PROJECT SUMMARY/ABSTRACT Among the “big three” infectious diseases worldwide, malaria stands out for the complexity of the Plasmodium life-cycle and biology. Malaria parasites breed mainly within red blood cells, and across their lifespan there are dramatic shifts in protein expression and metabolism that alter their appearance, behavior, and susceptibility to clearance by the host immune system or antimalarial drugs. Because it is an infection of the blood, a biopsy can be taken with a simple finger prick, and the ability to derive histopathological information via light microscopy is a critical tool in the study of, and ultimately control and treatment of, malaria. Manual review is painstaking and imperfect. Neural network-based computer vision (CV) approaches can accelerate data acquisition from light microscopy and innovate new methods of extracting data currently only possible through costly, labor-intensive benchtop molecular methods or time-consuming review by a small number of malaria microscopy experts with the necessary training and experience to distinguish subtle differences between parasite forms. This R21 proposal builds on 12 months of preparatory work supported by a pilot grant from The Johns Hopkins University Institute for Data Intensive Engineering and Science, a collaborative pursuit of the Schools of Medicine and Engineering. The co-principal investigators developed a deep learning-based CV algorithm trained on a public dataset of >10,000 images of Plasmodium falciparum ring stage parasites that can detect and quantify parasites with >0.97 accuracy. However, significantly more information is ripe for extraction from malaria smears beyond the simple detection of parasites. We built an early prototype of a 2nd-generation CV algorithm capable of identifying the correct parasite stage to the level of early, middle or late ring stage with >0.80 accuracy, and in this proposal we aim to refine the performance and extend the capabilities of the malaria CV system to wider applications while pioneering new computational methods in multiple domain adaptation and weakly- and semi-supervised learning. The proposed project would result in the development of a next-generation malaria CV system that can derive molecular data from brightfield images for use by investigators at the bench or in the clinic. We will build out the prototype CV system to optimize performance, develop higher-order classifiers (e.g., differentiating viable from nonviable circulating parasites, finding once-infected cells for the prognosis of delayed hemolysis after treatment), and run the algorithm against different tissue backgrounds (e.g., liver, spleen). The product of this work will be a cutting-edge neural network-based malaria CV system that provides a multiplex readout of parasite biological parameters and cellular pathology to help propel the fields of malaria research and biomedical CV analysis forward.

PROJECT NARRATIVE Machine learning-based computer vision technology can supplement or even supplant some of the traditional benchtop methods that malariologists currently rely on for interrogating the biology of Plasmodium, a persistant cause of death and disease in much of the tropical and sub-tropical world where one-third of the world population resides. Microscopy-based methods may seem on the surface mundane or archaic, but they remain a crucial clinical and investigative tool for malaria where a tissue “biopsy” is as simple as a prick of the finger. We aim to modernize microscopy applications in basic science and pre-clinical malaria research by leveraging recent advances in computer vision and artificial intelligence to accelerate and innovate the means by which information is derived from malaria blood smears and, in the process, establish novel methods in weakly- and semi-supervised learning and domain adaptation to overcome key challenges in biomedical image analysis.