Background
Parasitemia measurement are crucial for management of parasites/pathogens of human and veterinary relevance. In this context, egg enumeration represents an important parasitic helminth surveillance tool and an evaluating parameter for the efficacy of in vitro and in vivo anthelmintic drug treatments. Egg counting has been for years a time-consuming process, requiring technical expertise of appropriately trained individuals using (predominantly) microscopic methods. The development of a system to automate parasitic helminth egg counts would represent a substantial improvement to the current process. Most recently, neural network-based object detection was applied to detect eggs of common soil-transmitted helminths (e.g. Ascaris lumbricoides and Trichuris trichiura).
Methodology
Here, we present a successful application of AI-based automated egg counting of Schistosoma mansoni eggs. Firstly, a neural network (RetinaNet) was trained using a small subset of images (about 200). This preliminary study aimed to investigate the effects of the main hyperparameters (e.g. mini-batch size, epochs and steps number, learning rate and augmentation) on the accuracy of the resulting model. Based on this, a second training study was undertaken on a larger library of images (1.14k images) and the resulting model was evaluated for interpretive accuracy on 200 unseen images. The working protype achieved promising performances with a sensitivity and precision above 88% and 91%, respectively. Preliminary data have been collected on the application of this image detection tool on other S. mansoni life cycle stages such as the miracidia-sporocyst transformation enabling automatic enumeration and morphological-based scoring of this intra-molluscan stage.
Conclusion
This AI-based automated egg counting procedure could represent a great asset for parasitologists for speeding up and standardising the otherwise lengthy procedure of egg enumeration by microscopy. Indeed, the current prototype could achieve even better performance with continued expansion of the image database by the users and improvement of machine learning technology. Moreover, this process opens up other possible applications including quantifying the differentiation of S. mansoni egg development as well as discriminating between other schistosome or helminth species.
