Deep learning-based parasite detection for early and comprehensive diagnosis of animal trypanosomosis from thin blood smears.

The aim of this study was to assess the possibility of deep learning-based object detection models for the early and comprehensive detection of animal trypanosomosis. We constructed deep learning models for the early detection of Trypanosoma parasites including Trypanosoma congolense from in vitro and in vivo thin blood smears. Our models were based on YOLO (You Look at Once), one of the most common and high-performance object detection models. The method was applied to 14, 380 thin blood smear images with 47, 276 parasites (cells). For the T. congolense, the possibility for the early detection of the parasites was assessed through different concentration levels of cultured parasites (sparse to dense) and time-course blood sampling from infected mice. The in vitro model trained by T. congolense was applied to different species of T. brucei brucei and T. evansi in order to investigate the comprehensiveness of our approach. Our deep learning models successfully identified Trypanosoma parasites even for the settings of early detection. Our models also showed high precision (>0.90) for the dense and late predictions, not only for the same species and same sample source (in vitro / in vivo) of trypanosomes but also for the different species and different sample source (comprehensive prediction from in vitro to in vivo). The results showed that our methods are applicable for the purpose of early detection, not only for a specific Trypanosoma parasite spp. and the same sample source, but also for other spp. and sample source.