images, the resulting model can simultaneously detect and count Biomphalaria and Radix snails with average precision of 98.1% and 98.8%, respectively. The object detection model also agreed with the expert’s decision, on average, for 98.8% of the test images and can be run in real time (24.6 images per second). We conclude that the automatic and instant detection can rapidly and reliably validate data submitted by CSs in the field, ultimately minimizing expert validation efforts and thereby facilitating the mapping of putative schistosomiasis transmission sites. An extension to a mobile application could equip CSs in remote areas with instant learning opportunities and expert-like identification skills, overcoming the need for on-site training and extensive expert intervention.